fs/dax.c

   1 /*
   2  * fs/dax.c - Direct Access filesystem code
   3  * Copyright (c) 2013-2014 Intel Corporation
   4  * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
   5  * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms and conditions of the GNU General Public License,
   9  * version 2, as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  */
  16
  17 #include <linux/atomic.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/buffer_head.h>
  20 #include <linux/dax.h>
  21 #include <linux/fs.h>
  22 #include <linux/genhd.h>
  23 #include <linux/highmem.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm.h>
  26 #include <linux/mutex.h>
  27 #include <linux/pmem.h>
  28 #include <linux/sched.h>
  29 #include <linux/uio.h>
  30 #include <linux/vmstat.h>
  31 #include <linux/pfn_t.h>
  32 #include <linux/sizes.h>
  33
  34 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
  35 {
  36         struct request_queue *q = bdev->bd_queue;
  37         long rc = -EIO;
  38
  39         dax->addr = (void __pmem *) ERR_PTR(-EIO);
  40         if (blk_queue_enter(q, true) != 0)
  41                 return rc;
  42
  43         rc = bdev_direct_access(bdev, dax);
  44         if (rc < 0) {
  45                 dax->addr = (void __pmem *) ERR_PTR(rc);
  46                 blk_queue_exit(q);
  47                 return rc;
  48         }
  49         return rc;
  50 }
  51
  52 static void dax_unmap_atomic(struct block_device *bdev,
  53                 const struct blk_dax_ctl *dax)
  54 {
  55         if (IS_ERR(dax->addr))
  56                 return;
  57         blk_queue_exit(bdev->bd_queue);
  58 }
  59
  60 /*
  61  * dax_clear_blocks() is called from within transaction context from XFS,
  62  * and hence this means the stack from this point must follow GFP_NOFS
  63  * semantics for all operations.
  64  */
  65 int dax_clear_blocks(struct inode *inode, sector_t block, long _size)
  66 {
  67         struct block_device *bdev = inode->i_sb->s_bdev;
  68         struct blk_dax_ctl dax = {
  69                 .sector = block << (inode->i_blkbits - 9),
  70                 .size = _size,
  71         };
  72
  73         might_sleep();
  74         do {
  75                 long count, sz;
  76
  77                 count = dax_map_atomic(bdev, &dax);
  78                 if (count < 0)
  79                         return count;
  80                 sz = min_t(long, count, SZ_128K);
  81                 clear_pmem(dax.addr, sz);
  82                 dax.size -= sz;
  83                 dax.sector += sz / 512;
  84                 dax_unmap_atomic(bdev, &dax);
  85                 cond_resched();
  86         } while (dax.size);
  87
  88         wmb_pmem();
  89         return 0;
  90 }
  91 EXPORT_SYMBOL_GPL(dax_clear_blocks);
  92
  93 /* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
  94 static void dax_new_buf(void __pmem *addr, unsigned size, unsigned first,
  95                 loff_t pos, loff_t end)
  96 {
  97         loff_t final = end - pos + first; /* The final byte of the buffer */
  98
  99         if (first > 0)
 100                 clear_pmem(addr, first);
 101         if (final < size)
 102                 clear_pmem(addr + final, size - final);
 103 }
 104
 105 static bool buffer_written(struct buffer_head *bh)
 106 {
 107         return buffer_mapped(bh) && !buffer_unwritten(bh);
 108 }
 109
 110 /*
 111  * When ext4 encounters a hole, it returns without modifying the buffer_head
 112  * which means that we can't trust b_size.  To cope with this, we set b_state
 113  * to 0 before calling get_block and, if any bit is set, we know we can trust
 114  * b_size.  Unfortunate, really, since ext4 knows precisely how long a hole is
 115  * and would save us time calling get_block repeatedly.
 116  */
 117 static bool buffer_size_valid(struct buffer_head *bh)
 118 {
 119         return bh->b_state != 0;
 120 }
 121
 122
 123 static sector_t to_sector(const struct buffer_head *bh,
 124                 const struct inode *inode)
 125 {
 126         sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
 127
 128         return sector;
 129 }
 130
 131 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 132                       loff_t start, loff_t end, get_block_t get_block,
 133                       struct buffer_head *bh)
 134 {
 135         loff_t pos = start, max = start, bh_max = start;
 136         bool hole = false, need_wmb = false;
 137         struct block_device *bdev = NULL;
 138         int rw = iov_iter_rw(iter), rc;
 139         long map_len = 0;
 140         struct blk_dax_ctl dax = {
 141                 .addr = (void __pmem *) ERR_PTR(-EIO),
 142         };
 143
 144         if (rw == READ)
 145                 end = min(end, i_size_read(inode));
 146
 147         while (pos < end) {
 148                 size_t len;
 149                 if (pos == max) {
 150                         unsigned blkbits = inode->i_blkbits;
 151                         long page = pos >> PAGE_SHIFT;
 152                         sector_t block = page << (PAGE_SHIFT - blkbits);
 153                         unsigned first = pos - (block << blkbits);
 154                         long size;
 155
 156                         if (pos == bh_max) {
 157                                 bh->b_size = PAGE_ALIGN(end - pos);
 158                                 bh->b_state = 0;
 159                                 rc = get_block(inode, block, bh, rw == WRITE);
 160                                 if (rc)
 161                                         break;
 162                                 if (!buffer_size_valid(bh))
 163                                         bh->b_size = 1 << blkbits;
 164                                 bh_max = pos - first + bh->b_size;
 165                                 bdev = bh->b_bdev;
 166                         } else {
 167                                 unsigned done = bh->b_size -
 168                                                 (bh_max - (pos - first));
 169                                 bh->b_blocknr += done >> blkbits;
 170                                 bh->b_size -= done;
 171                         }
 172
 173                         hole = rw == READ && !buffer_written(bh);
 174                         if (hole) {
 175                                 size = bh->b_size - first;
 176                         } else {
 177                                 dax_unmap_atomic(bdev, &dax);
 178                                 dax.sector = to_sector(bh, inode);
 179                                 dax.size = bh->b_size;
 180                                 map_len = dax_map_atomic(bdev, &dax);
 181                                 if (map_len < 0) {
 182                                         rc = map_len;
 183                                         break;
 184                                 }
 185                                 if (buffer_unwritten(bh) || buffer_new(bh)) {
 186                                         dax_new_buf(dax.addr, map_len, first,
 187                                                         pos, end);
 188                                         need_wmb = true;
 189                                 }
 190                                 dax.addr += first;
 191                                 size = map_len - first;
 192                         }
 193                         max = min(pos + size, end);
 194                 }
 195
 196                 if (iov_iter_rw(iter) == WRITE) {
 197                         len = copy_from_iter_pmem(dax.addr, max - pos, iter);
 198                         need_wmb = true;
 199                 } else if (!hole)
 200                         len = copy_to_iter((void __force *) dax.addr, max - pos,
 201                                         iter);
 202                 else
 203                         len = iov_iter_zero(max - pos, iter);
 204
 205                 if (!len) {
 206                         rc = -EFAULT;
 207                         break;
 208                 }
 209
 210                 pos += len;
 211                 if (!IS_ERR(dax.addr))
 212                         dax.addr += len;
 213         }
 214
 215         if (need_wmb)
 216                 wmb_pmem();
 217         dax_unmap_atomic(bdev, &dax);
 218
 219         return (pos == start) ? rc : pos - start;
 220 }
 221
 222 /**
 223  * dax_do_io - Perform I/O to a DAX file
 224  * @iocb: The control block for this I/O
 225  * @inode: The file which the I/O is directed at
 226  * @iter: The addresses to do I/O from or to
 227  * @pos: The file offset where the I/O starts
 228  * @get_block: The filesystem method used to translate file offsets to blocks
 229  * @end_io: A filesystem callback for I/O completion
 230  * @flags: See below
 231  *
 232  * This function uses the same locking scheme as do_blockdev_direct_IO:
 233  * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
 234  * caller for writes.  For reads, we take and release the i_mutex ourselves.
 235  * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
 236  * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
 237  * is in progress.
 238  */
 239 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
 240                   struct iov_iter *iter, loff_t pos, get_block_t get_block,
 241                   dio_iodone_t end_io, int flags)
 242 {
 243         struct buffer_head bh;
 244         ssize_t retval = -EINVAL;
 245         loff_t end = pos + iov_iter_count(iter);
 246
 247         memset(&bh, 0, sizeof(bh));
 248
 249         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) {
 250                 struct address_space *mapping = inode->i_mapping;
 251                 mutex_lock(&inode->i_mutex);
 252                 retval = filemap_write_and_wait_range(mapping, pos, end - 1);
 253                 if (retval) {
 254                         mutex_unlock(&inode->i_mutex);
 255                         goto out;
 256                 }
 257         }
 258
 259         /* Protects against truncate */
 260         if (!(flags & DIO_SKIP_DIO_COUNT))
 261                 inode_dio_begin(inode);
 262
 263         retval = dax_io(inode, iter, pos, end, get_block, &bh);
 264
 265         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 266                 mutex_unlock(&inode->i_mutex);
 267
 268         if ((retval > 0) && end_io)
 269                 end_io(iocb, pos, retval, bh.b_private);
 270
 271         if (!(flags & DIO_SKIP_DIO_COUNT))
 272                 inode_dio_end(inode);
 273  out:
 274         return retval;
 275 }
 276 EXPORT_SYMBOL_GPL(dax_do_io);
 277
 278 /*
 279  * The user has performed a load from a hole in the file.  Allocating
 280  * a new page in the file would cause excessive storage usage for
 281  * workloads with sparse files.  We allocate a page cache page instead.
 282  * We'll kick it out of the page cache if it's ever written to,
 283  * otherwise it will simply fall out of the page cache under memory
 284  * pressure without ever having been dirtied.
 285  */
 286 static int dax_load_hole(struct address_space *mapping, struct page *page,
 287                                                         struct vm_fault *vmf)
 288 {
 289         unsigned long size;
 290         struct inode *inode = mapping->host;
 291         if (!page)
 292                 page = find_or_create_page(mapping, vmf->pgoff,
 293                                                 GFP_KERNEL | __GFP_ZERO);
 294         if (!page)
 295                 return VM_FAULT_OOM;
 296         /* Recheck i_size under page lock to avoid truncate race */
 297         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 298         if (vmf->pgoff >= size) {
 299                 unlock_page(page);
 300                 page_cache_release(page);
 301                 return VM_FAULT_SIGBUS;
 302         }
 303
 304         vmf->page = page;
 305         return VM_FAULT_LOCKED;
 306 }
 307
 308 static int copy_user_bh(struct page *to, struct inode *inode,
 309                 struct buffer_head *bh, unsigned long vaddr)
 310 {
 311         struct blk_dax_ctl dax = {
 312                 .sector = to_sector(bh, inode),
 313                 .size = bh->b_size,
 314         };
 315         struct block_device *bdev = bh->b_bdev;
 316         void *vto;
 317
 318         if (dax_map_atomic(bdev, &dax) < 0)
 319                 return PTR_ERR(dax.addr);
 320         vto = kmap_atomic(to);
 321         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 322         kunmap_atomic(vto);
 323         dax_unmap_atomic(bdev, &dax);
 324         return 0;
 325 }
 326
 327 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 328                         struct vm_area_struct *vma, struct vm_fault *vmf)
 329 {
 330         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 331         struct address_space *mapping = inode->i_mapping;
 332         struct block_device *bdev = bh->b_bdev;
 333         struct blk_dax_ctl dax = {
 334                 .sector = to_sector(bh, inode),
 335                 .size = bh->b_size,
 336         };
 337         pgoff_t size;
 338         int error;
 339
 340         i_mmap_lock_read(mapping);
 341
 342         /*
 343          * Check truncate didn't happen while we were allocating a block.
 344          * If it did, this block may or may not be still allocated to the
 345          * file.  We can't tell the filesystem to free it because we can't
 346          * take i_mutex here.  In the worst case, the file still has blocks
 347          * allocated past the end of the file.
 348          */
 349         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 350         if (unlikely(vmf->pgoff >= size)) {
 351                 error = -EIO;
 352                 goto out;
 353         }
 354
 355         if (dax_map_atomic(bdev, &dax) < 0) {
 356                 error = PTR_ERR(dax.addr);
 357                 goto out;
 358         }
 359
 360         if (buffer_unwritten(bh) || buffer_new(bh)) {
 361                 clear_pmem(dax.addr, PAGE_SIZE);
 362                 wmb_pmem();
 363         }
 364         dax_unmap_atomic(bdev, &dax);
 365
 366         error = vm_insert_mixed(vma, vaddr, dax.pfn);
 367
 368  out:
 369         i_mmap_unlock_read(mapping);
 370
 371         return error;
 372 }
 373
 374 /**
 375  * __dax_fault - handle a page fault on a DAX file
 376  * @vma: The virtual memory area where the fault occurred
 377  * @vmf: The description of the fault
 378  * @get_block: The filesystem method used to translate file offsets to blocks
 379  * @complete_unwritten: The filesystem method used to convert unwritten blocks
 380  *      to written so the data written to them is exposed. This is required for
 381  *      required by write faults for filesystems that will return unwritten
 382  *      extent mappings from @get_block, but it is optional for reads as
 383  *      dax_insert_mapping() will always zero unwritten blocks. If the fs does
 384  *      not support unwritten extents, the it should pass NULL.
 385  *
 386  * When a page fault occurs, filesystems may call this helper in their
 387  * fault handler for DAX files. __dax_fault() assumes the caller has done all
 388  * the necessary locking for the page fault to proceed successfully.
 389  */
 390 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 391                         get_block_t get_block, dax_iodone_t complete_unwritten)
 392 {
 393         struct file *file = vma->vm_file;
 394         struct address_space *mapping = file->f_mapping;
 395         struct inode *inode = mapping->host;
 396         struct page *page;
 397         struct buffer_head bh;
 398         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 399         unsigned blkbits = inode->i_blkbits;
 400         sector_t block;
 401         pgoff_t size;
 402         int error;
 403         int major = 0;
 404
 405         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 406         if (vmf->pgoff >= size)
 407                 return VM_FAULT_SIGBUS;
 408
 409         memset(&bh, 0, sizeof(bh));
 410         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 411         bh.b_size = PAGE_SIZE;
 412
 413  repeat:
 414         page = find_get_page(mapping, vmf->pgoff);
 415         if (page) {
 416                 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
 417                         page_cache_release(page);
 418                         return VM_FAULT_RETRY;
 419                 }
 420                 if (unlikely(page->mapping != mapping)) {
 421                         unlock_page(page);
 422                         page_cache_release(page);
 423                         goto repeat;
 424                 }
 425                 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 426                 if (unlikely(vmf->pgoff >= size)) {
 427                         /*
 428                          * We have a struct page covering a hole in the file
 429                          * from a read fault and we've raced with a truncate
 430                          */
 431                         error = -EIO;
 432                         goto unlock_page;
 433                 }
 434         }
 435
 436         error = get_block(inode, block, &bh, 0);
 437         if (!error && (bh.b_size < PAGE_SIZE))
 438                 error = -EIO;           /* fs corruption? */
 439         if (error)
 440                 goto unlock_page;
 441
 442         if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
 443                 if (vmf->flags & FAULT_FLAG_WRITE) {
 444                         error = get_block(inode, block, &bh, 1);
 445                         count_vm_event(PGMAJFAULT);
 446                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 447                         major = VM_FAULT_MAJOR;
 448                         if (!error && (bh.b_size < PAGE_SIZE))
 449                                 error = -EIO;
 450                         if (error)
 451                                 goto unlock_page;
 452                 } else {
 453                         return dax_load_hole(mapping, page, vmf);
 454                 }
 455         }
 456
 457         if (vmf->cow_page) {
 458                 struct page *new_page = vmf->cow_page;
 459                 if (buffer_written(&bh))
 460                         error = copy_user_bh(new_page, inode, &bh, vaddr);
 461                 else
 462                         clear_user_highpage(new_page, vaddr);
 463                 if (error)
 464                         goto unlock_page;
 465                 vmf->page = page;
 466                 if (!page) {
 467                         i_mmap_lock_read(mapping);
 468                         /* Check we didn't race with truncate */
 469                         size = (i_size_read(inode) + PAGE_SIZE - 1) >>
 470                                                                 PAGE_SHIFT;
 471                         if (vmf->pgoff >= size) {
 472                                 i_mmap_unlock_read(mapping);
 473                                 error = -EIO;
 474                                 goto out;
 475                         }
 476                 }
 477                 return VM_FAULT_LOCKED;
 478         }
 479
 480         /* Check we didn't race with a read fault installing a new page */
 481         if (!page && major)
 482                 page = find_lock_page(mapping, vmf->pgoff);
 483
 484         if (page) {
 485                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 486                                                         PAGE_CACHE_SIZE, 0);
 487                 delete_from_page_cache(page);
 488                 unlock_page(page);
 489                 page_cache_release(page);
 490         }
 491
 492         /*
 493          * If we successfully insert the new mapping over an unwritten extent,
 494          * we need to ensure we convert the unwritten extent. If there is an
 495          * error inserting the mapping, the filesystem needs to leave it as
 496          * unwritten to prevent exposure of the stale underlying data to
 497          * userspace, but we still need to call the completion function so
 498          * the private resources on the mapping buffer can be released. We
 499          * indicate what the callback should do via the uptodate variable, same
 500          * as for normal BH based IO completions.
 501          */
 502         error = dax_insert_mapping(inode, &bh, vma, vmf);
 503         if (buffer_unwritten(&bh)) {
 504                 if (complete_unwritten)
 505                         complete_unwritten(&bh, !error);
 506                 else
 507                         WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
 508         }
 509
 510  out:
 511         if (error == -ENOMEM)
 512                 return VM_FAULT_OOM | major;
 513         /* -EBUSY is fine, somebody else faulted on the same PTE */
 514         if ((error < 0) && (error != -EBUSY))
 515                 return VM_FAULT_SIGBUS | major;
 516         return VM_FAULT_NOPAGE | major;
 517
 518  unlock_page:
 519         if (page) {
 520                 unlock_page(page);
 521                 page_cache_release(page);
 522         }
 523         goto out;
 524 }
 525 EXPORT_SYMBOL(__dax_fault);
 526
 527 /**
 528  * dax_fault - handle a page fault on a DAX file
 529  * @vma: The virtual memory area where the fault occurred
 530  * @vmf: The description of the fault
 531  * @get_block: The filesystem method used to translate file offsets to blocks
 532  *
 533  * When a page fault occurs, filesystems may call this helper in their
 534  * fault handler for DAX files.
 535  */
 536 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 537               get_block_t get_block, dax_iodone_t complete_unwritten)
 538 {
 539         int result;
 540         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 541
 542         if (vmf->flags & FAULT_FLAG_WRITE) {
 543                 sb_start_pagefault(sb);
 544                 file_update_time(vma->vm_file);
 545         }
 546         result = __dax_fault(vma, vmf, get_block, complete_unwritten);
 547         if (vmf->flags & FAULT_FLAG_WRITE)
 548                 sb_end_pagefault(sb);
 549
 550         return result;
 551 }
 552 EXPORT_SYMBOL_GPL(dax_fault);
 553
 554 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 555 /*
 556  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 557  * more often than one might expect in the below function.
 558  */
 559 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 560
 561 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 562                 const char *reason, const char *fn)
 563 {
 564         if (bh) {
 565                 char bname[BDEVNAME_SIZE];
 566                 bdevname(bh->b_bdev, bname);
 567                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 568                         "length %zd fallback: %s\n", fn, current->comm,
 569                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 570                         bh->b_size, reason);
 571         } else {
 572                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 573                         current->comm, address, reason);
 574         }
 575 }
 576
 577 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 578
 579 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 580                 pmd_t *pmd, unsigned int flags, get_block_t get_block,
 581                 dax_iodone_t complete_unwritten)
 582 {
 583         struct file *file = vma->vm_file;
 584         struct address_space *mapping = file->f_mapping;
 585         struct inode *inode = mapping->host;
 586         struct buffer_head bh;
 587         unsigned blkbits = inode->i_blkbits;
 588         unsigned long pmd_addr = address & PMD_MASK;
 589         bool write = flags & FAULT_FLAG_WRITE;
 590         struct block_device *bdev;
 591         pgoff_t size, pgoff;
 592         loff_t lstart, lend;
 593         sector_t block;
 594         int result = 0;
 595
 596         /* dax pmd mappings require pfn_t_devmap() */
 597         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 598                 return VM_FAULT_FALLBACK;
 599
 600         /* Fall back to PTEs if we're going to COW */
 601         if (write && !(vma->vm_flags & VM_SHARED)) {
 602                 split_huge_pmd(vma, pmd, address);
 603                 dax_pmd_dbg(NULL, address, "cow write");
 604                 return VM_FAULT_FALLBACK;
 605         }
 606         /* If the PMD would extend outside the VMA */
 607         if (pmd_addr < vma->vm_start) {
 608                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 609                 return VM_FAULT_FALLBACK;
 610         }
 611         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 612                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 613                 return VM_FAULT_FALLBACK;
 614         }
 615
 616         pgoff = linear_page_index(vma, pmd_addr);
 617         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 618         if (pgoff >= size)
 619                 return VM_FAULT_SIGBUS;
 620         /* If the PMD would cover blocks out of the file */
 621         if ((pgoff | PG_PMD_COLOUR) >= size) {
 622                 dax_pmd_dbg(NULL, address,
 623                                 "offset + huge page size > file size");
 624                 return VM_FAULT_FALLBACK;
 625         }
 626
 627         memset(&bh, 0, sizeof(bh));
 628         bh.b_bdev = inode->i_sb->s_bdev;
 629         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 630
 631         bh.b_size = PMD_SIZE;
 632         if (get_block(inode, block, &bh, write) != 0)
 633                 return VM_FAULT_SIGBUS;
 634         bdev = bh.b_bdev;
 635         i_mmap_lock_read(mapping);
 636
 637         /*
 638          * If the filesystem isn't willing to tell us the length of a hole,
 639          * just fall back to PTEs.  Calling get_block 512 times in a loop
 640          * would be silly.
 641          */
 642         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
 643                 dax_pmd_dbg(&bh, address, "allocated block too small");
 644                 goto fallback;
 645         }
 646
 647         /* make sure no process has any zero pages covering this hole */
 648         lstart = pgoff << PAGE_SHIFT;
 649         lend = lstart + PMD_SIZE - 1; /* inclusive */
 650         i_mmap_unlock_read(mapping);
 651         unmap_mapping_range(mapping, lstart, PMD_SIZE, 0);
 652         truncate_inode_pages_range(mapping, lstart, lend);
 653         i_mmap_lock_read(mapping);
 654
 655         /*
 656          * If a truncate happened while we were allocating blocks, we may
 657          * leave blocks allocated to the file that are beyond EOF.  We can't
 658          * take i_mutex here, so just leave them hanging; they'll be freed
 659          * when the file is deleted.
 660          */
 661         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 662         if (pgoff >= size) {
 663                 result = VM_FAULT_SIGBUS;
 664                 goto out;
 665         }
 666         if ((pgoff | PG_PMD_COLOUR) >= size) {
 667                 dax_pmd_dbg(&bh, address,
 668                                 "offset + huge page size > file size");
 669                 goto fallback;
 670         }
 671
 672         if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
 673                 spinlock_t *ptl;
 674                 pmd_t entry;
 675                 struct page *zero_page = get_huge_zero_page();
 676
 677                 if (unlikely(!zero_page)) {
 678                         dax_pmd_dbg(&bh, address, "no zero page");
 679                         goto fallback;
 680                 }
 681
 682                 ptl = pmd_lock(vma->vm_mm, pmd);
 683                 if (!pmd_none(*pmd)) {
 684                         spin_unlock(ptl);
 685                         dax_pmd_dbg(&bh, address, "pmd already present");
 686                         goto fallback;
 687                 }
 688
 689                 dev_dbg(part_to_dev(bdev->bd_part),
 690                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
 691                                 __func__, current->comm, address,
 692                                 (unsigned long long) to_sector(&bh, inode));
 693
 694                 entry = mk_pmd(zero_page, vma->vm_page_prot);
 695                 entry = pmd_mkhuge(entry);
 696                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
 697                 result = VM_FAULT_NOPAGE;
 698                 spin_unlock(ptl);
 699         } else {
 700                 struct blk_dax_ctl dax = {
 701                         .sector = to_sector(&bh, inode),
 702                         .size = PMD_SIZE,
 703                 };
 704                 long length = dax_map_atomic(bdev, &dax);
 705
 706                 if (length < 0) {
 707                         result = VM_FAULT_SIGBUS;
 708                         goto out;
 709                 }
 710                 if (length < PMD_SIZE) {
 711                         dax_pmd_dbg(&bh, address, "dax-length too small");
 712                         dax_unmap_atomic(bdev, &dax);
 713                         goto fallback;
 714                 }
 715                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
 716                         dax_pmd_dbg(&bh, address, "pfn unaligned");
 717                         dax_unmap_atomic(bdev, &dax);
 718                         goto fallback;
 719                 }
 720
 721                 if (!pfn_t_devmap(dax.pfn)) {
 722                         dax_unmap_atomic(bdev, &dax);
 723                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
 724                         goto fallback;
 725                 }
 726
 727                 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
 728                         clear_pmem(dax.addr, PMD_SIZE);
 729                         wmb_pmem();
 730                         count_vm_event(PGMAJFAULT);
 731                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 732                         result |= VM_FAULT_MAJOR;
 733                 }
 734                 dax_unmap_atomic(bdev, &dax);
 735
 736                 dev_dbg(part_to_dev(bdev->bd_part),
 737                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
 738                                 __func__, current->comm, address,
 739                                 pfn_t_to_pfn(dax.pfn),
 740                                 (unsigned long long) dax.sector);
 741                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
 742                                 dax.pfn, write);
 743         }
 744
 745  out:
 746         i_mmap_unlock_read(mapping);
 747
 748         if (buffer_unwritten(&bh))
 749                 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
 750
 751         return result;
 752
 753  fallback:
 754         count_vm_event(THP_FAULT_FALLBACK);
 755         result = VM_FAULT_FALLBACK;
 756         goto out;
 757 }
 758 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
 759
 760 /**
 761  * dax_pmd_fault - handle a PMD fault on a DAX file
 762  * @vma: The virtual memory area where the fault occurred
 763  * @vmf: The description of the fault
 764  * @get_block: The filesystem method used to translate file offsets to blocks
 765  *
 766  * When a page fault occurs, filesystems may call this helper in their
 767  * pmd_fault handler for DAX files.
 768  */
 769 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 770                         pmd_t *pmd, unsigned int flags, get_block_t get_block,
 771                         dax_iodone_t complete_unwritten)
 772 {
 773         int result;
 774         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 775
 776         if (flags & FAULT_FLAG_WRITE) {
 777                 sb_start_pagefault(sb);
 778                 file_update_time(vma->vm_file);
 779         }
 780         result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
 781                                 complete_unwritten);
 782         if (flags & FAULT_FLAG_WRITE)
 783                 sb_end_pagefault(sb);
 784
 785         return result;
 786 }
 787 EXPORT_SYMBOL_GPL(dax_pmd_fault);
 788 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 789
 790 /**
 791  * dax_pfn_mkwrite - handle first write to DAX page
 792  * @vma: The virtual memory area where the fault occurred
 793  * @vmf: The description of the fault
 794  *
 795  */
 796 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
 797 {
 798         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 799
 800         sb_start_pagefault(sb);
 801         file_update_time(vma->vm_file);
 802         sb_end_pagefault(sb);
 803         return VM_FAULT_NOPAGE;
 804 }
 805 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
 806
 807 /**
 808  * dax_zero_page_range - zero a range within a page of a DAX file
 809  * @inode: The file being truncated
 810  * @from: The file offset that is being truncated to
 811  * @length: The number of bytes to zero
 812  * @get_block: The filesystem method used to translate file offsets to blocks
 813  *
 814  * This function can be called by a filesystem when it is zeroing part of a
 815  * page in a DAX file.  This is intended for hole-punch operations.  If
 816  * you are truncating a file, the helper function dax_truncate_page() may be
 817  * more convenient.
 818  *
 819  * We work in terms of PAGE_CACHE_SIZE here for commonality with
 820  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
 821  * took care of disposing of the unnecessary blocks.  Even if the filesystem
 822  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
 823  * since the file might be mmapped.
 824  */
 825 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
 826                                                         get_block_t get_block)
 827 {
 828         struct buffer_head bh;
 829         pgoff_t index = from >> PAGE_CACHE_SHIFT;
 830         unsigned offset = from & (PAGE_CACHE_SIZE-1);
 831         int err;
 832
 833         /* Block boundary? Nothing to do */
 834         if (!length)
 835                 return 0;
 836         BUG_ON((offset + length) > PAGE_CACHE_SIZE);
 837
 838         memset(&bh, 0, sizeof(bh));
 839         bh.b_size = PAGE_CACHE_SIZE;
 840         err = get_block(inode, index, &bh, 0);
 841         if (err < 0)
 842                 return err;
 843         if (buffer_written(&bh)) {
 844                 struct block_device *bdev = bh.b_bdev;
 845                 struct blk_dax_ctl dax = {
 846                         .sector = to_sector(&bh, inode),
 847                         .size = PAGE_CACHE_SIZE,
 848                 };
 849
 850                 if (dax_map_atomic(bdev, &dax) < 0)
 851                         return PTR_ERR(dax.addr);
 852                 clear_pmem(dax.addr + offset, length);
 853                 wmb_pmem();
 854                 dax_unmap_atomic(bdev, &dax);
 855         }
 856
 857         return 0;
 858 }
 859 EXPORT_SYMBOL_GPL(dax_zero_page_range);
 860
 861 /**
 862  * dax_truncate_page - handle a partial page being truncated in a DAX file
 863  * @inode: The file being truncated
 864  * @from: The file offset that is being truncated to
 865  * @get_block: The filesystem method used to translate file offsets to blocks
 866  *
 867  * Similar to block_truncate_page(), this function can be called by a
 868  * filesystem when it is truncating a DAX file to handle the partial page.
 869  *
 870  * We work in terms of PAGE_CACHE_SIZE here for commonality with
 871  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
 872  * took care of disposing of the unnecessary blocks.  Even if the filesystem
 873  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
 874  * since the file might be mmapped.
 875  */
 876 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
 877 {
 878         unsigned length = PAGE_CACHE_ALIGN(from) - from;
 879         return dax_zero_page_range(inode, from, length, get_block);
 880 }
 881 EXPORT_SYMBOL_GPL(dax_truncate_page);