fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 struct btrfs_dio_data {
  70         u64 outstanding_extents;
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74 };
  75
  76 static const struct inode_operations btrfs_dir_inode_operations;
  77 static const struct inode_operations btrfs_symlink_inode_operations;
  78 static const struct inode_operations btrfs_dir_ro_inode_operations;
  79 static const struct inode_operations btrfs_special_inode_operations;
  80 static const struct inode_operations btrfs_file_inode_operations;
  81 static const struct address_space_operations btrfs_aops;
  82 static const struct address_space_operations btrfs_symlink_aops;
  83 static const struct file_operations btrfs_dir_file_operations;
  84 static const struct extent_io_ops btrfs_extent_io_ops;
  85
  86 static struct kmem_cache *btrfs_inode_cachep;
  87 struct kmem_cache *btrfs_trans_handle_cachep;
  88 struct kmem_cache *btrfs_transaction_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, int *page_started,
 109                                    unsigned long *nr_written, int unlock);
 110 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 111                                            u64 len, u64 orig_start,
 112                                            u64 block_start, u64 block_len,
 113                                            u64 orig_block_len, u64 ram_bytes,
 114                                            int type);
 115
 116 static int btrfs_dirty_inode(struct inode *inode);
 117
 118 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 119 void btrfs_test_inode_set_ops(struct inode *inode)
 120 {
 121         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 122 }
 123 #endif
 124
 125 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 126                                      struct inode *inode,  struct inode *dir,
 127                                      const struct qstr *qstr)
 128 {
 129         int err;
 130
 131         err = btrfs_init_acl(trans, inode, dir);
 132         if (!err)
 133                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 134         return err;
 135 }
 136
 137 /*
 138  * this does all the hard work for inserting an inline extent into
 139  * the btree.  The caller should have done a btrfs_drop_extents so that
 140  * no overlapping inline items exist in the btree
 141  */
 142 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 143                                 struct btrfs_path *path, int extent_inserted,
 144                                 struct btrfs_root *root, struct inode *inode,
 145                                 u64 start, size_t size, size_t compressed_size,
 146                                 int compress_type,
 147                                 struct page **compressed_pages)
 148 {
 149         struct extent_buffer *leaf;
 150         struct page *page = NULL;
 151         char *kaddr;
 152         unsigned long ptr;
 153         struct btrfs_file_extent_item *ei;
 154         int err = 0;
 155         int ret;
 156         size_t cur_size = size;
 157         unsigned long offset;
 158
 159         if (compressed_size && compressed_pages)
 160                 cur_size = compressed_size;
 161
 162         inode_add_bytes(inode, size);
 163
 164         if (!extent_inserted) {
 165                 struct btrfs_key key;
 166                 size_t datasize;
 167
 168                 key.objectid = btrfs_ino(inode);
 169                 key.offset = start;
 170                 key.type = BTRFS_EXTENT_DATA_KEY;
 171
 172                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 173                 path->leave_spinning = 1;
 174                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                               datasize);
 176                 if (ret) {
 177                         err = ret;
 178                         goto fail;
 179                 }
 180         }
 181         leaf = path->nodes[0];
 182         ei = btrfs_item_ptr(leaf, path->slots[0],
 183                             struct btrfs_file_extent_item);
 184         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 185         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 186         btrfs_set_file_extent_encryption(leaf, ei, 0);
 187         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 188         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 189         ptr = btrfs_file_extent_inline_start(ei);
 190
 191         if (compress_type != BTRFS_COMPRESS_NONE) {
 192                 struct page *cpage;
 193                 int i = 0;
 194                 while (compressed_size > 0) {
 195                         cpage = compressed_pages[i];
 196                         cur_size = min_t(unsigned long, compressed_size,
 197                                        PAGE_CACHE_SIZE);
 198
 199                         kaddr = kmap_atomic(cpage);
 200                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 201                         kunmap_atomic(kaddr);
 202
 203                         i++;
 204                         ptr += cur_size;
 205                         compressed_size -= cur_size;
 206                 }
 207                 btrfs_set_file_extent_compression(leaf, ei,
 208                                                   compress_type);
 209         } else {
 210                 page = find_get_page(inode->i_mapping,
 211                                      start >> PAGE_CACHE_SHIFT);
 212                 btrfs_set_file_extent_compression(leaf, ei, 0);
 213                 kaddr = kmap_atomic(page);
 214                 offset = start & (PAGE_CACHE_SIZE - 1);
 215                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 216                 kunmap_atomic(kaddr);
 217                 page_cache_release(page);
 218         }
 219         btrfs_mark_buffer_dirty(leaf);
 220         btrfs_release_path(path);
 221
 222         /*
 223          * we're an inline extent, so nobody can
 224          * extend the file past i_size without locking
 225          * a page we already have locked.
 226          *
 227          * We must do any isize and inode updates
 228          * before we unlock the pages.  Otherwise we
 229          * could end up racing with unlink.
 230          */
 231         BTRFS_I(inode)->disk_i_size = inode->i_size;
 232         ret = btrfs_update_inode(trans, root, inode);
 233
 234         return ret;
 235 fail:
 236         return err;
 237 }
 238
 239
 240 /*
 241  * conditionally insert an inline extent into the file.  This
 242  * does the checks required to make sure the data is small enough
 243  * to fit as an inline extent.
 244  */
 245 static noinline int cow_file_range_inline(struct btrfs_root *root,
 246                                           struct inode *inode, u64 start,
 247                                           u64 end, size_t compressed_size,
 248                                           int compress_type,
 249                                           struct page **compressed_pages)
 250 {
 251         struct btrfs_trans_handle *trans;
 252         u64 isize = i_size_read(inode);
 253         u64 actual_end = min(end + 1, isize);
 254         u64 inline_len = actual_end - start;
 255         u64 aligned_end = ALIGN(end, root->sectorsize);
 256         u64 data_len = inline_len;
 257         int ret;
 258         struct btrfs_path *path;
 259         int extent_inserted = 0;
 260         u32 extent_item_size;
 261
 262         if (compressed_size)
 263                 data_len = compressed_size;
 264
 265         if (start > 0 ||
 266             actual_end > PAGE_CACHE_SIZE ||
 267             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 268             (!compressed_size &&
 269             (actual_end & (root->sectorsize - 1)) == 0) ||
 270             end + 1 < isize ||
 271             data_len > root->fs_info->max_inline) {
 272                 return 1;
 273         }
 274
 275         path = btrfs_alloc_path();
 276         if (!path)
 277                 return -ENOMEM;
 278
 279         trans = btrfs_join_transaction(root);
 280         if (IS_ERR(trans)) {
 281                 btrfs_free_path(path);
 282                 return PTR_ERR(trans);
 283         }
 284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 285
 286         if (compressed_size && compressed_pages)
 287                 extent_item_size = btrfs_file_extent_calc_inline_size(
 288                    compressed_size);
 289         else
 290                 extent_item_size = btrfs_file_extent_calc_inline_size(
 291                     inline_len);
 292
 293         ret = __btrfs_drop_extents(trans, root, inode, path,
 294                                    start, aligned_end, NULL,
 295                                    1, 1, extent_item_size, &extent_inserted);
 296         if (ret) {
 297                 btrfs_abort_transaction(trans, root, ret);
 298                 goto out;
 299         }
 300
 301         if (isize > actual_end)
 302                 inline_len = min_t(u64, isize, actual_end);
 303         ret = insert_inline_extent(trans, path, extent_inserted,
 304                                    root, inode, start,
 305                                    inline_len, compressed_size,
 306                                    compress_type, compressed_pages);
 307         if (ret && ret != -ENOSPC) {
 308                 btrfs_abort_transaction(trans, root, ret);
 309                 goto out;
 310         } else if (ret == -ENOSPC) {
 311                 ret = 1;
 312                 goto out;
 313         }
 314
 315         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 316         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 317         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 318 out:
 319         /*
 320          * Don't forget to free the reserved space, as for inlined extent
 321          * it won't count as data extent, free them directly here.
 322          * And at reserve time, it's always aligned to page size, so
 323          * just free one page here.
 324          */
 325         btrfs_qgroup_free_data(inode, 0, PAGE_CACHE_SIZE);
 326         btrfs_free_path(path);
 327         btrfs_end_transaction(trans, root);
 328         return ret;
 329 }
 330
 331 struct async_extent {
 332         u64 start;
 333         u64 ram_size;
 334         u64 compressed_size;
 335         struct page **pages;
 336         unsigned long nr_pages;
 337         int compress_type;
 338         struct list_head list;
 339 };
 340
 341 struct async_cow {
 342         struct inode *inode;
 343         struct btrfs_root *root;
 344         struct page *locked_page;
 345         u64 start;
 346         u64 end;
 347         struct list_head extents;
 348         struct btrfs_work work;
 349 };
 350
 351 static noinline int add_async_extent(struct async_cow *cow,
 352                                      u64 start, u64 ram_size,
 353                                      u64 compressed_size,
 354                                      struct page **pages,
 355                                      unsigned long nr_pages,
 356                                      int compress_type)
 357 {
 358         struct async_extent *async_extent;
 359
 360         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 361         BUG_ON(!async_extent); /* -ENOMEM */
 362         async_extent->start = start;
 363         async_extent->ram_size = ram_size;
 364         async_extent->compressed_size = compressed_size;
 365         async_extent->pages = pages;
 366         async_extent->nr_pages = nr_pages;
 367         async_extent->compress_type = compress_type;
 368         list_add_tail(&async_extent->list, &cow->extents);
 369         return 0;
 370 }
 371
 372 static inline int inode_need_compress(struct inode *inode)
 373 {
 374         struct btrfs_root *root = BTRFS_I(inode)->root;
 375
 376         /* force compress */
 377         if (btrfs_test_opt(root, FORCE_COMPRESS))
 378                 return 1;
 379         /* bad compression ratios */
 380         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 381                 return 0;
 382         if (btrfs_test_opt(root, COMPRESS) ||
 383             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 384             BTRFS_I(inode)->force_compress)
 385                 return 1;
 386         return 0;
 387 }
 388
 389 /*
 390  * we create compressed extents in two phases.  The first
 391  * phase compresses a range of pages that have already been
 392  * locked (both pages and state bits are locked).
 393  *
 394  * This is done inside an ordered work queue, and the compression
 395  * is spread across many cpus.  The actual IO submission is step
 396  * two, and the ordered work queue takes care of making sure that
 397  * happens in the same order things were put onto the queue by
 398  * writepages and friends.
 399  *
 400  * If this code finds it can't get good compression, it puts an
 401  * entry onto the work queue to write the uncompressed bytes.  This
 402  * makes sure that both compressed inodes and uncompressed inodes
 403  * are written in the same order that the flusher thread sent them
 404  * down.
 405  */
 406 static noinline void compress_file_range(struct inode *inode,
 407                                         struct page *locked_page,
 408                                         u64 start, u64 end,
 409                                         struct async_cow *async_cow,
 410                                         int *num_added)
 411 {
 412         struct btrfs_root *root = BTRFS_I(inode)->root;
 413         u64 num_bytes;
 414         u64 blocksize = root->sectorsize;
 415         u64 actual_end;
 416         u64 isize = i_size_read(inode);
 417         int ret = 0;
 418         struct page **pages = NULL;
 419         unsigned long nr_pages;
 420         unsigned long nr_pages_ret = 0;
 421         unsigned long total_compressed = 0;
 422         unsigned long total_in = 0;
 423         unsigned long max_compressed = SZ_128K;
 424         unsigned long max_uncompressed = SZ_128K;
 425         int i;
 426         int will_compress;
 427         int compress_type = root->fs_info->compress_type;
 428         int redirty = 0;
 429
 430         /* if this is a small write inside eof, kick off a defrag */
 431         if ((end - start + 1) < SZ_16K &&
 432             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434
 435         actual_end = min_t(u64, isize, end + 1);
 436 again:
 437         will_compress = 0;
 438         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
 439         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_CACHE_SIZE);
 440
 441         /*
 442          * we don't want to send crud past the end of i_size through
 443          * compression, that's just a waste of CPU time.  So, if the
 444          * end of the file is before the start of our current
 445          * requested range of bytes, we bail out to the uncompressed
 446          * cleanup code that can deal with all of this.
 447          *
 448          * It isn't really the fastest way to fix things, but this is a
 449          * very uncommon corner.
 450          */
 451         if (actual_end <= start)
 452                 goto cleanup_and_bail_uncompressed;
 453
 454         total_compressed = actual_end - start;
 455
 456         /*
 457          * skip compression for a small file range(<=blocksize) that
 458          * isn't an inline extent, since it dosen't save disk space at all.
 459          */
 460         if (total_compressed <= blocksize &&
 461            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 462                 goto cleanup_and_bail_uncompressed;
 463
 464         /* we want to make sure that amount of ram required to uncompress
 465          * an extent is reasonable, so we limit the total size in ram
 466          * of a compressed extent to 128k.  This is a crucial number
 467          * because it also controls how easily we can spread reads across
 468          * cpus for decompression.
 469          *
 470          * We also want to make sure the amount of IO required to do
 471          * a random read is reasonably small, so we limit the size of
 472          * a compressed extent to 128k.
 473          */
 474         total_compressed = min(total_compressed, max_uncompressed);
 475         num_bytes = ALIGN(end - start + 1, blocksize);
 476         num_bytes = max(blocksize,  num_bytes);
 477         total_in = 0;
 478         ret = 0;
 479
 480         /*
 481          * we do compression for mount -o compress and when the
 482          * inode has not been flagged as nocompress.  This flag can
 483          * change at any time if we discover bad compression ratios.
 484          */
 485         if (inode_need_compress(inode)) {
 486                 WARN_ON(pages);
 487                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 488                 if (!pages) {
 489                         /* just bail out to the uncompressed code */
 490                         goto cont;
 491                 }
 492
 493                 if (BTRFS_I(inode)->force_compress)
 494                         compress_type = BTRFS_I(inode)->force_compress;
 495
 496                 /*
 497                  * we need to call clear_page_dirty_for_io on each
 498                  * page in the range.  Otherwise applications with the file
 499                  * mmap'd can wander in and change the page contents while
 500                  * we are compressing them.
 501                  *
 502                  * If the compression fails for any reason, we set the pages
 503                  * dirty again later on.
 504                  */
 505                 extent_range_clear_dirty_for_io(inode, start, end);
 506                 redirty = 1;
 507                 ret = btrfs_compress_pages(compress_type,
 508                                            inode->i_mapping, start,
 509                                            total_compressed, pages,
 510                                            nr_pages, &nr_pages_ret,
 511                                            &total_in,
 512                                            &total_compressed,
 513                                            max_compressed);
 514
 515                 if (!ret) {
 516                         unsigned long offset = total_compressed &
 517                                 (PAGE_CACHE_SIZE - 1);
 518                         struct page *page = pages[nr_pages_ret - 1];
 519                         char *kaddr;
 520
 521                         /* zero the tail end of the last page, we might be
 522                          * sending it down to disk
 523                          */
 524                         if (offset) {
 525                                 kaddr = kmap_atomic(page);
 526                                 memset(kaddr + offset, 0,
 527                                        PAGE_CACHE_SIZE - offset);
 528                                 kunmap_atomic(kaddr);
 529                         }
 530                         will_compress = 1;
 531                 }
 532         }
 533 cont:
 534         if (start == 0) {
 535                 /* lets try to make an inline extent */
 536                 if (ret || total_in < (actual_end - start)) {
 537                         /* we didn't compress the entire range, try
 538                          * to make an uncompressed inline extent.
 539                          */
 540                         ret = cow_file_range_inline(root, inode, start, end,
 541                                                     0, 0, NULL);
 542                 } else {
 543                         /* try making a compressed inline extent */
 544                         ret = cow_file_range_inline(root, inode, start, end,
 545                                                     total_compressed,
 546                                                     compress_type, pages);
 547                 }
 548                 if (ret <= 0) {
 549                         unsigned long clear_flags = EXTENT_DELALLOC |
 550                                 EXTENT_DEFRAG;
 551                         unsigned long page_error_op;
 552
 553                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 554                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 555
 556                         /*
 557                          * inline extent creation worked or returned error,
 558                          * we don't need to create any more async work items.
 559                          * Unlock and free up our temp pages.
 560                          */
 561                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 562                                                      clear_flags, PAGE_UNLOCK |
 563                                                      PAGE_CLEAR_DIRTY |
 564                                                      PAGE_SET_WRITEBACK |
 565                                                      page_error_op |
 566                                                      PAGE_END_WRITEBACK);
 567                         goto free_pages_out;
 568                 }
 569         }
 570
 571         if (will_compress) {
 572                 /*
 573                  * we aren't doing an inline extent round the compressed size
 574                  * up to a block size boundary so the allocator does sane
 575                  * things
 576                  */
 577                 total_compressed = ALIGN(total_compressed, blocksize);
 578
 579                 /*
 580                  * one last check to make sure the compression is really a
 581                  * win, compare the page count read with the blocks on disk
 582                  */
 583                 total_in = ALIGN(total_in, PAGE_CACHE_SIZE);
 584                 if (total_compressed >= total_in) {
 585                         will_compress = 0;
 586                 } else {
 587                         num_bytes = total_in;
 588                 }
 589         }
 590         if (!will_compress && pages) {
 591                 /*
 592                  * the compression code ran but failed to make things smaller,
 593                  * free any pages it allocated and our page pointer array
 594                  */
 595                 for (i = 0; i < nr_pages_ret; i++) {
 596                         WARN_ON(pages[i]->mapping);
 597                         page_cache_release(pages[i]);
 598                 }
 599                 kfree(pages);
 600                 pages = NULL;
 601                 total_compressed = 0;
 602                 nr_pages_ret = 0;
 603
 604                 /* flag the file so we don't compress in the future */
 605                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 606                     !(BTRFS_I(inode)->force_compress)) {
 607                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 608                 }
 609         }
 610         if (will_compress) {
 611                 *num_added += 1;
 612
 613                 /* the async work queues will take care of doing actual
 614                  * allocation on disk for these compressed pages,
 615                  * and will submit them to the elevator.
 616                  */
 617                 add_async_extent(async_cow, start, num_bytes,
 618                                  total_compressed, pages, nr_pages_ret,
 619                                  compress_type);
 620
 621                 if (start + num_bytes < end) {
 622                         start += num_bytes;
 623                         pages = NULL;
 624                         cond_resched();
 625                         goto again;
 626                 }
 627         } else {
 628 cleanup_and_bail_uncompressed:
 629                 /*
 630                  * No compression, but we still need to write the pages in
 631                  * the file we've been given so far.  redirty the locked
 632                  * page if it corresponds to our extent and set things up
 633                  * for the async work queue to run cow_file_range to do
 634                  * the normal delalloc dance
 635                  */
 636                 if (page_offset(locked_page) >= start &&
 637                     page_offset(locked_page) <= end) {
 638                         __set_page_dirty_nobuffers(locked_page);
 639                         /* unlocked later on in the async handlers */
 640                 }
 641                 if (redirty)
 642                         extent_range_redirty_for_io(inode, start, end);
 643                 add_async_extent(async_cow, start, end - start + 1,
 644                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 645                 *num_added += 1;
 646         }
 647
 648         return;
 649
 650 free_pages_out:
 651         for (i = 0; i < nr_pages_ret; i++) {
 652                 WARN_ON(pages[i]->mapping);
 653                 page_cache_release(pages[i]);
 654         }
 655         kfree(pages);
 656 }
 657
 658 static void free_async_extent_pages(struct async_extent *async_extent)
 659 {
 660         int i;
 661
 662         if (!async_extent->pages)
 663                 return;
 664
 665         for (i = 0; i < async_extent->nr_pages; i++) {
 666                 WARN_ON(async_extent->pages[i]->mapping);
 667                 page_cache_release(async_extent->pages[i]);
 668         }
 669         kfree(async_extent->pages);
 670         async_extent->nr_pages = 0;
 671         async_extent->pages = NULL;
 672 }
 673
 674 /*
 675  * phase two of compressed writeback.  This is the ordered portion
 676  * of the code, which only gets called in the order the work was
 677  * queued.  We walk all the async extents created by compress_file_range
 678  * and send them down to the disk.
 679  */
 680 static noinline void submit_compressed_extents(struct inode *inode,
 681                                               struct async_cow *async_cow)
 682 {
 683         struct async_extent *async_extent;
 684         u64 alloc_hint = 0;
 685         struct btrfs_key ins;
 686         struct extent_map *em;
 687         struct btrfs_root *root = BTRFS_I(inode)->root;
 688         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 689         struct extent_io_tree *io_tree;
 690         int ret = 0;
 691
 692 again:
 693         while (!list_empty(&async_cow->extents)) {
 694                 async_extent = list_entry(async_cow->extents.next,
 695                                           struct async_extent, list);
 696                 list_del(&async_extent->list);
 697
 698                 io_tree = &BTRFS_I(inode)->io_tree;
 699
 700 retry:
 701                 /* did the compression code fall back to uncompressed IO? */
 702                 if (!async_extent->pages) {
 703                         int page_started = 0;
 704                         unsigned long nr_written = 0;
 705
 706                         lock_extent(io_tree, async_extent->start,
 707                                          async_extent->start +
 708                                          async_extent->ram_size - 1);
 709
 710                         /* allocate blocks */
 711                         ret = cow_file_range(inode, async_cow->locked_page,
 712                                              async_extent->start,
 713                                              async_extent->start +
 714                                              async_extent->ram_size - 1,
 715                                              &page_started, &nr_written, 0);
 716
 717                         /* JDM XXX */
 718
 719                         /*
 720                          * if page_started, cow_file_range inserted an
 721                          * inline extent and took care of all the unlocking
 722                          * and IO for us.  Otherwise, we need to submit
 723                          * all those pages down to the drive.
 724                          */
 725                         if (!page_started && !ret)
 726                                 extent_write_locked_range(io_tree,
 727                                                   inode, async_extent->start,
 728                                                   async_extent->start +
 729                                                   async_extent->ram_size - 1,
 730                                                   btrfs_get_extent,
 731                                                   WB_SYNC_ALL);
 732                         else if (ret)
 733                                 unlock_page(async_cow->locked_page);
 734                         kfree(async_extent);
 735                         cond_resched();
 736                         continue;
 737                 }
 738
 739                 lock_extent(io_tree, async_extent->start,
 740                             async_extent->start + async_extent->ram_size - 1);
 741
 742                 ret = btrfs_reserve_extent(root,
 743                                            async_extent->compressed_size,
 744                                            async_extent->compressed_size,
 745                                            0, alloc_hint, &ins, 1, 1);
 746                 if (ret) {
 747                         free_async_extent_pages(async_extent);
 748
 749                         if (ret == -ENOSPC) {
 750                                 unlock_extent(io_tree, async_extent->start,
 751                                               async_extent->start +
 752                                               async_extent->ram_size - 1);
 753
 754                                 /*
 755                                  * we need to redirty the pages if we decide to
 756                                  * fallback to uncompressed IO, otherwise we
 757                                  * will not submit these pages down to lower
 758                                  * layers.
 759                                  */
 760                                 extent_range_redirty_for_io(inode,
 761                                                 async_extent->start,
 762                                                 async_extent->start +
 763                                                 async_extent->ram_size - 1);
 764
 765                                 goto retry;
 766                         }
 767                         goto out_free;
 768                 }
 769                 /*
 770                  * here we're doing allocation and writeback of the
 771                  * compressed pages
 772                  */
 773                 btrfs_drop_extent_cache(inode, async_extent->start,
 774                                         async_extent->start +
 775                                         async_extent->ram_size - 1, 0);
 776
 777                 em = alloc_extent_map();
 778                 if (!em) {
 779                         ret = -ENOMEM;
 780                         goto out_free_reserve;
 781                 }
 782                 em->start = async_extent->start;
 783                 em->len = async_extent->ram_size;
 784                 em->orig_start = em->start;
 785                 em->mod_start = em->start;
 786                 em->mod_len = em->len;
 787
 788                 em->block_start = ins.objectid;
 789                 em->block_len = ins.offset;
 790                 em->orig_block_len = ins.offset;
 791                 em->ram_bytes = async_extent->ram_size;
 792                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 793                 em->compress_type = async_extent->compress_type;
 794                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 795                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 796                 em->generation = -1;
 797
 798                 while (1) {
 799                         write_lock(&em_tree->lock);
 800                         ret = add_extent_mapping(em_tree, em, 1);
 801                         write_unlock(&em_tree->lock);
 802                         if (ret != -EEXIST) {
 803                                 free_extent_map(em);
 804                                 break;
 805                         }
 806                         btrfs_drop_extent_cache(inode, async_extent->start,
 807                                                 async_extent->start +
 808                                                 async_extent->ram_size - 1, 0);
 809                 }
 810
 811                 if (ret)
 812                         goto out_free_reserve;
 813
 814                 ret = btrfs_add_ordered_extent_compress(inode,
 815                                                 async_extent->start,
 816                                                 ins.objectid,
 817                                                 async_extent->ram_size,
 818                                                 ins.offset,
 819                                                 BTRFS_ORDERED_COMPRESSED,
 820                                                 async_extent->compress_type);
 821                 if (ret) {
 822                         btrfs_drop_extent_cache(inode, async_extent->start,
 823                                                 async_extent->start +
 824                                                 async_extent->ram_size - 1, 0);
 825                         goto out_free_reserve;
 826                 }
 827
 828                 /*
 829                  * clear dirty, set writeback and unlock the pages.
 830                  */
 831                 extent_clear_unlock_delalloc(inode, async_extent->start,
 832                                 async_extent->start +
 833                                 async_extent->ram_size - 1,
 834                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 835                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 836                                 PAGE_SET_WRITEBACK);
 837                 ret = btrfs_submit_compressed_write(inode,
 838                                     async_extent->start,
 839                                     async_extent->ram_size,
 840                                     ins.objectid,
 841                                     ins.offset, async_extent->pages,
 842                                     async_extent->nr_pages);
 843                 if (ret) {
 844                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 845                         struct page *p = async_extent->pages[0];
 846                         const u64 start = async_extent->start;
 847                         const u64 end = start + async_extent->ram_size - 1;
 848
 849                         p->mapping = inode->i_mapping;
 850                         tree->ops->writepage_end_io_hook(p, start, end,
 851                                                          NULL, 0);
 852                         p->mapping = NULL;
 853                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 854                                                      PAGE_END_WRITEBACK |
 855                                                      PAGE_SET_ERROR);
 856                         free_async_extent_pages(async_extent);
 857                 }
 858                 alloc_hint = ins.objectid + ins.offset;
 859                 kfree(async_extent);
 860                 cond_resched();
 861         }
 862         return;
 863 out_free_reserve:
 864         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 865 out_free:
 866         extent_clear_unlock_delalloc(inode, async_extent->start,
 867                                      async_extent->start +
 868                                      async_extent->ram_size - 1,
 869                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 870                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 871                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 872                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 873                                      PAGE_SET_ERROR);
 874         free_async_extent_pages(async_extent);
 875         kfree(async_extent);
 876         goto again;
 877 }
 878
 879 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 880                                       u64 num_bytes)
 881 {
 882         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 883         struct extent_map *em;
 884         u64 alloc_hint = 0;
 885
 886         read_lock(&em_tree->lock);
 887         em = search_extent_mapping(em_tree, start, num_bytes);
 888         if (em) {
 889                 /*
 890                  * if block start isn't an actual block number then find the
 891                  * first block in this inode and use that as a hint.  If that
 892                  * block is also bogus then just don't worry about it.
 893                  */
 894                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 895                         free_extent_map(em);
 896                         em = search_extent_mapping(em_tree, 0, 0);
 897                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 898                                 alloc_hint = em->block_start;
 899                         if (em)
 900                                 free_extent_map(em);
 901                 } else {
 902                         alloc_hint = em->block_start;
 903                         free_extent_map(em);
 904                 }
 905         }
 906         read_unlock(&em_tree->lock);
 907
 908         return alloc_hint;
 909 }
 910
 911 /*
 912  * when extent_io.c finds a delayed allocation range in the file,
 913  * the call backs end up in this code.  The basic idea is to
 914  * allocate extents on disk for the range, and create ordered data structs
 915  * in ram to track those extents.
 916  *
 917  * locked_page is the page that writepage had locked already.  We use
 918  * it to make sure we don't do extra locks or unlocks.
 919  *
 920  * *page_started is set to one if we unlock locked_page and do everything
 921  * required to start IO on it.  It may be clean and already done with
 922  * IO when we return.
 923  */
 924 static noinline int cow_file_range(struct inode *inode,
 925                                    struct page *locked_page,
 926                                    u64 start, u64 end, int *page_started,
 927                                    unsigned long *nr_written,
 928                                    int unlock)
 929 {
 930         struct btrfs_root *root = BTRFS_I(inode)->root;
 931         u64 alloc_hint = 0;
 932         u64 num_bytes;
 933         unsigned long ram_size;
 934         u64 disk_num_bytes;
 935         u64 cur_alloc_size;
 936         u64 blocksize = root->sectorsize;
 937         struct btrfs_key ins;
 938         struct extent_map *em;
 939         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 940         int ret = 0;
 941
 942         if (btrfs_is_free_space_inode(inode)) {
 943                 WARN_ON_ONCE(1);
 944                 ret = -EINVAL;
 945                 goto out_unlock;
 946         }
 947
 948         num_bytes = ALIGN(end - start + 1, blocksize);
 949         num_bytes = max(blocksize,  num_bytes);
 950         disk_num_bytes = num_bytes;
 951
 952         /* if this is a small write inside eof, kick off defrag */
 953         if (num_bytes < SZ_64K &&
 954             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 955                 btrfs_add_inode_defrag(NULL, inode);
 956
 957         if (start == 0) {
 958                 /* lets try to make an inline extent */
 959                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 960                                             NULL);
 961                 if (ret == 0) {
 962                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 963                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 964                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 965                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 966                                      PAGE_END_WRITEBACK);
 967
 968                         *nr_written = *nr_written +
 969                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
 970                         *page_started = 1;
 971                         goto out;
 972                 } else if (ret < 0) {
 973                         goto out_unlock;
 974                 }
 975         }
 976
 977         BUG_ON(disk_num_bytes >
 978                btrfs_super_total_bytes(root->fs_info->super_copy));
 979
 980         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 981         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 982
 983         while (disk_num_bytes > 0) {
 984                 unsigned long op;
 985
 986                 cur_alloc_size = disk_num_bytes;
 987                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 988                                            root->sectorsize, 0, alloc_hint,
 989                                            &ins, 1, 1);
 990                 if (ret < 0)
 991                         goto out_unlock;
 992
 993                 em = alloc_extent_map();
 994                 if (!em) {
 995                         ret = -ENOMEM;
 996                         goto out_reserve;
 997                 }
 998                 em->start = start;
 999                 em->orig_start = em->start;
1000                 ram_size = ins.offset;
1001                 em->len = ins.offset;
1002                 em->mod_start = em->start;
1003                 em->mod_len = em->len;
1004
1005                 em->block_start = ins.objectid;
1006                 em->block_len = ins.offset;
1007                 em->orig_block_len = ins.offset;
1008                 em->ram_bytes = ram_size;
1009                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1010                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1011                 em->generation = -1;
1012
1013                 while (1) {
1014                         write_lock(&em_tree->lock);
1015                         ret = add_extent_mapping(em_tree, em, 1);
1016                         write_unlock(&em_tree->lock);
1017                         if (ret != -EEXIST) {
1018                                 free_extent_map(em);
1019                                 break;
1020                         }
1021                         btrfs_drop_extent_cache(inode, start,
1022                                                 start + ram_size - 1, 0);
1023                 }
1024                 if (ret)
1025                         goto out_reserve;
1026
1027                 cur_alloc_size = ins.offset;
1028                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1029                                                ram_size, cur_alloc_size, 0);
1030                 if (ret)
1031                         goto out_drop_extent_cache;
1032
1033                 if (root->root_key.objectid ==
1034                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1035                         ret = btrfs_reloc_clone_csums(inode, start,
1036                                                       cur_alloc_size);
1037                         if (ret)
1038                                 goto out_drop_extent_cache;
1039                 }
1040
1041                 if (disk_num_bytes < cur_alloc_size)
1042                         break;
1043
1044                 /* we're not doing compressed IO, don't unlock the first
1045                  * page (which the caller expects to stay locked), don't
1046                  * clear any dirty bits and don't set any writeback bits
1047                  *
1048                  * Do set the Private2 bit so we know this page was properly
1049                  * setup for writepage
1050                  */
1051                 op = unlock ? PAGE_UNLOCK : 0;
1052                 op |= PAGE_SET_PRIVATE2;
1053
1054                 extent_clear_unlock_delalloc(inode, start,
1055                                              start + ram_size - 1, locked_page,
1056                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1057                                              op);
1058                 disk_num_bytes -= cur_alloc_size;
1059                 num_bytes -= cur_alloc_size;
1060                 alloc_hint = ins.objectid + ins.offset;
1061                 start += cur_alloc_size;
1062         }
1063 out:
1064         return ret;
1065
1066 out_drop_extent_cache:
1067         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1068 out_reserve:
1069         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1070 out_unlock:
1071         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1072                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1073                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1074                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1075                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1076         goto out;
1077 }
1078
1079 /*
1080  * work queue call back to started compression on a file and pages
1081  */
1082 static noinline void async_cow_start(struct btrfs_work *work)
1083 {
1084         struct async_cow *async_cow;
1085         int num_added = 0;
1086         async_cow = container_of(work, struct async_cow, work);
1087
1088         compress_file_range(async_cow->inode, async_cow->locked_page,
1089                             async_cow->start, async_cow->end, async_cow,
1090                             &num_added);
1091         if (num_added == 0) {
1092                 btrfs_add_delayed_iput(async_cow->inode);
1093                 async_cow->inode = NULL;
1094         }
1095 }
1096
1097 /*
1098  * work queue call back to submit previously compressed pages
1099  */
1100 static noinline void async_cow_submit(struct btrfs_work *work)
1101 {
1102         struct async_cow *async_cow;
1103         struct btrfs_root *root;
1104         unsigned long nr_pages;
1105
1106         async_cow = container_of(work, struct async_cow, work);
1107
1108         root = async_cow->root;
1109         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
1110                 PAGE_CACHE_SHIFT;
1111
1112         /*
1113          * atomic_sub_return implies a barrier for waitqueue_active
1114          */
1115         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1116             5 * SZ_1M &&
1117             waitqueue_active(&root->fs_info->async_submit_wait))
1118                 wake_up(&root->fs_info->async_submit_wait);
1119
1120         if (async_cow->inode)
1121                 submit_compressed_extents(async_cow->inode, async_cow);
1122 }
1123
1124 static noinline void async_cow_free(struct btrfs_work *work)
1125 {
1126         struct async_cow *async_cow;
1127         async_cow = container_of(work, struct async_cow, work);
1128         if (async_cow->inode)
1129                 btrfs_add_delayed_iput(async_cow->inode);
1130         kfree(async_cow);
1131 }
1132
1133 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1134                                 u64 start, u64 end, int *page_started,
1135                                 unsigned long *nr_written)
1136 {
1137         struct async_cow *async_cow;
1138         struct btrfs_root *root = BTRFS_I(inode)->root;
1139         unsigned long nr_pages;
1140         u64 cur_end;
1141         int limit = 10 * SZ_1M;
1142
1143         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1144                          1, 0, NULL, GFP_NOFS);
1145         while (start < end) {
1146                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1147                 BUG_ON(!async_cow); /* -ENOMEM */
1148                 async_cow->inode = igrab(inode);
1149                 async_cow->root = root;
1150                 async_cow->locked_page = locked_page;
1151                 async_cow->start = start;
1152
1153                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1154                     !btrfs_test_opt(root, FORCE_COMPRESS))
1155                         cur_end = end;
1156                 else
1157                         cur_end = min(end, start + SZ_512K - 1);
1158
1159                 async_cow->end = cur_end;
1160                 INIT_LIST_HEAD(&async_cow->extents);
1161
1162                 btrfs_init_work(&async_cow->work,
1163                                 btrfs_delalloc_helper,
1164                                 async_cow_start, async_cow_submit,
1165                                 async_cow_free);
1166
1167                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
1168                         PAGE_CACHE_SHIFT;
1169                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1170
1171                 btrfs_queue_work(root->fs_info->delalloc_workers,
1172                                  &async_cow->work);
1173
1174                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1175                         wait_event(root->fs_info->async_submit_wait,
1176                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1177                             limit));
1178                 }
1179
1180                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1181                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1182                         wait_event(root->fs_info->async_submit_wait,
1183                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1184                            0));
1185                 }
1186
1187                 *nr_written += nr_pages;
1188                 start = cur_end + 1;
1189         }
1190         *page_started = 1;
1191         return 0;
1192 }
1193
1194 static noinline int csum_exist_in_range(struct btrfs_root *root,
1195                                         u64 bytenr, u64 num_bytes)
1196 {
1197         int ret;
1198         struct btrfs_ordered_sum *sums;
1199         LIST_HEAD(list);
1200
1201         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1202                                        bytenr + num_bytes - 1, &list, 0);
1203         if (ret == 0 && list_empty(&list))
1204                 return 0;
1205
1206         while (!list_empty(&list)) {
1207                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1208                 list_del(&sums->list);
1209                 kfree(sums);
1210         }
1211         return 1;
1212 }
1213
1214 /*
1215  * when nowcow writeback call back.  This checks for snapshots or COW copies
1216  * of the extents that exist in the file, and COWs the file as required.
1217  *
1218  * If no cow copies or snapshots exist, we write directly to the existing
1219  * blocks on disk
1220  */
1221 static noinline int run_delalloc_nocow(struct inode *inode,
1222                                        struct page *locked_page,
1223                               u64 start, u64 end, int *page_started, int force,
1224                               unsigned long *nr_written)
1225 {
1226         struct btrfs_root *root = BTRFS_I(inode)->root;
1227         struct btrfs_trans_handle *trans;
1228         struct extent_buffer *leaf;
1229         struct btrfs_path *path;
1230         struct btrfs_file_extent_item *fi;
1231         struct btrfs_key found_key;
1232         u64 cow_start;
1233         u64 cur_offset;
1234         u64 extent_end;
1235         u64 extent_offset;
1236         u64 disk_bytenr;
1237         u64 num_bytes;
1238         u64 disk_num_bytes;
1239         u64 ram_bytes;
1240         int extent_type;
1241         int ret, err;
1242         int type;
1243         int nocow;
1244         int check_prev = 1;
1245         bool nolock;
1246         u64 ino = btrfs_ino(inode);
1247
1248         path = btrfs_alloc_path();
1249         if (!path) {
1250                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1251                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1252                                              EXTENT_DO_ACCOUNTING |
1253                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1254                                              PAGE_CLEAR_DIRTY |
1255                                              PAGE_SET_WRITEBACK |
1256                                              PAGE_END_WRITEBACK);
1257                 return -ENOMEM;
1258         }
1259
1260         nolock = btrfs_is_free_space_inode(inode);
1261
1262         if (nolock)
1263                 trans = btrfs_join_transaction_nolock(root);
1264         else
1265                 trans = btrfs_join_transaction(root);
1266
1267         if (IS_ERR(trans)) {
1268                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1269                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1270                                              EXTENT_DO_ACCOUNTING |
1271                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1272                                              PAGE_CLEAR_DIRTY |
1273                                              PAGE_SET_WRITEBACK |
1274                                              PAGE_END_WRITEBACK);
1275                 btrfs_free_path(path);
1276                 return PTR_ERR(trans);
1277         }
1278
1279         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1280
1281         cow_start = (u64)-1;
1282         cur_offset = start;
1283         while (1) {
1284                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1285                                                cur_offset, 0);
1286                 if (ret < 0)
1287                         goto error;
1288                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1289                         leaf = path->nodes[0];
1290                         btrfs_item_key_to_cpu(leaf, &found_key,
1291                                               path->slots[0] - 1);
1292                         if (found_key.objectid == ino &&
1293                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1294                                 path->slots[0]--;
1295                 }
1296                 check_prev = 0;
1297 next_slot:
1298                 leaf = path->nodes[0];
1299                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1300                         ret = btrfs_next_leaf(root, path);
1301                         if (ret < 0)
1302                                 goto error;
1303                         if (ret > 0)
1304                                 break;
1305                         leaf = path->nodes[0];
1306                 }
1307
1308                 nocow = 0;
1309                 disk_bytenr = 0;
1310                 num_bytes = 0;
1311                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1312
1313                 if (found_key.objectid > ino)
1314                         break;
1315                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1316                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1317                         path->slots[0]++;
1318                         goto next_slot;
1319                 }
1320                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1321                     found_key.offset > end)
1322                         break;
1323
1324                 if (found_key.offset > cur_offset) {
1325                         extent_end = found_key.offset;
1326                         extent_type = 0;
1327                         goto out_check;
1328                 }
1329
1330                 fi = btrfs_item_ptr(leaf, path->slots[0],
1331                                     struct btrfs_file_extent_item);
1332                 extent_type = btrfs_file_extent_type(leaf, fi);
1333
1334                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1335                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1336                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1337                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1338                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1339                         extent_end = found_key.offset +
1340                                 btrfs_file_extent_num_bytes(leaf, fi);
1341                         disk_num_bytes =
1342                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1343                         if (extent_end <= start) {
1344                                 path->slots[0]++;
1345                                 goto next_slot;
1346                         }
1347                         if (disk_bytenr == 0)
1348                                 goto out_check;
1349                         if (btrfs_file_extent_compression(leaf, fi) ||
1350                             btrfs_file_extent_encryption(leaf, fi) ||
1351                             btrfs_file_extent_other_encoding(leaf, fi))
1352                                 goto out_check;
1353                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1354                                 goto out_check;
1355                         if (btrfs_extent_readonly(root, disk_bytenr))
1356                                 goto out_check;
1357                         if (btrfs_cross_ref_exist(trans, root, ino,
1358                                                   found_key.offset -
1359                                                   extent_offset, disk_bytenr))
1360                                 goto out_check;
1361                         disk_bytenr += extent_offset;
1362                         disk_bytenr += cur_offset - found_key.offset;
1363                         num_bytes = min(end + 1, extent_end) - cur_offset;
1364                         /*
1365                          * if there are pending snapshots for this root,
1366                          * we fall into common COW way.
1367                          */
1368                         if (!nolock) {
1369                                 err = btrfs_start_write_no_snapshoting(root);
1370                                 if (!err)
1371                                         goto out_check;
1372                         }
1373                         /*
1374                          * force cow if csum exists in the range.
1375                          * this ensure that csum for a given extent are
1376                          * either valid or do not exist.
1377                          */
1378                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1379                                 goto out_check;
1380                         nocow = 1;
1381                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1382                         extent_end = found_key.offset +
1383                                 btrfs_file_extent_inline_len(leaf,
1384                                                      path->slots[0], fi);
1385                         extent_end = ALIGN(extent_end, root->sectorsize);
1386                 } else {
1387                         BUG_ON(1);
1388                 }
1389 out_check:
1390                 if (extent_end <= start) {
1391                         path->slots[0]++;
1392                         if (!nolock && nocow)
1393                                 btrfs_end_write_no_snapshoting(root);
1394                         goto next_slot;
1395                 }
1396                 if (!nocow) {
1397                         if (cow_start == (u64)-1)
1398                                 cow_start = cur_offset;
1399                         cur_offset = extent_end;
1400                         if (cur_offset > end)
1401                                 break;
1402                         path->slots[0]++;
1403                         goto next_slot;
1404                 }
1405
1406                 btrfs_release_path(path);
1407                 if (cow_start != (u64)-1) {
1408                         ret = cow_file_range(inode, locked_page,
1409                                              cow_start, found_key.offset - 1,
1410                                              page_started, nr_written, 1);
1411                         if (ret) {
1412                                 if (!nolock && nocow)
1413                                         btrfs_end_write_no_snapshoting(root);
1414                                 goto error;
1415                         }
1416                         cow_start = (u64)-1;
1417                 }
1418
1419                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1420                         struct extent_map *em;
1421                         struct extent_map_tree *em_tree;
1422                         em_tree = &BTRFS_I(inode)->extent_tree;
1423                         em = alloc_extent_map();
1424                         BUG_ON(!em); /* -ENOMEM */
1425                         em->start = cur_offset;
1426                         em->orig_start = found_key.offset - extent_offset;
1427                         em->len = num_bytes;
1428                         em->block_len = num_bytes;
1429                         em->block_start = disk_bytenr;
1430                         em->orig_block_len = disk_num_bytes;
1431                         em->ram_bytes = ram_bytes;
1432                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1433                         em->mod_start = em->start;
1434                         em->mod_len = em->len;
1435                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1436                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1437                         em->generation = -1;
1438                         while (1) {
1439                                 write_lock(&em_tree->lock);
1440                                 ret = add_extent_mapping(em_tree, em, 1);
1441                                 write_unlock(&em_tree->lock);
1442                                 if (ret != -EEXIST) {
1443                                         free_extent_map(em);
1444                                         break;
1445                                 }
1446                                 btrfs_drop_extent_cache(inode, em->start,
1447                                                 em->start + em->len - 1, 0);
1448                         }
1449                         type = BTRFS_ORDERED_PREALLOC;
1450                 } else {
1451                         type = BTRFS_ORDERED_NOCOW;
1452                 }
1453
1454                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1455                                                num_bytes, num_bytes, type);
1456                 BUG_ON(ret); /* -ENOMEM */
1457
1458                 if (root->root_key.objectid ==
1459                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1460                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1461                                                       num_bytes);
1462                         if (ret) {
1463                                 if (!nolock && nocow)
1464                                         btrfs_end_write_no_snapshoting(root);
1465                                 goto error;
1466                         }
1467                 }
1468
1469                 extent_clear_unlock_delalloc(inode, cur_offset,
1470                                              cur_offset + num_bytes - 1,
1471                                              locked_page, EXTENT_LOCKED |
1472                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1473                                              PAGE_SET_PRIVATE2);
1474                 if (!nolock && nocow)
1475                         btrfs_end_write_no_snapshoting(root);
1476                 cur_offset = extent_end;
1477                 if (cur_offset > end)
1478                         break;
1479         }
1480         btrfs_release_path(path);
1481
1482         if (cur_offset <= end && cow_start == (u64)-1) {
1483                 cow_start = cur_offset;
1484                 cur_offset = end;
1485         }
1486
1487         if (cow_start != (u64)-1) {
1488                 ret = cow_file_range(inode, locked_page, cow_start, end,
1489                                      page_started, nr_written, 1);
1490                 if (ret)
1491                         goto error;
1492         }
1493
1494 error:
1495         err = btrfs_end_transaction(trans, root);
1496         if (!ret)
1497                 ret = err;
1498
1499         if (ret && cur_offset < end)
1500                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1501                                              locked_page, EXTENT_LOCKED |
1502                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1503                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1504                                              PAGE_CLEAR_DIRTY |
1505                                              PAGE_SET_WRITEBACK |
1506                                              PAGE_END_WRITEBACK);
1507         btrfs_free_path(path);
1508         return ret;
1509 }
1510
1511 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1512 {
1513
1514         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1515             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1516                 return 0;
1517
1518         /*
1519          * @defrag_bytes is a hint value, no spinlock held here,
1520          * if is not zero, it means the file is defragging.
1521          * Force cow if given extent needs to be defragged.
1522          */
1523         if (BTRFS_I(inode)->defrag_bytes &&
1524             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1525                            EXTENT_DEFRAG, 0, NULL))
1526                 return 1;
1527
1528         return 0;
1529 }
1530
1531 /*
1532  * extent_io.c call back to do delayed allocation processing
1533  */
1534 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1535                               u64 start, u64 end, int *page_started,
1536                               unsigned long *nr_written)
1537 {
1538         int ret;
1539         int force_cow = need_force_cow(inode, start, end);
1540
1541         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1542                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1543                                          page_started, 1, nr_written);
1544         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1545                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1546                                          page_started, 0, nr_written);
1547         } else if (!inode_need_compress(inode)) {
1548                 ret = cow_file_range(inode, locked_page, start, end,
1549                                       page_started, nr_written, 1);
1550         } else {
1551                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1552                         &BTRFS_I(inode)->runtime_flags);
1553                 ret = cow_file_range_async(inode, locked_page, start, end,
1554                                            page_started, nr_written);
1555         }
1556         return ret;
1557 }
1558
1559 static void btrfs_split_extent_hook(struct inode *inode,
1560                                     struct extent_state *orig, u64 split)
1561 {
1562         u64 size;
1563
1564         /* not delalloc, ignore it */
1565         if (!(orig->state & EXTENT_DELALLOC))
1566                 return;
1567
1568         size = orig->end - orig->start + 1;
1569         if (size > BTRFS_MAX_EXTENT_SIZE) {
1570                 u64 num_extents;
1571                 u64 new_size;
1572
1573                 /*
1574                  * See the explanation in btrfs_merge_extent_hook, the same
1575                  * applies here, just in reverse.
1576                  */
1577                 new_size = orig->end - split + 1;
1578                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1579                                         BTRFS_MAX_EXTENT_SIZE);
1580                 new_size = split - orig->start;
1581                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1582                                         BTRFS_MAX_EXTENT_SIZE);
1583                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1584                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1585                         return;
1586         }
1587
1588         spin_lock(&BTRFS_I(inode)->lock);
1589         BTRFS_I(inode)->outstanding_extents++;
1590         spin_unlock(&BTRFS_I(inode)->lock);
1591 }
1592
1593 /*
1594  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1595  * extents so we can keep track of new extents that are just merged onto old
1596  * extents, such as when we are doing sequential writes, so we can properly
1597  * account for the metadata space we'll need.
1598  */
1599 static void btrfs_merge_extent_hook(struct inode *inode,
1600                                     struct extent_state *new,
1601                                     struct extent_state *other)
1602 {
1603         u64 new_size, old_size;
1604         u64 num_extents;
1605
1606         /* not delalloc, ignore it */
1607         if (!(other->state & EXTENT_DELALLOC))
1608                 return;
1609
1610         if (new->start > other->start)
1611                 new_size = new->end - other->start + 1;
1612         else
1613                 new_size = other->end - new->start + 1;
1614
1615         /* we're not bigger than the max, unreserve the space and go */
1616         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1617                 spin_lock(&BTRFS_I(inode)->lock);
1618                 BTRFS_I(inode)->outstanding_extents--;
1619                 spin_unlock(&BTRFS_I(inode)->lock);
1620                 return;
1621         }
1622
1623         /*
1624          * We have to add up either side to figure out how many extents were
1625          * accounted for before we merged into one big extent.  If the number of
1626          * extents we accounted for is <= the amount we need for the new range
1627          * then we can return, otherwise drop.  Think of it like this
1628          *
1629          * [ 4k][MAX_SIZE]
1630          *
1631          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1632          * need 2 outstanding extents, on one side we have 1 and the other side
1633          * we have 1 so they are == and we can return.  But in this case
1634          *
1635          * [MAX_SIZE+4k][MAX_SIZE+4k]
1636          *
1637          * Each range on their own accounts for 2 extents, but merged together
1638          * they are only 3 extents worth of accounting, so we need to drop in
1639          * this case.
1640          */
1641         old_size = other->end - other->start + 1;
1642         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1643                                 BTRFS_MAX_EXTENT_SIZE);
1644         old_size = new->end - new->start + 1;
1645         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1646                                  BTRFS_MAX_EXTENT_SIZE);
1647
1648         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1649                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1650                 return;
1651
1652         spin_lock(&BTRFS_I(inode)->lock);
1653         BTRFS_I(inode)->outstanding_extents--;
1654         spin_unlock(&BTRFS_I(inode)->lock);
1655 }
1656
1657 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1658                                       struct inode *inode)
1659 {
1660         spin_lock(&root->delalloc_lock);
1661         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1662                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1663                               &root->delalloc_inodes);
1664                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1665                         &BTRFS_I(inode)->runtime_flags);
1666                 root->nr_delalloc_inodes++;
1667                 if (root->nr_delalloc_inodes == 1) {
1668                         spin_lock(&root->fs_info->delalloc_root_lock);
1669                         BUG_ON(!list_empty(&root->delalloc_root));
1670                         list_add_tail(&root->delalloc_root,
1671                                       &root->fs_info->delalloc_roots);
1672                         spin_unlock(&root->fs_info->delalloc_root_lock);
1673                 }
1674         }
1675         spin_unlock(&root->delalloc_lock);
1676 }
1677
1678 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1679                                      struct inode *inode)
1680 {
1681         spin_lock(&root->delalloc_lock);
1682         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1683                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1684                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1685                           &BTRFS_I(inode)->runtime_flags);
1686                 root->nr_delalloc_inodes--;
1687                 if (!root->nr_delalloc_inodes) {
1688                         spin_lock(&root->fs_info->delalloc_root_lock);
1689                         BUG_ON(list_empty(&root->delalloc_root));
1690                         list_del_init(&root->delalloc_root);
1691                         spin_unlock(&root->fs_info->delalloc_root_lock);
1692                 }
1693         }
1694         spin_unlock(&root->delalloc_lock);
1695 }
1696
1697 /*
1698  * extent_io.c set_bit_hook, used to track delayed allocation
1699  * bytes in this file, and to maintain the list of inodes that
1700  * have pending delalloc work to be done.
1701  */
1702 static void btrfs_set_bit_hook(struct inode *inode,
1703                                struct extent_state *state, unsigned *bits)
1704 {
1705
1706         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1707                 WARN_ON(1);
1708         /*
1709          * set_bit and clear bit hooks normally require _irqsave/restore
1710          * but in this case, we are only testing for the DELALLOC
1711          * bit, which is only set or cleared with irqs on
1712          */
1713         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1714                 struct btrfs_root *root = BTRFS_I(inode)->root;
1715                 u64 len = state->end + 1 - state->start;
1716                 bool do_list = !btrfs_is_free_space_inode(inode);
1717
1718                 if (*bits & EXTENT_FIRST_DELALLOC) {
1719                         *bits &= ~EXTENT_FIRST_DELALLOC;
1720                 } else {
1721                         spin_lock(&BTRFS_I(inode)->lock);
1722                         BTRFS_I(inode)->outstanding_extents++;
1723                         spin_unlock(&BTRFS_I(inode)->lock);
1724                 }
1725
1726                 /* For sanity tests */
1727                 if (btrfs_test_is_dummy_root(root))
1728                         return;
1729
1730                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1731                                      root->fs_info->delalloc_batch);
1732                 spin_lock(&BTRFS_I(inode)->lock);
1733                 BTRFS_I(inode)->delalloc_bytes += len;
1734                 if (*bits & EXTENT_DEFRAG)
1735                         BTRFS_I(inode)->defrag_bytes += len;
1736                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1737                                          &BTRFS_I(inode)->runtime_flags))
1738                         btrfs_add_delalloc_inodes(root, inode);
1739                 spin_unlock(&BTRFS_I(inode)->lock);
1740         }
1741 }
1742
1743 /*
1744  * extent_io.c clear_bit_hook, see set_bit_hook for why
1745  */
1746 static void btrfs_clear_bit_hook(struct inode *inode,
1747                                  struct extent_state *state,
1748                                  unsigned *bits)
1749 {
1750         u64 len = state->end + 1 - state->start;
1751         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1752                                     BTRFS_MAX_EXTENT_SIZE);
1753
1754         spin_lock(&BTRFS_I(inode)->lock);
1755         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1756                 BTRFS_I(inode)->defrag_bytes -= len;
1757         spin_unlock(&BTRFS_I(inode)->lock);
1758
1759         /*
1760          * set_bit and clear bit hooks normally require _irqsave/restore
1761          * but in this case, we are only testing for the DELALLOC
1762          * bit, which is only set or cleared with irqs on
1763          */
1764         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1765                 struct btrfs_root *root = BTRFS_I(inode)->root;
1766                 bool do_list = !btrfs_is_free_space_inode(inode);
1767
1768                 if (*bits & EXTENT_FIRST_DELALLOC) {
1769                         *bits &= ~EXTENT_FIRST_DELALLOC;
1770                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1771                         spin_lock(&BTRFS_I(inode)->lock);
1772                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1773                         spin_unlock(&BTRFS_I(inode)->lock);
1774                 }
1775
1776                 /*
1777                  * We don't reserve metadata space for space cache inodes so we
1778                  * don't need to call dellalloc_release_metadata if there is an
1779                  * error.
1780                  */
1781                 if (*bits & EXTENT_DO_ACCOUNTING &&
1782                     root != root->fs_info->tree_root)
1783                         btrfs_delalloc_release_metadata(inode, len);
1784
1785                 /* For sanity tests. */
1786                 if (btrfs_test_is_dummy_root(root))
1787                         return;
1788
1789                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1790                     && do_list && !(state->state & EXTENT_NORESERVE))
1791                         btrfs_free_reserved_data_space_noquota(inode,
1792                                         state->start, len);
1793
1794                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1795                                      root->fs_info->delalloc_batch);
1796                 spin_lock(&BTRFS_I(inode)->lock);
1797                 BTRFS_I(inode)->delalloc_bytes -= len;
1798                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1799                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1800                              &BTRFS_I(inode)->runtime_flags))
1801                         btrfs_del_delalloc_inode(root, inode);
1802                 spin_unlock(&BTRFS_I(inode)->lock);
1803         }
1804 }
1805
1806 /*
1807  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1808  * we don't create bios that span stripes or chunks
1809  */
1810 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1811                          size_t size, struct bio *bio,
1812                          unsigned long bio_flags)
1813 {
1814         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1815         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1816         u64 length = 0;
1817         u64 map_length;
1818         int ret;
1819
1820         if (bio_flags & EXTENT_BIO_COMPRESSED)
1821                 return 0;
1822
1823         length = bio->bi_iter.bi_size;
1824         map_length = length;
1825         ret = btrfs_map_block(root->fs_info, rw, logical,
1826                               &map_length, NULL, 0);
1827         /* Will always return 0 with map_multi == NULL */
1828         BUG_ON(ret < 0);
1829         if (map_length < length + size)
1830                 return 1;
1831         return 0;
1832 }
1833
1834 /*
1835  * in order to insert checksums into the metadata in large chunks,
1836  * we wait until bio submission time.   All the pages in the bio are
1837  * checksummed and sums are attached onto the ordered extent record.
1838  *
1839  * At IO completion time the cums attached on the ordered extent record
1840  * are inserted into the btree
1841  */
1842 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1843                                     struct bio *bio, int mirror_num,
1844                                     unsigned long bio_flags,
1845                                     u64 bio_offset)
1846 {
1847         struct btrfs_root *root = BTRFS_I(inode)->root;
1848         int ret = 0;
1849
1850         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1851         BUG_ON(ret); /* -ENOMEM */
1852         return 0;
1853 }
1854
1855 /*
1856  * in order to insert checksums into the metadata in large chunks,
1857  * we wait until bio submission time.   All the pages in the bio are
1858  * checksummed and sums are attached onto the ordered extent record.
1859  *
1860  * At IO completion time the cums attached on the ordered extent record
1861  * are inserted into the btree
1862  */
1863 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1864                           int mirror_num, unsigned long bio_flags,
1865                           u64 bio_offset)
1866 {
1867         struct btrfs_root *root = BTRFS_I(inode)->root;
1868         int ret;
1869
1870         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1871         if (ret) {
1872                 bio->bi_error = ret;
1873                 bio_endio(bio);
1874         }
1875         return ret;
1876 }
1877
1878 /*
1879  * extent_io.c submission hook. This does the right thing for csum calculation
1880  * on write, or reading the csums from the tree before a read
1881  */
1882 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1883                           int mirror_num, unsigned long bio_flags,
1884                           u64 bio_offset)
1885 {
1886         struct btrfs_root *root = BTRFS_I(inode)->root;
1887         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1888         int ret = 0;
1889         int skip_sum;
1890         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1891
1892         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1893
1894         if (btrfs_is_free_space_inode(inode))
1895                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1896
1897         if (!(rw & REQ_WRITE)) {
1898                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1899                 if (ret)
1900                         goto out;
1901
1902                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1903                         ret = btrfs_submit_compressed_read(inode, bio,
1904                                                            mirror_num,
1905                                                            bio_flags);
1906                         goto out;
1907                 } else if (!skip_sum) {
1908                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1909                         if (ret)
1910                                 goto out;
1911                 }
1912                 goto mapit;
1913         } else if (async && !skip_sum) {
1914                 /* csum items have already been cloned */
1915                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1916                         goto mapit;
1917                 /* we're doing a write, do the async checksumming */
1918                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1919                                    inode, rw, bio, mirror_num,
1920                                    bio_flags, bio_offset,
1921                                    __btrfs_submit_bio_start,
1922                                    __btrfs_submit_bio_done);
1923                 goto out;
1924         } else if (!skip_sum) {
1925                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1926                 if (ret)
1927                         goto out;
1928         }
1929
1930 mapit:
1931         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1932
1933 out:
1934         if (ret < 0) {
1935                 bio->bi_error = ret;
1936                 bio_endio(bio);
1937         }
1938         return ret;
1939 }
1940
1941 /*
1942  * given a list of ordered sums record them in the inode.  This happens
1943  * at IO completion time based on sums calculated at bio submission time.
1944  */
1945 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1946                              struct inode *inode, u64 file_offset,
1947                              struct list_head *list)
1948 {
1949         struct btrfs_ordered_sum *sum;
1950
1951         list_for_each_entry(sum, list, list) {
1952                 trans->adding_csums = 1;
1953                 btrfs_csum_file_blocks(trans,
1954                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1955                 trans->adding_csums = 0;
1956         }
1957         return 0;
1958 }
1959
1960 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1961                               struct extent_state **cached_state)
1962 {
1963         WARN_ON((end & (PAGE_CACHE_SIZE - 1)) == 0);
1964         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1965                                    cached_state, GFP_NOFS);
1966 }
1967
1968 /* see btrfs_writepage_start_hook for details on why this is required */
1969 struct btrfs_writepage_fixup {
1970         struct page *page;
1971         struct btrfs_work work;
1972 };
1973
1974 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1975 {
1976         struct btrfs_writepage_fixup *fixup;
1977         struct btrfs_ordered_extent *ordered;
1978         struct extent_state *cached_state = NULL;
1979         struct page *page;
1980         struct inode *inode;
1981         u64 page_start;
1982         u64 page_end;
1983         int ret;
1984
1985         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1986         page = fixup->page;
1987 again:
1988         lock_page(page);
1989         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1990                 ClearPageChecked(page);
1991                 goto out_page;
1992         }
1993
1994         inode = page->mapping->host;
1995         page_start = page_offset(page);
1996         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1997
1998         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
1999                          &cached_state);
2000
2001         /* already ordered? We're done */
2002         if (PagePrivate2(page))
2003                 goto out;
2004
2005         ordered = btrfs_lookup_ordered_extent(inode, page_start);
2006         if (ordered) {
2007                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2008                                      page_end, &cached_state, GFP_NOFS);
2009                 unlock_page(page);
2010                 btrfs_start_ordered_extent(inode, ordered, 1);
2011                 btrfs_put_ordered_extent(ordered);
2012                 goto again;
2013         }
2014
2015         ret = btrfs_delalloc_reserve_space(inode, page_start,
2016                                            PAGE_CACHE_SIZE);
2017         if (ret) {
2018                 mapping_set_error(page->mapping, ret);
2019                 end_extent_writepage(page, ret, page_start, page_end);
2020                 ClearPageChecked(page);
2021                 goto out;
2022          }
2023
2024         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2025         ClearPageChecked(page);
2026         set_page_dirty(page);
2027 out:
2028         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2029                              &cached_state, GFP_NOFS);
2030 out_page:
2031         unlock_page(page);
2032         page_cache_release(page);
2033         kfree(fixup);
2034 }
2035
2036 /*
2037  * There are a few paths in the higher layers of the kernel that directly
2038  * set the page dirty bit without asking the filesystem if it is a
2039  * good idea.  This causes problems because we want to make sure COW
2040  * properly happens and the data=ordered rules are followed.
2041  *
2042  * In our case any range that doesn't have the ORDERED bit set
2043  * hasn't been properly setup for IO.  We kick off an async process
2044  * to fix it up.  The async helper will wait for ordered extents, set
2045  * the delalloc bit and make it safe to write the page.
2046  */
2047 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2048 {
2049         struct inode *inode = page->mapping->host;
2050         struct btrfs_writepage_fixup *fixup;
2051         struct btrfs_root *root = BTRFS_I(inode)->root;
2052
2053         /* this page is properly in the ordered list */
2054         if (TestClearPagePrivate2(page))
2055                 return 0;
2056
2057         if (PageChecked(page))
2058                 return -EAGAIN;
2059
2060         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2061         if (!fixup)
2062                 return -EAGAIN;
2063
2064         SetPageChecked(page);
2065         page_cache_get(page);
2066         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2067                         btrfs_writepage_fixup_worker, NULL, NULL);
2068         fixup->page = page;
2069         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2070         return -EBUSY;
2071 }
2072
2073 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2074                                        struct inode *inode, u64 file_pos,
2075                                        u64 disk_bytenr, u64 disk_num_bytes,
2076                                        u64 num_bytes, u64 ram_bytes,
2077                                        u8 compression, u8 encryption,
2078                                        u16 other_encoding, int extent_type)
2079 {
2080         struct btrfs_root *root = BTRFS_I(inode)->root;
2081         struct btrfs_file_extent_item *fi;
2082         struct btrfs_path *path;
2083         struct extent_buffer *leaf;
2084         struct btrfs_key ins;
2085         int extent_inserted = 0;
2086         int ret;
2087
2088         path = btrfs_alloc_path();
2089         if (!path)
2090                 return -ENOMEM;
2091
2092         /*
2093          * we may be replacing one extent in the tree with another.
2094          * The new extent is pinned in the extent map, and we don't want
2095          * to drop it from the cache until it is completely in the btree.
2096          *
2097          * So, tell btrfs_drop_extents to leave this extent in the cache.
2098          * the caller is expected to unpin it and allow it to be merged
2099          * with the others.
2100          */
2101         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2102                                    file_pos + num_bytes, NULL, 0,
2103                                    1, sizeof(*fi), &extent_inserted);
2104         if (ret)
2105                 goto out;
2106
2107         if (!extent_inserted) {
2108                 ins.objectid = btrfs_ino(inode);
2109                 ins.offset = file_pos;
2110                 ins.type = BTRFS_EXTENT_DATA_KEY;
2111
2112                 path->leave_spinning = 1;
2113                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2114                                               sizeof(*fi));
2115                 if (ret)
2116                         goto out;
2117         }
2118         leaf = path->nodes[0];
2119         fi = btrfs_item_ptr(leaf, path->slots[0],
2120                             struct btrfs_file_extent_item);
2121         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2122         btrfs_set_file_extent_type(leaf, fi, extent_type);
2123         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2124         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2125         btrfs_set_file_extent_offset(leaf, fi, 0);
2126         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2127         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2128         btrfs_set_file_extent_compression(leaf, fi, compression);
2129         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2130         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2131
2132         btrfs_mark_buffer_dirty(leaf);
2133         btrfs_release_path(path);
2134
2135         inode_add_bytes(inode, num_bytes);
2136
2137         ins.objectid = disk_bytenr;
2138         ins.offset = disk_num_bytes;
2139         ins.type = BTRFS_EXTENT_ITEM_KEY;
2140         ret = btrfs_alloc_reserved_file_extent(trans, root,
2141                                         root->root_key.objectid,
2142                                         btrfs_ino(inode), file_pos,
2143                                         ram_bytes, &ins);
2144         /*
2145          * Release the reserved range from inode dirty range map, as it is
2146          * already moved into delayed_ref_head
2147          */
2148         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2149 out:
2150         btrfs_free_path(path);
2151
2152         return ret;
2153 }
2154
2155 /* snapshot-aware defrag */
2156 struct sa_defrag_extent_backref {
2157         struct rb_node node;
2158         struct old_sa_defrag_extent *old;
2159         u64 root_id;
2160         u64 inum;
2161         u64 file_pos;
2162         u64 extent_offset;
2163         u64 num_bytes;
2164         u64 generation;
2165 };
2166
2167 struct old_sa_defrag_extent {
2168         struct list_head list;
2169         struct new_sa_defrag_extent *new;
2170
2171         u64 extent_offset;
2172         u64 bytenr;
2173         u64 offset;
2174         u64 len;
2175         int count;
2176 };
2177
2178 struct new_sa_defrag_extent {
2179         struct rb_root root;
2180         struct list_head head;
2181         struct btrfs_path *path;
2182         struct inode *inode;
2183         u64 file_pos;
2184         u64 len;
2185         u64 bytenr;
2186         u64 disk_len;
2187         u8 compress_type;
2188 };
2189
2190 static int backref_comp(struct sa_defrag_extent_backref *b1,
2191                         struct sa_defrag_extent_backref *b2)
2192 {
2193         if (b1->root_id < b2->root_id)
2194                 return -1;
2195         else if (b1->root_id > b2->root_id)
2196                 return 1;
2197
2198         if (b1->inum < b2->inum)
2199                 return -1;
2200         else if (b1->inum > b2->inum)
2201                 return 1;
2202
2203         if (b1->file_pos < b2->file_pos)
2204                 return -1;
2205         else if (b1->file_pos > b2->file_pos)
2206                 return 1;
2207
2208         /*
2209          * [------------------------------] ===> (a range of space)
2210          *     |<--->|   |<---->| =============> (fs/file tree A)
2211          * |<---------------------------->| ===> (fs/file tree B)
2212          *
2213          * A range of space can refer to two file extents in one tree while
2214          * refer to only one file extent in another tree.
2215          *
2216          * So we may process a disk offset more than one time(two extents in A)
2217          * and locate at the same extent(one extent in B), then insert two same
2218          * backrefs(both refer to the extent in B).
2219          */
2220         return 0;
2221 }
2222
2223 static void backref_insert(struct rb_root *root,
2224                            struct sa_defrag_extent_backref *backref)
2225 {
2226         struct rb_node **p = &root->rb_node;
2227         struct rb_node *parent = NULL;
2228         struct sa_defrag_extent_backref *entry;
2229         int ret;
2230
2231         while (*p) {
2232                 parent = *p;
2233                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2234
2235                 ret = backref_comp(backref, entry);
2236                 if (ret < 0)
2237                         p = &(*p)->rb_left;
2238                 else
2239                         p = &(*p)->rb_right;
2240         }
2241
2242         rb_link_node(&backref->node, parent, p);
2243         rb_insert_color(&backref->node, root);
2244 }
2245
2246 /*
2247  * Note the backref might has changed, and in this case we just return 0.
2248  */
2249 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2250                                        void *ctx)
2251 {
2252         struct btrfs_file_extent_item *extent;
2253         struct btrfs_fs_info *fs_info;
2254         struct old_sa_defrag_extent *old = ctx;
2255         struct new_sa_defrag_extent *new = old->new;
2256         struct btrfs_path *path = new->path;
2257         struct btrfs_key key;
2258         struct btrfs_root *root;
2259         struct sa_defrag_extent_backref *backref;
2260         struct extent_buffer *leaf;
2261         struct inode *inode = new->inode;
2262         int slot;
2263         int ret;
2264         u64 extent_offset;
2265         u64 num_bytes;
2266
2267         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2268             inum == btrfs_ino(inode))
2269                 return 0;
2270
2271         key.objectid = root_id;
2272         key.type = BTRFS_ROOT_ITEM_KEY;
2273         key.offset = (u64)-1;
2274
2275         fs_info = BTRFS_I(inode)->root->fs_info;
2276         root = btrfs_read_fs_root_no_name(fs_info, &key);
2277         if (IS_ERR(root)) {
2278                 if (PTR_ERR(root) == -ENOENT)
2279                         return 0;
2280                 WARN_ON(1);
2281                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2282                          inum, offset, root_id);
2283                 return PTR_ERR(root);
2284         }
2285
2286         key.objectid = inum;
2287         key.type = BTRFS_EXTENT_DATA_KEY;
2288         if (offset > (u64)-1 << 32)
2289                 key.offset = 0;
2290         else
2291                 key.offset = offset;
2292
2293         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2294         if (WARN_ON(ret < 0))
2295                 return ret;
2296         ret = 0;
2297
2298         while (1) {
2299                 cond_resched();
2300
2301                 leaf = path->nodes[0];
2302                 slot = path->slots[0];
2303
2304                 if (slot >= btrfs_header_nritems(leaf)) {
2305                         ret = btrfs_next_leaf(root, path);
2306                         if (ret < 0) {
2307                                 goto out;
2308                         } else if (ret > 0) {
2309                                 ret = 0;
2310                                 goto out;
2311                         }
2312                         continue;
2313                 }
2314
2315                 path->slots[0]++;
2316
2317                 btrfs_item_key_to_cpu(leaf, &key, slot);
2318
2319                 if (key.objectid > inum)
2320                         goto out;
2321
2322                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2323                         continue;
2324
2325                 extent = btrfs_item_ptr(leaf, slot,
2326                                         struct btrfs_file_extent_item);
2327
2328                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2329                         continue;
2330
2331                 /*
2332                  * 'offset' refers to the exact key.offset,
2333                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2334                  * (key.offset - extent_offset).
2335                  */
2336                 if (key.offset != offset)
2337                         continue;
2338
2339                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2340                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2341
2342                 if (extent_offset >= old->extent_offset + old->offset +
2343                     old->len || extent_offset + num_bytes <=
2344                     old->extent_offset + old->offset)
2345                         continue;
2346                 break;
2347         }
2348
2349         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2350         if (!backref) {
2351                 ret = -ENOENT;
2352                 goto out;
2353         }
2354
2355         backref->root_id = root_id;
2356         backref->inum = inum;
2357         backref->file_pos = offset;
2358         backref->num_bytes = num_bytes;
2359         backref->extent_offset = extent_offset;
2360         backref->generation = btrfs_file_extent_generation(leaf, extent);
2361         backref->old = old;
2362         backref_insert(&new->root, backref);
2363         old->count++;
2364 out:
2365         btrfs_release_path(path);
2366         WARN_ON(ret);
2367         return ret;
2368 }
2369
2370 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2371                                    struct new_sa_defrag_extent *new)
2372 {
2373         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2374         struct old_sa_defrag_extent *old, *tmp;
2375         int ret;
2376
2377         new->path = path;
2378
2379         list_for_each_entry_safe(old, tmp, &new->head, list) {
2380                 ret = iterate_inodes_from_logical(old->bytenr +
2381                                                   old->extent_offset, fs_info,
2382                                                   path, record_one_backref,
2383                                                   old);
2384                 if (ret < 0 && ret != -ENOENT)
2385                         return false;
2386
2387                 /* no backref to be processed for this extent */
2388                 if (!old->count) {
2389                         list_del(&old->list);
2390                         kfree(old);
2391                 }
2392         }
2393
2394         if (list_empty(&new->head))
2395                 return false;
2396
2397         return true;
2398 }
2399
2400 static int relink_is_mergable(struct extent_buffer *leaf,
2401                               struct btrfs_file_extent_item *fi,
2402                               struct new_sa_defrag_extent *new)
2403 {
2404         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2405                 return 0;
2406
2407         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2408                 return 0;
2409
2410         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2411                 return 0;
2412
2413         if (btrfs_file_extent_encryption(leaf, fi) ||
2414             btrfs_file_extent_other_encoding(leaf, fi))
2415                 return 0;
2416
2417         return 1;
2418 }
2419
2420 /*
2421  * Note the backref might has changed, and in this case we just return 0.
2422  */
2423 static noinline int relink_extent_backref(struct btrfs_path *path,
2424                                  struct sa_defrag_extent_backref *prev,
2425                                  struct sa_defrag_extent_backref *backref)
2426 {
2427         struct btrfs_file_extent_item *extent;
2428         struct btrfs_file_extent_item *item;
2429         struct btrfs_ordered_extent *ordered;
2430         struct btrfs_trans_handle *trans;
2431         struct btrfs_fs_info *fs_info;
2432         struct btrfs_root *root;
2433         struct btrfs_key key;
2434         struct extent_buffer *leaf;
2435         struct old_sa_defrag_extent *old = backref->old;
2436         struct new_sa_defrag_extent *new = old->new;
2437         struct inode *src_inode = new->inode;
2438         struct inode *inode;
2439         struct extent_state *cached = NULL;
2440         int ret = 0;
2441         u64 start;
2442         u64 len;
2443         u64 lock_start;
2444         u64 lock_end;
2445         bool merge = false;
2446         int index;
2447
2448         if (prev && prev->root_id == backref->root_id &&
2449             prev->inum == backref->inum &&
2450             prev->file_pos + prev->num_bytes == backref->file_pos)
2451                 merge = true;
2452
2453         /* step 1: get root */
2454         key.objectid = backref->root_id;
2455         key.type = BTRFS_ROOT_ITEM_KEY;
2456         key.offset = (u64)-1;
2457
2458         fs_info = BTRFS_I(src_inode)->root->fs_info;
2459         index = srcu_read_lock(&fs_info->subvol_srcu);
2460
2461         root = btrfs_read_fs_root_no_name(fs_info, &key);
2462         if (IS_ERR(root)) {
2463                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2464                 if (PTR_ERR(root) == -ENOENT)
2465                         return 0;
2466                 return PTR_ERR(root);
2467         }
2468
2469         if (btrfs_root_readonly(root)) {
2470                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2471                 return 0;
2472         }
2473
2474         /* step 2: get inode */
2475         key.objectid = backref->inum;
2476         key.type = BTRFS_INODE_ITEM_KEY;
2477         key.offset = 0;
2478
2479         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2480         if (IS_ERR(inode)) {
2481                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2482                 return 0;
2483         }
2484
2485         srcu_read_unlock(&fs_info->subvol_srcu, index);
2486
2487         /* step 3: relink backref */
2488         lock_start = backref->file_pos;
2489         lock_end = backref->file_pos + backref->num_bytes - 1;
2490         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2491                          &cached);
2492
2493         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2494         if (ordered) {
2495                 btrfs_put_ordered_extent(ordered);
2496                 goto out_unlock;
2497         }
2498
2499         trans = btrfs_join_transaction(root);
2500         if (IS_ERR(trans)) {
2501                 ret = PTR_ERR(trans);
2502                 goto out_unlock;
2503         }
2504
2505         key.objectid = backref->inum;
2506         key.type = BTRFS_EXTENT_DATA_KEY;
2507         key.offset = backref->file_pos;
2508
2509         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2510         if (ret < 0) {
2511                 goto out_free_path;
2512         } else if (ret > 0) {
2513                 ret = 0;
2514                 goto out_free_path;
2515         }
2516
2517         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2518                                 struct btrfs_file_extent_item);
2519
2520         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2521             backref->generation)
2522                 goto out_free_path;
2523
2524         btrfs_release_path(path);
2525
2526         start = backref->file_pos;
2527         if (backref->extent_offset < old->extent_offset + old->offset)
2528                 start += old->extent_offset + old->offset -
2529                          backref->extent_offset;
2530
2531         len = min(backref->extent_offset + backref->num_bytes,
2532                   old->extent_offset + old->offset + old->len);
2533         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2534
2535         ret = btrfs_drop_extents(trans, root, inode, start,
2536                                  start + len, 1);
2537         if (ret)
2538                 goto out_free_path;
2539 again:
2540         key.objectid = btrfs_ino(inode);
2541         key.type = BTRFS_EXTENT_DATA_KEY;
2542         key.offset = start;
2543
2544         path->leave_spinning = 1;
2545         if (merge) {
2546                 struct btrfs_file_extent_item *fi;
2547                 u64 extent_len;
2548                 struct btrfs_key found_key;
2549
2550                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2551                 if (ret < 0)
2552                         goto out_free_path;
2553
2554                 path->slots[0]--;
2555                 leaf = path->nodes[0];
2556                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2557
2558                 fi = btrfs_item_ptr(leaf, path->slots[0],
2559                                     struct btrfs_file_extent_item);
2560                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2561
2562                 if (extent_len + found_key.offset == start &&
2563                     relink_is_mergable(leaf, fi, new)) {
2564                         btrfs_set_file_extent_num_bytes(leaf, fi,
2565                                                         extent_len + len);
2566                         btrfs_mark_buffer_dirty(leaf);
2567                         inode_add_bytes(inode, len);
2568
2569                         ret = 1;
2570                         goto out_free_path;
2571                 } else {
2572                         merge = false;
2573                         btrfs_release_path(path);
2574                         goto again;
2575                 }
2576         }
2577
2578         ret = btrfs_insert_empty_item(trans, root, path, &key,
2579                                         sizeof(*extent));
2580         if (ret) {
2581                 btrfs_abort_transaction(trans, root, ret);
2582                 goto out_free_path;
2583         }
2584
2585         leaf = path->nodes[0];
2586         item = btrfs_item_ptr(leaf, path->slots[0],
2587                                 struct btrfs_file_extent_item);
2588         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2589         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2590         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2591         btrfs_set_file_extent_num_bytes(leaf, item, len);
2592         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2593         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2594         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2595         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2596         btrfs_set_file_extent_encryption(leaf, item, 0);
2597         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2598
2599         btrfs_mark_buffer_dirty(leaf);
2600         inode_add_bytes(inode, len);
2601         btrfs_release_path(path);
2602
2603         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2604                         new->disk_len, 0,
2605                         backref->root_id, backref->inum,
2606                         new->file_pos); /* start - extent_offset */
2607         if (ret) {
2608                 btrfs_abort_transaction(trans, root, ret);
2609                 goto out_free_path;
2610         }
2611
2612         ret = 1;
2613 out_free_path:
2614         btrfs_release_path(path);
2615         path->leave_spinning = 0;
2616         btrfs_end_transaction(trans, root);
2617 out_unlock:
2618         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2619                              &cached, GFP_NOFS);
2620         iput(inode);
2621         return ret;
2622 }
2623
2624 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2625 {
2626         struct old_sa_defrag_extent *old, *tmp;
2627
2628         if (!new)
2629                 return;
2630
2631         list_for_each_entry_safe(old, tmp, &new->head, list) {
2632                 kfree(old);
2633         }
2634         kfree(new);
2635 }
2636
2637 static void relink_file_extents(struct new_sa_defrag_extent *new)
2638 {
2639         struct btrfs_path *path;
2640         struct sa_defrag_extent_backref *backref;
2641         struct sa_defrag_extent_backref *prev = NULL;
2642         struct inode *inode;
2643         struct btrfs_root *root;
2644         struct rb_node *node;
2645         int ret;
2646
2647         inode = new->inode;
2648         root = BTRFS_I(inode)->root;
2649
2650         path = btrfs_alloc_path();
2651         if (!path)
2652                 return;
2653
2654         if (!record_extent_backrefs(path, new)) {
2655                 btrfs_free_path(path);
2656                 goto out;
2657         }
2658         btrfs_release_path(path);
2659
2660         while (1) {
2661                 node = rb_first(&new->root);
2662                 if (!node)
2663                         break;
2664                 rb_erase(node, &new->root);
2665
2666                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2667
2668                 ret = relink_extent_backref(path, prev, backref);
2669                 WARN_ON(ret < 0);
2670
2671                 kfree(prev);
2672
2673                 if (ret == 1)
2674                         prev = backref;
2675                 else
2676                         prev = NULL;
2677                 cond_resched();
2678         }
2679         kfree(prev);
2680
2681         btrfs_free_path(path);
2682 out:
2683         free_sa_defrag_extent(new);
2684
2685         atomic_dec(&root->fs_info->defrag_running);
2686         wake_up(&root->fs_info->transaction_wait);
2687 }
2688
2689 static struct new_sa_defrag_extent *
2690 record_old_file_extents(struct inode *inode,
2691                         struct btrfs_ordered_extent *ordered)
2692 {
2693         struct btrfs_root *root = BTRFS_I(inode)->root;
2694         struct btrfs_path *path;
2695         struct btrfs_key key;
2696         struct old_sa_defrag_extent *old;
2697         struct new_sa_defrag_extent *new;
2698         int ret;
2699
2700         new = kmalloc(sizeof(*new), GFP_NOFS);
2701         if (!new)
2702                 return NULL;
2703
2704         new->inode = inode;
2705         new->file_pos = ordered->file_offset;
2706         new->len = ordered->len;
2707         new->bytenr = ordered->start;
2708         new->disk_len = ordered->disk_len;
2709         new->compress_type = ordered->compress_type;
2710         new->root = RB_ROOT;
2711         INIT_LIST_HEAD(&new->head);
2712
2713         path = btrfs_alloc_path();
2714         if (!path)
2715                 goto out_kfree;
2716
2717         key.objectid = btrfs_ino(inode);
2718         key.type = BTRFS_EXTENT_DATA_KEY;
2719         key.offset = new->file_pos;
2720
2721         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2722         if (ret < 0)
2723                 goto out_free_path;
2724         if (ret > 0 && path->slots[0] > 0)
2725                 path->slots[0]--;
2726
2727         /* find out all the old extents for the file range */
2728         while (1) {
2729                 struct btrfs_file_extent_item *extent;
2730                 struct extent_buffer *l;
2731                 int slot;
2732                 u64 num_bytes;
2733                 u64 offset;
2734                 u64 end;
2735                 u64 disk_bytenr;
2736                 u64 extent_offset;
2737
2738                 l = path->nodes[0];
2739                 slot = path->slots[0];
2740
2741                 if (slot >= btrfs_header_nritems(l)) {
2742                         ret = btrfs_next_leaf(root, path);
2743                         if (ret < 0)
2744                                 goto out_free_path;
2745                         else if (ret > 0)
2746                                 break;
2747                         continue;
2748                 }
2749
2750                 btrfs_item_key_to_cpu(l, &key, slot);
2751
2752                 if (key.objectid != btrfs_ino(inode))
2753                         break;
2754                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2755                         break;
2756                 if (key.offset >= new->file_pos + new->len)
2757                         break;
2758
2759                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2760
2761                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2762                 if (key.offset + num_bytes < new->file_pos)
2763                         goto next;
2764
2765                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2766                 if (!disk_bytenr)
2767                         goto next;
2768
2769                 extent_offset = btrfs_file_extent_offset(l, extent);
2770
2771                 old = kmalloc(sizeof(*old), GFP_NOFS);
2772                 if (!old)
2773                         goto out_free_path;
2774
2775                 offset = max(new->file_pos, key.offset);
2776                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2777
2778                 old->bytenr = disk_bytenr;
2779                 old->extent_offset = extent_offset;
2780                 old->offset = offset - key.offset;
2781                 old->len = end - offset;
2782                 old->new = new;
2783                 old->count = 0;
2784                 list_add_tail(&old->list, &new->head);
2785 next:
2786                 path->slots[0]++;
2787                 cond_resched();
2788         }
2789
2790         btrfs_free_path(path);
2791         atomic_inc(&root->fs_info->defrag_running);
2792
2793         return new;
2794
2795 out_free_path:
2796         btrfs_free_path(path);
2797 out_kfree:
2798         free_sa_defrag_extent(new);
2799         return NULL;
2800 }
2801
2802 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2803                                          u64 start, u64 len)
2804 {
2805         struct btrfs_block_group_cache *cache;
2806
2807         cache = btrfs_lookup_block_group(root->fs_info, start);
2808         ASSERT(cache);
2809
2810         spin_lock(&cache->lock);
2811         cache->delalloc_bytes -= len;
2812         spin_unlock(&cache->lock);
2813
2814         btrfs_put_block_group(cache);
2815 }
2816
2817 /* as ordered data IO finishes, this gets called so we can finish
2818  * an ordered extent if the range of bytes in the file it covers are
2819  * fully written.
2820  */
2821 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2822 {
2823         struct inode *inode = ordered_extent->inode;
2824         struct btrfs_root *root = BTRFS_I(inode)->root;
2825         struct btrfs_trans_handle *trans = NULL;
2826         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2827         struct extent_state *cached_state = NULL;
2828         struct new_sa_defrag_extent *new = NULL;
2829         int compress_type = 0;
2830         int ret = 0;
2831         u64 logical_len = ordered_extent->len;
2832         bool nolock;
2833         bool truncated = false;
2834
2835         nolock = btrfs_is_free_space_inode(inode);
2836
2837         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2838                 ret = -EIO;
2839                 goto out;
2840         }
2841
2842         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2843                                      ordered_extent->file_offset +
2844                                      ordered_extent->len - 1);
2845
2846         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2847                 truncated = true;
2848                 logical_len = ordered_extent->truncated_len;
2849                 /* Truncated the entire extent, don't bother adding */
2850                 if (!logical_len)
2851                         goto out;
2852         }
2853
2854         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2855                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2856
2857                 /*
2858                  * For mwrite(mmap + memset to write) case, we still reserve
2859                  * space for NOCOW range.
2860                  * As NOCOW won't cause a new delayed ref, just free the space
2861                  */
2862                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2863                                        ordered_extent->len);
2864                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2865                 if (nolock)
2866                         trans = btrfs_join_transaction_nolock(root);
2867                 else
2868                         trans = btrfs_join_transaction(root);
2869                 if (IS_ERR(trans)) {
2870                         ret = PTR_ERR(trans);
2871                         trans = NULL;
2872                         goto out;
2873                 }
2874                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2875                 ret = btrfs_update_inode_fallback(trans, root, inode);
2876                 if (ret) /* -ENOMEM or corruption */
2877                         btrfs_abort_transaction(trans, root, ret);
2878                 goto out;
2879         }
2880
2881         lock_extent_bits(io_tree, ordered_extent->file_offset,
2882                          ordered_extent->file_offset + ordered_extent->len - 1,
2883                          &cached_state);
2884
2885         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2886                         ordered_extent->file_offset + ordered_extent->len - 1,
2887                         EXTENT_DEFRAG, 1, cached_state);
2888         if (ret) {
2889                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2890                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2891                         /* the inode is shared */
2892                         new = record_old_file_extents(inode, ordered_extent);
2893
2894                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2895                         ordered_extent->file_offset + ordered_extent->len - 1,
2896                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2897         }
2898
2899         if (nolock)
2900                 trans = btrfs_join_transaction_nolock(root);
2901         else
2902                 trans = btrfs_join_transaction(root);
2903         if (IS_ERR(trans)) {
2904                 ret = PTR_ERR(trans);
2905                 trans = NULL;
2906                 goto out_unlock;
2907         }
2908
2909         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2910
2911         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2912                 compress_type = ordered_extent->compress_type;
2913         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2914                 BUG_ON(compress_type);
2915                 ret = btrfs_mark_extent_written(trans, inode,
2916                                                 ordered_extent->file_offset,
2917                                                 ordered_extent->file_offset +
2918                                                 logical_len);
2919         } else {
2920                 BUG_ON(root == root->fs_info->tree_root);
2921                 ret = insert_reserved_file_extent(trans, inode,
2922                                                 ordered_extent->file_offset,
2923                                                 ordered_extent->start,
2924                                                 ordered_extent->disk_len,
2925                                                 logical_len, logical_len,
2926                                                 compress_type, 0, 0,
2927                                                 BTRFS_FILE_EXTENT_REG);
2928                 if (!ret)
2929                         btrfs_release_delalloc_bytes(root,
2930                                                      ordered_extent->start,
2931                                                      ordered_extent->disk_len);
2932         }
2933         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2934                            ordered_extent->file_offset, ordered_extent->len,
2935                            trans->transid);
2936         if (ret < 0) {
2937                 btrfs_abort_transaction(trans, root, ret);
2938                 goto out_unlock;
2939         }
2940
2941         add_pending_csums(trans, inode, ordered_extent->file_offset,
2942                           &ordered_extent->list);
2943
2944         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2945         ret = btrfs_update_inode_fallback(trans, root, inode);
2946         if (ret) { /* -ENOMEM or corruption */
2947                 btrfs_abort_transaction(trans, root, ret);
2948                 goto out_unlock;
2949         }
2950         ret = 0;
2951 out_unlock:
2952         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2953                              ordered_extent->file_offset +
2954                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2955 out:
2956         if (root != root->fs_info->tree_root)
2957                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2958         if (trans)
2959                 btrfs_end_transaction(trans, root);
2960
2961         if (ret || truncated) {
2962                 u64 start, end;
2963
2964                 if (truncated)
2965                         start = ordered_extent->file_offset + logical_len;
2966                 else
2967                         start = ordered_extent->file_offset;
2968                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2969                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2970
2971                 /* Drop the cache for the part of the extent we didn't write. */
2972                 btrfs_drop_extent_cache(inode, start, end, 0);
2973
2974                 /*
2975                  * If the ordered extent had an IOERR or something else went
2976                  * wrong we need to return the space for this ordered extent
2977                  * back to the allocator.  We only free the extent in the
2978                  * truncated case if we didn't write out the extent at all.
2979                  */
2980                 if ((ret || !logical_len) &&
2981                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2982                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
2983                         btrfs_free_reserved_extent(root, ordered_extent->start,
2984                                                    ordered_extent->disk_len, 1);
2985         }
2986
2987
2988         /*
2989          * This needs to be done to make sure anybody waiting knows we are done
2990          * updating everything for this ordered extent.
2991          */
2992         btrfs_remove_ordered_extent(inode, ordered_extent);
2993
2994         /* for snapshot-aware defrag */
2995         if (new) {
2996                 if (ret) {
2997                         free_sa_defrag_extent(new);
2998                         atomic_dec(&root->fs_info->defrag_running);
2999                 } else {
3000                         relink_file_extents(new);
3001                 }
3002         }
3003
3004         /* once for us */
3005         btrfs_put_ordered_extent(ordered_extent);
3006         /* once for the tree */
3007         btrfs_put_ordered_extent(ordered_extent);
3008
3009         return ret;
3010 }
3011
3012 static void finish_ordered_fn(struct btrfs_work *work)
3013 {
3014         struct btrfs_ordered_extent *ordered_extent;
3015         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3016         btrfs_finish_ordered_io(ordered_extent);
3017 }
3018
3019 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3020                                 struct extent_state *state, int uptodate)
3021 {
3022         struct inode *inode = page->mapping->host;
3023         struct btrfs_root *root = BTRFS_I(inode)->root;
3024         struct btrfs_ordered_extent *ordered_extent = NULL;
3025         struct btrfs_workqueue *wq;
3026         btrfs_work_func_t func;
3027
3028         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3029
3030         ClearPagePrivate2(page);
3031         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3032                                             end - start + 1, uptodate))
3033                 return 0;
3034
3035         if (btrfs_is_free_space_inode(inode)) {
3036                 wq = root->fs_info->endio_freespace_worker;
3037                 func = btrfs_freespace_write_helper;
3038         } else {
3039                 wq = root->fs_info->endio_write_workers;
3040                 func = btrfs_endio_write_helper;
3041         }
3042
3043         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3044                         NULL);
3045         btrfs_queue_work(wq, &ordered_extent->work);
3046
3047         return 0;
3048 }
3049
3050 static int __readpage_endio_check(struct inode *inode,
3051                                   struct btrfs_io_bio *io_bio,
3052                                   int icsum, struct page *page,
3053                                   int pgoff, u64 start, size_t len)
3054 {
3055         char *kaddr;
3056         u32 csum_expected;
3057         u32 csum = ~(u32)0;
3058
3059         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3060
3061         kaddr = kmap_atomic(page);
3062         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3063         btrfs_csum_final(csum, (char *)&csum);
3064         if (csum != csum_expected)
3065                 goto zeroit;
3066
3067         kunmap_atomic(kaddr);
3068         return 0;
3069 zeroit:
3070         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3071                 "csum failed ino %llu off %llu csum %u expected csum %u",
3072                            btrfs_ino(inode), start, csum, csum_expected);
3073         memset(kaddr + pgoff, 1, len);
3074         flush_dcache_page(page);
3075         kunmap_atomic(kaddr);
3076         if (csum_expected == 0)
3077                 return 0;
3078         return -EIO;
3079 }
3080
3081 /*
3082  * when reads are done, we need to check csums to verify the data is correct
3083  * if there's a match, we allow the bio to finish.  If not, the code in
3084  * extent_io.c will try to find good copies for us.
3085  */
3086 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3087                                       u64 phy_offset, struct page *page,
3088                                       u64 start, u64 end, int mirror)
3089 {
3090         size_t offset = start - page_offset(page);
3091         struct inode *inode = page->mapping->host;
3092         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3093         struct btrfs_root *root = BTRFS_I(inode)->root;
3094
3095         if (PageChecked(page)) {
3096                 ClearPageChecked(page);
3097                 return 0;
3098         }
3099
3100         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3101                 return 0;
3102
3103         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3104             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3105                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
3106                                   GFP_NOFS);
3107                 return 0;
3108         }
3109
3110         phy_offset >>= inode->i_sb->s_blocksize_bits;
3111         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3112                                       start, (size_t)(end - start + 1));
3113 }
3114
3115 void btrfs_add_delayed_iput(struct inode *inode)
3116 {
3117         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3118         struct btrfs_inode *binode = BTRFS_I(inode);
3119
3120         if (atomic_add_unless(&inode->i_count, -1, 1))
3121                 return;
3122
3123         spin_lock(&fs_info->delayed_iput_lock);
3124         if (binode->delayed_iput_count == 0) {
3125                 ASSERT(list_empty(&binode->delayed_iput));
3126                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3127         } else {
3128                 binode->delayed_iput_count++;
3129         }
3130         spin_unlock(&fs_info->delayed_iput_lock);
3131 }
3132
3133 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3134 {
3135         struct btrfs_fs_info *fs_info = root->fs_info;
3136
3137         spin_lock(&fs_info->delayed_iput_lock);
3138         while (!list_empty(&fs_info->delayed_iputs)) {
3139                 struct btrfs_inode *inode;
3140
3141                 inode = list_first_entry(&fs_info->delayed_iputs,
3142                                 struct btrfs_inode, delayed_iput);
3143                 if (inode->delayed_iput_count) {
3144                         inode->delayed_iput_count--;
3145                         list_move_tail(&inode->delayed_iput,
3146                                         &fs_info->delayed_iputs);
3147                 } else {
3148                         list_del_init(&inode->delayed_iput);
3149                 }
3150                 spin_unlock(&fs_info->delayed_iput_lock);
3151                 iput(&inode->vfs_inode);
3152                 spin_lock(&fs_info->delayed_iput_lock);
3153         }
3154         spin_unlock(&fs_info->delayed_iput_lock);
3155 }
3156
3157 /*
3158  * This is called in transaction commit time. If there are no orphan
3159  * files in the subvolume, it removes orphan item and frees block_rsv
3160  * structure.
3161  */
3162 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3163                               struct btrfs_root *root)
3164 {
3165         struct btrfs_block_rsv *block_rsv;
3166         int ret;
3167
3168         if (atomic_read(&root->orphan_inodes) ||
3169             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3170                 return;
3171
3172         spin_lock(&root->orphan_lock);
3173         if (atomic_read(&root->orphan_inodes)) {
3174                 spin_unlock(&root->orphan_lock);
3175                 return;
3176         }
3177
3178         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3179                 spin_unlock(&root->orphan_lock);
3180                 return;
3181         }
3182
3183         block_rsv = root->orphan_block_rsv;
3184         root->orphan_block_rsv = NULL;
3185         spin_unlock(&root->orphan_lock);
3186
3187         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3188             btrfs_root_refs(&root->root_item) > 0) {
3189                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3190                                             root->root_key.objectid);
3191                 if (ret)
3192                         btrfs_abort_transaction(trans, root, ret);
3193                 else
3194                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3195                                   &root->state);
3196         }
3197
3198         if (block_rsv) {
3199                 WARN_ON(block_rsv->size > 0);
3200                 btrfs_free_block_rsv(root, block_rsv);
3201         }
3202 }
3203
3204 /*
3205  * This creates an orphan entry for the given inode in case something goes
3206  * wrong in the middle of an unlink/truncate.
3207  *
3208  * NOTE: caller of this function should reserve 5 units of metadata for
3209  *       this function.
3210  */
3211 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3212 {
3213         struct btrfs_root *root = BTRFS_I(inode)->root;
3214         struct btrfs_block_rsv *block_rsv = NULL;
3215         int reserve = 0;
3216         int insert = 0;
3217         int ret;
3218
3219         if (!root->orphan_block_rsv) {
3220                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3221                 if (!block_rsv)
3222                         return -ENOMEM;
3223         }
3224
3225         spin_lock(&root->orphan_lock);
3226         if (!root->orphan_block_rsv) {
3227                 root->orphan_block_rsv = block_rsv;
3228         } else if (block_rsv) {
3229                 btrfs_free_block_rsv(root, block_rsv);
3230                 block_rsv = NULL;
3231         }
3232
3233         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3234                               &BTRFS_I(inode)->runtime_flags)) {
3235 #if 0
3236                 /*
3237                  * For proper ENOSPC handling, we should do orphan
3238                  * cleanup when mounting. But this introduces backward
3239                  * compatibility issue.
3240                  */
3241                 if (!xchg(&root->orphan_item_inserted, 1))
3242                         insert = 2;
3243                 else
3244                         insert = 1;
3245 #endif
3246                 insert = 1;
3247                 atomic_inc(&root->orphan_inodes);
3248         }
3249
3250         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3251                               &BTRFS_I(inode)->runtime_flags))
3252                 reserve = 1;
3253         spin_unlock(&root->orphan_lock);
3254
3255         /* grab metadata reservation from transaction handle */
3256         if (reserve) {
3257                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3258                 BUG_ON(ret); /* -ENOSPC in reservation; Logic error? JDM */
3259         }
3260
3261         /* insert an orphan item to track this unlinked/truncated file */
3262         if (insert >= 1) {
3263                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3264                 if (ret) {
3265                         atomic_dec(&root->orphan_inodes);
3266                         if (reserve) {
3267                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3268                                           &BTRFS_I(inode)->runtime_flags);
3269                                 btrfs_orphan_release_metadata(inode);
3270                         }
3271                         if (ret != -EEXIST) {
3272                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3273                                           &BTRFS_I(inode)->runtime_flags);
3274                                 btrfs_abort_transaction(trans, root, ret);
3275                                 return ret;
3276                         }
3277                 }
3278                 ret = 0;
3279         }
3280
3281         /* insert an orphan item to track subvolume contains orphan files */
3282         if (insert >= 2) {
3283                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3284                                                root->root_key.objectid);
3285                 if (ret && ret != -EEXIST) {
3286                         btrfs_abort_transaction(trans, root, ret);
3287                         return ret;
3288                 }
3289         }
3290         return 0;
3291 }
3292
3293 /*
3294  * We have done the truncate/delete so we can go ahead and remove the orphan
3295  * item for this particular inode.
3296  */
3297 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3298                             struct inode *inode)
3299 {
3300         struct btrfs_root *root = BTRFS_I(inode)->root;
3301         int delete_item = 0;
3302         int release_rsv = 0;
3303         int ret = 0;
3304
3305         spin_lock(&root->orphan_lock);
3306         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3307                                &BTRFS_I(inode)->runtime_flags))
3308                 delete_item = 1;
3309
3310         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3311                                &BTRFS_I(inode)->runtime_flags))
3312                 release_rsv = 1;
3313         spin_unlock(&root->orphan_lock);
3314
3315         if (delete_item) {
3316                 atomic_dec(&root->orphan_inodes);
3317                 if (trans)
3318                         ret = btrfs_del_orphan_item(trans, root,
3319                                                     btrfs_ino(inode));
3320         }
3321
3322         if (release_rsv)
3323                 btrfs_orphan_release_metadata(inode);
3324
3325         return ret;
3326 }
3327
3328 /*
3329  * this cleans up any orphans that may be left on the list from the last use
3330  * of this root.
3331  */
3332 int btrfs_orphan_cleanup(struct btrfs_root *root)
3333 {
3334         struct btrfs_path *path;
3335         struct extent_buffer *leaf;
3336         struct btrfs_key key, found_key;
3337         struct btrfs_trans_handle *trans;
3338         struct inode *inode;
3339         u64 last_objectid = 0;
3340         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3341
3342         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3343                 return 0;
3344
3345         path = btrfs_alloc_path();
3346         if (!path) {
3347                 ret = -ENOMEM;
3348                 goto out;
3349         }
3350         path->reada = READA_BACK;
3351
3352         key.objectid = BTRFS_ORPHAN_OBJECTID;
3353         key.type = BTRFS_ORPHAN_ITEM_KEY;
3354         key.offset = (u64)-1;
3355
3356         while (1) {
3357                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3358                 if (ret < 0)
3359                         goto out;
3360
3361                 /*
3362                  * if ret == 0 means we found what we were searching for, which
3363                  * is weird, but possible, so only screw with path if we didn't
3364                  * find the key and see if we have stuff that matches
3365                  */
3366                 if (ret > 0) {
3367                         ret = 0;
3368                         if (path->slots[0] == 0)
3369                                 break;
3370                         path->slots[0]--;
3371                 }
3372
3373                 /* pull out the item */
3374                 leaf = path->nodes[0];
3375                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3376
3377                 /* make sure the item matches what we want */
3378                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3379                         break;
3380                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3381                         break;
3382
3383                 /* release the path since we're done with it */
3384                 btrfs_release_path(path);
3385
3386                 /*
3387                  * this is where we are basically btrfs_lookup, without the
3388                  * crossing root thing.  we store the inode number in the
3389                  * offset of the orphan item.
3390                  */
3391
3392                 if (found_key.offset == last_objectid) {
3393                         btrfs_err(root->fs_info,
3394                                 "Error removing orphan entry, stopping orphan cleanup");
3395                         ret = -EINVAL;
3396                         goto out;
3397                 }
3398
3399                 last_objectid = found_key.offset;
3400
3401                 found_key.objectid = found_key.offset;
3402                 found_key.type = BTRFS_INODE_ITEM_KEY;
3403                 found_key.offset = 0;
3404                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3405                 ret = PTR_ERR_OR_ZERO(inode);
3406                 if (ret && ret != -ESTALE)
3407                         goto out;
3408
3409                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3410                         struct btrfs_root *dead_root;
3411                         struct btrfs_fs_info *fs_info = root->fs_info;
3412                         int is_dead_root = 0;
3413
3414                         /*
3415                          * this is an orphan in the tree root. Currently these
3416                          * could come from 2 sources:
3417                          *  a) a snapshot deletion in progress
3418                          *  b) a free space cache inode
3419                          * We need to distinguish those two, as the snapshot
3420                          * orphan must not get deleted.
3421                          * find_dead_roots already ran before us, so if this
3422                          * is a snapshot deletion, we should find the root
3423                          * in the dead_roots list
3424                          */
3425                         spin_lock(&fs_info->trans_lock);
3426                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3427                                             root_list) {
3428                                 if (dead_root->root_key.objectid ==
3429                                     found_key.objectid) {
3430                                         is_dead_root = 1;
3431                                         break;
3432                                 }
3433                         }
3434                         spin_unlock(&fs_info->trans_lock);
3435                         if (is_dead_root) {
3436                                 /* prevent this orphan from being found again */
3437                                 key.offset = found_key.objectid - 1;
3438                                 continue;
3439                         }
3440                 }
3441                 /*
3442                  * Inode is already gone but the orphan item is still there,
3443                  * kill the orphan item.
3444                  */
3445                 if (ret == -ESTALE) {
3446                         trans = btrfs_start_transaction(root, 1);
3447                         if (IS_ERR(trans)) {
3448                                 ret = PTR_ERR(trans);
3449                                 goto out;
3450                         }
3451                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3452                                 found_key.objectid);
3453                         ret = btrfs_del_orphan_item(trans, root,
3454                                                     found_key.objectid);
3455                         btrfs_end_transaction(trans, root);
3456                         if (ret)
3457                                 goto out;
3458                         continue;
3459                 }
3460
3461                 /*
3462                  * add this inode to the orphan list so btrfs_orphan_del does
3463                  * the proper thing when we hit it
3464                  */
3465                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3466                         &BTRFS_I(inode)->runtime_flags);
3467                 atomic_inc(&root->orphan_inodes);
3468
3469                 /* if we have links, this was a truncate, lets do that */
3470                 if (inode->i_nlink) {
3471                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3472                                 iput(inode);
3473                                 continue;
3474                         }
3475                         nr_truncate++;
3476
3477                         /* 1 for the orphan item deletion. */
3478                         trans = btrfs_start_transaction(root, 1);
3479                         if (IS_ERR(trans)) {
3480                                 iput(inode);
3481                                 ret = PTR_ERR(trans);
3482                                 goto out;
3483                         }
3484                         ret = btrfs_orphan_add(trans, inode);
3485                         btrfs_end_transaction(trans, root);
3486                         if (ret) {
3487                                 iput(inode);
3488                                 goto out;
3489                         }
3490
3491                         ret = btrfs_truncate(inode);
3492                         if (ret)
3493                                 btrfs_orphan_del(NULL, inode);
3494                 } else {
3495                         nr_unlink++;
3496                 }
3497
3498                 /* this will do delete_inode and everything for us */
3499                 iput(inode);
3500                 if (ret)
3501                         goto out;
3502         }
3503         /* release the path since we're done with it */
3504         btrfs_release_path(path);
3505
3506         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3507
3508         if (root->orphan_block_rsv)
3509                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3510                                         (u64)-1);
3511
3512         if (root->orphan_block_rsv ||
3513             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3514                 trans = btrfs_join_transaction(root);
3515                 if (!IS_ERR(trans))
3516                         btrfs_end_transaction(trans, root);
3517         }
3518
3519         if (nr_unlink)
3520                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3521         if (nr_truncate)
3522                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3523
3524 out:
3525         if (ret)
3526                 btrfs_err(root->fs_info,
3527                         "could not do orphan cleanup %d", ret);
3528         btrfs_free_path(path);
3529         return ret;
3530 }
3531
3532 /*
3533  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3534  * don't find any xattrs, we know there can't be any acls.
3535  *
3536  * slot is the slot the inode is in, objectid is the objectid of the inode
3537  */
3538 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3539                                           int slot, u64 objectid,
3540                                           int *first_xattr_slot)
3541 {
3542         u32 nritems = btrfs_header_nritems(leaf);
3543         struct btrfs_key found_key;
3544         static u64 xattr_access = 0;
3545         static u64 xattr_default = 0;
3546         int scanned = 0;
3547
3548         if (!xattr_access) {
3549                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3550                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3551                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3552                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3553         }
3554
3555         slot++;
3556         *first_xattr_slot = -1;
3557         while (slot < nritems) {
3558                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3559
3560                 /* we found a different objectid, there must not be acls */
3561                 if (found_key.objectid != objectid)
3562                         return 0;
3563
3564                 /* we found an xattr, assume we've got an acl */
3565                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3566                         if (*first_xattr_slot == -1)
3567                                 *first_xattr_slot = slot;
3568                         if (found_key.offset == xattr_access ||
3569                             found_key.offset == xattr_default)
3570                                 return 1;
3571                 }
3572
3573                 /*
3574                  * we found a key greater than an xattr key, there can't
3575                  * be any acls later on
3576                  */
3577                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3578                         return 0;
3579
3580                 slot++;
3581                 scanned++;
3582
3583                 /*
3584                  * it goes inode, inode backrefs, xattrs, extents,
3585                  * so if there are a ton of hard links to an inode there can
3586                  * be a lot of backrefs.  Don't waste time searching too hard,
3587                  * this is just an optimization
3588                  */
3589                 if (scanned >= 8)
3590                         break;
3591         }
3592         /* we hit the end of the leaf before we found an xattr or
3593          * something larger than an xattr.  We have to assume the inode
3594          * has acls
3595          */
3596         if (*first_xattr_slot == -1)
3597                 *first_xattr_slot = slot;
3598         return 1;
3599 }
3600
3601 /*
3602  * read an inode from the btree into the in-memory inode
3603  */
3604 static void btrfs_read_locked_inode(struct inode *inode)
3605 {
3606         struct btrfs_path *path;
3607         struct extent_buffer *leaf;
3608         struct btrfs_inode_item *inode_item;
3609         struct btrfs_root *root = BTRFS_I(inode)->root;
3610         struct btrfs_key location;
3611         unsigned long ptr;
3612         int maybe_acls;
3613         u32 rdev;
3614         int ret;
3615         bool filled = false;
3616         int first_xattr_slot;
3617
3618         ret = btrfs_fill_inode(inode, &rdev);
3619         if (!ret)
3620                 filled = true;
3621
3622         path = btrfs_alloc_path();
3623         if (!path)
3624                 goto make_bad;
3625
3626         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3627
3628         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3629         if (ret)
3630                 goto make_bad;
3631
3632         leaf = path->nodes[0];
3633
3634         if (filled)
3635                 goto cache_index;
3636
3637         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3638                                     struct btrfs_inode_item);
3639         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3640         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3641         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3642         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3643         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3644
3645         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3646         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3647
3648         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3649         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3650
3651         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3652         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3653
3654         BTRFS_I(inode)->i_otime.tv_sec =
3655                 btrfs_timespec_sec(leaf, &inode_item->otime);
3656         BTRFS_I(inode)->i_otime.tv_nsec =
3657                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3658
3659         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3660         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3661         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3662
3663         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3664         inode->i_generation = BTRFS_I(inode)->generation;
3665         inode->i_rdev = 0;
3666         rdev = btrfs_inode_rdev(leaf, inode_item);
3667
3668         BTRFS_I(inode)->index_cnt = (u64)-1;
3669         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3670
3671 cache_index:
3672         /*
3673          * If we were modified in the current generation and evicted from memory
3674          * and then re-read we need to do a full sync since we don't have any
3675          * idea about which extents were modified before we were evicted from
3676          * cache.
3677          *
3678          * This is required for both inode re-read from disk and delayed inode
3679          * in delayed_nodes_tree.
3680          */
3681         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3682                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3683                         &BTRFS_I(inode)->runtime_flags);
3684
3685         /*
3686          * We don't persist the id of the transaction where an unlink operation
3687          * against the inode was last made. So here we assume the inode might
3688          * have been evicted, and therefore the exact value of last_unlink_trans
3689          * lost, and set it to last_trans to avoid metadata inconsistencies
3690          * between the inode and its parent if the inode is fsync'ed and the log
3691          * replayed. For example, in the scenario:
3692          *
3693          * touch mydir/foo
3694          * ln mydir/foo mydir/bar
3695          * sync
3696          * unlink mydir/bar
3697          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3698          * xfs_io -c fsync mydir/foo
3699          * <power failure>
3700          * mount fs, triggers fsync log replay
3701          *
3702          * We must make sure that when we fsync our inode foo we also log its
3703          * parent inode, otherwise after log replay the parent still has the
3704          * dentry with the "bar" name but our inode foo has a link count of 1
3705          * and doesn't have an inode ref with the name "bar" anymore.
3706          *
3707          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3708          * but it guarantees correctness at the expense of ocassional full
3709          * transaction commits on fsync if our inode is a directory, or if our
3710          * inode is not a directory, logging its parent unnecessarily.
3711          */
3712         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3713
3714         path->slots[0]++;
3715         if (inode->i_nlink != 1 ||
3716             path->slots[0] >= btrfs_header_nritems(leaf))
3717                 goto cache_acl;
3718
3719         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3720         if (location.objectid != btrfs_ino(inode))
3721                 goto cache_acl;
3722
3723         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3724         if (location.type == BTRFS_INODE_REF_KEY) {
3725                 struct btrfs_inode_ref *ref;
3726
3727                 ref = (struct btrfs_inode_ref *)ptr;
3728                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3729         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3730                 struct btrfs_inode_extref *extref;
3731
3732                 extref = (struct btrfs_inode_extref *)ptr;
3733                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3734                                                                      extref);
3735         }
3736 cache_acl:
3737         /*
3738          * try to precache a NULL acl entry for files that don't have
3739          * any xattrs or acls
3740          */
3741         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3742                                            btrfs_ino(inode), &first_xattr_slot);
3743         if (first_xattr_slot != -1) {
3744                 path->slots[0] = first_xattr_slot;
3745                 ret = btrfs_load_inode_props(inode, path);
3746                 if (ret)
3747                         btrfs_err(root->fs_info,
3748                                   "error loading props for ino %llu (root %llu): %d",
3749                                   btrfs_ino(inode),
3750                                   root->root_key.objectid, ret);
3751         }
3752         btrfs_free_path(path);
3753
3754         if (!maybe_acls)
3755                 cache_no_acl(inode);
3756
3757         switch (inode->i_mode & S_IFMT) {
3758         case S_IFREG:
3759                 inode->i_mapping->a_ops = &btrfs_aops;
3760                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3761                 inode->i_fop = &btrfs_file_operations;
3762                 inode->i_op = &btrfs_file_inode_operations;
3763                 break;
3764         case S_IFDIR:
3765                 inode->i_fop = &btrfs_dir_file_operations;
3766                 if (root == root->fs_info->tree_root)
3767                         inode->i_op = &btrfs_dir_ro_inode_operations;
3768                 else
3769                         inode->i_op = &btrfs_dir_inode_operations;
3770                 break;
3771         case S_IFLNK:
3772                 inode->i_op = &btrfs_symlink_inode_operations;
3773                 inode_nohighmem(inode);
3774                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3775                 break;
3776         default:
3777                 inode->i_op = &btrfs_special_inode_operations;
3778                 init_special_inode(inode, inode->i_mode, rdev);
3779                 break;
3780         }
3781
3782         btrfs_update_iflags(inode);
3783         return;
3784
3785 make_bad:
3786         btrfs_free_path(path);
3787         make_bad_inode(inode);
3788 }
3789
3790 /*
3791  * given a leaf and an inode, copy the inode fields into the leaf
3792  */
3793 static void fill_inode_item(struct btrfs_trans_handle *trans,
3794                             struct extent_buffer *leaf,
3795                             struct btrfs_inode_item *item,
3796                             struct inode *inode)
3797 {
3798         struct btrfs_map_token token;
3799
3800         btrfs_init_map_token(&token);
3801
3802         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3803         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3804         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3805                                    &token);
3806         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3807         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3808
3809         btrfs_set_token_timespec_sec(leaf, &item->atime,
3810                                      inode->i_atime.tv_sec, &token);
3811         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3812                                       inode->i_atime.tv_nsec, &token);
3813
3814         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3815                                      inode->i_mtime.tv_sec, &token);
3816         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3817                                       inode->i_mtime.tv_nsec, &token);
3818
3819         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3820                                      inode->i_ctime.tv_sec, &token);
3821         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3822                                       inode->i_ctime.tv_nsec, &token);
3823
3824         btrfs_set_token_timespec_sec(leaf, &item->otime,
3825                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3826         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3827                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3828
3829         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3830                                      &token);
3831         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3832                                          &token);
3833         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3834         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3835         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3836         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3837         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3838 }
3839
3840 /*
3841  * copy everything in the in-memory inode into the btree.
3842  */
3843 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3844                                 struct btrfs_root *root, struct inode *inode)
3845 {
3846         struct btrfs_inode_item *inode_item;
3847         struct btrfs_path *path;
3848         struct extent_buffer *leaf;
3849         int ret;
3850
3851         path = btrfs_alloc_path();
3852         if (!path)
3853                 return -ENOMEM;
3854
3855         path->leave_spinning = 1;
3856         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3857                                  1);
3858         if (ret) {
3859                 if (ret > 0)
3860                         ret = -ENOENT;
3861                 goto failed;
3862         }
3863
3864         leaf = path->nodes[0];
3865         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3866                                     struct btrfs_inode_item);
3867
3868         fill_inode_item(trans, leaf, inode_item, inode);
3869         btrfs_mark_buffer_dirty(leaf);
3870         btrfs_set_inode_last_trans(trans, inode);
3871         ret = 0;
3872 failed:
3873         btrfs_free_path(path);
3874         return ret;
3875 }
3876
3877 /*
3878  * copy everything in the in-memory inode into the btree.
3879  */
3880 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3881                                 struct btrfs_root *root, struct inode *inode)
3882 {
3883         int ret;
3884
3885         /*
3886          * If the inode is a free space inode, we can deadlock during commit
3887          * if we put it into the delayed code.
3888          *
3889          * The data relocation inode should also be directly updated
3890          * without delay
3891          */
3892         if (!btrfs_is_free_space_inode(inode)
3893             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3894             && !root->fs_info->log_root_recovering) {
3895                 btrfs_update_root_times(trans, root);
3896
3897                 ret = btrfs_delayed_update_inode(trans, root, inode);
3898                 if (!ret)
3899                         btrfs_set_inode_last_trans(trans, inode);
3900                 return ret;
3901         }
3902
3903         return btrfs_update_inode_item(trans, root, inode);
3904 }
3905
3906 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3907                                          struct btrfs_root *root,
3908                                          struct inode *inode)
3909 {
3910         int ret;
3911
3912         ret = btrfs_update_inode(trans, root, inode);
3913         if (ret == -ENOSPC)
3914                 return btrfs_update_inode_item(trans, root, inode);
3915         return ret;
3916 }
3917
3918 /*
3919  * unlink helper that gets used here in inode.c and in the tree logging
3920  * recovery code.  It remove a link in a directory with a given name, and
3921  * also drops the back refs in the inode to the directory
3922  */
3923 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3924                                 struct btrfs_root *root,
3925                                 struct inode *dir, struct inode *inode,
3926                                 const char *name, int name_len)
3927 {
3928         struct btrfs_path *path;
3929         int ret = 0;
3930         struct extent_buffer *leaf;
3931         struct btrfs_dir_item *di;
3932         struct btrfs_key key;
3933         u64 index;
3934         u64 ino = btrfs_ino(inode);
3935         u64 dir_ino = btrfs_ino(dir);
3936
3937         path = btrfs_alloc_path();
3938         if (!path) {
3939                 ret = -ENOMEM;
3940                 goto out;
3941         }
3942
3943         path->leave_spinning = 1;
3944         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3945                                     name, name_len, -1);
3946         if (IS_ERR(di)) {
3947                 ret = PTR_ERR(di);
3948                 goto err;
3949         }
3950         if (!di) {
3951                 ret = -ENOENT;
3952                 goto err;
3953         }
3954         leaf = path->nodes[0];
3955         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3956         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3957         if (ret)
3958                 goto err;
3959         btrfs_release_path(path);
3960
3961         /*
3962          * If we don't have dir index, we have to get it by looking up
3963          * the inode ref, since we get the inode ref, remove it directly,
3964          * it is unnecessary to do delayed deletion.
3965          *
3966          * But if we have dir index, needn't search inode ref to get it.
3967          * Since the inode ref is close to the inode item, it is better
3968          * that we delay to delete it, and just do this deletion when
3969          * we update the inode item.
3970          */
3971         if (BTRFS_I(inode)->dir_index) {
3972                 ret = btrfs_delayed_delete_inode_ref(inode);
3973                 if (!ret) {
3974                         index = BTRFS_I(inode)->dir_index;
3975                         goto skip_backref;
3976                 }
3977         }
3978
3979         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3980                                   dir_ino, &index);
3981         if (ret) {
3982                 btrfs_info(root->fs_info,
3983                         "failed to delete reference to %.*s, inode %llu parent %llu",
3984                         name_len, name, ino, dir_ino);
3985                 btrfs_abort_transaction(trans, root, ret);
3986                 goto err;
3987         }
3988 skip_backref:
3989         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3990         if (ret) {
3991                 btrfs_abort_transaction(trans, root, ret);
3992                 goto err;
3993         }
3994
3995         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
3996                                          inode, dir_ino);
3997         if (ret != 0 && ret != -ENOENT) {
3998                 btrfs_abort_transaction(trans, root, ret);
3999                 goto err;
4000         }
4001
4002         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4003                                            dir, index);
4004         if (ret == -ENOENT)
4005                 ret = 0;
4006         else if (ret)
4007                 btrfs_abort_transaction(trans, root, ret);
4008 err:
4009         btrfs_free_path(path);
4010         if (ret)
4011                 goto out;
4012
4013         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4014         inode_inc_iversion(inode);
4015         inode_inc_iversion(dir);
4016         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4017         ret = btrfs_update_inode(trans, root, dir);
4018 out:
4019         return ret;
4020 }
4021
4022 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4023                        struct btrfs_root *root,
4024                        struct inode *dir, struct inode *inode,
4025                        const char *name, int name_len)
4026 {
4027         int ret;
4028         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4029         if (!ret) {
4030                 drop_nlink(inode);
4031                 ret = btrfs_update_inode(trans, root, inode);
4032         }
4033         return ret;
4034 }
4035
4036 /*
4037  * helper to start transaction for unlink and rmdir.
4038  *
4039  * unlink and rmdir are special in btrfs, they do not always free space, so
4040  * if we cannot make our reservations the normal way try and see if there is
4041  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4042  * allow the unlink to occur.
4043  */
4044 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4045 {
4046         struct btrfs_root *root = BTRFS_I(dir)->root;
4047
4048         /*
4049          * 1 for the possible orphan item
4050          * 1 for the dir item
4051          * 1 for the dir index
4052          * 1 for the inode ref
4053          * 1 for the inode
4054          */
4055         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4056 }
4057
4058 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4059 {
4060         struct btrfs_root *root = BTRFS_I(dir)->root;
4061         struct btrfs_trans_handle *trans;
4062         struct inode *inode = d_inode(dentry);
4063         int ret;
4064
4065         trans = __unlink_start_trans(dir);
4066         if (IS_ERR(trans))
4067                 return PTR_ERR(trans);
4068
4069         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4070
4071         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4072                                  dentry->d_name.name, dentry->d_name.len);
4073         if (ret)
4074                 goto out;
4075
4076         if (inode->i_nlink == 0) {
4077                 ret = btrfs_orphan_add(trans, inode);
4078                 if (ret)
4079                         goto out;
4080         }
4081
4082 out:
4083         btrfs_end_transaction(trans, root);
4084         btrfs_btree_balance_dirty(root);
4085         return ret;
4086 }
4087
4088 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4089                         struct btrfs_root *root,
4090                         struct inode *dir, u64 objectid,
4091                         const char *name, int name_len)
4092 {
4093         struct btrfs_path *path;
4094         struct extent_buffer *leaf;
4095         struct btrfs_dir_item *di;
4096         struct btrfs_key key;
4097         u64 index;
4098         int ret;
4099         u64 dir_ino = btrfs_ino(dir);
4100
4101         path = btrfs_alloc_path();
4102         if (!path)
4103                 return -ENOMEM;
4104
4105         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4106                                    name, name_len, -1);
4107         if (IS_ERR_OR_NULL(di)) {
4108                 if (!di)
4109                         ret = -ENOENT;
4110                 else
4111                         ret = PTR_ERR(di);
4112                 goto out;
4113         }
4114
4115         leaf = path->nodes[0];
4116         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4117         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4118         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4119         if (ret) {
4120                 btrfs_abort_transaction(trans, root, ret);
4121                 goto out;
4122         }
4123         btrfs_release_path(path);
4124
4125         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4126                                  objectid, root->root_key.objectid,
4127                                  dir_ino, &index, name, name_len);
4128         if (ret < 0) {
4129                 if (ret != -ENOENT) {
4130                         btrfs_abort_transaction(trans, root, ret);
4131                         goto out;
4132                 }
4133                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4134                                                  name, name_len);
4135                 if (IS_ERR_OR_NULL(di)) {
4136                         if (!di)
4137                                 ret = -ENOENT;
4138                         else
4139                                 ret = PTR_ERR(di);
4140                         btrfs_abort_transaction(trans, root, ret);
4141                         goto out;
4142                 }
4143
4144                 leaf = path->nodes[0];
4145                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4146                 btrfs_release_path(path);
4147                 index = key.offset;
4148         }
4149         btrfs_release_path(path);
4150
4151         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4152         if (ret) {
4153                 btrfs_abort_transaction(trans, root, ret);
4154                 goto out;
4155         }
4156
4157         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4158         inode_inc_iversion(dir);
4159         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4160         ret = btrfs_update_inode_fallback(trans, root, dir);
4161         if (ret)
4162                 btrfs_abort_transaction(trans, root, ret);
4163 out:
4164         btrfs_free_path(path);
4165         return ret;
4166 }
4167
4168 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4169 {
4170         struct inode *inode = d_inode(dentry);
4171         int err = 0;
4172         struct btrfs_root *root = BTRFS_I(dir)->root;
4173         struct btrfs_trans_handle *trans;
4174
4175         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4176                 return -ENOTEMPTY;
4177         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4178                 return -EPERM;
4179
4180         trans = __unlink_start_trans(dir);
4181         if (IS_ERR(trans))
4182                 return PTR_ERR(trans);
4183
4184         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4185                 err = btrfs_unlink_subvol(trans, root, dir,
4186                                           BTRFS_I(inode)->location.objectid,
4187                                           dentry->d_name.name,
4188                                           dentry->d_name.len);
4189                 goto out;
4190         }
4191
4192         err = btrfs_orphan_add(trans, inode);
4193         if (err)
4194                 goto out;
4195
4196         /* now the directory is empty */
4197         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4198                                  dentry->d_name.name, dentry->d_name.len);
4199         if (!err)
4200                 btrfs_i_size_write(inode, 0);
4201 out:
4202         btrfs_end_transaction(trans, root);
4203         btrfs_btree_balance_dirty(root);
4204
4205         return err;
4206 }
4207
4208 static int truncate_space_check(struct btrfs_trans_handle *trans,
4209                                 struct btrfs_root *root,
4210                                 u64 bytes_deleted)
4211 {
4212         int ret;
4213
4214         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4215         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4216                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4217         if (!ret)
4218                 trans->bytes_reserved += bytes_deleted;
4219         return ret;
4220
4221 }
4222
4223 static int truncate_inline_extent(struct inode *inode,
4224                                   struct btrfs_path *path,
4225                                   struct btrfs_key *found_key,
4226                                   const u64 item_end,
4227                                   const u64 new_size)
4228 {
4229         struct extent_buffer *leaf = path->nodes[0];
4230         int slot = path->slots[0];
4231         struct btrfs_file_extent_item *fi;
4232         u32 size = (u32)(new_size - found_key->offset);
4233         struct btrfs_root *root = BTRFS_I(inode)->root;
4234
4235         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4236
4237         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4238                 loff_t offset = new_size;
4239                 loff_t page_end = ALIGN(offset, PAGE_CACHE_SIZE);
4240
4241                 /*
4242                  * Zero out the remaining of the last page of our inline extent,
4243                  * instead of directly truncating our inline extent here - that
4244                  * would be much more complex (decompressing all the data, then
4245                  * compressing the truncated data, which might be bigger than
4246                  * the size of the inline extent, resize the extent, etc).
4247                  * We release the path because to get the page we might need to
4248                  * read the extent item from disk (data not in the page cache).
4249                  */
4250                 btrfs_release_path(path);
4251                 return btrfs_truncate_page(inode, offset, page_end - offset, 0);
4252         }
4253
4254         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4255         size = btrfs_file_extent_calc_inline_size(size);
4256         btrfs_truncate_item(root, path, size, 1);
4257
4258         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4259                 inode_sub_bytes(inode, item_end + 1 - new_size);
4260
4261         return 0;
4262 }
4263
4264 /*
4265  * this can truncate away extent items, csum items and directory items.
4266  * It starts at a high offset and removes keys until it can't find
4267  * any higher than new_size
4268  *
4269  * csum items that cross the new i_size are truncated to the new size
4270  * as well.
4271  *
4272  * min_type is the minimum key type to truncate down to.  If set to 0, this
4273  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4274  */
4275 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4276                                struct btrfs_root *root,
4277                                struct inode *inode,
4278                                u64 new_size, u32 min_type)
4279 {
4280         struct btrfs_path *path;
4281         struct extent_buffer *leaf;
4282         struct btrfs_file_extent_item *fi;
4283         struct btrfs_key key;
4284         struct btrfs_key found_key;
4285         u64 extent_start = 0;
4286         u64 extent_num_bytes = 0;
4287         u64 extent_offset = 0;
4288         u64 item_end = 0;
4289         u64 last_size = new_size;
4290         u32 found_type = (u8)-1;
4291         int found_extent;
4292         int del_item;
4293         int pending_del_nr = 0;
4294         int pending_del_slot = 0;
4295         int extent_type = -1;
4296         int ret;
4297         int err = 0;
4298         u64 ino = btrfs_ino(inode);
4299         u64 bytes_deleted = 0;
4300         bool be_nice = 0;
4301         bool should_throttle = 0;
4302         bool should_end = 0;
4303
4304         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4305
4306         /*
4307          * for non-free space inodes and ref cows, we want to back off from
4308          * time to time
4309          */
4310         if (!btrfs_is_free_space_inode(inode) &&
4311             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4312                 be_nice = 1;
4313
4314         path = btrfs_alloc_path();
4315         if (!path)
4316                 return -ENOMEM;
4317         path->reada = READA_BACK;
4318
4319         /*
4320          * We want to drop from the next block forward in case this new size is
4321          * not block aligned since we will be keeping the last block of the
4322          * extent just the way it is.
4323          */
4324         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4325             root == root->fs_info->tree_root)
4326                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4327                                         root->sectorsize), (u64)-1, 0);
4328
4329         /*
4330          * This function is also used to drop the items in the log tree before
4331          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4332          * it is used to drop the loged items. So we shouldn't kill the delayed
4333          * items.
4334          */
4335         if (min_type == 0 && root == BTRFS_I(inode)->root)
4336                 btrfs_kill_delayed_inode_items(inode);
4337
4338         key.objectid = ino;
4339         key.offset = (u64)-1;
4340         key.type = (u8)-1;
4341
4342 search_again:
4343         /*
4344          * with a 16K leaf size and 128MB extents, you can actually queue
4345          * up a huge file in a single leaf.  Most of the time that
4346          * bytes_deleted is > 0, it will be huge by the time we get here
4347          */
4348         if (be_nice && bytes_deleted > SZ_32M) {
4349                 if (btrfs_should_end_transaction(trans, root)) {
4350                         err = -EAGAIN;
4351                         goto error;
4352                 }
4353         }
4354
4355
4356         path->leave_spinning = 1;
4357         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4358         if (ret < 0) {
4359                 err = ret;
4360                 goto out;
4361         }
4362
4363         if (ret > 0) {
4364                 /* there are no items in the tree for us to truncate, we're
4365                  * done
4366                  */
4367                 if (path->slots[0] == 0)
4368                         goto out;
4369                 path->slots[0]--;
4370         }
4371
4372         while (1) {
4373                 fi = NULL;
4374                 leaf = path->nodes[0];
4375                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4376                 found_type = found_key.type;
4377
4378                 if (found_key.objectid != ino)
4379                         break;
4380
4381                 if (found_type < min_type)
4382                         break;
4383
4384                 item_end = found_key.offset;
4385                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4386                         fi = btrfs_item_ptr(leaf, path->slots[0],
4387                                             struct btrfs_file_extent_item);
4388                         extent_type = btrfs_file_extent_type(leaf, fi);
4389                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4390                                 item_end +=
4391                                     btrfs_file_extent_num_bytes(leaf, fi);
4392                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4393                                 item_end += btrfs_file_extent_inline_len(leaf,
4394                                                          path->slots[0], fi);
4395                         }
4396                         item_end--;
4397                 }
4398                 if (found_type > min_type) {
4399                         del_item = 1;
4400                 } else {
4401                         if (item_end < new_size)
4402                                 break;
4403                         if (found_key.offset >= new_size)
4404                                 del_item = 1;
4405                         else
4406                                 del_item = 0;
4407                 }
4408                 found_extent = 0;
4409                 /* FIXME, shrink the extent if the ref count is only 1 */
4410                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4411                         goto delete;
4412
4413                 if (del_item)
4414                         last_size = found_key.offset;
4415                 else
4416                         last_size = new_size;
4417
4418                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4419                         u64 num_dec;
4420                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4421                         if (!del_item) {
4422                                 u64 orig_num_bytes =
4423                                         btrfs_file_extent_num_bytes(leaf, fi);
4424                                 extent_num_bytes = ALIGN(new_size -
4425                                                 found_key.offset,
4426                                                 root->sectorsize);
4427                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4428                                                          extent_num_bytes);
4429                                 num_dec = (orig_num_bytes -
4430                                            extent_num_bytes);
4431                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4432                                              &root->state) &&
4433                                     extent_start != 0)
4434                                         inode_sub_bytes(inode, num_dec);
4435                                 btrfs_mark_buffer_dirty(leaf);
4436                         } else {
4437                                 extent_num_bytes =
4438                                         btrfs_file_extent_disk_num_bytes(leaf,
4439                                                                          fi);
4440                                 extent_offset = found_key.offset -
4441                                         btrfs_file_extent_offset(leaf, fi);
4442
4443                                 /* FIXME blocksize != 4096 */
4444                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4445                                 if (extent_start != 0) {
4446                                         found_extent = 1;
4447                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4448                                                      &root->state))
4449                                                 inode_sub_bytes(inode, num_dec);
4450                                 }
4451                         }
4452                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4453                         /*
4454                          * we can't truncate inline items that have had
4455                          * special encodings
4456                          */
4457                         if (!del_item &&
4458                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4459                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4460
4461                                 /*
4462                                  * Need to release path in order to truncate a
4463                                  * compressed extent. So delete any accumulated
4464                                  * extent items so far.
4465                                  */
4466                                 if (btrfs_file_extent_compression(leaf, fi) !=
4467                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4468                                         err = btrfs_del_items(trans, root, path,
4469                                                               pending_del_slot,
4470                                                               pending_del_nr);
4471                                         if (err) {
4472                                                 btrfs_abort_transaction(trans,
4473                                                                         root,
4474                                                                         err);
4475                                                 goto error;
4476                                         }
4477                                         pending_del_nr = 0;
4478                                 }
4479
4480                                 err = truncate_inline_extent(inode, path,
4481                                                              &found_key,
4482                                                              item_end,
4483                                                              new_size);
4484                                 if (err) {
4485                                         btrfs_abort_transaction(trans,
4486                                                                 root, err);
4487                                         goto error;
4488                                 }
4489                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4490                                             &root->state)) {
4491                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4492                         }
4493                 }
4494 delete:
4495                 if (del_item) {
4496                         if (!pending_del_nr) {
4497                                 /* no pending yet, add ourselves */
4498                                 pending_del_slot = path->slots[0];
4499                                 pending_del_nr = 1;
4500                         } else if (pending_del_nr &&
4501                                    path->slots[0] + 1 == pending_del_slot) {
4502                                 /* hop on the pending chunk */
4503                                 pending_del_nr++;
4504                                 pending_del_slot = path->slots[0];
4505                         } else {
4506                                 BUG();
4507                         }
4508                 } else {
4509                         break;
4510                 }
4511                 should_throttle = 0;
4512
4513                 if (found_extent &&
4514                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4515                      root == root->fs_info->tree_root)) {
4516                         btrfs_set_path_blocking(path);
4517                         bytes_deleted += extent_num_bytes;
4518                         ret = btrfs_free_extent(trans, root, extent_start,
4519                                                 extent_num_bytes, 0,
4520                                                 btrfs_header_owner(leaf),
4521                                                 ino, extent_offset);
4522                         BUG_ON(ret);
4523                         if (btrfs_should_throttle_delayed_refs(trans, root))
4524                                 btrfs_async_run_delayed_refs(root,
4525                                         trans->delayed_ref_updates * 2, 0);
4526                         if (be_nice) {
4527                                 if (truncate_space_check(trans, root,
4528                                                          extent_num_bytes)) {
4529                                         should_end = 1;
4530                                 }
4531                                 if (btrfs_should_throttle_delayed_refs(trans,
4532                                                                        root)) {
4533                                         should_throttle = 1;
4534                                 }
4535                         }
4536                 }
4537
4538                 if (found_type == BTRFS_INODE_ITEM_KEY)
4539                         break;
4540
4541                 if (path->slots[0] == 0 ||
4542                     path->slots[0] != pending_del_slot ||
4543                     should_throttle || should_end) {
4544                         if (pending_del_nr) {
4545                                 ret = btrfs_del_items(trans, root, path,
4546                                                 pending_del_slot,
4547                                                 pending_del_nr);
4548                                 if (ret) {
4549                                         btrfs_abort_transaction(trans,
4550                                                                 root, ret);
4551                                         goto error;
4552                                 }
4553                                 pending_del_nr = 0;
4554                         }
4555                         btrfs_release_path(path);
4556                         if (should_throttle) {
4557                                 unsigned long updates = trans->delayed_ref_updates;
4558                                 if (updates) {
4559                                         trans->delayed_ref_updates = 0;
4560                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4561                                         if (ret && !err)
4562                                                 err = ret;
4563                                 }
4564                         }
4565                         /*
4566                          * if we failed to refill our space rsv, bail out
4567                          * and let the transaction restart
4568                          */
4569                         if (should_end) {
4570                                 err = -EAGAIN;
4571                                 goto error;
4572                         }
4573                         goto search_again;
4574                 } else {
4575                         path->slots[0]--;
4576                 }
4577         }
4578 out:
4579         if (pending_del_nr) {
4580                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4581                                       pending_del_nr);
4582                 if (ret)
4583                         btrfs_abort_transaction(trans, root, ret);
4584         }
4585 error:
4586         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4587                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4588
4589         btrfs_free_path(path);
4590
4591         if (be_nice && bytes_deleted > SZ_32M) {
4592                 unsigned long updates = trans->delayed_ref_updates;
4593                 if (updates) {
4594                         trans->delayed_ref_updates = 0;
4595                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4596                         if (ret && !err)
4597                                 err = ret;
4598                 }
4599         }
4600         return err;
4601 }
4602
4603 /*
4604  * btrfs_truncate_page - read, zero a chunk and write a page
4605  * @inode - inode that we're zeroing
4606  * @from - the offset to start zeroing
4607  * @len - the length to zero, 0 to zero the entire range respective to the
4608  *      offset
4609  * @front - zero up to the offset instead of from the offset on
4610  *
4611  * This will find the page for the "from" offset and cow the page and zero the
4612  * part we want to zero.  This is used with truncate and hole punching.
4613  */
4614 int btrfs_truncate_page(struct inode *inode, loff_t from, loff_t len,
4615                         int front)
4616 {
4617         struct address_space *mapping = inode->i_mapping;
4618         struct btrfs_root *root = BTRFS_I(inode)->root;
4619         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4620         struct btrfs_ordered_extent *ordered;
4621         struct extent_state *cached_state = NULL;
4622         char *kaddr;
4623         u32 blocksize = root->sectorsize;
4624         pgoff_t index = from >> PAGE_CACHE_SHIFT;
4625         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4626         struct page *page;
4627         gfp_t mask = btrfs_alloc_write_mask(mapping);
4628         int ret = 0;
4629         u64 page_start;
4630         u64 page_end;
4631
4632         if ((offset & (blocksize - 1)) == 0 &&
4633             (!len || ((len & (blocksize - 1)) == 0)))
4634                 goto out;
4635         ret = btrfs_delalloc_reserve_space(inode,
4636                         round_down(from, PAGE_CACHE_SIZE), PAGE_CACHE_SIZE);
4637         if (ret)
4638                 goto out;
4639
4640 again:
4641         page = find_or_create_page(mapping, index, mask);
4642         if (!page) {
4643                 btrfs_delalloc_release_space(inode,
4644                                 round_down(from, PAGE_CACHE_SIZE),
4645                                 PAGE_CACHE_SIZE);
4646                 ret = -ENOMEM;
4647                 goto out;
4648         }
4649
4650         page_start = page_offset(page);
4651         page_end = page_start + PAGE_CACHE_SIZE - 1;
4652
4653         if (!PageUptodate(page)) {
4654                 ret = btrfs_readpage(NULL, page);
4655                 lock_page(page);
4656                 if (page->mapping != mapping) {
4657                         unlock_page(page);
4658                         page_cache_release(page);
4659                         goto again;
4660                 }
4661                 if (!PageUptodate(page)) {
4662                         ret = -EIO;
4663                         goto out_unlock;
4664                 }
4665         }
4666         wait_on_page_writeback(page);
4667
4668         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
4669         set_page_extent_mapped(page);
4670
4671         ordered = btrfs_lookup_ordered_extent(inode, page_start);
4672         if (ordered) {
4673                 unlock_extent_cached(io_tree, page_start, page_end,
4674                                      &cached_state, GFP_NOFS);
4675                 unlock_page(page);
4676                 page_cache_release(page);
4677                 btrfs_start_ordered_extent(inode, ordered, 1);
4678                 btrfs_put_ordered_extent(ordered);
4679                 goto again;
4680         }
4681
4682         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
4683                           EXTENT_DIRTY | EXTENT_DELALLOC |
4684                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4685                           0, 0, &cached_state, GFP_NOFS);
4686
4687         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
4688                                         &cached_state);
4689         if (ret) {
4690                 unlock_extent_cached(io_tree, page_start, page_end,
4691                                      &cached_state, GFP_NOFS);
4692                 goto out_unlock;
4693         }
4694
4695         if (offset != PAGE_CACHE_SIZE) {
4696                 if (!len)
4697                         len = PAGE_CACHE_SIZE - offset;
4698                 kaddr = kmap(page);
4699                 if (front)
4700                         memset(kaddr, 0, offset);
4701                 else
4702                         memset(kaddr + offset, 0, len);
4703                 flush_dcache_page(page);
4704                 kunmap(page);
4705         }
4706         ClearPageChecked(page);
4707         set_page_dirty(page);
4708         unlock_extent_cached(io_tree, page_start, page_end, &cached_state,
4709                              GFP_NOFS);
4710
4711 out_unlock:
4712         if (ret)
4713                 btrfs_delalloc_release_space(inode, page_start,
4714                                              PAGE_CACHE_SIZE);
4715         unlock_page(page);
4716         page_cache_release(page);
4717 out:
4718         return ret;
4719 }
4720
4721 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4722                              u64 offset, u64 len)
4723 {
4724         struct btrfs_trans_handle *trans;
4725         int ret;
4726
4727         /*
4728          * Still need to make sure the inode looks like it's been updated so
4729          * that any holes get logged if we fsync.
4730          */
4731         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4732                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4733                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4734                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4735                 return 0;
4736         }
4737
4738         /*
4739          * 1 - for the one we're dropping
4740          * 1 - for the one we're adding
4741          * 1 - for updating the inode.
4742          */
4743         trans = btrfs_start_transaction(root, 3);
4744         if (IS_ERR(trans))
4745                 return PTR_ERR(trans);
4746
4747         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4748         if (ret) {
4749                 btrfs_abort_transaction(trans, root, ret);
4750                 btrfs_end_transaction(trans, root);
4751                 return ret;
4752         }
4753
4754         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4755                                        0, 0, len, 0, len, 0, 0, 0);
4756         if (ret)
4757                 btrfs_abort_transaction(trans, root, ret);
4758         else
4759                 btrfs_update_inode(trans, root, inode);
4760         btrfs_end_transaction(trans, root);
4761         return ret;
4762 }
4763
4764 /*
4765  * This function puts in dummy file extents for the area we're creating a hole
4766  * for.  So if we are truncating this file to a larger size we need to insert
4767  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4768  * the range between oldsize and size
4769  */
4770 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4771 {
4772         struct btrfs_root *root = BTRFS_I(inode)->root;
4773         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4774         struct extent_map *em = NULL;
4775         struct extent_state *cached_state = NULL;
4776         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4777         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4778         u64 block_end = ALIGN(size, root->sectorsize);
4779         u64 last_byte;
4780         u64 cur_offset;
4781         u64 hole_size;
4782         int err = 0;
4783
4784         /*
4785          * If our size started in the middle of a page we need to zero out the
4786          * rest of the page before we expand the i_size, otherwise we could
4787          * expose stale data.
4788          */
4789         err = btrfs_truncate_page(inode, oldsize, 0, 0);
4790         if (err)
4791                 return err;
4792
4793         if (size <= hole_start)
4794                 return 0;
4795
4796         while (1) {
4797                 struct btrfs_ordered_extent *ordered;
4798
4799                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4800                                  &cached_state);
4801                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4802                                                      block_end - hole_start);
4803                 if (!ordered)
4804                         break;
4805                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4806                                      &cached_state, GFP_NOFS);
4807                 btrfs_start_ordered_extent(inode, ordered, 1);
4808                 btrfs_put_ordered_extent(ordered);
4809         }
4810
4811         cur_offset = hole_start;
4812         while (1) {
4813                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4814                                 block_end - cur_offset, 0);
4815                 if (IS_ERR(em)) {
4816                         err = PTR_ERR(em);
4817                         em = NULL;
4818                         break;
4819                 }
4820                 last_byte = min(extent_map_end(em), block_end);
4821                 last_byte = ALIGN(last_byte , root->sectorsize);
4822                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4823                         struct extent_map *hole_em;
4824                         hole_size = last_byte - cur_offset;
4825
4826                         err = maybe_insert_hole(root, inode, cur_offset,
4827                                                 hole_size);
4828                         if (err)
4829                                 break;
4830                         btrfs_drop_extent_cache(inode, cur_offset,
4831                                                 cur_offset + hole_size - 1, 0);
4832                         hole_em = alloc_extent_map();
4833                         if (!hole_em) {
4834                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4835                                         &BTRFS_I(inode)->runtime_flags);
4836                                 goto next;
4837                         }
4838                         hole_em->start = cur_offset;
4839                         hole_em->len = hole_size;
4840                         hole_em->orig_start = cur_offset;
4841
4842                         hole_em->block_start = EXTENT_MAP_HOLE;
4843                         hole_em->block_len = 0;
4844                         hole_em->orig_block_len = 0;
4845                         hole_em->ram_bytes = hole_size;
4846                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4847                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4848                         hole_em->generation = root->fs_info->generation;
4849
4850                         while (1) {
4851                                 write_lock(&em_tree->lock);
4852                                 err = add_extent_mapping(em_tree, hole_em, 1);
4853                                 write_unlock(&em_tree->lock);
4854                                 if (err != -EEXIST)
4855                                         break;
4856                                 btrfs_drop_extent_cache(inode, cur_offset,
4857                                                         cur_offset +
4858                                                         hole_size - 1, 0);
4859                         }
4860                         free_extent_map(hole_em);
4861                 }
4862 next:
4863                 free_extent_map(em);
4864                 em = NULL;
4865                 cur_offset = last_byte;
4866                 if (cur_offset >= block_end)
4867                         break;
4868         }
4869         free_extent_map(em);
4870         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4871                              GFP_NOFS);
4872         return err;
4873 }
4874
4875 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4876 {
4877         struct btrfs_root *root = BTRFS_I(inode)->root;
4878         struct btrfs_trans_handle *trans;
4879         loff_t oldsize = i_size_read(inode);
4880         loff_t newsize = attr->ia_size;
4881         int mask = attr->ia_valid;
4882         int ret;
4883
4884         /*
4885          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4886          * special case where we need to update the times despite not having
4887          * these flags set.  For all other operations the VFS set these flags
4888          * explicitly if it wants a timestamp update.
4889          */
4890         if (newsize != oldsize) {
4891                 inode_inc_iversion(inode);
4892                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4893                         inode->i_ctime = inode->i_mtime =
4894                                 current_fs_time(inode->i_sb);
4895         }
4896
4897         if (newsize > oldsize) {
4898                 truncate_pagecache(inode, newsize);
4899                 /*
4900                  * Don't do an expanding truncate while snapshoting is ongoing.
4901                  * This is to ensure the snapshot captures a fully consistent
4902                  * state of this file - if the snapshot captures this expanding
4903                  * truncation, it must capture all writes that happened before
4904                  * this truncation.
4905                  */
4906                 btrfs_wait_for_snapshot_creation(root);
4907                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4908                 if (ret) {
4909                         btrfs_end_write_no_snapshoting(root);
4910                         return ret;
4911                 }
4912
4913                 trans = btrfs_start_transaction(root, 1);
4914                 if (IS_ERR(trans)) {
4915                         btrfs_end_write_no_snapshoting(root);
4916                         return PTR_ERR(trans);
4917                 }
4918
4919                 i_size_write(inode, newsize);
4920                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4921                 ret = btrfs_update_inode(trans, root, inode);
4922                 btrfs_end_write_no_snapshoting(root);
4923                 btrfs_end_transaction(trans, root);
4924         } else {
4925
4926                 /*
4927                  * We're truncating a file that used to have good data down to
4928                  * zero. Make sure it gets into the ordered flush list so that
4929                  * any new writes get down to disk quickly.
4930                  */
4931                 if (newsize == 0)
4932                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4933                                 &BTRFS_I(inode)->runtime_flags);
4934
4935                 /*
4936                  * 1 for the orphan item we're going to add
4937                  * 1 for the orphan item deletion.
4938                  */
4939                 trans = btrfs_start_transaction(root, 2);
4940                 if (IS_ERR(trans))
4941                         return PTR_ERR(trans);
4942
4943                 /*
4944                  * We need to do this in case we fail at _any_ point during the
4945                  * actual truncate.  Once we do the truncate_setsize we could
4946                  * invalidate pages which forces any outstanding ordered io to
4947                  * be instantly completed which will give us extents that need
4948                  * to be truncated.  If we fail to get an orphan inode down we
4949                  * could have left over extents that were never meant to live,
4950                  * so we need to garuntee from this point on that everything
4951                  * will be consistent.
4952                  */
4953                 ret = btrfs_orphan_add(trans, inode);
4954                 btrfs_end_transaction(trans, root);
4955                 if (ret)
4956                         return ret;
4957
4958                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4959                 truncate_setsize(inode, newsize);
4960
4961                 /* Disable nonlocked read DIO to avoid the end less truncate */
4962                 btrfs_inode_block_unlocked_dio(inode);
4963                 inode_dio_wait(inode);
4964                 btrfs_inode_resume_unlocked_dio(inode);
4965
4966                 ret = btrfs_truncate(inode);
4967                 if (ret && inode->i_nlink) {
4968                         int err;
4969
4970                         /*
4971                          * failed to truncate, disk_i_size is only adjusted down
4972                          * as we remove extents, so it should represent the true
4973                          * size of the inode, so reset the in memory size and
4974                          * delete our orphan entry.
4975                          */
4976                         trans = btrfs_join_transaction(root);
4977                         if (IS_ERR(trans)) {
4978                                 btrfs_orphan_del(NULL, inode);
4979                                 return ret;
4980                         }
4981                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
4982                         err = btrfs_orphan_del(trans, inode);
4983                         if (err)
4984                                 btrfs_abort_transaction(trans, root, err);
4985                         btrfs_end_transaction(trans, root);
4986                 }
4987         }
4988
4989         return ret;
4990 }
4991
4992 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
4993 {
4994         struct inode *inode = d_inode(dentry);
4995         struct btrfs_root *root = BTRFS_I(inode)->root;
4996         int err;
4997
4998         if (btrfs_root_readonly(root))
4999                 return -EROFS;
5000
5001         err = inode_change_ok(inode, attr);
5002         if (err)
5003                 return err;
5004
5005         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5006                 err = btrfs_setsize(inode, attr);
5007                 if (err)
5008                         return err;
5009         }
5010
5011         if (attr->ia_valid) {
5012                 setattr_copy(inode, attr);
5013                 inode_inc_iversion(inode);
5014                 err = btrfs_dirty_inode(inode);
5015
5016                 if (!err && attr->ia_valid & ATTR_MODE)
5017                         err = posix_acl_chmod(inode, inode->i_mode);
5018         }
5019
5020         return err;
5021 }
5022
5023 /*
5024  * While truncating the inode pages during eviction, we get the VFS calling
5025  * btrfs_invalidatepage() against each page of the inode. This is slow because
5026  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5027  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5028  * extent_state structures over and over, wasting lots of time.
5029  *
5030  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5031  * those expensive operations on a per page basis and do only the ordered io
5032  * finishing, while we release here the extent_map and extent_state structures,
5033  * without the excessive merging and splitting.
5034  */
5035 static void evict_inode_truncate_pages(struct inode *inode)
5036 {
5037         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5038         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5039         struct rb_node *node;
5040
5041         ASSERT(inode->i_state & I_FREEING);
5042         truncate_inode_pages_final(&inode->i_data);
5043
5044         write_lock(&map_tree->lock);
5045         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5046                 struct extent_map *em;
5047
5048                 node = rb_first(&map_tree->map);
5049                 em = rb_entry(node, struct extent_map, rb_node);
5050                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5051                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5052                 remove_extent_mapping(map_tree, em);
5053                 free_extent_map(em);
5054                 if (need_resched()) {
5055                         write_unlock(&map_tree->lock);
5056                         cond_resched();
5057                         write_lock(&map_tree->lock);
5058                 }
5059         }
5060         write_unlock(&map_tree->lock);
5061
5062         /*
5063          * Keep looping until we have no more ranges in the io tree.
5064          * We can have ongoing bios started by readpages (called from readahead)
5065          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5066          * still in progress (unlocked the pages in the bio but did not yet
5067          * unlocked the ranges in the io tree). Therefore this means some
5068          * ranges can still be locked and eviction started because before
5069          * submitting those bios, which are executed by a separate task (work
5070          * queue kthread), inode references (inode->i_count) were not taken
5071          * (which would be dropped in the end io callback of each bio).
5072          * Therefore here we effectively end up waiting for those bios and
5073          * anyone else holding locked ranges without having bumped the inode's
5074          * reference count - if we don't do it, when they access the inode's
5075          * io_tree to unlock a range it may be too late, leading to an
5076          * use-after-free issue.
5077          */
5078         spin_lock(&io_tree->lock);
5079         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5080                 struct extent_state *state;
5081                 struct extent_state *cached_state = NULL;
5082                 u64 start;
5083                 u64 end;
5084
5085                 node = rb_first(&io_tree->state);
5086                 state = rb_entry(node, struct extent_state, rb_node);
5087                 start = state->start;
5088                 end = state->end;
5089                 spin_unlock(&io_tree->lock);
5090
5091                 lock_extent_bits(io_tree, start, end, &cached_state);
5092
5093                 /*
5094                  * If still has DELALLOC flag, the extent didn't reach disk,
5095                  * and its reserved space won't be freed by delayed_ref.
5096                  * So we need to free its reserved space here.
5097                  * (Refer to comment in btrfs_invalidatepage, case 2)
5098                  *
5099                  * Note, end is the bytenr of last byte, so we need + 1 here.
5100                  */
5101                 if (state->state & EXTENT_DELALLOC)
5102                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5103
5104                 clear_extent_bit(io_tree, start, end,
5105                                  EXTENT_LOCKED | EXTENT_DIRTY |
5106                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5107                                  EXTENT_DEFRAG, 1, 1,
5108                                  &cached_state, GFP_NOFS);
5109
5110                 cond_resched();
5111                 spin_lock(&io_tree->lock);
5112         }
5113         spin_unlock(&io_tree->lock);
5114 }
5115
5116 void btrfs_evict_inode(struct inode *inode)
5117 {
5118         struct btrfs_trans_handle *trans;
5119         struct btrfs_root *root = BTRFS_I(inode)->root;
5120         struct btrfs_block_rsv *rsv, *global_rsv;
5121         int steal_from_global = 0;
5122         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5123         int ret;
5124
5125         trace_btrfs_inode_evict(inode);
5126
5127         evict_inode_truncate_pages(inode);
5128
5129         if (inode->i_nlink &&
5130             ((btrfs_root_refs(&root->root_item) != 0 &&
5131               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5132              btrfs_is_free_space_inode(inode)))
5133                 goto no_delete;
5134
5135         if (is_bad_inode(inode)) {
5136                 btrfs_orphan_del(NULL, inode);
5137                 goto no_delete;
5138         }
5139         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5140         if (!special_file(inode->i_mode))
5141                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5142
5143         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5144
5145         if (root->fs_info->log_root_recovering) {
5146                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5147                                  &BTRFS_I(inode)->runtime_flags));
5148                 goto no_delete;
5149         }
5150
5151         if (inode->i_nlink > 0) {
5152                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5153                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5154                 goto no_delete;
5155         }
5156
5157         ret = btrfs_commit_inode_delayed_inode(inode);
5158         if (ret) {
5159                 btrfs_orphan_del(NULL, inode);
5160                 goto no_delete;
5161         }
5162
5163         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5164         if (!rsv) {
5165                 btrfs_orphan_del(NULL, inode);
5166                 goto no_delete;
5167         }
5168         rsv->size = min_size;
5169         rsv->failfast = 1;
5170         global_rsv = &root->fs_info->global_block_rsv;
5171
5172         btrfs_i_size_write(inode, 0);
5173
5174         /*
5175          * This is a bit simpler than btrfs_truncate since we've already
5176          * reserved our space for our orphan item in the unlink, so we just
5177          * need to reserve some slack space in case we add bytes and update
5178          * inode item when doing the truncate.
5179          */
5180         while (1) {
5181                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5182                                              BTRFS_RESERVE_FLUSH_LIMIT);
5183
5184                 /*
5185                  * Try and steal from the global reserve since we will
5186                  * likely not use this space anyway, we want to try as
5187                  * hard as possible to get this to work.
5188                  */
5189                 if (ret)
5190                         steal_from_global++;
5191                 else
5192                         steal_from_global = 0;
5193                 ret = 0;
5194
5195                 /*
5196                  * steal_from_global == 0: we reserved stuff, hooray!
5197                  * steal_from_global == 1: we didn't reserve stuff, boo!
5198                  * steal_from_global == 2: we've committed, still not a lot of
5199                  * room but maybe we'll have room in the global reserve this
5200                  * time.
5201                  * steal_from_global == 3: abandon all hope!
5202                  */
5203                 if (steal_from_global > 2) {
5204                         btrfs_warn(root->fs_info,
5205                                 "Could not get space for a delete, will truncate on mount %d",
5206                                 ret);
5207                         btrfs_orphan_del(NULL, inode);
5208                         btrfs_free_block_rsv(root, rsv);
5209                         goto no_delete;
5210                 }
5211
5212                 trans = btrfs_join_transaction(root);
5213                 if (IS_ERR(trans)) {
5214                         btrfs_orphan_del(NULL, inode);
5215                         btrfs_free_block_rsv(root, rsv);
5216                         goto no_delete;
5217                 }
5218
5219                 /*
5220                  * We can't just steal from the global reserve, we need tomake
5221                  * sure there is room to do it, if not we need to commit and try
5222                  * again.
5223                  */
5224                 if (steal_from_global) {
5225                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5226                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5227                                                               min_size);
5228                         else
5229                                 ret = -ENOSPC;
5230                 }
5231
5232                 /*
5233                  * Couldn't steal from the global reserve, we have too much
5234                  * pending stuff built up, commit the transaction and try it
5235                  * again.
5236                  */
5237                 if (ret) {
5238                         ret = btrfs_commit_transaction(trans, root);
5239                         if (ret) {
5240                                 btrfs_orphan_del(NULL, inode);
5241                                 btrfs_free_block_rsv(root, rsv);
5242                                 goto no_delete;
5243                         }
5244                         continue;
5245                 } else {
5246                         steal_from_global = 0;
5247                 }
5248
5249                 trans->block_rsv = rsv;
5250
5251                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5252                 if (ret != -ENOSPC && ret != -EAGAIN)
5253                         break;
5254
5255                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5256                 btrfs_end_transaction(trans, root);
5257                 trans = NULL;
5258                 btrfs_btree_balance_dirty(root);
5259         }
5260
5261         btrfs_free_block_rsv(root, rsv);
5262
5263         /*
5264          * Errors here aren't a big deal, it just means we leave orphan items
5265          * in the tree.  They will be cleaned up on the next mount.
5266          */
5267         if (ret == 0) {
5268                 trans->block_rsv = root->orphan_block_rsv;
5269                 btrfs_orphan_del(trans, inode);
5270         } else {
5271                 btrfs_orphan_del(NULL, inode);
5272         }
5273
5274         trans->block_rsv = &root->fs_info->trans_block_rsv;
5275         if (!(root == root->fs_info->tree_root ||
5276               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5277                 btrfs_return_ino(root, btrfs_ino(inode));
5278
5279         btrfs_end_transaction(trans, root);
5280         btrfs_btree_balance_dirty(root);
5281 no_delete:
5282         btrfs_remove_delayed_node(inode);
5283         clear_inode(inode);
5284 }
5285
5286 /*
5287  * this returns the key found in the dir entry in the location pointer.
5288  * If no dir entries were found, location->objectid is 0.
5289  */
5290 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5291                                struct btrfs_key *location)
5292 {
5293         const char *name = dentry->d_name.name;
5294         int namelen = dentry->d_name.len;
5295         struct btrfs_dir_item *di;
5296         struct btrfs_path *path;
5297         struct btrfs_root *root = BTRFS_I(dir)->root;
5298         int ret = 0;
5299
5300         path = btrfs_alloc_path();
5301         if (!path)
5302                 return -ENOMEM;
5303
5304         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5305                                     namelen, 0);
5306         if (IS_ERR(di))
5307                 ret = PTR_ERR(di);
5308
5309         if (IS_ERR_OR_NULL(di))
5310                 goto out_err;
5311
5312         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5313 out:
5314         btrfs_free_path(path);
5315         return ret;
5316 out_err:
5317         location->objectid = 0;
5318         goto out;
5319 }
5320
5321 /*
5322  * when we hit a tree root in a directory, the btrfs part of the inode
5323  * needs to be changed to reflect the root directory of the tree root.  This
5324  * is kind of like crossing a mount point.
5325  */
5326 static int fixup_tree_root_location(struct btrfs_root *root,
5327                                     struct inode *dir,
5328                                     struct dentry *dentry,
5329                                     struct btrfs_key *location,
5330                                     struct btrfs_root **sub_root)
5331 {
5332         struct btrfs_path *path;
5333         struct btrfs_root *new_root;
5334         struct btrfs_root_ref *ref;
5335         struct extent_buffer *leaf;
5336         struct btrfs_key key;
5337         int ret;
5338         int err = 0;
5339
5340         path = btrfs_alloc_path();
5341         if (!path) {
5342                 err = -ENOMEM;
5343                 goto out;
5344         }
5345
5346         err = -ENOENT;
5347         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5348         key.type = BTRFS_ROOT_REF_KEY;
5349         key.offset = location->objectid;
5350
5351         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5352                                 0, 0);
5353         if (ret) {
5354                 if (ret < 0)
5355                         err = ret;
5356                 goto out;
5357         }
5358
5359         leaf = path->nodes[0];
5360         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5361         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5362             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5363                 goto out;
5364
5365         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5366                                    (unsigned long)(ref + 1),
5367                                    dentry->d_name.len);
5368         if (ret)
5369                 goto out;
5370
5371         btrfs_release_path(path);
5372
5373         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5374         if (IS_ERR(new_root)) {
5375                 err = PTR_ERR(new_root);
5376                 goto out;
5377         }
5378
5379         *sub_root = new_root;
5380         location->objectid = btrfs_root_dirid(&new_root->root_item);
5381         location->type = BTRFS_INODE_ITEM_KEY;
5382         location->offset = 0;
5383         err = 0;
5384 out:
5385         btrfs_free_path(path);
5386         return err;
5387 }
5388
5389 static void inode_tree_add(struct inode *inode)
5390 {
5391         struct btrfs_root *root = BTRFS_I(inode)->root;
5392         struct btrfs_inode *entry;
5393         struct rb_node **p;
5394         struct rb_node *parent;
5395         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5396         u64 ino = btrfs_ino(inode);
5397
5398         if (inode_unhashed(inode))
5399                 return;
5400         parent = NULL;
5401         spin_lock(&root->inode_lock);
5402         p = &root->inode_tree.rb_node;
5403         while (*p) {
5404                 parent = *p;
5405                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5406
5407                 if (ino < btrfs_ino(&entry->vfs_inode))
5408                         p = &parent->rb_left;
5409                 else if (ino > btrfs_ino(&entry->vfs_inode))
5410                         p = &parent->rb_right;
5411                 else {
5412                         WARN_ON(!(entry->vfs_inode.i_state &
5413                                   (I_WILL_FREE | I_FREEING)));
5414                         rb_replace_node(parent, new, &root->inode_tree);
5415                         RB_CLEAR_NODE(parent);
5416                         spin_unlock(&root->inode_lock);
5417                         return;
5418                 }
5419         }
5420         rb_link_node(new, parent, p);
5421         rb_insert_color(new, &root->inode_tree);
5422         spin_unlock(&root->inode_lock);
5423 }
5424
5425 static void inode_tree_del(struct inode *inode)
5426 {
5427         struct btrfs_root *root = BTRFS_I(inode)->root;
5428         int empty = 0;
5429
5430         spin_lock(&root->inode_lock);
5431         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5432                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5433                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5434                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5435         }
5436         spin_unlock(&root->inode_lock);
5437
5438         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5439                 synchronize_srcu(&root->fs_info->subvol_srcu);
5440                 spin_lock(&root->inode_lock);
5441                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5442                 spin_unlock(&root->inode_lock);
5443                 if (empty)
5444                         btrfs_add_dead_root(root);
5445         }
5446 }
5447
5448 void btrfs_invalidate_inodes(struct btrfs_root *root)
5449 {
5450         struct rb_node *node;
5451         struct rb_node *prev;
5452         struct btrfs_inode *entry;
5453         struct inode *inode;
5454         u64 objectid = 0;
5455
5456         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5457                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5458
5459         spin_lock(&root->inode_lock);
5460 again:
5461         node = root->inode_tree.rb_node;
5462         prev = NULL;
5463         while (node) {
5464                 prev = node;
5465                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5466
5467                 if (objectid < btrfs_ino(&entry->vfs_inode))
5468                         node = node->rb_left;
5469                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5470                         node = node->rb_right;
5471                 else
5472                         break;
5473         }
5474         if (!node) {
5475                 while (prev) {
5476                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5477                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5478                                 node = prev;
5479                                 break;
5480                         }
5481                         prev = rb_next(prev);
5482                 }
5483         }
5484         while (node) {
5485                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5486                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5487                 inode = igrab(&entry->vfs_inode);
5488                 if (inode) {
5489                         spin_unlock(&root->inode_lock);
5490                         if (atomic_read(&inode->i_count) > 1)
5491                                 d_prune_aliases(inode);
5492                         /*
5493                          * btrfs_drop_inode will have it removed from
5494                          * the inode cache when its usage count
5495                          * hits zero.
5496                          */
5497                         iput(inode);
5498                         cond_resched();
5499                         spin_lock(&root->inode_lock);
5500                         goto again;
5501                 }
5502
5503                 if (cond_resched_lock(&root->inode_lock))
5504                         goto again;
5505
5506                 node = rb_next(node);
5507         }
5508         spin_unlock(&root->inode_lock);
5509 }
5510
5511 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5512 {
5513         struct btrfs_iget_args *args = p;
5514         inode->i_ino = args->location->objectid;
5515         memcpy(&BTRFS_I(inode)->location, args->location,
5516                sizeof(*args->location));
5517         BTRFS_I(inode)->root = args->root;
5518         return 0;
5519 }
5520
5521 static int btrfs_find_actor(struct inode *inode, void *opaque)
5522 {
5523         struct btrfs_iget_args *args = opaque;
5524         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5525                 args->root == BTRFS_I(inode)->root;
5526 }
5527
5528 static struct inode *btrfs_iget_locked(struct super_block *s,
5529                                        struct btrfs_key *location,
5530                                        struct btrfs_root *root)
5531 {
5532         struct inode *inode;
5533         struct btrfs_iget_args args;
5534         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5535
5536         args.location = location;
5537         args.root = root;
5538
5539         inode = iget5_locked(s, hashval, btrfs_find_actor,
5540                              btrfs_init_locked_inode,
5541                              (void *)&args);
5542         return inode;
5543 }
5544
5545 /* Get an inode object given its location and corresponding root.
5546  * Returns in *is_new if the inode was read from disk
5547  */
5548 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5549                          struct btrfs_root *root, int *new)
5550 {
5551         struct inode *inode;
5552
5553         inode = btrfs_iget_locked(s, location, root);
5554         if (!inode)
5555                 return ERR_PTR(-ENOMEM);
5556
5557         if (inode->i_state & I_NEW) {
5558                 btrfs_read_locked_inode(inode);
5559                 if (!is_bad_inode(inode)) {
5560                         inode_tree_add(inode);
5561                         unlock_new_inode(inode);
5562                         if (new)
5563                                 *new = 1;
5564                 } else {
5565                         unlock_new_inode(inode);
5566                         iput(inode);
5567                         inode = ERR_PTR(-ESTALE);
5568                 }
5569         }
5570
5571         return inode;
5572 }
5573
5574 static struct inode *new_simple_dir(struct super_block *s,
5575                                     struct btrfs_key *key,
5576                                     struct btrfs_root *root)
5577 {
5578         struct inode *inode = new_inode(s);
5579
5580         if (!inode)
5581                 return ERR_PTR(-ENOMEM);
5582
5583         BTRFS_I(inode)->root = root;
5584         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5585         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5586
5587         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5588         inode->i_op = &btrfs_dir_ro_inode_operations;
5589         inode->i_fop = &simple_dir_operations;
5590         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5591         inode->i_mtime = CURRENT_TIME;
5592         inode->i_atime = inode->i_mtime;
5593         inode->i_ctime = inode->i_mtime;
5594         BTRFS_I(inode)->i_otime = inode->i_mtime;
5595
5596         return inode;
5597 }
5598
5599 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5600 {
5601         struct inode *inode;
5602         struct btrfs_root *root = BTRFS_I(dir)->root;
5603         struct btrfs_root *sub_root = root;
5604         struct btrfs_key location;
5605         int index;
5606         int ret = 0;
5607
5608         if (dentry->d_name.len > BTRFS_NAME_LEN)
5609                 return ERR_PTR(-ENAMETOOLONG);
5610
5611         ret = btrfs_inode_by_name(dir, dentry, &location);
5612         if (ret < 0)
5613                 return ERR_PTR(ret);
5614
5615         if (location.objectid == 0)
5616                 return ERR_PTR(-ENOENT);
5617
5618         if (location.type == BTRFS_INODE_ITEM_KEY) {
5619                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5620                 return inode;
5621         }
5622
5623         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5624
5625         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5626         ret = fixup_tree_root_location(root, dir, dentry,
5627                                        &location, &sub_root);
5628         if (ret < 0) {
5629                 if (ret != -ENOENT)
5630                         inode = ERR_PTR(ret);
5631                 else
5632                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5633         } else {
5634                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5635         }
5636         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5637
5638         if (!IS_ERR(inode) && root != sub_root) {
5639                 down_read(&root->fs_info->cleanup_work_sem);
5640                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5641                         ret = btrfs_orphan_cleanup(sub_root);
5642                 up_read(&root->fs_info->cleanup_work_sem);
5643                 if (ret) {
5644                         iput(inode);
5645                         inode = ERR_PTR(ret);
5646                 }
5647         }
5648
5649         return inode;
5650 }
5651
5652 static int btrfs_dentry_delete(const struct dentry *dentry)
5653 {
5654         struct btrfs_root *root;
5655         struct inode *inode = d_inode(dentry);
5656
5657         if (!inode && !IS_ROOT(dentry))
5658                 inode = d_inode(dentry->d_parent);
5659
5660         if (inode) {
5661                 root = BTRFS_I(inode)->root;
5662                 if (btrfs_root_refs(&root->root_item) == 0)
5663                         return 1;
5664
5665                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5666                         return 1;
5667         }
5668         return 0;
5669 }
5670
5671 static void btrfs_dentry_release(struct dentry *dentry)
5672 {
5673         kfree(dentry->d_fsdata);
5674 }
5675
5676 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5677                                    unsigned int flags)
5678 {
5679         struct inode *inode;
5680
5681         inode = btrfs_lookup_dentry(dir, dentry);
5682         if (IS_ERR(inode)) {
5683                 if (PTR_ERR(inode) == -ENOENT)
5684                         inode = NULL;
5685                 else
5686                         return ERR_CAST(inode);
5687         }
5688
5689         return d_splice_alias(inode, dentry);
5690 }
5691
5692 unsigned char btrfs_filetype_table[] = {
5693         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5694 };
5695
5696 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5697 {
5698         struct inode *inode = file_inode(file);
5699         struct btrfs_root *root = BTRFS_I(inode)->root;
5700         struct btrfs_item *item;
5701         struct btrfs_dir_item *di;
5702         struct btrfs_key key;
5703         struct btrfs_key found_key;
5704         struct btrfs_path *path;
5705         struct list_head ins_list;
5706         struct list_head del_list;
5707         int ret;
5708         struct extent_buffer *leaf;
5709         int slot;
5710         unsigned char d_type;
5711         int over = 0;
5712         u32 di_cur;
5713         u32 di_total;
5714         u32 di_len;
5715         int key_type = BTRFS_DIR_INDEX_KEY;
5716         char tmp_name[32];
5717         char *name_ptr;
5718         int name_len;
5719         int is_curr = 0;        /* ctx->pos points to the current index? */
5720
5721         /* FIXME, use a real flag for deciding about the key type */
5722         if (root->fs_info->tree_root == root)
5723                 key_type = BTRFS_DIR_ITEM_KEY;
5724
5725         if (!dir_emit_dots(file, ctx))
5726                 return 0;
5727
5728         path = btrfs_alloc_path();
5729         if (!path)
5730                 return -ENOMEM;
5731
5732         path->reada = READA_FORWARD;
5733
5734         if (key_type == BTRFS_DIR_INDEX_KEY) {
5735                 INIT_LIST_HEAD(&ins_list);
5736                 INIT_LIST_HEAD(&del_list);
5737                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5738         }
5739
5740         key.type = key_type;
5741         key.offset = ctx->pos;
5742         key.objectid = btrfs_ino(inode);
5743
5744         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5745         if (ret < 0)
5746                 goto err;
5747
5748         while (1) {
5749                 leaf = path->nodes[0];
5750                 slot = path->slots[0];
5751                 if (slot >= btrfs_header_nritems(leaf)) {
5752                         ret = btrfs_next_leaf(root, path);
5753                         if (ret < 0)
5754                                 goto err;
5755                         else if (ret > 0)
5756                                 break;
5757                         continue;
5758                 }
5759
5760                 item = btrfs_item_nr(slot);
5761                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5762
5763                 if (found_key.objectid != key.objectid)
5764                         break;
5765                 if (found_key.type != key_type)
5766                         break;
5767                 if (found_key.offset < ctx->pos)
5768                         goto next;
5769                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5770                     btrfs_should_delete_dir_index(&del_list,
5771                                                   found_key.offset))
5772                         goto next;
5773
5774                 ctx->pos = found_key.offset;
5775                 is_curr = 1;
5776
5777                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5778                 di_cur = 0;
5779                 di_total = btrfs_item_size(leaf, item);
5780
5781                 while (di_cur < di_total) {
5782                         struct btrfs_key location;
5783
5784                         if (verify_dir_item(root, leaf, di))
5785                                 break;
5786
5787                         name_len = btrfs_dir_name_len(leaf, di);
5788                         if (name_len <= sizeof(tmp_name)) {
5789                                 name_ptr = tmp_name;
5790                         } else {
5791                                 name_ptr = kmalloc(name_len, GFP_NOFS);
5792                                 if (!name_ptr) {
5793                                         ret = -ENOMEM;
5794                                         goto err;
5795                                 }
5796                         }
5797                         read_extent_buffer(leaf, name_ptr,
5798                                            (unsigned long)(di + 1), name_len);
5799
5800                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5801                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5802
5803
5804                         /* is this a reference to our own snapshot? If so
5805                          * skip it.
5806                          *
5807                          * In contrast to old kernels, we insert the snapshot's
5808                          * dir item and dir index after it has been created, so
5809                          * we won't find a reference to our own snapshot. We
5810                          * still keep the following code for backward
5811                          * compatibility.
5812                          */
5813                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5814                             location.objectid == root->root_key.objectid) {
5815                                 over = 0;
5816                                 goto skip;
5817                         }
5818                         over = !dir_emit(ctx, name_ptr, name_len,
5819                                        location.objectid, d_type);
5820
5821 skip:
5822                         if (name_ptr != tmp_name)
5823                                 kfree(name_ptr);
5824
5825                         if (over)
5826                                 goto nopos;
5827                         di_len = btrfs_dir_name_len(leaf, di) +
5828                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5829                         di_cur += di_len;
5830                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5831                 }
5832 next:
5833                 path->slots[0]++;
5834         }
5835
5836         if (key_type == BTRFS_DIR_INDEX_KEY) {
5837                 if (is_curr)
5838                         ctx->pos++;
5839                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5840                 if (ret)
5841                         goto nopos;
5842         }
5843
5844         /* Reached end of directory/root. Bump pos past the last item. */
5845         ctx->pos++;
5846
5847         /*
5848          * Stop new entries from being returned after we return the last
5849          * entry.
5850          *
5851          * New directory entries are assigned a strictly increasing
5852          * offset.  This means that new entries created during readdir
5853          * are *guaranteed* to be seen in the future by that readdir.
5854          * This has broken buggy programs which operate on names as
5855          * they're returned by readdir.  Until we re-use freed offsets
5856          * we have this hack to stop new entries from being returned
5857          * under the assumption that they'll never reach this huge
5858          * offset.
5859          *
5860          * This is being careful not to overflow 32bit loff_t unless the
5861          * last entry requires it because doing so has broken 32bit apps
5862          * in the past.
5863          */
5864         if (key_type == BTRFS_DIR_INDEX_KEY) {
5865                 if (ctx->pos >= INT_MAX)
5866                         ctx->pos = LLONG_MAX;
5867                 else
5868                         ctx->pos = INT_MAX;
5869         }
5870 nopos:
5871         ret = 0;
5872 err:
5873         if (key_type == BTRFS_DIR_INDEX_KEY)
5874                 btrfs_put_delayed_items(&ins_list, &del_list);
5875         btrfs_free_path(path);
5876         return ret;
5877 }
5878
5879 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5880 {
5881         struct btrfs_root *root = BTRFS_I(inode)->root;
5882         struct btrfs_trans_handle *trans;
5883         int ret = 0;
5884         bool nolock = false;
5885
5886         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5887                 return 0;
5888
5889         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5890                 nolock = true;
5891
5892         if (wbc->sync_mode == WB_SYNC_ALL) {
5893                 if (nolock)
5894                         trans = btrfs_join_transaction_nolock(root);
5895                 else
5896                         trans = btrfs_join_transaction(root);
5897                 if (IS_ERR(trans))
5898                         return PTR_ERR(trans);
5899                 ret = btrfs_commit_transaction(trans, root);
5900         }
5901         return ret;
5902 }
5903
5904 /*
5905  * This is somewhat expensive, updating the tree every time the
5906  * inode changes.  But, it is most likely to find the inode in cache.
5907  * FIXME, needs more benchmarking...there are no reasons other than performance
5908  * to keep or drop this code.
5909  */
5910 static int btrfs_dirty_inode(struct inode *inode)
5911 {
5912         struct btrfs_root *root = BTRFS_I(inode)->root;
5913         struct btrfs_trans_handle *trans;
5914         int ret;
5915
5916         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5917                 return 0;
5918
5919         trans = btrfs_join_transaction(root);
5920         if (IS_ERR(trans))
5921                 return PTR_ERR(trans);
5922
5923         ret = btrfs_update_inode(trans, root, inode);
5924         if (ret && ret == -ENOSPC) {
5925                 /* whoops, lets try again with the full transaction */
5926                 btrfs_end_transaction(trans, root);
5927                 trans = btrfs_start_transaction(root, 1);
5928                 if (IS_ERR(trans))
5929                         return PTR_ERR(trans);
5930
5931                 ret = btrfs_update_inode(trans, root, inode);
5932         }
5933         btrfs_end_transaction(trans, root);
5934         if (BTRFS_I(inode)->delayed_node)
5935                 btrfs_balance_delayed_items(root);
5936
5937         return ret;
5938 }
5939
5940 /*
5941  * This is a copy of file_update_time.  We need this so we can return error on
5942  * ENOSPC for updating the inode in the case of file write and mmap writes.
5943  */
5944 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5945                              int flags)
5946 {
5947         struct btrfs_root *root = BTRFS_I(inode)->root;
5948
5949         if (btrfs_root_readonly(root))
5950                 return -EROFS;
5951
5952         if (flags & S_VERSION)
5953                 inode_inc_iversion(inode);
5954         if (flags & S_CTIME)
5955                 inode->i_ctime = *now;
5956         if (flags & S_MTIME)
5957                 inode->i_mtime = *now;
5958         if (flags & S_ATIME)
5959                 inode->i_atime = *now;
5960         return btrfs_dirty_inode(inode);
5961 }
5962
5963 /*
5964  * find the highest existing sequence number in a directory
5965  * and then set the in-memory index_cnt variable to reflect
5966  * free sequence numbers
5967  */
5968 static int btrfs_set_inode_index_count(struct inode *inode)
5969 {
5970         struct btrfs_root *root = BTRFS_I(inode)->root;
5971         struct btrfs_key key, found_key;
5972         struct btrfs_path *path;
5973         struct extent_buffer *leaf;
5974         int ret;
5975
5976         key.objectid = btrfs_ino(inode);
5977         key.type = BTRFS_DIR_INDEX_KEY;
5978         key.offset = (u64)-1;
5979
5980         path = btrfs_alloc_path();
5981         if (!path)
5982                 return -ENOMEM;
5983
5984         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5985         if (ret < 0)
5986                 goto out;
5987         /* FIXME: we should be able to handle this */
5988         if (ret == 0)
5989                 goto out;
5990         ret = 0;
5991
5992         /*
5993          * MAGIC NUMBER EXPLANATION:
5994          * since we search a directory based on f_pos we have to start at 2
5995          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
5996          * else has to start at 2
5997          */
5998         if (path->slots[0] == 0) {
5999                 BTRFS_I(inode)->index_cnt = 2;
6000                 goto out;
6001         }
6002
6003         path->slots[0]--;
6004
6005         leaf = path->nodes[0];
6006         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6007
6008         if (found_key.objectid != btrfs_ino(inode) ||
6009             found_key.type != BTRFS_DIR_INDEX_KEY) {
6010                 BTRFS_I(inode)->index_cnt = 2;
6011                 goto out;
6012         }
6013
6014         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6015 out:
6016         btrfs_free_path(path);
6017         return ret;
6018 }
6019
6020 /*
6021  * helper to find a free sequence number in a given directory.  This current
6022  * code is very simple, later versions will do smarter things in the btree
6023  */
6024 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6025 {
6026         int ret = 0;
6027
6028         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6029                 ret = btrfs_inode_delayed_dir_index_count(dir);
6030                 if (ret) {
6031                         ret = btrfs_set_inode_index_count(dir);
6032                         if (ret)
6033                                 return ret;
6034                 }
6035         }
6036
6037         *index = BTRFS_I(dir)->index_cnt;
6038         BTRFS_I(dir)->index_cnt++;
6039
6040         return ret;
6041 }
6042
6043 static int btrfs_insert_inode_locked(struct inode *inode)
6044 {
6045         struct btrfs_iget_args args;
6046         args.location = &BTRFS_I(inode)->location;
6047         args.root = BTRFS_I(inode)->root;
6048
6049         return insert_inode_locked4(inode,
6050                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6051                    btrfs_find_actor, &args);
6052 }
6053
6054 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6055                                      struct btrfs_root *root,
6056                                      struct inode *dir,
6057                                      const char *name, int name_len,
6058                                      u64 ref_objectid, u64 objectid,
6059                                      umode_t mode, u64 *index)
6060 {
6061         struct inode *inode;
6062         struct btrfs_inode_item *inode_item;
6063         struct btrfs_key *location;
6064         struct btrfs_path *path;
6065         struct btrfs_inode_ref *ref;
6066         struct btrfs_key key[2];
6067         u32 sizes[2];
6068         int nitems = name ? 2 : 1;
6069         unsigned long ptr;
6070         int ret;
6071
6072         path = btrfs_alloc_path();
6073         if (!path)
6074                 return ERR_PTR(-ENOMEM);
6075
6076         inode = new_inode(root->fs_info->sb);
6077         if (!inode) {
6078                 btrfs_free_path(path);
6079                 return ERR_PTR(-ENOMEM);
6080         }
6081
6082         /*
6083          * O_TMPFILE, set link count to 0, so that after this point,
6084          * we fill in an inode item with the correct link count.
6085          */
6086         if (!name)
6087                 set_nlink(inode, 0);
6088
6089         /*
6090          * we have to initialize this early, so we can reclaim the inode
6091          * number if we fail afterwards in this function.
6092          */
6093         inode->i_ino = objectid;
6094
6095         if (dir && name) {
6096                 trace_btrfs_inode_request(dir);
6097
6098                 ret = btrfs_set_inode_index(dir, index);
6099                 if (ret) {
6100                         btrfs_free_path(path);
6101                         iput(inode);
6102                         return ERR_PTR(ret);
6103                 }
6104         } else if (dir) {
6105                 *index = 0;
6106         }
6107         /*
6108          * index_cnt is ignored for everything but a dir,
6109          * btrfs_get_inode_index_count has an explanation for the magic
6110          * number
6111          */
6112         BTRFS_I(inode)->index_cnt = 2;
6113         BTRFS_I(inode)->dir_index = *index;
6114         BTRFS_I(inode)->root = root;
6115         BTRFS_I(inode)->generation = trans->transid;
6116         inode->i_generation = BTRFS_I(inode)->generation;
6117
6118         /*
6119          * We could have gotten an inode number from somebody who was fsynced
6120          * and then removed in this same transaction, so let's just set full
6121          * sync since it will be a full sync anyway and this will blow away the
6122          * old info in the log.
6123          */
6124         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6125
6126         key[0].objectid = objectid;
6127         key[0].type = BTRFS_INODE_ITEM_KEY;
6128         key[0].offset = 0;
6129
6130         sizes[0] = sizeof(struct btrfs_inode_item);
6131
6132         if (name) {
6133                 /*
6134                  * Start new inodes with an inode_ref. This is slightly more
6135                  * efficient for small numbers of hard links since they will
6136                  * be packed into one item. Extended refs will kick in if we
6137                  * add more hard links than can fit in the ref item.
6138                  */
6139                 key[1].objectid = objectid;
6140                 key[1].type = BTRFS_INODE_REF_KEY;
6141                 key[1].offset = ref_objectid;
6142
6143                 sizes[1] = name_len + sizeof(*ref);
6144         }
6145
6146         location = &BTRFS_I(inode)->location;
6147         location->objectid = objectid;
6148         location->offset = 0;
6149         location->type = BTRFS_INODE_ITEM_KEY;
6150
6151         ret = btrfs_insert_inode_locked(inode);
6152         if (ret < 0)
6153                 goto fail;
6154
6155         path->leave_spinning = 1;
6156         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6157         if (ret != 0)
6158                 goto fail_unlock;
6159
6160         inode_init_owner(inode, dir, mode);
6161         inode_set_bytes(inode, 0);
6162
6163         inode->i_mtime = CURRENT_TIME;
6164         inode->i_atime = inode->i_mtime;
6165         inode->i_ctime = inode->i_mtime;
6166         BTRFS_I(inode)->i_otime = inode->i_mtime;
6167
6168         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6169                                   struct btrfs_inode_item);
6170         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6171                              sizeof(*inode_item));
6172         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6173
6174         if (name) {
6175                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6176                                      struct btrfs_inode_ref);
6177                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6178                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6179                 ptr = (unsigned long)(ref + 1);
6180                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6181         }
6182
6183         btrfs_mark_buffer_dirty(path->nodes[0]);
6184         btrfs_free_path(path);
6185
6186         btrfs_inherit_iflags(inode, dir);
6187
6188         if (S_ISREG(mode)) {
6189                 if (btrfs_test_opt(root, NODATASUM))
6190                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6191                 if (btrfs_test_opt(root, NODATACOW))
6192                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6193                                 BTRFS_INODE_NODATASUM;
6194         }
6195
6196         inode_tree_add(inode);
6197
6198         trace_btrfs_inode_new(inode);
6199         btrfs_set_inode_last_trans(trans, inode);
6200
6201         btrfs_update_root_times(trans, root);
6202
6203         ret = btrfs_inode_inherit_props(trans, inode, dir);
6204         if (ret)
6205                 btrfs_err(root->fs_info,
6206                           "error inheriting props for ino %llu (root %llu): %d",
6207                           btrfs_ino(inode), root->root_key.objectid, ret);
6208
6209         return inode;
6210
6211 fail_unlock:
6212         unlock_new_inode(inode);
6213 fail:
6214         if (dir && name)
6215                 BTRFS_I(dir)->index_cnt--;
6216         btrfs_free_path(path);
6217         iput(inode);
6218         return ERR_PTR(ret);
6219 }
6220
6221 static inline u8 btrfs_inode_type(struct inode *inode)
6222 {
6223         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6224 }
6225
6226 /*
6227  * utility function to add 'inode' into 'parent_inode' with
6228  * a give name and a given sequence number.
6229  * if 'add_backref' is true, also insert a backref from the
6230  * inode to the parent directory.
6231  */
6232 int btrfs_add_link(struct btrfs_trans_handle *trans,
6233                    struct inode *parent_inode, struct inode *inode,
6234                    const char *name, int name_len, int add_backref, u64 index)
6235 {
6236         int ret = 0;
6237         struct btrfs_key key;
6238         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6239         u64 ino = btrfs_ino(inode);
6240         u64 parent_ino = btrfs_ino(parent_inode);
6241
6242         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6243                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6244         } else {
6245                 key.objectid = ino;
6246                 key.type = BTRFS_INODE_ITEM_KEY;
6247                 key.offset = 0;
6248         }
6249
6250         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6251                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6252                                          key.objectid, root->root_key.objectid,
6253                                          parent_ino, index, name, name_len);
6254         } else if (add_backref) {
6255                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6256                                              parent_ino, index);
6257         }
6258
6259         /* Nothing to clean up yet */
6260         if (ret)
6261                 return ret;
6262
6263         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6264                                     parent_inode, &key,
6265                                     btrfs_inode_type(inode), index);
6266         if (ret == -EEXIST || ret == -EOVERFLOW)
6267                 goto fail_dir_item;
6268         else if (ret) {
6269                 btrfs_abort_transaction(trans, root, ret);
6270                 return ret;
6271         }
6272
6273         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6274                            name_len * 2);
6275         inode_inc_iversion(parent_inode);
6276         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
6277         ret = btrfs_update_inode(trans, root, parent_inode);
6278         if (ret)
6279                 btrfs_abort_transaction(trans, root, ret);
6280         return ret;
6281
6282 fail_dir_item:
6283         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6284                 u64 local_index;
6285                 int err;
6286                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6287                                  key.objectid, root->root_key.objectid,
6288                                  parent_ino, &local_index, name, name_len);
6289
6290         } else if (add_backref) {
6291                 u64 local_index;
6292                 int err;
6293
6294                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6295                                           ino, parent_ino, &local_index);
6296         }
6297         return ret;
6298 }
6299
6300 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6301                             struct inode *dir, struct dentry *dentry,
6302                             struct inode *inode, int backref, u64 index)
6303 {
6304         int err = btrfs_add_link(trans, dir, inode,
6305                                  dentry->d_name.name, dentry->d_name.len,
6306                                  backref, index);
6307         if (err > 0)
6308                 err = -EEXIST;
6309         return err;
6310 }
6311
6312 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6313                         umode_t mode, dev_t rdev)
6314 {
6315         struct btrfs_trans_handle *trans;
6316         struct btrfs_root *root = BTRFS_I(dir)->root;
6317         struct inode *inode = NULL;
6318         int err;
6319         int drop_inode = 0;
6320         u64 objectid;
6321         u64 index = 0;
6322
6323         /*
6324          * 2 for inode item and ref
6325          * 2 for dir items
6326          * 1 for xattr if selinux is on
6327          */
6328         trans = btrfs_start_transaction(root, 5);
6329         if (IS_ERR(trans))
6330                 return PTR_ERR(trans);
6331
6332         err = btrfs_find_free_ino(root, &objectid);
6333         if (err)
6334                 goto out_unlock;
6335
6336         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6337                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6338                                 mode, &index);
6339         if (IS_ERR(inode)) {
6340                 err = PTR_ERR(inode);
6341                 goto out_unlock;
6342         }
6343
6344         /*
6345         * If the active LSM wants to access the inode during
6346         * d_instantiate it needs these. Smack checks to see
6347         * if the filesystem supports xattrs by looking at the
6348         * ops vector.
6349         */
6350         inode->i_op = &btrfs_special_inode_operations;
6351         init_special_inode(inode, inode->i_mode, rdev);
6352
6353         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6354         if (err)
6355                 goto out_unlock_inode;
6356
6357         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6358         if (err) {
6359                 goto out_unlock_inode;
6360         } else {
6361                 btrfs_update_inode(trans, root, inode);
6362                 unlock_new_inode(inode);
6363                 d_instantiate(dentry, inode);
6364         }
6365
6366 out_unlock:
6367         btrfs_end_transaction(trans, root);
6368         btrfs_balance_delayed_items(root);
6369         btrfs_btree_balance_dirty(root);
6370         if (drop_inode) {
6371                 inode_dec_link_count(inode);
6372                 iput(inode);
6373         }
6374         return err;
6375
6376 out_unlock_inode:
6377         drop_inode = 1;
6378         unlock_new_inode(inode);
6379         goto out_unlock;
6380
6381 }
6382
6383 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6384                         umode_t mode, bool excl)
6385 {
6386         struct btrfs_trans_handle *trans;
6387         struct btrfs_root *root = BTRFS_I(dir)->root;
6388         struct inode *inode = NULL;
6389         int drop_inode_on_err = 0;
6390         int err;
6391         u64 objectid;
6392         u64 index = 0;
6393
6394         /*
6395          * 2 for inode item and ref
6396          * 2 for dir items
6397          * 1 for xattr if selinux is on
6398          */
6399         trans = btrfs_start_transaction(root, 5);
6400         if (IS_ERR(trans))
6401                 return PTR_ERR(trans);
6402
6403         err = btrfs_find_free_ino(root, &objectid);
6404         if (err)
6405                 goto out_unlock;
6406
6407         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6408                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6409                                 mode, &index);
6410         if (IS_ERR(inode)) {
6411                 err = PTR_ERR(inode);
6412                 goto out_unlock;
6413         }
6414         drop_inode_on_err = 1;
6415         /*
6416         * If the active LSM wants to access the inode during
6417         * d_instantiate it needs these. Smack checks to see
6418         * if the filesystem supports xattrs by looking at the
6419         * ops vector.
6420         */
6421         inode->i_fop = &btrfs_file_operations;
6422         inode->i_op = &btrfs_file_inode_operations;
6423         inode->i_mapping->a_ops = &btrfs_aops;
6424
6425         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6426         if (err)
6427                 goto out_unlock_inode;
6428
6429         err = btrfs_update_inode(trans, root, inode);
6430         if (err)
6431                 goto out_unlock_inode;
6432
6433         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6434         if (err)
6435                 goto out_unlock_inode;
6436
6437         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6438         unlock_new_inode(inode);
6439         d_instantiate(dentry, inode);
6440
6441 out_unlock:
6442         btrfs_end_transaction(trans, root);
6443         if (err && drop_inode_on_err) {
6444                 inode_dec_link_count(inode);
6445                 iput(inode);
6446         }
6447         btrfs_balance_delayed_items(root);
6448         btrfs_btree_balance_dirty(root);
6449         return err;
6450
6451 out_unlock_inode:
6452         unlock_new_inode(inode);
6453         goto out_unlock;
6454
6455 }
6456
6457 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6458                       struct dentry *dentry)
6459 {
6460         struct btrfs_trans_handle *trans = NULL;
6461         struct btrfs_root *root = BTRFS_I(dir)->root;
6462         struct inode *inode = d_inode(old_dentry);
6463         u64 index;
6464         int err;
6465         int drop_inode = 0;
6466
6467         /* do not allow sys_link's with other subvols of the same device */
6468         if (root->objectid != BTRFS_I(inode)->root->objectid)
6469                 return -EXDEV;
6470
6471         if (inode->i_nlink >= BTRFS_LINK_MAX)
6472                 return -EMLINK;
6473
6474         err = btrfs_set_inode_index(dir, &index);
6475         if (err)
6476                 goto fail;
6477
6478         /*
6479          * 2 items for inode and inode ref
6480          * 2 items for dir items
6481          * 1 item for parent inode
6482          */
6483         trans = btrfs_start_transaction(root, 5);
6484         if (IS_ERR(trans)) {
6485                 err = PTR_ERR(trans);
6486                 trans = NULL;
6487                 goto fail;
6488         }
6489
6490         /* There are several dir indexes for this inode, clear the cache. */
6491         BTRFS_I(inode)->dir_index = 0ULL;
6492         inc_nlink(inode);
6493         inode_inc_iversion(inode);
6494         inode->i_ctime = CURRENT_TIME;
6495         ihold(inode);
6496         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6497
6498         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6499
6500         if (err) {
6501                 drop_inode = 1;
6502         } else {
6503                 struct dentry *parent = dentry->d_parent;
6504                 err = btrfs_update_inode(trans, root, inode);
6505                 if (err)
6506                         goto fail;
6507                 if (inode->i_nlink == 1) {
6508                         /*
6509                          * If new hard link count is 1, it's a file created
6510                          * with open(2) O_TMPFILE flag.
6511                          */
6512                         err = btrfs_orphan_del(trans, inode);
6513                         if (err)
6514                                 goto fail;
6515                 }
6516                 d_instantiate(dentry, inode);
6517                 btrfs_log_new_name(trans, inode, NULL, parent);
6518         }
6519
6520         btrfs_balance_delayed_items(root);
6521 fail:
6522         if (trans)
6523                 btrfs_end_transaction(trans, root);
6524         if (drop_inode) {
6525                 inode_dec_link_count(inode);
6526                 iput(inode);
6527         }
6528         btrfs_btree_balance_dirty(root);
6529         return err;
6530 }
6531
6532 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6533 {
6534         struct inode *inode = NULL;
6535         struct btrfs_trans_handle *trans;
6536         struct btrfs_root *root = BTRFS_I(dir)->root;
6537         int err = 0;
6538         int drop_on_err = 0;
6539         u64 objectid = 0;
6540         u64 index = 0;
6541
6542         /*
6543          * 2 items for inode and ref
6544          * 2 items for dir items
6545          * 1 for xattr if selinux is on
6546          */
6547         trans = btrfs_start_transaction(root, 5);
6548         if (IS_ERR(trans))
6549                 return PTR_ERR(trans);
6550
6551         err = btrfs_find_free_ino(root, &objectid);
6552         if (err)
6553                 goto out_fail;
6554
6555         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6556                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6557                                 S_IFDIR | mode, &index);
6558         if (IS_ERR(inode)) {
6559                 err = PTR_ERR(inode);
6560                 goto out_fail;
6561         }
6562
6563         drop_on_err = 1;
6564         /* these must be set before we unlock the inode */
6565         inode->i_op = &btrfs_dir_inode_operations;
6566         inode->i_fop = &btrfs_dir_file_operations;
6567
6568         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6569         if (err)
6570                 goto out_fail_inode;
6571
6572         btrfs_i_size_write(inode, 0);
6573         err = btrfs_update_inode(trans, root, inode);
6574         if (err)
6575                 goto out_fail_inode;
6576
6577         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6578                              dentry->d_name.len, 0, index);
6579         if (err)
6580                 goto out_fail_inode;
6581
6582         d_instantiate(dentry, inode);
6583         /*
6584          * mkdir is special.  We're unlocking after we call d_instantiate
6585          * to avoid a race with nfsd calling d_instantiate.
6586          */
6587         unlock_new_inode(inode);
6588         drop_on_err = 0;
6589
6590 out_fail:
6591         btrfs_end_transaction(trans, root);
6592         if (drop_on_err) {
6593                 inode_dec_link_count(inode);
6594                 iput(inode);
6595         }
6596         btrfs_balance_delayed_items(root);
6597         btrfs_btree_balance_dirty(root);
6598         return err;
6599
6600 out_fail_inode:
6601         unlock_new_inode(inode);
6602         goto out_fail;
6603 }
6604
6605 /* Find next extent map of a given extent map, caller needs to ensure locks */
6606 static struct extent_map *next_extent_map(struct extent_map *em)
6607 {
6608         struct rb_node *next;
6609
6610         next = rb_next(&em->rb_node);
6611         if (!next)
6612                 return NULL;
6613         return container_of(next, struct extent_map, rb_node);
6614 }
6615
6616 static struct extent_map *prev_extent_map(struct extent_map *em)
6617 {
6618         struct rb_node *prev;
6619
6620         prev = rb_prev(&em->rb_node);
6621         if (!prev)
6622                 return NULL;
6623         return container_of(prev, struct extent_map, rb_node);
6624 }
6625
6626 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6627  * the existing extent is the nearest extent to map_start,
6628  * and an extent that you want to insert, deal with overlap and insert
6629  * the best fitted new extent into the tree.
6630  */
6631 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6632                                 struct extent_map *existing,
6633                                 struct extent_map *em,
6634                                 u64 map_start)
6635 {
6636         struct extent_map *prev;
6637         struct extent_map *next;
6638         u64 start;
6639         u64 end;
6640         u64 start_diff;
6641
6642         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6643
6644         if (existing->start > map_start) {
6645                 next = existing;
6646                 prev = prev_extent_map(next);
6647         } else {
6648                 prev = existing;
6649                 next = next_extent_map(prev);
6650         }
6651
6652         start = prev ? extent_map_end(prev) : em->start;
6653         start = max_t(u64, start, em->start);
6654         end = next ? next->start : extent_map_end(em);
6655         end = min_t(u64, end, extent_map_end(em));
6656         start_diff = start - em->start;
6657         em->start = start;
6658         em->len = end - start;
6659         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6660             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6661                 em->block_start += start_diff;
6662                 em->block_len -= start_diff;
6663         }
6664         return add_extent_mapping(em_tree, em, 0);
6665 }
6666
6667 static noinline int uncompress_inline(struct btrfs_path *path,
6668                                       struct page *page,
6669                                       size_t pg_offset, u64 extent_offset,
6670                                       struct btrfs_file_extent_item *item)
6671 {
6672         int ret;
6673         struct extent_buffer *leaf = path->nodes[0];
6674         char *tmp;
6675         size_t max_size;
6676         unsigned long inline_size;
6677         unsigned long ptr;
6678         int compress_type;
6679
6680         WARN_ON(pg_offset != 0);
6681         compress_type = btrfs_file_extent_compression(leaf, item);
6682         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6683         inline_size = btrfs_file_extent_inline_item_len(leaf,
6684                                         btrfs_item_nr(path->slots[0]));
6685         tmp = kmalloc(inline_size, GFP_NOFS);
6686         if (!tmp)
6687                 return -ENOMEM;
6688         ptr = btrfs_file_extent_inline_start(item);
6689
6690         read_extent_buffer(leaf, tmp, ptr, inline_size);
6691
6692         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
6693         ret = btrfs_decompress(compress_type, tmp, page,
6694                                extent_offset, inline_size, max_size);
6695         kfree(tmp);
6696         return ret;
6697 }
6698
6699 /*
6700  * a bit scary, this does extent mapping from logical file offset to the disk.
6701  * the ugly parts come from merging extents from the disk with the in-ram
6702  * representation.  This gets more complex because of the data=ordered code,
6703  * where the in-ram extents might be locked pending data=ordered completion.
6704  *
6705  * This also copies inline extents directly into the page.
6706  */
6707
6708 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6709                                     size_t pg_offset, u64 start, u64 len,
6710                                     int create)
6711 {
6712         int ret;
6713         int err = 0;
6714         u64 extent_start = 0;
6715         u64 extent_end = 0;
6716         u64 objectid = btrfs_ino(inode);
6717         u32 found_type;
6718         struct btrfs_path *path = NULL;
6719         struct btrfs_root *root = BTRFS_I(inode)->root;
6720         struct btrfs_file_extent_item *item;
6721         struct extent_buffer *leaf;
6722         struct btrfs_key found_key;
6723         struct extent_map *em = NULL;
6724         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6725         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6726         struct btrfs_trans_handle *trans = NULL;
6727         const bool new_inline = !page || create;
6728
6729 again:
6730         read_lock(&em_tree->lock);
6731         em = lookup_extent_mapping(em_tree, start, len);
6732         if (em)
6733                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6734         read_unlock(&em_tree->lock);
6735
6736         if (em) {
6737                 if (em->start > start || em->start + em->len <= start)
6738                         free_extent_map(em);
6739                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6740                         free_extent_map(em);
6741                 else
6742                         goto out;
6743         }
6744         em = alloc_extent_map();
6745         if (!em) {
6746                 err = -ENOMEM;
6747                 goto out;
6748         }
6749         em->bdev = root->fs_info->fs_devices->latest_bdev;
6750         em->start = EXTENT_MAP_HOLE;
6751         em->orig_start = EXTENT_MAP_HOLE;
6752         em->len = (u64)-1;
6753         em->block_len = (u64)-1;
6754
6755         if (!path) {
6756                 path = btrfs_alloc_path();
6757                 if (!path) {
6758                         err = -ENOMEM;
6759                         goto out;
6760                 }
6761                 /*
6762                  * Chances are we'll be called again, so go ahead and do
6763                  * readahead
6764                  */
6765                 path->reada = READA_FORWARD;
6766         }
6767
6768         ret = btrfs_lookup_file_extent(trans, root, path,
6769                                        objectid, start, trans != NULL);
6770         if (ret < 0) {
6771                 err = ret;
6772                 goto out;
6773         }
6774
6775         if (ret != 0) {
6776                 if (path->slots[0] == 0)
6777                         goto not_found;
6778                 path->slots[0]--;
6779         }
6780
6781         leaf = path->nodes[0];
6782         item = btrfs_item_ptr(leaf, path->slots[0],
6783                               struct btrfs_file_extent_item);
6784         /* are we inside the extent that was found? */
6785         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6786         found_type = found_key.type;
6787         if (found_key.objectid != objectid ||
6788             found_type != BTRFS_EXTENT_DATA_KEY) {
6789                 /*
6790                  * If we backup past the first extent we want to move forward
6791                  * and see if there is an extent in front of us, otherwise we'll
6792                  * say there is a hole for our whole search range which can
6793                  * cause problems.
6794                  */
6795                 extent_end = start;
6796                 goto next;
6797         }
6798
6799         found_type = btrfs_file_extent_type(leaf, item);
6800         extent_start = found_key.offset;
6801         if (found_type == BTRFS_FILE_EXTENT_REG ||
6802             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6803                 extent_end = extent_start +
6804                        btrfs_file_extent_num_bytes(leaf, item);
6805         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6806                 size_t size;
6807                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6808                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6809         }
6810 next:
6811         if (start >= extent_end) {
6812                 path->slots[0]++;
6813                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6814                         ret = btrfs_next_leaf(root, path);
6815                         if (ret < 0) {
6816                                 err = ret;
6817                                 goto out;
6818                         }
6819                         if (ret > 0)
6820                                 goto not_found;
6821                         leaf = path->nodes[0];
6822                 }
6823                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6824                 if (found_key.objectid != objectid ||
6825                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6826                         goto not_found;
6827                 if (start + len <= found_key.offset)
6828                         goto not_found;
6829                 if (start > found_key.offset)
6830                         goto next;
6831                 em->start = start;
6832                 em->orig_start = start;
6833                 em->len = found_key.offset - start;
6834                 goto not_found_em;
6835         }
6836
6837         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6838
6839         if (found_type == BTRFS_FILE_EXTENT_REG ||
6840             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6841                 goto insert;
6842         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6843                 unsigned long ptr;
6844                 char *map;
6845                 size_t size;
6846                 size_t extent_offset;
6847                 size_t copy_size;
6848
6849                 if (new_inline)
6850                         goto out;
6851
6852                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6853                 extent_offset = page_offset(page) + pg_offset - extent_start;
6854                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
6855                                 size - extent_offset);
6856                 em->start = extent_start + extent_offset;
6857                 em->len = ALIGN(copy_size, root->sectorsize);
6858                 em->orig_block_len = em->len;
6859                 em->orig_start = em->start;
6860                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6861                 if (create == 0 && !PageUptodate(page)) {
6862                         if (btrfs_file_extent_compression(leaf, item) !=
6863                             BTRFS_COMPRESS_NONE) {
6864                                 ret = uncompress_inline(path, page, pg_offset,
6865                                                         extent_offset, item);
6866                                 if (ret) {
6867                                         err = ret;
6868                                         goto out;
6869                                 }
6870                         } else {
6871                                 map = kmap(page);
6872                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6873                                                    copy_size);
6874                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
6875                                         memset(map + pg_offset + copy_size, 0,
6876                                                PAGE_CACHE_SIZE - pg_offset -
6877                                                copy_size);
6878                                 }
6879                                 kunmap(page);
6880                         }
6881                         flush_dcache_page(page);
6882                 } else if (create && PageUptodate(page)) {
6883                         BUG();
6884                         if (!trans) {
6885                                 kunmap(page);
6886                                 free_extent_map(em);
6887                                 em = NULL;
6888
6889                                 btrfs_release_path(path);
6890                                 trans = btrfs_join_transaction(root);
6891
6892                                 if (IS_ERR(trans))
6893                                         return ERR_CAST(trans);
6894                                 goto again;
6895                         }
6896                         map = kmap(page);
6897                         write_extent_buffer(leaf, map + pg_offset, ptr,
6898                                             copy_size);
6899                         kunmap(page);
6900                         btrfs_mark_buffer_dirty(leaf);
6901                 }
6902                 set_extent_uptodate(io_tree, em->start,
6903                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6904                 goto insert;
6905         }
6906 not_found:
6907         em->start = start;
6908         em->orig_start = start;
6909         em->len = len;
6910 not_found_em:
6911         em->block_start = EXTENT_MAP_HOLE;
6912         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6913 insert:
6914         btrfs_release_path(path);
6915         if (em->start > start || extent_map_end(em) <= start) {
6916                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6917                         em->start, em->len, start, len);
6918                 err = -EIO;
6919                 goto out;
6920         }
6921
6922         err = 0;
6923         write_lock(&em_tree->lock);
6924         ret = add_extent_mapping(em_tree, em, 0);
6925         /* it is possible that someone inserted the extent into the tree
6926          * while we had the lock dropped.  It is also possible that
6927          * an overlapping map exists in the tree
6928          */
6929         if (ret == -EEXIST) {
6930                 struct extent_map *existing;
6931
6932                 ret = 0;
6933
6934                 existing = search_extent_mapping(em_tree, start, len);
6935                 /*
6936                  * existing will always be non-NULL, since there must be
6937                  * extent causing the -EEXIST.
6938                  */
6939                 if (start >= extent_map_end(existing) ||
6940                     start <= existing->start) {
6941                         /*
6942                          * The existing extent map is the one nearest to
6943                          * the [start, start + len) range which overlaps
6944                          */
6945                         err = merge_extent_mapping(em_tree, existing,
6946                                                    em, start);
6947                         free_extent_map(existing);
6948                         if (err) {
6949                                 free_extent_map(em);
6950                                 em = NULL;
6951                         }
6952                 } else {
6953                         free_extent_map(em);
6954                         em = existing;
6955                         err = 0;
6956                 }
6957         }
6958         write_unlock(&em_tree->lock);
6959 out:
6960
6961         trace_btrfs_get_extent(root, em);
6962
6963         btrfs_free_path(path);
6964         if (trans) {
6965                 ret = btrfs_end_transaction(trans, root);
6966                 if (!err)
6967                         err = ret;
6968         }
6969         if (err) {
6970                 free_extent_map(em);
6971                 return ERR_PTR(err);
6972         }
6973         BUG_ON(!em); /* Error is always set */
6974         return em;
6975 }
6976
6977 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
6978                                            size_t pg_offset, u64 start, u64 len,
6979                                            int create)
6980 {
6981         struct extent_map *em;
6982         struct extent_map *hole_em = NULL;
6983         u64 range_start = start;
6984         u64 end;
6985         u64 found;
6986         u64 found_end;
6987         int err = 0;
6988
6989         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
6990         if (IS_ERR(em))
6991                 return em;
6992         if (em) {
6993                 /*
6994                  * if our em maps to
6995                  * -  a hole or
6996                  * -  a pre-alloc extent,
6997                  * there might actually be delalloc bytes behind it.
6998                  */
6999                 if (em->block_start != EXTENT_MAP_HOLE &&
7000                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7001                         return em;
7002                 else
7003                         hole_em = em;
7004         }
7005
7006         /* check to see if we've wrapped (len == -1 or similar) */
7007         end = start + len;
7008         if (end < start)
7009                 end = (u64)-1;
7010         else
7011                 end -= 1;
7012
7013         em = NULL;
7014
7015         /* ok, we didn't find anything, lets look for delalloc */
7016         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7017                                  end, len, EXTENT_DELALLOC, 1);
7018         found_end = range_start + found;
7019         if (found_end < range_start)
7020                 found_end = (u64)-1;
7021
7022         /*
7023          * we didn't find anything useful, return
7024          * the original results from get_extent()
7025          */
7026         if (range_start > end || found_end <= start) {
7027                 em = hole_em;
7028                 hole_em = NULL;
7029                 goto out;
7030         }
7031
7032         /* adjust the range_start to make sure it doesn't
7033          * go backwards from the start they passed in
7034          */
7035         range_start = max(start, range_start);
7036         found = found_end - range_start;
7037
7038         if (found > 0) {
7039                 u64 hole_start = start;
7040                 u64 hole_len = len;
7041
7042                 em = alloc_extent_map();
7043                 if (!em) {
7044                         err = -ENOMEM;
7045                         goto out;
7046                 }
7047                 /*
7048                  * when btrfs_get_extent can't find anything it
7049                  * returns one huge hole
7050                  *
7051                  * make sure what it found really fits our range, and
7052                  * adjust to make sure it is based on the start from
7053                  * the caller
7054                  */
7055                 if (hole_em) {
7056                         u64 calc_end = extent_map_end(hole_em);
7057
7058                         if (calc_end <= start || (hole_em->start > end)) {
7059                                 free_extent_map(hole_em);
7060                                 hole_em = NULL;
7061                         } else {
7062                                 hole_start = max(hole_em->start, start);
7063                                 hole_len = calc_end - hole_start;
7064                         }
7065                 }
7066                 em->bdev = NULL;
7067                 if (hole_em && range_start > hole_start) {
7068                         /* our hole starts before our delalloc, so we
7069                          * have to return just the parts of the hole
7070                          * that go until  the delalloc starts
7071                          */
7072                         em->len = min(hole_len,
7073                                       range_start - hole_start);
7074                         em->start = hole_start;
7075                         em->orig_start = hole_start;
7076                         /*
7077                          * don't adjust block start at all,
7078                          * it is fixed at EXTENT_MAP_HOLE
7079                          */
7080                         em->block_start = hole_em->block_start;
7081                         em->block_len = hole_len;
7082                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7083                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7084                 } else {
7085                         em->start = range_start;
7086                         em->len = found;
7087                         em->orig_start = range_start;
7088                         em->block_start = EXTENT_MAP_DELALLOC;
7089                         em->block_len = found;
7090                 }
7091         } else if (hole_em) {
7092                 return hole_em;
7093         }
7094 out:
7095
7096         free_extent_map(hole_em);
7097         if (err) {
7098                 free_extent_map(em);
7099                 return ERR_PTR(err);
7100         }
7101         return em;
7102 }
7103
7104 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7105                                                   u64 start, u64 len)
7106 {
7107         struct btrfs_root *root = BTRFS_I(inode)->root;
7108         struct extent_map *em;
7109         struct btrfs_key ins;
7110         u64 alloc_hint;
7111         int ret;
7112
7113         alloc_hint = get_extent_allocation_hint(inode, start, len);
7114         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7115                                    alloc_hint, &ins, 1, 1);
7116         if (ret)
7117                 return ERR_PTR(ret);
7118
7119         em = create_pinned_em(inode, start, ins.offset, start, ins.objectid,
7120                               ins.offset, ins.offset, ins.offset, 0);
7121         if (IS_ERR(em)) {
7122                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7123                 return em;
7124         }
7125
7126         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
7127                                            ins.offset, ins.offset, 0);
7128         if (ret) {
7129                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7130                 free_extent_map(em);
7131                 return ERR_PTR(ret);
7132         }
7133
7134         return em;
7135 }
7136
7137 /*
7138  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7139  * block must be cow'd
7140  */
7141 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7142                               u64 *orig_start, u64 *orig_block_len,
7143                               u64 *ram_bytes)
7144 {
7145         struct btrfs_trans_handle *trans;
7146         struct btrfs_path *path;
7147         int ret;
7148         struct extent_buffer *leaf;
7149         struct btrfs_root *root = BTRFS_I(inode)->root;
7150         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7151         struct btrfs_file_extent_item *fi;
7152         struct btrfs_key key;
7153         u64 disk_bytenr;
7154         u64 backref_offset;
7155         u64 extent_end;
7156         u64 num_bytes;
7157         int slot;
7158         int found_type;
7159         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7160
7161         path = btrfs_alloc_path();
7162         if (!path)
7163                 return -ENOMEM;
7164
7165         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7166                                        offset, 0);
7167         if (ret < 0)
7168                 goto out;
7169
7170         slot = path->slots[0];
7171         if (ret == 1) {
7172                 if (slot == 0) {
7173                         /* can't find the item, must cow */
7174                         ret = 0;
7175                         goto out;
7176                 }
7177                 slot--;
7178         }
7179         ret = 0;
7180         leaf = path->nodes[0];
7181         btrfs_item_key_to_cpu(leaf, &key, slot);
7182         if (key.objectid != btrfs_ino(inode) ||
7183             key.type != BTRFS_EXTENT_DATA_KEY) {
7184                 /* not our file or wrong item type, must cow */
7185                 goto out;
7186         }
7187
7188         if (key.offset > offset) {
7189                 /* Wrong offset, must cow */
7190                 goto out;
7191         }
7192
7193         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7194         found_type = btrfs_file_extent_type(leaf, fi);
7195         if (found_type != BTRFS_FILE_EXTENT_REG &&
7196             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7197                 /* not a regular extent, must cow */
7198                 goto out;
7199         }
7200
7201         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7202                 goto out;
7203
7204         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7205         if (extent_end <= offset)
7206                 goto out;
7207
7208         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7209         if (disk_bytenr == 0)
7210                 goto out;
7211
7212         if (btrfs_file_extent_compression(leaf, fi) ||
7213             btrfs_file_extent_encryption(leaf, fi) ||
7214             btrfs_file_extent_other_encoding(leaf, fi))
7215                 goto out;
7216
7217         backref_offset = btrfs_file_extent_offset(leaf, fi);
7218
7219         if (orig_start) {
7220                 *orig_start = key.offset - backref_offset;
7221                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7222                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7223         }
7224
7225         if (btrfs_extent_readonly(root, disk_bytenr))
7226                 goto out;
7227
7228         num_bytes = min(offset + *len, extent_end) - offset;
7229         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7230                 u64 range_end;
7231
7232                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7233                 ret = test_range_bit(io_tree, offset, range_end,
7234                                      EXTENT_DELALLOC, 0, NULL);
7235                 if (ret) {
7236                         ret = -EAGAIN;
7237                         goto out;
7238                 }
7239         }
7240
7241         btrfs_release_path(path);
7242
7243         /*
7244          * look for other files referencing this extent, if we
7245          * find any we must cow
7246          */
7247         trans = btrfs_join_transaction(root);
7248         if (IS_ERR(trans)) {
7249                 ret = 0;
7250                 goto out;
7251         }
7252
7253         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7254                                     key.offset - backref_offset, disk_bytenr);
7255         btrfs_end_transaction(trans, root);
7256         if (ret) {
7257                 ret = 0;
7258                 goto out;
7259         }
7260
7261         /*
7262          * adjust disk_bytenr and num_bytes to cover just the bytes
7263          * in this extent we are about to write.  If there
7264          * are any csums in that range we have to cow in order
7265          * to keep the csums correct
7266          */
7267         disk_bytenr += backref_offset;
7268         disk_bytenr += offset - key.offset;
7269         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7270                                 goto out;
7271         /*
7272          * all of the above have passed, it is safe to overwrite this extent
7273          * without cow
7274          */
7275         *len = num_bytes;
7276         ret = 1;
7277 out:
7278         btrfs_free_path(path);
7279         return ret;
7280 }
7281
7282 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7283 {
7284         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7285         int found = false;
7286         void **pagep = NULL;
7287         struct page *page = NULL;
7288         int start_idx;
7289         int end_idx;
7290
7291         start_idx = start >> PAGE_CACHE_SHIFT;
7292
7293         /*
7294          * end is the last byte in the last page.  end == start is legal
7295          */
7296         end_idx = end >> PAGE_CACHE_SHIFT;
7297
7298         rcu_read_lock();
7299
7300         /* Most of the code in this while loop is lifted from
7301          * find_get_page.  It's been modified to begin searching from a
7302          * page and return just the first page found in that range.  If the
7303          * found idx is less than or equal to the end idx then we know that
7304          * a page exists.  If no pages are found or if those pages are
7305          * outside of the range then we're fine (yay!) */
7306         while (page == NULL &&
7307                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7308                 page = radix_tree_deref_slot(pagep);
7309                 if (unlikely(!page))
7310                         break;
7311
7312                 if (radix_tree_exception(page)) {
7313                         if (radix_tree_deref_retry(page)) {
7314                                 page = NULL;
7315                                 continue;
7316                         }
7317                         /*
7318                          * Otherwise, shmem/tmpfs must be storing a swap entry
7319                          * here as an exceptional entry: so return it without
7320                          * attempting to raise page count.
7321                          */
7322                         page = NULL;
7323                         break; /* TODO: Is this relevant for this use case? */
7324                 }
7325
7326                 if (!page_cache_get_speculative(page)) {
7327                         page = NULL;
7328                         continue;
7329                 }
7330
7331                 /*
7332                  * Has the page moved?
7333                  * This is part of the lockless pagecache protocol. See
7334                  * include/linux/pagemap.h for details.
7335                  */
7336                 if (unlikely(page != *pagep)) {
7337                         page_cache_release(page);
7338                         page = NULL;
7339                 }
7340         }
7341
7342         if (page) {
7343                 if (page->index <= end_idx)
7344                         found = true;
7345                 page_cache_release(page);
7346         }
7347
7348         rcu_read_unlock();
7349         return found;
7350 }
7351
7352 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7353                               struct extent_state **cached_state, int writing)
7354 {
7355         struct btrfs_ordered_extent *ordered;
7356         int ret = 0;
7357
7358         while (1) {
7359                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7360                                  cached_state);
7361                 /*
7362                  * We're concerned with the entire range that we're going to be
7363                  * doing DIO to, so we need to make sure theres no ordered
7364                  * extents in this range.
7365                  */
7366                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7367                                                      lockend - lockstart + 1);
7368
7369                 /*
7370                  * We need to make sure there are no buffered pages in this
7371                  * range either, we could have raced between the invalidate in
7372                  * generic_file_direct_write and locking the extent.  The
7373                  * invalidate needs to happen so that reads after a write do not
7374                  * get stale data.
7375                  */
7376                 if (!ordered &&
7377                     (!writing ||
7378                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7379                         break;
7380
7381                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7382                                      cached_state, GFP_NOFS);
7383
7384                 if (ordered) {
7385                         btrfs_start_ordered_extent(inode, ordered, 1);
7386                         btrfs_put_ordered_extent(ordered);
7387                 } else {
7388                         /*
7389                          * We could trigger writeback for this range (and wait
7390                          * for it to complete) and then invalidate the pages for
7391                          * this range (through invalidate_inode_pages2_range()),
7392                          * but that can lead us to a deadlock with a concurrent
7393                          * call to readpages() (a buffered read or a defrag call
7394                          * triggered a readahead) on a page lock due to an
7395                          * ordered dio extent we created before but did not have
7396                          * yet a corresponding bio submitted (whence it can not
7397                          * complete), which makes readpages() wait for that
7398                          * ordered extent to complete while holding a lock on
7399                          * that page.
7400                          */
7401                         ret = -ENOTBLK;
7402                         break;
7403                 }
7404
7405                 cond_resched();
7406         }
7407
7408         return ret;
7409 }
7410
7411 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7412                                            u64 len, u64 orig_start,
7413                                            u64 block_start, u64 block_len,
7414                                            u64 orig_block_len, u64 ram_bytes,
7415                                            int type)
7416 {
7417         struct extent_map_tree *em_tree;
7418         struct extent_map *em;
7419         struct btrfs_root *root = BTRFS_I(inode)->root;
7420         int ret;
7421
7422         em_tree = &BTRFS_I(inode)->extent_tree;
7423         em = alloc_extent_map();
7424         if (!em)
7425                 return ERR_PTR(-ENOMEM);
7426
7427         em->start = start;
7428         em->orig_start = orig_start;
7429         em->mod_start = start;
7430         em->mod_len = len;
7431         em->len = len;
7432         em->block_len = block_len;
7433         em->block_start = block_start;
7434         em->bdev = root->fs_info->fs_devices->latest_bdev;
7435         em->orig_block_len = orig_block_len;
7436         em->ram_bytes = ram_bytes;
7437         em->generation = -1;
7438         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7439         if (type == BTRFS_ORDERED_PREALLOC)
7440                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7441
7442         do {
7443                 btrfs_drop_extent_cache(inode, em->start,
7444                                 em->start + em->len - 1, 0);
7445                 write_lock(&em_tree->lock);
7446                 ret = add_extent_mapping(em_tree, em, 1);
7447                 write_unlock(&em_tree->lock);
7448         } while (ret == -EEXIST);
7449
7450         if (ret) {
7451                 free_extent_map(em);
7452                 return ERR_PTR(ret);
7453         }
7454
7455         return em;
7456 }
7457
7458 static void adjust_dio_outstanding_extents(struct inode *inode,
7459                                            struct btrfs_dio_data *dio_data,
7460                                            const u64 len)
7461 {
7462         unsigned num_extents;
7463
7464         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7465                                            BTRFS_MAX_EXTENT_SIZE);
7466         /*
7467          * If we have an outstanding_extents count still set then we're
7468          * within our reservation, otherwise we need to adjust our inode
7469          * counter appropriately.
7470          */
7471         if (dio_data->outstanding_extents) {
7472                 dio_data->outstanding_extents -= num_extents;
7473         } else {
7474                 spin_lock(&BTRFS_I(inode)->lock);
7475                 BTRFS_I(inode)->outstanding_extents += num_extents;
7476                 spin_unlock(&BTRFS_I(inode)->lock);
7477         }
7478 }
7479
7480 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7481                                    struct buffer_head *bh_result, int create)
7482 {
7483         struct extent_map *em;
7484         struct btrfs_root *root = BTRFS_I(inode)->root;
7485         struct extent_state *cached_state = NULL;
7486         struct btrfs_dio_data *dio_data = NULL;
7487         u64 start = iblock << inode->i_blkbits;
7488         u64 lockstart, lockend;
7489         u64 len = bh_result->b_size;
7490         int unlock_bits = EXTENT_LOCKED;
7491         int ret = 0;
7492
7493         if (create)
7494                 unlock_bits |= EXTENT_DIRTY;
7495         else
7496                 len = min_t(u64, len, root->sectorsize);
7497
7498         lockstart = start;
7499         lockend = start + len - 1;
7500
7501         if (current->journal_info) {
7502                 /*
7503                  * Need to pull our outstanding extents and set journal_info to NULL so
7504                  * that anything that needs to check if there's a transction doesn't get
7505                  * confused.
7506                  */
7507                 dio_data = current->journal_info;
7508                 current->journal_info = NULL;
7509         }
7510
7511         /*
7512          * If this errors out it's because we couldn't invalidate pagecache for
7513          * this range and we need to fallback to buffered.
7514          */
7515         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7516                                create)) {
7517                 ret = -ENOTBLK;
7518                 goto err;
7519         }
7520
7521         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7522         if (IS_ERR(em)) {
7523                 ret = PTR_ERR(em);
7524                 goto unlock_err;
7525         }
7526
7527         /*
7528          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7529          * io.  INLINE is special, and we could probably kludge it in here, but
7530          * it's still buffered so for safety lets just fall back to the generic
7531          * buffered path.
7532          *
7533          * For COMPRESSED we _have_ to read the entire extent in so we can
7534          * decompress it, so there will be buffering required no matter what we
7535          * do, so go ahead and fallback to buffered.
7536          *
7537          * We return -ENOTBLK because thats what makes DIO go ahead and go back
7538          * to buffered IO.  Don't blame me, this is the price we pay for using
7539          * the generic code.
7540          */
7541         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7542             em->block_start == EXTENT_MAP_INLINE) {
7543                 free_extent_map(em);
7544                 ret = -ENOTBLK;
7545                 goto unlock_err;
7546         }
7547
7548         /* Just a good old fashioned hole, return */
7549         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7550                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7551                 free_extent_map(em);
7552                 goto unlock_err;
7553         }
7554
7555         /*
7556          * We don't allocate a new extent in the following cases
7557          *
7558          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7559          * existing extent.
7560          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7561          * just use the extent.
7562          *
7563          */
7564         if (!create) {
7565                 len = min(len, em->len - (start - em->start));
7566                 lockstart = start + len;
7567                 goto unlock;
7568         }
7569
7570         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7571             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7572              em->block_start != EXTENT_MAP_HOLE)) {
7573                 int type;
7574                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7575
7576                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7577                         type = BTRFS_ORDERED_PREALLOC;
7578                 else
7579                         type = BTRFS_ORDERED_NOCOW;
7580                 len = min(len, em->len - (start - em->start));
7581                 block_start = em->block_start + (start - em->start);
7582
7583                 if (can_nocow_extent(inode, start, &len, &orig_start,
7584                                      &orig_block_len, &ram_bytes) == 1) {
7585                         if (type == BTRFS_ORDERED_PREALLOC) {
7586                                 free_extent_map(em);
7587                                 em = create_pinned_em(inode, start, len,
7588                                                        orig_start,
7589                                                        block_start, len,
7590                                                        orig_block_len,
7591                                                        ram_bytes, type);
7592                                 if (IS_ERR(em)) {
7593                                         ret = PTR_ERR(em);
7594                                         goto unlock_err;
7595                                 }
7596                         }
7597
7598                         ret = btrfs_add_ordered_extent_dio(inode, start,
7599                                            block_start, len, len, type);
7600                         if (ret) {
7601                                 free_extent_map(em);
7602                                 goto unlock_err;
7603                         }
7604                         goto unlock;
7605                 }
7606         }
7607
7608         /*
7609          * this will cow the extent, reset the len in case we changed
7610          * it above
7611          */
7612         len = bh_result->b_size;
7613         free_extent_map(em);
7614         em = btrfs_new_extent_direct(inode, start, len);
7615         if (IS_ERR(em)) {
7616                 ret = PTR_ERR(em);
7617                 goto unlock_err;
7618         }
7619         len = min(len, em->len - (start - em->start));
7620 unlock:
7621         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7622                 inode->i_blkbits;
7623         bh_result->b_size = len;
7624         bh_result->b_bdev = em->bdev;
7625         set_buffer_mapped(bh_result);
7626         if (create) {
7627                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7628                         set_buffer_new(bh_result);
7629
7630                 /*
7631                  * Need to update the i_size under the extent lock so buffered
7632                  * readers will get the updated i_size when we unlock.
7633                  */
7634                 if (start + len > i_size_read(inode))
7635                         i_size_write(inode, start + len);
7636
7637                 adjust_dio_outstanding_extents(inode, dio_data, len);
7638                 btrfs_free_reserved_data_space(inode, start, len);
7639                 WARN_ON(dio_data->reserve < len);
7640                 dio_data->reserve -= len;
7641                 dio_data->unsubmitted_oe_range_end = start + len;
7642                 current->journal_info = dio_data;
7643         }
7644
7645         /*
7646          * In the case of write we need to clear and unlock the entire range,
7647          * in the case of read we need to unlock only the end area that we
7648          * aren't using if there is any left over space.
7649          */
7650         if (lockstart < lockend) {
7651                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7652                                  lockend, unlock_bits, 1, 0,
7653                                  &cached_state, GFP_NOFS);
7654         } else {
7655                 free_extent_state(cached_state);
7656         }
7657
7658         free_extent_map(em);
7659
7660         return 0;
7661
7662 unlock_err:
7663         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7664                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7665 err:
7666         if (dio_data)
7667                 current->journal_info = dio_data;
7668         /*
7669          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7670          * write less data then expected, so that we don't underflow our inode's
7671          * outstanding extents counter.
7672          */
7673         if (create && dio_data)
7674                 adjust_dio_outstanding_extents(inode, dio_data, len);
7675
7676         return ret;
7677 }
7678
7679 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7680                                         int rw, int mirror_num)
7681 {
7682         struct btrfs_root *root = BTRFS_I(inode)->root;
7683         int ret;
7684
7685         BUG_ON(rw & REQ_WRITE);
7686
7687         bio_get(bio);
7688
7689         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7690                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7691         if (ret)
7692                 goto err;
7693
7694         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7695 err:
7696         bio_put(bio);
7697         return ret;
7698 }
7699
7700 static int btrfs_check_dio_repairable(struct inode *inode,
7701                                       struct bio *failed_bio,
7702                                       struct io_failure_record *failrec,
7703                                       int failed_mirror)
7704 {
7705         int num_copies;
7706
7707         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7708                                       failrec->logical, failrec->len);
7709         if (num_copies == 1) {
7710                 /*
7711                  * we only have a single copy of the data, so don't bother with
7712                  * all the retry and error correction code that follows. no
7713                  * matter what the error is, it is very likely to persist.
7714                  */
7715                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7716                          num_copies, failrec->this_mirror, failed_mirror);
7717                 return 0;
7718         }
7719
7720         failrec->failed_mirror = failed_mirror;
7721         failrec->this_mirror++;
7722         if (failrec->this_mirror == failed_mirror)
7723                 failrec->this_mirror++;
7724
7725         if (failrec->this_mirror > num_copies) {
7726                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7727                          num_copies, failrec->this_mirror, failed_mirror);
7728                 return 0;
7729         }
7730
7731         return 1;
7732 }
7733
7734 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7735                           struct page *page, u64 start, u64 end,
7736                           int failed_mirror, bio_end_io_t *repair_endio,
7737                           void *repair_arg)
7738 {
7739         struct io_failure_record *failrec;
7740         struct bio *bio;
7741         int isector;
7742         int read_mode;
7743         int ret;
7744
7745         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7746
7747         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7748         if (ret)
7749                 return ret;
7750
7751         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7752                                          failed_mirror);
7753         if (!ret) {
7754                 free_io_failure(inode, failrec);
7755                 return -EIO;
7756         }
7757
7758         if (failed_bio->bi_vcnt > 1)
7759                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7760         else
7761                 read_mode = READ_SYNC;
7762
7763         isector = start - btrfs_io_bio(failed_bio)->logical;
7764         isector >>= inode->i_sb->s_blocksize_bits;
7765         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7766                                       0, isector, repair_endio, repair_arg);
7767         if (!bio) {
7768                 free_io_failure(inode, failrec);
7769                 return -EIO;
7770         }
7771
7772         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7773                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7774                     read_mode, failrec->this_mirror, failrec->in_validation);
7775
7776         ret = submit_dio_repair_bio(inode, bio, read_mode,
7777                                     failrec->this_mirror);
7778         if (ret) {
7779                 free_io_failure(inode, failrec);
7780                 bio_put(bio);
7781         }
7782
7783         return ret;
7784 }
7785
7786 struct btrfs_retry_complete {
7787         struct completion done;
7788         struct inode *inode;
7789         u64 start;
7790         int uptodate;
7791 };
7792
7793 static void btrfs_retry_endio_nocsum(struct bio *bio)
7794 {
7795         struct btrfs_retry_complete *done = bio->bi_private;
7796         struct bio_vec *bvec;
7797         int i;
7798
7799         if (bio->bi_error)
7800                 goto end;
7801
7802         done->uptodate = 1;
7803         bio_for_each_segment_all(bvec, bio, i)
7804                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7805 end:
7806         complete(&done->done);
7807         bio_put(bio);
7808 }
7809
7810 static int __btrfs_correct_data_nocsum(struct inode *inode,
7811                                        struct btrfs_io_bio *io_bio)
7812 {
7813         struct bio_vec *bvec;
7814         struct btrfs_retry_complete done;
7815         u64 start;
7816         int i;
7817         int ret;
7818
7819         start = io_bio->logical;
7820         done.inode = inode;
7821
7822         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7823 try_again:
7824                 done.uptodate = 0;
7825                 done.start = start;
7826                 init_completion(&done.done);
7827
7828                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7829                                      start + bvec->bv_len - 1,
7830                                      io_bio->mirror_num,
7831                                      btrfs_retry_endio_nocsum, &done);
7832                 if (ret)
7833                         return ret;
7834
7835                 wait_for_completion(&done.done);
7836
7837                 if (!done.uptodate) {
7838                         /* We might have another mirror, so try again */
7839                         goto try_again;
7840                 }
7841
7842                 start += bvec->bv_len;
7843         }
7844
7845         return 0;
7846 }
7847
7848 static void btrfs_retry_endio(struct bio *bio)
7849 {
7850         struct btrfs_retry_complete *done = bio->bi_private;
7851         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7852         struct bio_vec *bvec;
7853         int uptodate;
7854         int ret;
7855         int i;
7856
7857         if (bio->bi_error)
7858                 goto end;
7859
7860         uptodate = 1;
7861         bio_for_each_segment_all(bvec, bio, i) {
7862                 ret = __readpage_endio_check(done->inode, io_bio, i,
7863                                              bvec->bv_page, 0,
7864                                              done->start, bvec->bv_len);
7865                 if (!ret)
7866                         clean_io_failure(done->inode, done->start,
7867                                          bvec->bv_page, 0);
7868                 else
7869                         uptodate = 0;
7870         }
7871
7872         done->uptodate = uptodate;
7873 end:
7874         complete(&done->done);
7875         bio_put(bio);
7876 }
7877
7878 static int __btrfs_subio_endio_read(struct inode *inode,
7879                                     struct btrfs_io_bio *io_bio, int err)
7880 {
7881         struct bio_vec *bvec;
7882         struct btrfs_retry_complete done;
7883         u64 start;
7884         u64 offset = 0;
7885         int i;
7886         int ret;
7887
7888         err = 0;
7889         start = io_bio->logical;
7890         done.inode = inode;
7891
7892         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7893                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7894                                              0, start, bvec->bv_len);
7895                 if (likely(!ret))
7896                         goto next;
7897 try_again:
7898                 done.uptodate = 0;
7899                 done.start = start;
7900                 init_completion(&done.done);
7901
7902                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7903                                      start + bvec->bv_len - 1,
7904                                      io_bio->mirror_num,
7905                                      btrfs_retry_endio, &done);
7906                 if (ret) {
7907                         err = ret;
7908                         goto next;
7909                 }
7910
7911                 wait_for_completion(&done.done);
7912
7913                 if (!done.uptodate) {
7914                         /* We might have another mirror, so try again */
7915                         goto try_again;
7916                 }
7917 next:
7918                 offset += bvec->bv_len;
7919                 start += bvec->bv_len;
7920         }
7921
7922         return err;
7923 }
7924
7925 static int btrfs_subio_endio_read(struct inode *inode,
7926                                   struct btrfs_io_bio *io_bio, int err)
7927 {
7928         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7929
7930         if (skip_csum) {
7931                 if (unlikely(err))
7932                         return __btrfs_correct_data_nocsum(inode, io_bio);
7933                 else
7934                         return 0;
7935         } else {
7936                 return __btrfs_subio_endio_read(inode, io_bio, err);
7937         }
7938 }
7939
7940 static void btrfs_endio_direct_read(struct bio *bio)
7941 {
7942         struct btrfs_dio_private *dip = bio->bi_private;
7943         struct inode *inode = dip->inode;
7944         struct bio *dio_bio;
7945         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7946         int err = bio->bi_error;
7947
7948         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
7949                 err = btrfs_subio_endio_read(inode, io_bio, err);
7950
7951         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
7952                       dip->logical_offset + dip->bytes - 1);
7953         dio_bio = dip->dio_bio;
7954
7955         kfree(dip);
7956
7957         dio_end_io(dio_bio, bio->bi_error);
7958
7959         if (io_bio->end_io)
7960                 io_bio->end_io(io_bio, err);
7961         bio_put(bio);
7962 }
7963
7964 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
7965                                                     const u64 offset,
7966                                                     const u64 bytes,
7967                                                     const int uptodate)
7968 {
7969         struct btrfs_root *root = BTRFS_I(inode)->root;
7970         struct btrfs_ordered_extent *ordered = NULL;
7971         u64 ordered_offset = offset;
7972         u64 ordered_bytes = bytes;
7973         int ret;
7974
7975 again:
7976         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
7977                                                    &ordered_offset,
7978                                                    ordered_bytes,
7979                                                    uptodate);
7980         if (!ret)
7981                 goto out_test;
7982
7983         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
7984                         finish_ordered_fn, NULL, NULL);
7985         btrfs_queue_work(root->fs_info->endio_write_workers,
7986                          &ordered->work);
7987 out_test:
7988         /*
7989          * our bio might span multiple ordered extents.  If we haven't
7990          * completed the accounting for the whole dio, go back and try again
7991          */
7992         if (ordered_offset < offset + bytes) {
7993                 ordered_bytes = offset + bytes - ordered_offset;
7994                 ordered = NULL;
7995                 goto again;
7996         }
7997 }
7998
7999 static void btrfs_endio_direct_write(struct bio *bio)
8000 {
8001         struct btrfs_dio_private *dip = bio->bi_private;
8002         struct bio *dio_bio = dip->dio_bio;
8003
8004         btrfs_endio_direct_write_update_ordered(dip->inode,
8005                                                 dip->logical_offset,
8006                                                 dip->bytes,
8007                                                 !bio->bi_error);
8008
8009         kfree(dip);
8010
8011         dio_end_io(dio_bio, bio->bi_error);
8012         bio_put(bio);
8013 }
8014
8015 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8016                                     struct bio *bio, int mirror_num,
8017                                     unsigned long bio_flags, u64 offset)
8018 {
8019         int ret;
8020         struct btrfs_root *root = BTRFS_I(inode)->root;
8021         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8022         BUG_ON(ret); /* -ENOMEM */
8023         return 0;
8024 }
8025
8026 static void btrfs_end_dio_bio(struct bio *bio)
8027 {
8028         struct btrfs_dio_private *dip = bio->bi_private;
8029         int err = bio->bi_error;
8030
8031         if (err)
8032                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8033                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8034                            btrfs_ino(dip->inode), bio->bi_rw,
8035                            (unsigned long long)bio->bi_iter.bi_sector,
8036                            bio->bi_iter.bi_size, err);
8037
8038         if (dip->subio_endio)
8039                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8040
8041         if (err) {
8042                 dip->errors = 1;
8043
8044                 /*
8045                  * before atomic variable goto zero, we must make sure
8046                  * dip->errors is perceived to be set.
8047                  */
8048                 smp_mb__before_atomic();
8049         }
8050
8051         /* if there are more bios still pending for this dio, just exit */
8052         if (!atomic_dec_and_test(&dip->pending_bios))
8053                 goto out;
8054
8055         if (dip->errors) {
8056                 bio_io_error(dip->orig_bio);
8057         } else {
8058                 dip->dio_bio->bi_error = 0;
8059                 bio_endio(dip->orig_bio);
8060         }
8061 out:
8062         bio_put(bio);
8063 }
8064
8065 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8066                                        u64 first_sector, gfp_t gfp_flags)
8067 {
8068         struct bio *bio;
8069         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8070         if (bio)
8071                 bio_associate_current(bio);
8072         return bio;
8073 }
8074
8075 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8076                                                  struct inode *inode,
8077                                                  struct btrfs_dio_private *dip,
8078                                                  struct bio *bio,
8079                                                  u64 file_offset)
8080 {
8081         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8082         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8083         int ret;
8084
8085         /*
8086          * We load all the csum data we need when we submit
8087          * the first bio to reduce the csum tree search and
8088          * contention.
8089          */
8090         if (dip->logical_offset == file_offset) {
8091                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8092                                                 file_offset);
8093                 if (ret)
8094                         return ret;
8095         }
8096
8097         if (bio == dip->orig_bio)
8098                 return 0;
8099
8100         file_offset -= dip->logical_offset;
8101         file_offset >>= inode->i_sb->s_blocksize_bits;
8102         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8103
8104         return 0;
8105 }
8106
8107 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8108                                          int rw, u64 file_offset, int skip_sum,
8109                                          int async_submit)
8110 {
8111         struct btrfs_dio_private *dip = bio->bi_private;
8112         int write = rw & REQ_WRITE;
8113         struct btrfs_root *root = BTRFS_I(inode)->root;
8114         int ret;
8115
8116         if (async_submit)
8117                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8118
8119         bio_get(bio);
8120
8121         if (!write) {
8122                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8123                                 BTRFS_WQ_ENDIO_DATA);
8124                 if (ret)
8125                         goto err;
8126         }
8127
8128         if (skip_sum)
8129                 goto map;
8130
8131         if (write && async_submit) {
8132                 ret = btrfs_wq_submit_bio(root->fs_info,
8133                                    inode, rw, bio, 0, 0,
8134                                    file_offset,
8135                                    __btrfs_submit_bio_start_direct_io,
8136                                    __btrfs_submit_bio_done);
8137                 goto err;
8138         } else if (write) {
8139                 /*
8140                  * If we aren't doing async submit, calculate the csum of the
8141                  * bio now.
8142                  */
8143                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8144                 if (ret)
8145                         goto err;
8146         } else {
8147                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8148                                                      file_offset);
8149                 if (ret)
8150                         goto err;
8151         }
8152 map:
8153         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8154 err:
8155         bio_put(bio);
8156         return ret;
8157 }
8158
8159 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8160                                     int skip_sum)
8161 {
8162         struct inode *inode = dip->inode;
8163         struct btrfs_root *root = BTRFS_I(inode)->root;
8164         struct bio *bio;
8165         struct bio *orig_bio = dip->orig_bio;
8166         struct bio_vec *bvec = orig_bio->bi_io_vec;
8167         u64 start_sector = orig_bio->bi_iter.bi_sector;
8168         u64 file_offset = dip->logical_offset;
8169         u64 submit_len = 0;
8170         u64 map_length;
8171         int nr_pages = 0;
8172         int ret;
8173         int async_submit = 0;
8174
8175         map_length = orig_bio->bi_iter.bi_size;
8176         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8177                               &map_length, NULL, 0);
8178         if (ret)
8179                 return -EIO;
8180
8181         if (map_length >= orig_bio->bi_iter.bi_size) {
8182                 bio = orig_bio;
8183                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8184                 goto submit;
8185         }
8186
8187         /* async crcs make it difficult to collect full stripe writes. */
8188         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8189                 async_submit = 0;
8190         else
8191                 async_submit = 1;
8192
8193         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8194         if (!bio)
8195                 return -ENOMEM;
8196
8197         bio->bi_private = dip;
8198         bio->bi_end_io = btrfs_end_dio_bio;
8199         btrfs_io_bio(bio)->logical = file_offset;
8200         atomic_inc(&dip->pending_bios);
8201
8202         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8203                 if (map_length < submit_len + bvec->bv_len ||
8204                     bio_add_page(bio, bvec->bv_page, bvec->bv_len,
8205                                  bvec->bv_offset) < bvec->bv_len) {
8206                         /*
8207                          * inc the count before we submit the bio so
8208                          * we know the end IO handler won't happen before
8209                          * we inc the count. Otherwise, the dip might get freed
8210                          * before we're done setting it up
8211                          */
8212                         atomic_inc(&dip->pending_bios);
8213                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8214                                                      file_offset, skip_sum,
8215                                                      async_submit);
8216                         if (ret) {
8217                                 bio_put(bio);
8218                                 atomic_dec(&dip->pending_bios);
8219                                 goto out_err;
8220                         }
8221
8222                         start_sector += submit_len >> 9;
8223                         file_offset += submit_len;
8224
8225                         submit_len = 0;
8226                         nr_pages = 0;
8227
8228                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8229                                                   start_sector, GFP_NOFS);
8230                         if (!bio)
8231                                 goto out_err;
8232                         bio->bi_private = dip;
8233                         bio->bi_end_io = btrfs_end_dio_bio;
8234                         btrfs_io_bio(bio)->logical = file_offset;
8235
8236                         map_length = orig_bio->bi_iter.bi_size;
8237                         ret = btrfs_map_block(root->fs_info, rw,
8238                                               start_sector << 9,
8239                                               &map_length, NULL, 0);
8240                         if (ret) {
8241                                 bio_put(bio);
8242                                 goto out_err;
8243                         }
8244                 } else {
8245                         submit_len += bvec->bv_len;
8246                         nr_pages++;
8247                         bvec++;
8248                 }
8249         }
8250
8251 submit:
8252         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8253                                      async_submit);
8254         if (!ret)
8255                 return 0;
8256
8257         bio_put(bio);
8258 out_err:
8259         dip->errors = 1;
8260         /*
8261          * before atomic variable goto zero, we must
8262          * make sure dip->errors is perceived to be set.
8263          */
8264         smp_mb__before_atomic();
8265         if (atomic_dec_and_test(&dip->pending_bios))
8266                 bio_io_error(dip->orig_bio);
8267
8268         /* bio_end_io() will handle error, so we needn't return it */
8269         return 0;
8270 }
8271
8272 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8273                                 struct inode *inode, loff_t file_offset)
8274 {
8275         struct btrfs_dio_private *dip = NULL;
8276         struct bio *io_bio = NULL;
8277         struct btrfs_io_bio *btrfs_bio;
8278         int skip_sum;
8279         int write = rw & REQ_WRITE;
8280         int ret = 0;
8281
8282         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8283
8284         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8285         if (!io_bio) {
8286                 ret = -ENOMEM;
8287                 goto free_ordered;
8288         }
8289
8290         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8291         if (!dip) {
8292                 ret = -ENOMEM;
8293                 goto free_ordered;
8294         }
8295
8296         dip->private = dio_bio->bi_private;
8297         dip->inode = inode;
8298         dip->logical_offset = file_offset;
8299         dip->bytes = dio_bio->bi_iter.bi_size;
8300         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8301         io_bio->bi_private = dip;
8302         dip->orig_bio = io_bio;
8303         dip->dio_bio = dio_bio;
8304         atomic_set(&dip->pending_bios, 0);
8305         btrfs_bio = btrfs_io_bio(io_bio);
8306         btrfs_bio->logical = file_offset;
8307
8308         if (write) {
8309                 io_bio->bi_end_io = btrfs_endio_direct_write;
8310         } else {
8311                 io_bio->bi_end_io = btrfs_endio_direct_read;
8312                 dip->subio_endio = btrfs_subio_endio_read;
8313         }
8314
8315         /*
8316          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8317          * even if we fail to submit a bio, because in such case we do the
8318          * corresponding error handling below and it must not be done a second
8319          * time by btrfs_direct_IO().
8320          */
8321         if (write) {
8322                 struct btrfs_dio_data *dio_data = current->journal_info;
8323
8324                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8325                         dip->bytes;
8326                 dio_data->unsubmitted_oe_range_start =
8327                         dio_data->unsubmitted_oe_range_end;
8328         }
8329
8330         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8331         if (!ret)
8332                 return;
8333
8334         if (btrfs_bio->end_io)
8335                 btrfs_bio->end_io(btrfs_bio, ret);
8336
8337 free_ordered:
8338         /*
8339          * If we arrived here it means either we failed to submit the dip
8340          * or we either failed to clone the dio_bio or failed to allocate the
8341          * dip. If we cloned the dio_bio and allocated the dip, we can just
8342          * call bio_endio against our io_bio so that we get proper resource
8343          * cleanup if we fail to submit the dip, otherwise, we must do the
8344          * same as btrfs_endio_direct_[write|read] because we can't call these
8345          * callbacks - they require an allocated dip and a clone of dio_bio.
8346          */
8347         if (io_bio && dip) {
8348                 io_bio->bi_error = -EIO;
8349                 bio_endio(io_bio);
8350                 /*
8351                  * The end io callbacks free our dip, do the final put on io_bio
8352                  * and all the cleanup and final put for dio_bio (through
8353                  * dio_end_io()).
8354                  */
8355                 dip = NULL;
8356                 io_bio = NULL;
8357         } else {
8358                 if (write)
8359                         btrfs_endio_direct_write_update_ordered(inode,
8360                                                 file_offset,
8361                                                 dio_bio->bi_iter.bi_size,
8362                                                 0);
8363                 else
8364                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8365                               file_offset + dio_bio->bi_iter.bi_size - 1);
8366
8367                 dio_bio->bi_error = -EIO;
8368                 /*
8369                  * Releases and cleans up our dio_bio, no need to bio_put()
8370                  * nor bio_endio()/bio_io_error() against dio_bio.
8371                  */
8372                 dio_end_io(dio_bio, ret);
8373         }
8374         if (io_bio)
8375                 bio_put(io_bio);
8376         kfree(dip);
8377 }
8378
8379 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8380                         const struct iov_iter *iter, loff_t offset)
8381 {
8382         int seg;
8383         int i;
8384         unsigned blocksize_mask = root->sectorsize - 1;
8385         ssize_t retval = -EINVAL;
8386
8387         if (offset & blocksize_mask)
8388                 goto out;
8389
8390         if (iov_iter_alignment(iter) & blocksize_mask)
8391                 goto out;
8392
8393         /* If this is a write we don't need to check anymore */
8394         if (iov_iter_rw(iter) == WRITE)
8395                 return 0;
8396         /*
8397          * Check to make sure we don't have duplicate iov_base's in this
8398          * iovec, if so return EINVAL, otherwise we'll get csum errors
8399          * when reading back.
8400          */
8401         for (seg = 0; seg < iter->nr_segs; seg++) {
8402                 for (i = seg + 1; i < iter->nr_segs; i++) {
8403                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8404                                 goto out;
8405                 }
8406         }
8407         retval = 0;
8408 out:
8409         return retval;
8410 }
8411
8412 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
8413                                loff_t offset)
8414 {
8415         struct file *file = iocb->ki_filp;
8416         struct inode *inode = file->f_mapping->host;
8417         struct btrfs_root *root = BTRFS_I(inode)->root;
8418         struct btrfs_dio_data dio_data = { 0 };
8419         size_t count = 0;
8420         int flags = 0;
8421         bool wakeup = true;
8422         bool relock = false;
8423         ssize_t ret;
8424
8425         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8426                 return 0;
8427
8428         inode_dio_begin(inode);
8429         smp_mb__after_atomic();
8430
8431         /*
8432          * The generic stuff only does filemap_write_and_wait_range, which
8433          * isn't enough if we've written compressed pages to this area, so
8434          * we need to flush the dirty pages again to make absolutely sure
8435          * that any outstanding dirty pages are on disk.
8436          */
8437         count = iov_iter_count(iter);
8438         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8439                      &BTRFS_I(inode)->runtime_flags))
8440                 filemap_fdatawrite_range(inode->i_mapping, offset,
8441                                          offset + count - 1);
8442
8443         if (iov_iter_rw(iter) == WRITE) {
8444                 /*
8445                  * If the write DIO is beyond the EOF, we need update
8446                  * the isize, but it is protected by i_mutex. So we can
8447                  * not unlock the i_mutex at this case.
8448                  */
8449                 if (offset + count <= inode->i_size) {
8450                         mutex_unlock(&inode->i_mutex);
8451                         relock = true;
8452                 }
8453                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8454                 if (ret)
8455                         goto out;
8456                 dio_data.outstanding_extents = div64_u64(count +
8457                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8458                                                 BTRFS_MAX_EXTENT_SIZE);
8459
8460                 /*
8461                  * We need to know how many extents we reserved so that we can
8462                  * do the accounting properly if we go over the number we
8463                  * originally calculated.  Abuse current->journal_info for this.
8464                  */
8465                 dio_data.reserve = round_up(count, root->sectorsize);
8466                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8467                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8468                 current->journal_info = &dio_data;
8469         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8470                                      &BTRFS_I(inode)->runtime_flags)) {
8471                 inode_dio_end(inode);
8472                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8473                 wakeup = false;
8474         }
8475
8476         ret = __blockdev_direct_IO(iocb, inode,
8477                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8478                                    iter, offset, btrfs_get_blocks_direct, NULL,
8479                                    btrfs_submit_direct, flags);
8480         if (iov_iter_rw(iter) == WRITE) {
8481                 current->journal_info = NULL;
8482                 if (ret < 0 && ret != -EIOCBQUEUED) {
8483                         if (dio_data.reserve)
8484                                 btrfs_delalloc_release_space(inode, offset,
8485                                                              dio_data.reserve);
8486                         /*
8487                          * On error we might have left some ordered extents
8488                          * without submitting corresponding bios for them, so
8489                          * cleanup them up to avoid other tasks getting them
8490                          * and waiting for them to complete forever.
8491                          */
8492                         if (dio_data.unsubmitted_oe_range_start <
8493                             dio_data.unsubmitted_oe_range_end)
8494                                 btrfs_endio_direct_write_update_ordered(inode,
8495                                         dio_data.unsubmitted_oe_range_start,
8496                                         dio_data.unsubmitted_oe_range_end -
8497                                         dio_data.unsubmitted_oe_range_start,
8498                                         0);
8499                 } else if (ret >= 0 && (size_t)ret < count)
8500                         btrfs_delalloc_release_space(inode, offset,
8501                                                      count - (size_t)ret);
8502         }
8503 out:
8504         if (wakeup)
8505                 inode_dio_end(inode);
8506         if (relock)
8507                 mutex_lock(&inode->i_mutex);
8508
8509         return ret;
8510 }
8511
8512 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8513
8514 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8515                 __u64 start, __u64 len)
8516 {
8517         int     ret;
8518
8519         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8520         if (ret)
8521                 return ret;
8522
8523         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8524 }
8525
8526 int btrfs_readpage(struct file *file, struct page *page)
8527 {
8528         struct extent_io_tree *tree;
8529         tree = &BTRFS_I(page->mapping->host)->io_tree;
8530         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8531 }
8532
8533 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8534 {
8535         struct extent_io_tree *tree;
8536         struct inode *inode = page->mapping->host;
8537         int ret;
8538
8539         if (current->flags & PF_MEMALLOC) {
8540                 redirty_page_for_writepage(wbc, page);
8541                 unlock_page(page);
8542                 return 0;
8543         }
8544
8545         /*
8546          * If we are under memory pressure we will call this directly from the
8547          * VM, we need to make sure we have the inode referenced for the ordered
8548          * extent.  If not just return like we didn't do anything.
8549          */
8550         if (!igrab(inode)) {
8551                 redirty_page_for_writepage(wbc, page);
8552                 return AOP_WRITEPAGE_ACTIVATE;
8553         }
8554         tree = &BTRFS_I(page->mapping->host)->io_tree;
8555         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8556         btrfs_add_delayed_iput(inode);
8557         return ret;
8558 }
8559
8560 static int btrfs_writepages(struct address_space *mapping,
8561                             struct writeback_control *wbc)
8562 {
8563         struct extent_io_tree *tree;
8564
8565         tree = &BTRFS_I(mapping->host)->io_tree;
8566         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8567 }
8568
8569 static int
8570 btrfs_readpages(struct file *file, struct address_space *mapping,
8571                 struct list_head *pages, unsigned nr_pages)
8572 {
8573         struct extent_io_tree *tree;
8574         tree = &BTRFS_I(mapping->host)->io_tree;
8575         return extent_readpages(tree, mapping, pages, nr_pages,
8576                                 btrfs_get_extent);
8577 }
8578 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8579 {
8580         struct extent_io_tree *tree;
8581         struct extent_map_tree *map;
8582         int ret;
8583
8584         tree = &BTRFS_I(page->mapping->host)->io_tree;
8585         map = &BTRFS_I(page->mapping->host)->extent_tree;
8586         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8587         if (ret == 1) {
8588                 ClearPagePrivate(page);
8589                 set_page_private(page, 0);
8590                 page_cache_release(page);
8591         }
8592         return ret;
8593 }
8594
8595 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8596 {
8597         if (PageWriteback(page) || PageDirty(page))
8598                 return 0;
8599         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8600 }
8601
8602 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8603                                  unsigned int length)
8604 {
8605         struct inode *inode = page->mapping->host;
8606         struct extent_io_tree *tree;
8607         struct btrfs_ordered_extent *ordered;
8608         struct extent_state *cached_state = NULL;
8609         u64 page_start = page_offset(page);
8610         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
8611         int inode_evicting = inode->i_state & I_FREEING;
8612
8613         /*
8614          * we have the page locked, so new writeback can't start,
8615          * and the dirty bit won't be cleared while we are here.
8616          *
8617          * Wait for IO on this page so that we can safely clear
8618          * the PagePrivate2 bit and do ordered accounting
8619          */
8620         wait_on_page_writeback(page);
8621
8622         tree = &BTRFS_I(inode)->io_tree;
8623         if (offset) {
8624                 btrfs_releasepage(page, GFP_NOFS);
8625                 return;
8626         }
8627
8628         if (!inode_evicting)
8629                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8630         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8631         if (ordered) {
8632                 /*
8633                  * IO on this page will never be started, so we need
8634                  * to account for any ordered extents now
8635                  */
8636                 if (!inode_evicting)
8637                         clear_extent_bit(tree, page_start, page_end,
8638                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8639                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8640                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8641                                          GFP_NOFS);
8642                 /*
8643                  * whoever cleared the private bit is responsible
8644                  * for the finish_ordered_io
8645                  */
8646                 if (TestClearPagePrivate2(page)) {
8647                         struct btrfs_ordered_inode_tree *tree;
8648                         u64 new_len;
8649
8650                         tree = &BTRFS_I(inode)->ordered_tree;
8651
8652                         spin_lock_irq(&tree->lock);
8653                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8654                         new_len = page_start - ordered->file_offset;
8655                         if (new_len < ordered->truncated_len)
8656                                 ordered->truncated_len = new_len;
8657                         spin_unlock_irq(&tree->lock);
8658
8659                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8660                                                            page_start,
8661                                                            PAGE_CACHE_SIZE, 1))
8662                                 btrfs_finish_ordered_io(ordered);
8663                 }
8664                 btrfs_put_ordered_extent(ordered);
8665                 if (!inode_evicting) {
8666                         cached_state = NULL;
8667                         lock_extent_bits(tree, page_start, page_end,
8668                                          &cached_state);
8669                 }
8670         }
8671
8672         /*
8673          * Qgroup reserved space handler
8674          * Page here will be either
8675          * 1) Already written to disk
8676          *    In this case, its reserved space is released from data rsv map
8677          *    and will be freed by delayed_ref handler finally.
8678          *    So even we call qgroup_free_data(), it won't decrease reserved
8679          *    space.
8680          * 2) Not written to disk
8681          *    This means the reserved space should be freed here.
8682          */
8683         btrfs_qgroup_free_data(inode, page_start, PAGE_CACHE_SIZE);
8684         if (!inode_evicting) {
8685                 clear_extent_bit(tree, page_start, page_end,
8686                                  EXTENT_LOCKED | EXTENT_DIRTY |
8687                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8688                                  EXTENT_DEFRAG, 1, 1,
8689                                  &cached_state, GFP_NOFS);
8690
8691                 __btrfs_releasepage(page, GFP_NOFS);
8692         }
8693
8694         ClearPageChecked(page);
8695         if (PagePrivate(page)) {
8696                 ClearPagePrivate(page);
8697                 set_page_private(page, 0);
8698                 page_cache_release(page);
8699         }
8700 }
8701
8702 /*
8703  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8704  * called from a page fault handler when a page is first dirtied. Hence we must
8705  * be careful to check for EOF conditions here. We set the page up correctly
8706  * for a written page which means we get ENOSPC checking when writing into
8707  * holes and correct delalloc and unwritten extent mapping on filesystems that
8708  * support these features.
8709  *
8710  * We are not allowed to take the i_mutex here so we have to play games to
8711  * protect against truncate races as the page could now be beyond EOF.  Because
8712  * vmtruncate() writes the inode size before removing pages, once we have the
8713  * page lock we can determine safely if the page is beyond EOF. If it is not
8714  * beyond EOF, then the page is guaranteed safe against truncation until we
8715  * unlock the page.
8716  */
8717 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8718 {
8719         struct page *page = vmf->page;
8720         struct inode *inode = file_inode(vma->vm_file);
8721         struct btrfs_root *root = BTRFS_I(inode)->root;
8722         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8723         struct btrfs_ordered_extent *ordered;
8724         struct extent_state *cached_state = NULL;
8725         char *kaddr;
8726         unsigned long zero_start;
8727         loff_t size;
8728         int ret;
8729         int reserved = 0;
8730         u64 page_start;
8731         u64 page_end;
8732
8733         sb_start_pagefault(inode->i_sb);
8734         page_start = page_offset(page);
8735         page_end = page_start + PAGE_CACHE_SIZE - 1;
8736
8737         ret = btrfs_delalloc_reserve_space(inode, page_start,
8738                                            PAGE_CACHE_SIZE);
8739         if (!ret) {
8740                 ret = file_update_time(vma->vm_file);
8741                 reserved = 1;
8742         }
8743         if (ret) {
8744                 if (ret == -ENOMEM)
8745                         ret = VM_FAULT_OOM;
8746                 else /* -ENOSPC, -EIO, etc */
8747                         ret = VM_FAULT_SIGBUS;
8748                 if (reserved)
8749                         goto out;
8750                 goto out_noreserve;
8751         }
8752
8753         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8754 again:
8755         lock_page(page);
8756         size = i_size_read(inode);
8757
8758         if ((page->mapping != inode->i_mapping) ||
8759             (page_start >= size)) {
8760                 /* page got truncated out from underneath us */
8761                 goto out_unlock;
8762         }
8763         wait_on_page_writeback(page);
8764
8765         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8766         set_page_extent_mapped(page);
8767
8768         /*
8769          * we can't set the delalloc bits if there are pending ordered
8770          * extents.  Drop our locks and wait for them to finish
8771          */
8772         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8773         if (ordered) {
8774                 unlock_extent_cached(io_tree, page_start, page_end,
8775                                      &cached_state, GFP_NOFS);
8776                 unlock_page(page);
8777                 btrfs_start_ordered_extent(inode, ordered, 1);
8778                 btrfs_put_ordered_extent(ordered);
8779                 goto again;
8780         }
8781
8782         /*
8783          * XXX - page_mkwrite gets called every time the page is dirtied, even
8784          * if it was already dirty, so for space accounting reasons we need to
8785          * clear any delalloc bits for the range we are fixing to save.  There
8786          * is probably a better way to do this, but for now keep consistent with
8787          * prepare_pages in the normal write path.
8788          */
8789         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
8790                           EXTENT_DIRTY | EXTENT_DELALLOC |
8791                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8792                           0, 0, &cached_state, GFP_NOFS);
8793
8794         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
8795                                         &cached_state);
8796         if (ret) {
8797                 unlock_extent_cached(io_tree, page_start, page_end,
8798                                      &cached_state, GFP_NOFS);
8799                 ret = VM_FAULT_SIGBUS;
8800                 goto out_unlock;
8801         }
8802         ret = 0;
8803
8804         /* page is wholly or partially inside EOF */
8805         if (page_start + PAGE_CACHE_SIZE > size)
8806                 zero_start = size & ~PAGE_CACHE_MASK;
8807         else
8808                 zero_start = PAGE_CACHE_SIZE;
8809
8810         if (zero_start != PAGE_CACHE_SIZE) {
8811                 kaddr = kmap(page);
8812                 memset(kaddr + zero_start, 0, PAGE_CACHE_SIZE - zero_start);
8813                 flush_dcache_page(page);
8814                 kunmap(page);
8815         }
8816         ClearPageChecked(page);
8817         set_page_dirty(page);
8818         SetPageUptodate(page);
8819
8820         BTRFS_I(inode)->last_trans = root->fs_info->generation;
8821         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8822         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8823
8824         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
8825
8826 out_unlock:
8827         if (!ret) {
8828                 sb_end_pagefault(inode->i_sb);
8829                 return VM_FAULT_LOCKED;
8830         }
8831         unlock_page(page);
8832 out:
8833         btrfs_delalloc_release_space(inode, page_start, PAGE_CACHE_SIZE);
8834 out_noreserve:
8835         sb_end_pagefault(inode->i_sb);
8836         return ret;
8837 }
8838
8839 static int btrfs_truncate(struct inode *inode)
8840 {
8841         struct btrfs_root *root = BTRFS_I(inode)->root;
8842         struct btrfs_block_rsv *rsv;
8843         int ret = 0;
8844         int err = 0;
8845         struct btrfs_trans_handle *trans;
8846         u64 mask = root->sectorsize - 1;
8847         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
8848
8849         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8850                                        (u64)-1);
8851         if (ret)
8852                 return ret;
8853
8854         /*
8855          * Yes ladies and gentelment, this is indeed ugly.  The fact is we have
8856          * 3 things going on here
8857          *
8858          * 1) We need to reserve space for our orphan item and the space to
8859          * delete our orphan item.  Lord knows we don't want to have a dangling
8860          * orphan item because we didn't reserve space to remove it.
8861          *
8862          * 2) We need to reserve space to update our inode.
8863          *
8864          * 3) We need to have something to cache all the space that is going to
8865          * be free'd up by the truncate operation, but also have some slack
8866          * space reserved in case it uses space during the truncate (thank you
8867          * very much snapshotting).
8868          *
8869          * And we need these to all be seperate.  The fact is we can use alot of
8870          * space doing the truncate, and we have no earthly idea how much space
8871          * we will use, so we need the truncate reservation to be seperate so it
8872          * doesn't end up using space reserved for updating the inode or
8873          * removing the orphan item.  We also need to be able to stop the
8874          * transaction and start a new one, which means we need to be able to
8875          * update the inode several times, and we have no idea of knowing how
8876          * many times that will be, so we can't just reserve 1 item for the
8877          * entirety of the opration, so that has to be done seperately as well.
8878          * Then there is the orphan item, which does indeed need to be held on
8879          * to for the whole operation, and we need nobody to touch this reserved
8880          * space except the orphan code.
8881          *
8882          * So that leaves us with
8883          *
8884          * 1) root->orphan_block_rsv - for the orphan deletion.
8885          * 2) rsv - for the truncate reservation, which we will steal from the
8886          * transaction reservation.
8887          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
8888          * updating the inode.
8889          */
8890         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
8891         if (!rsv)
8892                 return -ENOMEM;
8893         rsv->size = min_size;
8894         rsv->failfast = 1;
8895
8896         /*
8897          * 1 for the truncate slack space
8898          * 1 for updating the inode.
8899          */
8900         trans = btrfs_start_transaction(root, 2);
8901         if (IS_ERR(trans)) {
8902                 err = PTR_ERR(trans);
8903                 goto out;
8904         }
8905
8906         /* Migrate the slack space for the truncate to our reserve */
8907         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
8908                                       min_size);
8909         BUG_ON(ret);
8910
8911         /*
8912          * So if we truncate and then write and fsync we normally would just
8913          * write the extents that changed, which is a problem if we need to
8914          * first truncate that entire inode.  So set this flag so we write out
8915          * all of the extents in the inode to the sync log so we're completely
8916          * safe.
8917          */
8918         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8919         trans->block_rsv = rsv;
8920
8921         while (1) {
8922                 ret = btrfs_truncate_inode_items(trans, root, inode,
8923                                                  inode->i_size,
8924                                                  BTRFS_EXTENT_DATA_KEY);
8925                 if (ret != -ENOSPC && ret != -EAGAIN) {
8926                         err = ret;
8927                         break;
8928                 }
8929
8930                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8931                 ret = btrfs_update_inode(trans, root, inode);
8932                 if (ret) {
8933                         err = ret;
8934                         break;
8935                 }
8936
8937                 btrfs_end_transaction(trans, root);
8938                 btrfs_btree_balance_dirty(root);
8939
8940                 trans = btrfs_start_transaction(root, 2);
8941                 if (IS_ERR(trans)) {
8942                         ret = err = PTR_ERR(trans);
8943                         trans = NULL;
8944                         break;
8945                 }
8946
8947                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
8948                                               rsv, min_size);
8949                 BUG_ON(ret);    /* shouldn't happen */
8950                 trans->block_rsv = rsv;
8951         }
8952
8953         if (ret == 0 && inode->i_nlink > 0) {
8954                 trans->block_rsv = root->orphan_block_rsv;
8955                 ret = btrfs_orphan_del(trans, inode);
8956                 if (ret)
8957                         err = ret;
8958         }
8959
8960         if (trans) {
8961                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8962                 ret = btrfs_update_inode(trans, root, inode);
8963                 if (ret && !err)
8964                         err = ret;
8965
8966                 ret = btrfs_end_transaction(trans, root);
8967                 btrfs_btree_balance_dirty(root);
8968         }
8969
8970 out:
8971         btrfs_free_block_rsv(root, rsv);
8972
8973         if (ret && !err)
8974                 err = ret;
8975
8976         return err;
8977 }
8978
8979 /*
8980  * create a new subvolume directory/inode (helper for the ioctl).
8981  */
8982 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
8983                              struct btrfs_root *new_root,
8984                              struct btrfs_root *parent_root,
8985                              u64 new_dirid)
8986 {
8987         struct inode *inode;
8988         int err;
8989         u64 index = 0;
8990
8991         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
8992                                 new_dirid, new_dirid,
8993                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
8994                                 &index);
8995         if (IS_ERR(inode))
8996                 return PTR_ERR(inode);
8997         inode->i_op = &btrfs_dir_inode_operations;
8998         inode->i_fop = &btrfs_dir_file_operations;
8999
9000         set_nlink(inode, 1);
9001         btrfs_i_size_write(inode, 0);
9002         unlock_new_inode(inode);
9003
9004         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9005         if (err)
9006                 btrfs_err(new_root->fs_info,
9007                           "error inheriting subvolume %llu properties: %d",
9008                           new_root->root_key.objectid, err);
9009
9010         err = btrfs_update_inode(trans, new_root, inode);
9011
9012         iput(inode);
9013         return err;
9014 }
9015
9016 struct inode *btrfs_alloc_inode(struct super_block *sb)
9017 {
9018         struct btrfs_inode *ei;
9019         struct inode *inode;
9020
9021         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9022         if (!ei)
9023                 return NULL;
9024
9025         ei->root = NULL;
9026         ei->generation = 0;
9027         ei->last_trans = 0;
9028         ei->last_sub_trans = 0;
9029         ei->logged_trans = 0;
9030         ei->delalloc_bytes = 0;
9031         ei->defrag_bytes = 0;
9032         ei->disk_i_size = 0;
9033         ei->flags = 0;
9034         ei->csum_bytes = 0;
9035         ei->index_cnt = (u64)-1;
9036         ei->dir_index = 0;
9037         ei->last_unlink_trans = 0;
9038         ei->last_log_commit = 0;
9039         ei->delayed_iput_count = 0;
9040
9041         spin_lock_init(&ei->lock);
9042         ei->outstanding_extents = 0;
9043         ei->reserved_extents = 0;
9044
9045         ei->runtime_flags = 0;
9046         ei->force_compress = BTRFS_COMPRESS_NONE;
9047
9048         ei->delayed_node = NULL;
9049
9050         ei->i_otime.tv_sec = 0;
9051         ei->i_otime.tv_nsec = 0;
9052
9053         inode = &ei->vfs_inode;
9054         extent_map_tree_init(&ei->extent_tree);
9055         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9056         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9057         ei->io_tree.track_uptodate = 1;
9058         ei->io_failure_tree.track_uptodate = 1;
9059         atomic_set(&ei->sync_writers, 0);
9060         mutex_init(&ei->log_mutex);
9061         mutex_init(&ei->delalloc_mutex);
9062         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9063         INIT_LIST_HEAD(&ei->delalloc_inodes);
9064         INIT_LIST_HEAD(&ei->delayed_iput);
9065         RB_CLEAR_NODE(&ei->rb_node);
9066
9067         return inode;
9068 }
9069
9070 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9071 void btrfs_test_destroy_inode(struct inode *inode)
9072 {
9073         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9074         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9075 }
9076 #endif
9077
9078 static void btrfs_i_callback(struct rcu_head *head)
9079 {
9080         struct inode *inode = container_of(head, struct inode, i_rcu);
9081         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9082 }
9083
9084 void btrfs_destroy_inode(struct inode *inode)
9085 {
9086         struct btrfs_ordered_extent *ordered;
9087         struct btrfs_root *root = BTRFS_I(inode)->root;
9088
9089         WARN_ON(!hlist_empty(&inode->i_dentry));
9090         WARN_ON(inode->i_data.nrpages);
9091         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9092         WARN_ON(BTRFS_I(inode)->reserved_extents);
9093         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9094         WARN_ON(BTRFS_I(inode)->csum_bytes);
9095         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9096
9097         /*
9098          * This can happen where we create an inode, but somebody else also
9099          * created the same inode and we need to destroy the one we already
9100          * created.
9101          */
9102         if (!root)
9103                 goto free;
9104
9105         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9106                      &BTRFS_I(inode)->runtime_flags)) {
9107                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9108                         btrfs_ino(inode));
9109                 atomic_dec(&root->orphan_inodes);
9110         }
9111
9112         while (1) {
9113                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9114                 if (!ordered)
9115                         break;
9116                 else {
9117                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9118                                 ordered->file_offset, ordered->len);
9119                         btrfs_remove_ordered_extent(inode, ordered);
9120                         btrfs_put_ordered_extent(ordered);
9121                         btrfs_put_ordered_extent(ordered);
9122                 }
9123         }
9124         btrfs_qgroup_check_reserved_leak(inode);
9125         inode_tree_del(inode);
9126         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9127 free:
9128         call_rcu(&inode->i_rcu, btrfs_i_callback);
9129 }
9130
9131 int btrfs_drop_inode(struct inode *inode)
9132 {
9133         struct btrfs_root *root = BTRFS_I(inode)->root;
9134
9135         if (root == NULL)
9136                 return 1;
9137
9138         /* the snap/subvol tree is on deleting */
9139         if (btrfs_root_refs(&root->root_item) == 0)
9140                 return 1;
9141         else
9142                 return generic_drop_inode(inode);
9143 }
9144
9145 static void init_once(void *foo)
9146 {
9147         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9148
9149         inode_init_once(&ei->vfs_inode);
9150 }
9151
9152 void btrfs_destroy_cachep(void)
9153 {
9154         /*
9155          * Make sure all delayed rcu free inodes are flushed before we
9156          * destroy cache.
9157          */
9158         rcu_barrier();
9159         if (btrfs_inode_cachep)
9160                 kmem_cache_destroy(btrfs_inode_cachep);
9161         if (btrfs_trans_handle_cachep)
9162                 kmem_cache_destroy(btrfs_trans_handle_cachep);
9163         if (btrfs_transaction_cachep)
9164                 kmem_cache_destroy(btrfs_transaction_cachep);
9165         if (btrfs_path_cachep)
9166                 kmem_cache_destroy(btrfs_path_cachep);
9167         if (btrfs_free_space_cachep)
9168                 kmem_cache_destroy(btrfs_free_space_cachep);
9169 }
9170
9171 int btrfs_init_cachep(void)
9172 {
9173         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9174                         sizeof(struct btrfs_inode), 0,
9175                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9176                         init_once);
9177         if (!btrfs_inode_cachep)
9178                 goto fail;
9179
9180         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9181                         sizeof(struct btrfs_trans_handle), 0,
9182                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9183         if (!btrfs_trans_handle_cachep)
9184                 goto fail;
9185
9186         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9187                         sizeof(struct btrfs_transaction), 0,
9188                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9189         if (!btrfs_transaction_cachep)
9190                 goto fail;
9191
9192         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9193                         sizeof(struct btrfs_path), 0,
9194                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9195         if (!btrfs_path_cachep)
9196                 goto fail;
9197
9198         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9199                         sizeof(struct btrfs_free_space), 0,
9200                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9201         if (!btrfs_free_space_cachep)
9202                 goto fail;
9203
9204         return 0;
9205 fail:
9206         btrfs_destroy_cachep();
9207         return -ENOMEM;
9208 }
9209
9210 static int btrfs_getattr(struct vfsmount *mnt,
9211                          struct dentry *dentry, struct kstat *stat)
9212 {
9213         u64 delalloc_bytes;
9214         struct inode *inode = d_inode(dentry);
9215         u32 blocksize = inode->i_sb->s_blocksize;
9216
9217         generic_fillattr(inode, stat);
9218         stat->dev = BTRFS_I(inode)->root->anon_dev;
9219         stat->blksize = PAGE_CACHE_SIZE;
9220
9221         spin_lock(&BTRFS_I(inode)->lock);
9222         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9223         spin_unlock(&BTRFS_I(inode)->lock);
9224         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9225                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9226         return 0;
9227 }
9228
9229 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9230                            struct inode *new_dir, struct dentry *new_dentry)
9231 {
9232         struct btrfs_trans_handle *trans;
9233         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9234         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9235         struct inode *new_inode = d_inode(new_dentry);
9236         struct inode *old_inode = d_inode(old_dentry);
9237         struct timespec ctime = CURRENT_TIME;
9238         u64 index = 0;
9239         u64 root_objectid;
9240         int ret;
9241         u64 old_ino = btrfs_ino(old_inode);
9242
9243         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9244                 return -EPERM;
9245
9246         /* we only allow rename subvolume link between subvolumes */
9247         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9248                 return -EXDEV;
9249
9250         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9251             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9252                 return -ENOTEMPTY;
9253
9254         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9255             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9256                 return -ENOTEMPTY;
9257
9258
9259         /* check for collisions, even if the  name isn't there */
9260         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9261                              new_dentry->d_name.name,
9262                              new_dentry->d_name.len);
9263
9264         if (ret) {
9265                 if (ret == -EEXIST) {
9266                         /* we shouldn't get
9267                          * eexist without a new_inode */
9268                         if (WARN_ON(!new_inode)) {
9269                                 return ret;
9270                         }
9271                 } else {
9272                         /* maybe -EOVERFLOW */
9273                         return ret;
9274                 }
9275         }
9276         ret = 0;
9277
9278         /*
9279          * we're using rename to replace one file with another.  Start IO on it
9280          * now so  we don't add too much work to the end of the transaction
9281          */
9282         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9283                 filemap_flush(old_inode->i_mapping);
9284
9285         /* close the racy window with snapshot create/destroy ioctl */
9286         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9287                 down_read(&root->fs_info->subvol_sem);
9288         /*
9289          * We want to reserve the absolute worst case amount of items.  So if
9290          * both inodes are subvols and we need to unlink them then that would
9291          * require 4 item modifications, but if they are both normal inodes it
9292          * would require 5 item modifications, so we'll assume their normal
9293          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9294          * should cover the worst case number of items we'll modify.
9295          */
9296         trans = btrfs_start_transaction(root, 11);
9297         if (IS_ERR(trans)) {
9298                 ret = PTR_ERR(trans);
9299                 goto out_notrans;
9300         }
9301
9302         if (dest != root)
9303                 btrfs_record_root_in_trans(trans, dest);
9304
9305         ret = btrfs_set_inode_index(new_dir, &index);
9306         if (ret)
9307                 goto out_fail;
9308
9309         BTRFS_I(old_inode)->dir_index = 0ULL;
9310         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9311                 /* force full log commit if subvolume involved. */
9312                 btrfs_set_log_full_commit(root->fs_info, trans);
9313         } else {
9314                 ret = btrfs_insert_inode_ref(trans, dest,
9315                                              new_dentry->d_name.name,
9316                                              new_dentry->d_name.len,
9317                                              old_ino,
9318                                              btrfs_ino(new_dir), index);
9319                 if (ret)
9320                         goto out_fail;
9321                 /*
9322                  * this is an ugly little race, but the rename is required
9323                  * to make sure that if we crash, the inode is either at the
9324                  * old name or the new one.  pinning the log transaction lets
9325                  * us make sure we don't allow a log commit to come in after
9326                  * we unlink the name but before we add the new name back in.
9327                  */
9328                 btrfs_pin_log_trans(root);
9329         }
9330
9331         inode_inc_iversion(old_dir);
9332         inode_inc_iversion(new_dir);
9333         inode_inc_iversion(old_inode);
9334         old_dir->i_ctime = old_dir->i_mtime = ctime;
9335         new_dir->i_ctime = new_dir->i_mtime = ctime;
9336         old_inode->i_ctime = ctime;
9337
9338         if (old_dentry->d_parent != new_dentry->d_parent)
9339                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9340
9341         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9342                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9343                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9344                                         old_dentry->d_name.name,
9345                                         old_dentry->d_name.len);
9346         } else {
9347                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9348                                         d_inode(old_dentry),
9349                                         old_dentry->d_name.name,
9350                                         old_dentry->d_name.len);
9351                 if (!ret)
9352                         ret = btrfs_update_inode(trans, root, old_inode);
9353         }
9354         if (ret) {
9355                 btrfs_abort_transaction(trans, root, ret);
9356                 goto out_fail;
9357         }
9358
9359         if (new_inode) {
9360                 inode_inc_iversion(new_inode);
9361                 new_inode->i_ctime = CURRENT_TIME;
9362                 if (unlikely(btrfs_ino(new_inode) ==
9363                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9364                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9365                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9366                                                 root_objectid,
9367                                                 new_dentry->d_name.name,
9368                                                 new_dentry->d_name.len);
9369                         BUG_ON(new_inode->i_nlink == 0);
9370                 } else {
9371                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9372                                                  d_inode(new_dentry),
9373                                                  new_dentry->d_name.name,
9374                                                  new_dentry->d_name.len);
9375                 }
9376                 if (!ret && new_inode->i_nlink == 0)
9377                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9378                 if (ret) {
9379                         btrfs_abort_transaction(trans, root, ret);
9380                         goto out_fail;
9381                 }
9382         }
9383
9384         ret = btrfs_add_link(trans, new_dir, old_inode,
9385                              new_dentry->d_name.name,
9386                              new_dentry->d_name.len, 0, index);
9387         if (ret) {
9388                 btrfs_abort_transaction(trans, root, ret);
9389                 goto out_fail;
9390         }
9391
9392         if (old_inode->i_nlink == 1)
9393                 BTRFS_I(old_inode)->dir_index = index;
9394
9395         if (old_ino != BTRFS_FIRST_FREE_OBJECTID) {
9396                 struct dentry *parent = new_dentry->d_parent;
9397                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9398                 btrfs_end_log_trans(root);
9399         }
9400 out_fail:
9401         btrfs_end_transaction(trans, root);
9402 out_notrans:
9403         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9404                 up_read(&root->fs_info->subvol_sem);
9405
9406         return ret;
9407 }
9408
9409 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9410                          struct inode *new_dir, struct dentry *new_dentry,
9411                          unsigned int flags)
9412 {
9413         if (flags & ~RENAME_NOREPLACE)
9414                 return -EINVAL;
9415
9416         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry);
9417 }
9418
9419 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9420 {
9421         struct btrfs_delalloc_work *delalloc_work;
9422         struct inode *inode;
9423
9424         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9425                                      work);
9426         inode = delalloc_work->inode;
9427         filemap_flush(inode->i_mapping);
9428         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9429                                 &BTRFS_I(inode)->runtime_flags))
9430                 filemap_flush(inode->i_mapping);
9431
9432         if (delalloc_work->delay_iput)
9433                 btrfs_add_delayed_iput(inode);
9434         else
9435                 iput(inode);
9436         complete(&delalloc_work->completion);
9437 }
9438
9439 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9440                                                     int delay_iput)
9441 {
9442         struct btrfs_delalloc_work *work;
9443
9444         work = kmalloc(sizeof(*work), GFP_NOFS);
9445         if (!work)
9446                 return NULL;
9447
9448         init_completion(&work->completion);
9449         INIT_LIST_HEAD(&work->list);
9450         work->inode = inode;
9451         work->delay_iput = delay_iput;
9452         WARN_ON_ONCE(!inode);
9453         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9454                         btrfs_run_delalloc_work, NULL, NULL);
9455
9456         return work;
9457 }
9458
9459 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9460 {
9461         wait_for_completion(&work->completion);
9462         kfree(work);
9463 }
9464
9465 /*
9466  * some fairly slow code that needs optimization. This walks the list
9467  * of all the inodes with pending delalloc and forces them to disk.
9468  */
9469 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9470                                    int nr)
9471 {
9472         struct btrfs_inode *binode;
9473         struct inode *inode;
9474         struct btrfs_delalloc_work *work, *next;
9475         struct list_head works;
9476         struct list_head splice;
9477         int ret = 0;
9478
9479         INIT_LIST_HEAD(&works);
9480         INIT_LIST_HEAD(&splice);
9481
9482         mutex_lock(&root->delalloc_mutex);
9483         spin_lock(&root->delalloc_lock);
9484         list_splice_init(&root->delalloc_inodes, &splice);
9485         while (!list_empty(&splice)) {
9486                 binode = list_entry(splice.next, struct btrfs_inode,
9487                                     delalloc_inodes);
9488
9489                 list_move_tail(&binode->delalloc_inodes,
9490                                &root->delalloc_inodes);
9491                 inode = igrab(&binode->vfs_inode);
9492                 if (!inode) {
9493                         cond_resched_lock(&root->delalloc_lock);
9494                         continue;
9495                 }
9496                 spin_unlock(&root->delalloc_lock);
9497
9498                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
9499                 if (!work) {
9500                         if (delay_iput)
9501                                 btrfs_add_delayed_iput(inode);
9502                         else
9503                                 iput(inode);
9504                         ret = -ENOMEM;
9505                         goto out;
9506                 }
9507                 list_add_tail(&work->list, &works);
9508                 btrfs_queue_work(root->fs_info->flush_workers,
9509                                  &work->work);
9510                 ret++;
9511                 if (nr != -1 && ret >= nr)
9512                         goto out;
9513                 cond_resched();
9514                 spin_lock(&root->delalloc_lock);
9515         }
9516         spin_unlock(&root->delalloc_lock);
9517
9518 out:
9519         list_for_each_entry_safe(work, next, &works, list) {
9520                 list_del_init(&work->list);
9521                 btrfs_wait_and_free_delalloc_work(work);
9522         }
9523
9524         if (!list_empty_careful(&splice)) {
9525                 spin_lock(&root->delalloc_lock);
9526                 list_splice_tail(&splice, &root->delalloc_inodes);
9527                 spin_unlock(&root->delalloc_lock);
9528         }
9529         mutex_unlock(&root->delalloc_mutex);
9530         return ret;
9531 }
9532
9533 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
9534 {
9535         int ret;
9536
9537         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
9538                 return -EROFS;
9539
9540         ret = __start_delalloc_inodes(root, delay_iput, -1);
9541         if (ret > 0)
9542                 ret = 0;
9543         /*
9544          * the filemap_flush will queue IO into the worker threads, but
9545          * we have to make sure the IO is actually started and that
9546          * ordered extents get created before we return
9547          */
9548         atomic_inc(&root->fs_info->async_submit_draining);
9549         while (atomic_read(&root->fs_info->nr_async_submits) ||
9550               atomic_read(&root->fs_info->async_delalloc_pages)) {
9551                 wait_event(root->fs_info->async_submit_wait,
9552                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
9553                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
9554         }
9555         atomic_dec(&root->fs_info->async_submit_draining);
9556         return ret;
9557 }
9558
9559 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
9560                                int nr)
9561 {
9562         struct btrfs_root *root;
9563         struct list_head splice;
9564         int ret;
9565
9566         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9567                 return -EROFS;
9568
9569         INIT_LIST_HEAD(&splice);
9570
9571         mutex_lock(&fs_info->delalloc_root_mutex);
9572         spin_lock(&fs_info->delalloc_root_lock);
9573         list_splice_init(&fs_info->delalloc_roots, &splice);
9574         while (!list_empty(&splice) && nr) {
9575                 root = list_first_entry(&splice, struct btrfs_root,
9576                                         delalloc_root);
9577                 root = btrfs_grab_fs_root(root);
9578                 BUG_ON(!root);
9579                 list_move_tail(&root->delalloc_root,
9580                                &fs_info->delalloc_roots);
9581                 spin_unlock(&fs_info->delalloc_root_lock);
9582
9583                 ret = __start_delalloc_inodes(root, delay_iput, nr);
9584                 btrfs_put_fs_root(root);
9585                 if (ret < 0)
9586                         goto out;
9587
9588                 if (nr != -1) {
9589                         nr -= ret;
9590                         WARN_ON(nr < 0);
9591                 }
9592                 spin_lock(&fs_info->delalloc_root_lock);
9593         }
9594         spin_unlock(&fs_info->delalloc_root_lock);
9595
9596         ret = 0;
9597         atomic_inc(&fs_info->async_submit_draining);
9598         while (atomic_read(&fs_info->nr_async_submits) ||
9599               atomic_read(&fs_info->async_delalloc_pages)) {
9600                 wait_event(fs_info->async_submit_wait,
9601                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
9602                     atomic_read(&fs_info->async_delalloc_pages) == 0));
9603         }
9604         atomic_dec(&fs_info->async_submit_draining);
9605 out:
9606         if (!list_empty_careful(&splice)) {
9607                 spin_lock(&fs_info->delalloc_root_lock);
9608                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9609                 spin_unlock(&fs_info->delalloc_root_lock);
9610         }
9611         mutex_unlock(&fs_info->delalloc_root_mutex);
9612         return ret;
9613 }
9614
9615 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9616                          const char *symname)
9617 {
9618         struct btrfs_trans_handle *trans;
9619         struct btrfs_root *root = BTRFS_I(dir)->root;
9620         struct btrfs_path *path;
9621         struct btrfs_key key;
9622         struct inode *inode = NULL;
9623         int err;
9624         int drop_inode = 0;
9625         u64 objectid;
9626         u64 index = 0;
9627         int name_len;
9628         int datasize;
9629         unsigned long ptr;
9630         struct btrfs_file_extent_item *ei;
9631         struct extent_buffer *leaf;
9632
9633         name_len = strlen(symname);
9634         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
9635                 return -ENAMETOOLONG;
9636
9637         /*
9638          * 2 items for inode item and ref
9639          * 2 items for dir items
9640          * 1 item for updating parent inode item
9641          * 1 item for the inline extent item
9642          * 1 item for xattr if selinux is on
9643          */
9644         trans = btrfs_start_transaction(root, 7);
9645         if (IS_ERR(trans))
9646                 return PTR_ERR(trans);
9647
9648         err = btrfs_find_free_ino(root, &objectid);
9649         if (err)
9650                 goto out_unlock;
9651
9652         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9653                                 dentry->d_name.len, btrfs_ino(dir), objectid,
9654                                 S_IFLNK|S_IRWXUGO, &index);
9655         if (IS_ERR(inode)) {
9656                 err = PTR_ERR(inode);
9657                 goto out_unlock;
9658         }
9659
9660         /*
9661         * If the active LSM wants to access the inode during
9662         * d_instantiate it needs these. Smack checks to see
9663         * if the filesystem supports xattrs by looking at the
9664         * ops vector.
9665         */
9666         inode->i_fop = &btrfs_file_operations;
9667         inode->i_op = &btrfs_file_inode_operations;
9668         inode->i_mapping->a_ops = &btrfs_aops;
9669         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9670
9671         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9672         if (err)
9673                 goto out_unlock_inode;
9674
9675         path = btrfs_alloc_path();
9676         if (!path) {
9677                 err = -ENOMEM;
9678                 goto out_unlock_inode;
9679         }
9680         key.objectid = btrfs_ino(inode);
9681         key.offset = 0;
9682         key.type = BTRFS_EXTENT_DATA_KEY;
9683         datasize = btrfs_file_extent_calc_inline_size(name_len);
9684         err = btrfs_insert_empty_item(trans, root, path, &key,
9685                                       datasize);
9686         if (err) {
9687                 btrfs_free_path(path);
9688                 goto out_unlock_inode;
9689         }
9690         leaf = path->nodes[0];
9691         ei = btrfs_item_ptr(leaf, path->slots[0],
9692                             struct btrfs_file_extent_item);
9693         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9694         btrfs_set_file_extent_type(leaf, ei,
9695                                    BTRFS_FILE_EXTENT_INLINE);
9696         btrfs_set_file_extent_encryption(leaf, ei, 0);
9697         btrfs_set_file_extent_compression(leaf, ei, 0);
9698         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9699         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9700
9701         ptr = btrfs_file_extent_inline_start(ei);
9702         write_extent_buffer(leaf, symname, ptr, name_len);
9703         btrfs_mark_buffer_dirty(leaf);
9704         btrfs_free_path(path);
9705
9706         inode->i_op = &btrfs_symlink_inode_operations;
9707         inode_nohighmem(inode);
9708         inode->i_mapping->a_ops = &btrfs_symlink_aops;
9709         inode_set_bytes(inode, name_len);
9710         btrfs_i_size_write(inode, name_len);
9711         err = btrfs_update_inode(trans, root, inode);
9712         /*
9713          * Last step, add directory indexes for our symlink inode. This is the
9714          * last step to avoid extra cleanup of these indexes if an error happens
9715          * elsewhere above.
9716          */
9717         if (!err)
9718                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
9719         if (err) {
9720                 drop_inode = 1;
9721                 goto out_unlock_inode;
9722         }
9723
9724         unlock_new_inode(inode);
9725         d_instantiate(dentry, inode);
9726
9727 out_unlock:
9728         btrfs_end_transaction(trans, root);
9729         if (drop_inode) {
9730                 inode_dec_link_count(inode);
9731                 iput(inode);
9732         }
9733         btrfs_btree_balance_dirty(root);
9734         return err;
9735
9736 out_unlock_inode:
9737         drop_inode = 1;
9738         unlock_new_inode(inode);
9739         goto out_unlock;
9740 }
9741
9742 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9743                                        u64 start, u64 num_bytes, u64 min_size,
9744                                        loff_t actual_len, u64 *alloc_hint,
9745                                        struct btrfs_trans_handle *trans)
9746 {
9747         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9748         struct extent_map *em;
9749         struct btrfs_root *root = BTRFS_I(inode)->root;
9750         struct btrfs_key ins;
9751         u64 cur_offset = start;
9752         u64 i_size;
9753         u64 cur_bytes;
9754         u64 last_alloc = (u64)-1;
9755         int ret = 0;
9756         bool own_trans = true;
9757
9758         if (trans)
9759                 own_trans = false;
9760         while (num_bytes > 0) {
9761                 if (own_trans) {
9762                         trans = btrfs_start_transaction(root, 3);
9763                         if (IS_ERR(trans)) {
9764                                 ret = PTR_ERR(trans);
9765                                 break;
9766                         }
9767                 }
9768
9769                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
9770                 cur_bytes = max(cur_bytes, min_size);
9771                 /*
9772                  * If we are severely fragmented we could end up with really
9773                  * small allocations, so if the allocator is returning small
9774                  * chunks lets make its job easier by only searching for those
9775                  * sized chunks.
9776                  */
9777                 cur_bytes = min(cur_bytes, last_alloc);
9778                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
9779                                            *alloc_hint, &ins, 1, 0);
9780                 if (ret) {
9781                         if (own_trans)
9782                                 btrfs_end_transaction(trans, root);
9783                         break;
9784                 }
9785
9786                 last_alloc = ins.offset;
9787                 ret = insert_reserved_file_extent(trans, inode,
9788                                                   cur_offset, ins.objectid,
9789                                                   ins.offset, ins.offset,
9790                                                   ins.offset, 0, 0, 0,
9791                                                   BTRFS_FILE_EXTENT_PREALLOC);
9792                 if (ret) {
9793                         btrfs_free_reserved_extent(root, ins.objectid,
9794                                                    ins.offset, 0);
9795                         btrfs_abort_transaction(trans, root, ret);
9796                         if (own_trans)
9797                                 btrfs_end_transaction(trans, root);
9798                         break;
9799                 }
9800
9801                 btrfs_drop_extent_cache(inode, cur_offset,
9802                                         cur_offset + ins.offset -1, 0);
9803
9804                 em = alloc_extent_map();
9805                 if (!em) {
9806                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9807                                 &BTRFS_I(inode)->runtime_flags);
9808                         goto next;
9809                 }
9810
9811                 em->start = cur_offset;
9812                 em->orig_start = cur_offset;
9813                 em->len = ins.offset;
9814                 em->block_start = ins.objectid;
9815                 em->block_len = ins.offset;
9816                 em->orig_block_len = ins.offset;
9817                 em->ram_bytes = ins.offset;
9818                 em->bdev = root->fs_info->fs_devices->latest_bdev;
9819                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9820                 em->generation = trans->transid;
9821
9822                 while (1) {
9823                         write_lock(&em_tree->lock);
9824                         ret = add_extent_mapping(em_tree, em, 1);
9825                         write_unlock(&em_tree->lock);
9826                         if (ret != -EEXIST)
9827                                 break;
9828                         btrfs_drop_extent_cache(inode, cur_offset,
9829                                                 cur_offset + ins.offset - 1,
9830                                                 0);
9831                 }
9832                 free_extent_map(em);
9833 next:
9834                 num_bytes -= ins.offset;
9835                 cur_offset += ins.offset;
9836                 *alloc_hint = ins.objectid + ins.offset;
9837
9838                 inode_inc_iversion(inode);
9839                 inode->i_ctime = CURRENT_TIME;
9840                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
9841                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
9842                     (actual_len > inode->i_size) &&
9843                     (cur_offset > inode->i_size)) {
9844                         if (cur_offset > actual_len)
9845                                 i_size = actual_len;
9846                         else
9847                                 i_size = cur_offset;
9848                         i_size_write(inode, i_size);
9849                         btrfs_ordered_update_i_size(inode, i_size, NULL);
9850                 }
9851
9852                 ret = btrfs_update_inode(trans, root, inode);
9853
9854                 if (ret) {
9855                         btrfs_abort_transaction(trans, root, ret);
9856                         if (own_trans)
9857                                 btrfs_end_transaction(trans, root);
9858                         break;
9859                 }
9860
9861                 if (own_trans)
9862                         btrfs_end_transaction(trans, root);
9863         }
9864         return ret;
9865 }
9866
9867 int btrfs_prealloc_file_range(struct inode *inode, int mode,
9868                               u64 start, u64 num_bytes, u64 min_size,
9869                               loff_t actual_len, u64 *alloc_hint)
9870 {
9871         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9872                                            min_size, actual_len, alloc_hint,
9873                                            NULL);
9874 }
9875
9876 int btrfs_prealloc_file_range_trans(struct inode *inode,
9877                                     struct btrfs_trans_handle *trans, int mode,
9878                                     u64 start, u64 num_bytes, u64 min_size,
9879                                     loff_t actual_len, u64 *alloc_hint)
9880 {
9881         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9882                                            min_size, actual_len, alloc_hint, trans);
9883 }
9884
9885 static int btrfs_set_page_dirty(struct page *page)
9886 {
9887         return __set_page_dirty_nobuffers(page);
9888 }
9889
9890 static int btrfs_permission(struct inode *inode, int mask)
9891 {
9892         struct btrfs_root *root = BTRFS_I(inode)->root;
9893         umode_t mode = inode->i_mode;
9894
9895         if (mask & MAY_WRITE &&
9896             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
9897                 if (btrfs_root_readonly(root))
9898                         return -EROFS;
9899                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
9900                         return -EACCES;
9901         }
9902         return generic_permission(inode, mask);
9903 }
9904
9905 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
9906 {
9907         struct btrfs_trans_handle *trans;
9908         struct btrfs_root *root = BTRFS_I(dir)->root;
9909         struct inode *inode = NULL;
9910         u64 objectid;
9911         u64 index;
9912         int ret = 0;
9913
9914         /*
9915          * 5 units required for adding orphan entry
9916          */
9917         trans = btrfs_start_transaction(root, 5);
9918         if (IS_ERR(trans))
9919                 return PTR_ERR(trans);
9920
9921         ret = btrfs_find_free_ino(root, &objectid);
9922         if (ret)
9923                 goto out;
9924
9925         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
9926                                 btrfs_ino(dir), objectid, mode, &index);
9927         if (IS_ERR(inode)) {
9928                 ret = PTR_ERR(inode);
9929                 inode = NULL;
9930                 goto out;
9931         }
9932
9933         inode->i_fop = &btrfs_file_operations;
9934         inode->i_op = &btrfs_file_inode_operations;
9935
9936         inode->i_mapping->a_ops = &btrfs_aops;
9937         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9938
9939         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
9940         if (ret)
9941                 goto out_inode;
9942
9943         ret = btrfs_update_inode(trans, root, inode);
9944         if (ret)
9945                 goto out_inode;
9946         ret = btrfs_orphan_add(trans, inode);
9947         if (ret)
9948                 goto out_inode;
9949
9950         /*
9951          * We set number of links to 0 in btrfs_new_inode(), and here we set
9952          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
9953          * through:
9954          *
9955          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
9956          */
9957         set_nlink(inode, 1);
9958         unlock_new_inode(inode);
9959         d_tmpfile(dentry, inode);
9960         mark_inode_dirty(inode);
9961
9962 out:
9963         btrfs_end_transaction(trans, root);
9964         if (ret)
9965                 iput(inode);
9966         btrfs_balance_delayed_items(root);
9967         btrfs_btree_balance_dirty(root);
9968         return ret;
9969
9970 out_inode:
9971         unlock_new_inode(inode);
9972         goto out;
9973
9974 }
9975
9976 /* Inspired by filemap_check_errors() */
9977 int btrfs_inode_check_errors(struct inode *inode)
9978 {
9979         int ret = 0;
9980
9981         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
9982             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
9983                 ret = -ENOSPC;
9984         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
9985             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
9986                 ret = -EIO;
9987
9988         return ret;
9989 }
9990
9991 static const struct inode_operations btrfs_dir_inode_operations = {
9992         .getattr        = btrfs_getattr,
9993         .lookup         = btrfs_lookup,
9994         .create         = btrfs_create,
9995         .unlink         = btrfs_unlink,
9996         .link           = btrfs_link,
9997         .mkdir          = btrfs_mkdir,
9998         .rmdir          = btrfs_rmdir,
9999         .rename2        = btrfs_rename2,
10000         .symlink        = btrfs_symlink,
10001         .setattr        = btrfs_setattr,
10002         .mknod          = btrfs_mknod,
10003         .setxattr       = btrfs_setxattr,
10004         .getxattr       = generic_getxattr,
10005         .listxattr      = btrfs_listxattr,
10006         .removexattr    = btrfs_removexattr,
10007         .permission     = btrfs_permission,
10008         .get_acl        = btrfs_get_acl,
10009         .set_acl        = btrfs_set_acl,
10010         .update_time    = btrfs_update_time,
10011         .tmpfile        = btrfs_tmpfile,
10012 };
10013 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10014         .lookup         = btrfs_lookup,
10015         .permission     = btrfs_permission,
10016         .get_acl        = btrfs_get_acl,
10017         .set_acl        = btrfs_set_acl,
10018         .update_time    = btrfs_update_time,
10019 };
10020
10021 static const struct file_operations btrfs_dir_file_operations = {
10022         .llseek         = generic_file_llseek,
10023         .read           = generic_read_dir,
10024         .iterate        = btrfs_real_readdir,
10025         .unlocked_ioctl = btrfs_ioctl,
10026 #ifdef CONFIG_COMPAT
10027         .compat_ioctl   = btrfs_ioctl,
10028 #endif
10029         .release        = btrfs_release_file,
10030         .fsync          = btrfs_sync_file,
10031 };
10032
10033 static const struct extent_io_ops btrfs_extent_io_ops = {
10034         .fill_delalloc = run_delalloc_range,
10035         .submit_bio_hook = btrfs_submit_bio_hook,
10036         .merge_bio_hook = btrfs_merge_bio_hook,
10037         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10038         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10039         .writepage_start_hook = btrfs_writepage_start_hook,
10040         .set_bit_hook = btrfs_set_bit_hook,
10041         .clear_bit_hook = btrfs_clear_bit_hook,
10042         .merge_extent_hook = btrfs_merge_extent_hook,
10043         .split_extent_hook = btrfs_split_extent_hook,
10044 };
10045
10046 /*
10047  * btrfs doesn't support the bmap operation because swapfiles
10048  * use bmap to make a mapping of extents in the file.  They assume
10049  * these extents won't change over the life of the file and they
10050  * use the bmap result to do IO directly to the drive.
10051  *
10052  * the btrfs bmap call would return logical addresses that aren't
10053  * suitable for IO and they also will change frequently as COW
10054  * operations happen.  So, swapfile + btrfs == corruption.
10055  *
10056  * For now we're avoiding this by dropping bmap.
10057  */
10058 static const struct address_space_operations btrfs_aops = {
10059         .readpage       = btrfs_readpage,
10060         .writepage      = btrfs_writepage,
10061         .writepages     = btrfs_writepages,
10062         .readpages      = btrfs_readpages,
10063         .direct_IO      = btrfs_direct_IO,
10064         .invalidatepage = btrfs_invalidatepage,
10065         .releasepage    = btrfs_releasepage,
10066         .set_page_dirty = btrfs_set_page_dirty,
10067         .error_remove_page = generic_error_remove_page,
10068 };
10069
10070 static const struct address_space_operations btrfs_symlink_aops = {
10071         .readpage       = btrfs_readpage,
10072         .writepage      = btrfs_writepage,
10073         .invalidatepage = btrfs_invalidatepage,
10074         .releasepage    = btrfs_releasepage,
10075 };
10076
10077 static const struct inode_operations btrfs_file_inode_operations = {
10078         .getattr        = btrfs_getattr,
10079         .setattr        = btrfs_setattr,
10080         .setxattr       = btrfs_setxattr,
10081         .getxattr       = generic_getxattr,
10082         .listxattr      = btrfs_listxattr,
10083         .removexattr    = btrfs_removexattr,
10084         .permission     = btrfs_permission,
10085         .fiemap         = btrfs_fiemap,
10086         .get_acl        = btrfs_get_acl,
10087         .set_acl        = btrfs_set_acl,
10088         .update_time    = btrfs_update_time,
10089 };
10090 static const struct inode_operations btrfs_special_inode_operations = {
10091         .getattr        = btrfs_getattr,
10092         .setattr        = btrfs_setattr,
10093         .permission     = btrfs_permission,
10094         .setxattr       = btrfs_setxattr,
10095         .getxattr       = generic_getxattr,
10096         .listxattr      = btrfs_listxattr,
10097         .removexattr    = btrfs_removexattr,
10098         .get_acl        = btrfs_get_acl,
10099         .set_acl        = btrfs_set_acl,
10100         .update_time    = btrfs_update_time,
10101 };
10102 static const struct inode_operations btrfs_symlink_inode_operations = {
10103         .readlink       = generic_readlink,
10104         .get_link       = page_get_link,
10105         .getattr        = btrfs_getattr,
10106         .setattr        = btrfs_setattr,
10107         .permission     = btrfs_permission,
10108         .setxattr       = btrfs_setxattr,
10109         .getxattr       = generic_getxattr,
10110         .listxattr      = btrfs_listxattr,
10111         .removexattr    = btrfs_removexattr,
10112         .update_time    = btrfs_update_time,
10113 };
10114
10115 const struct dentry_operations btrfs_dentry_operations = {
10116         .d_delete       = btrfs_dentry_delete,
10117         .d_release      = btrfs_dentry_release,
10118 };