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