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/pagevec.h>
  28 #include <linux/pmem.h>
  29 #include <linux/sched.h>
  30 #include <linux/uio.h>
  31 #include <linux/vmstat.h>
  32 #include <linux/pfn_t.h>
  33 #include <linux/sizes.h>
  34
  35 /*
  36  * We use lowest available bit in exceptional entry for locking, other two
  37  * bits to determine entry type. In total 3 special bits.
  38  */
  39 #define RADIX_DAX_SHIFT (RADIX_TREE_EXCEPTIONAL_SHIFT + 3)
  40 #define RADIX_DAX_PTE (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1))
  41 #define RADIX_DAX_PMD (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2))
  42 #define RADIX_DAX_TYPE_MASK (RADIX_DAX_PTE | RADIX_DAX_PMD)
  43 #define RADIX_DAX_TYPE(entry) ((unsigned long)entry & RADIX_DAX_TYPE_MASK)
  44 #define RADIX_DAX_SECTOR(entry) (((unsigned long)entry >> RADIX_DAX_SHIFT))
  45 #define RADIX_DAX_ENTRY(sector, pmd) ((void *)((unsigned long)sector << \
  46                 RADIX_DAX_SHIFT | (pmd ? RADIX_DAX_PMD : RADIX_DAX_PTE) | \
  47                 RADIX_TREE_EXCEPTIONAL_ENTRY))
  48
  49 /* We choose 4096 entries - same as per-zone page wait tables */
  50 #define DAX_WAIT_TABLE_BITS 12
  51 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
  52
  53 wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
  54
  55 static int __init init_dax_wait_table(void)
  56 {
  57         int i;
  58
  59         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
  60                 init_waitqueue_head(wait_table + i);
  61         return 0;
  62 }
  63 fs_initcall(init_dax_wait_table);
  64
  65 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
  66                                               pgoff_t index)
  67 {
  68         unsigned long hash = hash_long((unsigned long)mapping ^ index,
  69                                        DAX_WAIT_TABLE_BITS);
  70         return wait_table + hash;
  71 }
  72
  73 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
  74 {
  75         struct request_queue *q = bdev->bd_queue;
  76         long rc = -EIO;
  77
  78         dax->addr = ERR_PTR(-EIO);
  79         if (blk_queue_enter(q, true) != 0)
  80                 return rc;
  81
  82         rc = bdev_direct_access(bdev, dax);
  83         if (rc < 0) {
  84                 dax->addr = ERR_PTR(rc);
  85                 blk_queue_exit(q);
  86                 return rc;
  87         }
  88         return rc;
  89 }
  90
  91 static void dax_unmap_atomic(struct block_device *bdev,
  92                 const struct blk_dax_ctl *dax)
  93 {
  94         if (IS_ERR(dax->addr))
  95                 return;
  96         blk_queue_exit(bdev->bd_queue);
  97 }
  98
  99 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
 100 {
 101         struct page *page = alloc_pages(GFP_KERNEL, 0);
 102         struct blk_dax_ctl dax = {
 103                 .size = PAGE_SIZE,
 104                 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
 105         };
 106         long rc;
 107
 108         if (!page)
 109                 return ERR_PTR(-ENOMEM);
 110
 111         rc = dax_map_atomic(bdev, &dax);
 112         if (rc < 0)
 113                 return ERR_PTR(rc);
 114         memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
 115         dax_unmap_atomic(bdev, &dax);
 116         return page;
 117 }
 118
 119 static bool buffer_written(struct buffer_head *bh)
 120 {
 121         return buffer_mapped(bh) && !buffer_unwritten(bh);
 122 }
 123
 124 /*
 125  * When ext4 encounters a hole, it returns without modifying the buffer_head
 126  * which means that we can't trust b_size.  To cope with this, we set b_state
 127  * to 0 before calling get_block and, if any bit is set, we know we can trust
 128  * b_size.  Unfortunate, really, since ext4 knows precisely how long a hole is
 129  * and would save us time calling get_block repeatedly.
 130  */
 131 static bool buffer_size_valid(struct buffer_head *bh)
 132 {
 133         return bh->b_state != 0;
 134 }
 135
 136
 137 static sector_t to_sector(const struct buffer_head *bh,
 138                 const struct inode *inode)
 139 {
 140         sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
 141
 142         return sector;
 143 }
 144
 145 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 146                       loff_t start, loff_t end, get_block_t get_block,
 147                       struct buffer_head *bh)
 148 {
 149         loff_t pos = start, max = start, bh_max = start;
 150         bool hole = false;
 151         struct block_device *bdev = NULL;
 152         int rw = iov_iter_rw(iter), rc;
 153         long map_len = 0;
 154         struct blk_dax_ctl dax = {
 155                 .addr = ERR_PTR(-EIO),
 156         };
 157         unsigned blkbits = inode->i_blkbits;
 158         sector_t file_blks = (i_size_read(inode) + (1 << blkbits) - 1)
 159                                                                 >> blkbits;
 160
 161         if (rw == READ)
 162                 end = min(end, i_size_read(inode));
 163
 164         while (pos < end) {
 165                 size_t len;
 166                 if (pos == max) {
 167                         long page = pos >> PAGE_SHIFT;
 168                         sector_t block = page << (PAGE_SHIFT - blkbits);
 169                         unsigned first = pos - (block << blkbits);
 170                         long size;
 171
 172                         if (pos == bh_max) {
 173                                 bh->b_size = PAGE_ALIGN(end - pos);
 174                                 bh->b_state = 0;
 175                                 rc = get_block(inode, block, bh, rw == WRITE);
 176                                 if (rc)
 177                                         break;
 178                                 if (!buffer_size_valid(bh))
 179                                         bh->b_size = 1 << blkbits;
 180                                 bh_max = pos - first + bh->b_size;
 181                                 bdev = bh->b_bdev;
 182                                 /*
 183                                  * We allow uninitialized buffers for writes
 184                                  * beyond EOF as those cannot race with faults
 185                                  */
 186                                 WARN_ON_ONCE(
 187                                         (buffer_new(bh) && block < file_blks) ||
 188                                         (rw == WRITE && buffer_unwritten(bh)));
 189                         } else {
 190                                 unsigned done = bh->b_size -
 191                                                 (bh_max - (pos - first));
 192                                 bh->b_blocknr += done >> blkbits;
 193                                 bh->b_size -= done;
 194                         }
 195
 196                         hole = rw == READ && !buffer_written(bh);
 197                         if (hole) {
 198                                 size = bh->b_size - first;
 199                         } else {
 200                                 dax_unmap_atomic(bdev, &dax);
 201                                 dax.sector = to_sector(bh, inode);
 202                                 dax.size = bh->b_size;
 203                                 map_len = dax_map_atomic(bdev, &dax);
 204                                 if (map_len < 0) {
 205                                         rc = map_len;
 206                                         break;
 207                                 }
 208                                 dax.addr += first;
 209                                 size = map_len - first;
 210                         }
 211                         /*
 212                          * pos + size is one past the last offset for IO,
 213                          * so pos + size can overflow loff_t at extreme offsets.
 214                          * Cast to u64 to catch this and get the true minimum.
 215                          */
 216                         max = min_t(u64, pos + size, end);
 217                 }
 218
 219                 if (iov_iter_rw(iter) == WRITE) {
 220                         len = copy_from_iter_pmem(dax.addr, max - pos, iter);
 221                 } else if (!hole)
 222                         len = copy_to_iter((void __force *) dax.addr, max - pos,
 223                                         iter);
 224                 else
 225                         len = iov_iter_zero(max - pos, iter);
 226
 227                 if (!len) {
 228                         rc = -EFAULT;
 229                         break;
 230                 }
 231
 232                 pos += len;
 233                 if (!IS_ERR(dax.addr))
 234                         dax.addr += len;
 235         }
 236
 237         dax_unmap_atomic(bdev, &dax);
 238
 239         return (pos == start) ? rc : pos - start;
 240 }
 241
 242 /**
 243  * dax_do_io - Perform I/O to a DAX file
 244  * @iocb: The control block for this I/O
 245  * @inode: The file which the I/O is directed at
 246  * @iter: The addresses to do I/O from or to
 247  * @get_block: The filesystem method used to translate file offsets to blocks
 248  * @end_io: A filesystem callback for I/O completion
 249  * @flags: See below
 250  *
 251  * This function uses the same locking scheme as do_blockdev_direct_IO:
 252  * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
 253  * caller for writes.  For reads, we take and release the i_mutex ourselves.
 254  * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
 255  * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
 256  * is in progress.
 257  */
 258 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
 259                   struct iov_iter *iter, get_block_t get_block,
 260                   dio_iodone_t end_io, int flags)
 261 {
 262         struct buffer_head bh;
 263         ssize_t retval = -EINVAL;
 264         loff_t pos = iocb->ki_pos;
 265         loff_t end = pos + iov_iter_count(iter);
 266
 267         memset(&bh, 0, sizeof(bh));
 268         bh.b_bdev = inode->i_sb->s_bdev;
 269
 270         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 271                 inode_lock(inode);
 272
 273         /* Protects against truncate */
 274         if (!(flags & DIO_SKIP_DIO_COUNT))
 275                 inode_dio_begin(inode);
 276
 277         retval = dax_io(inode, iter, pos, end, get_block, &bh);
 278
 279         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 280                 inode_unlock(inode);
 281
 282         if (end_io) {
 283                 int err;
 284
 285                 err = end_io(iocb, pos, retval, bh.b_private);
 286                 if (err)
 287                         retval = err;
 288         }
 289
 290         if (!(flags & DIO_SKIP_DIO_COUNT))
 291                 inode_dio_end(inode);
 292         return retval;
 293 }
 294 EXPORT_SYMBOL_GPL(dax_do_io);
 295
 296 /*
 297  * DAX radix tree locking
 298  */
 299 struct exceptional_entry_key {
 300         struct address_space *mapping;
 301         unsigned long index;
 302 };
 303
 304 struct wait_exceptional_entry_queue {
 305         wait_queue_t wait;
 306         struct exceptional_entry_key key;
 307 };
 308
 309 static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
 310                                        int sync, void *keyp)
 311 {
 312         struct exceptional_entry_key *key = keyp;
 313         struct wait_exceptional_entry_queue *ewait =
 314                 container_of(wait, struct wait_exceptional_entry_queue, wait);
 315
 316         if (key->mapping != ewait->key.mapping ||
 317             key->index != ewait->key.index)
 318                 return 0;
 319         return autoremove_wake_function(wait, mode, sync, NULL);
 320 }
 321
 322 /*
 323  * Check whether the given slot is locked. The function must be called with
 324  * mapping->tree_lock held
 325  */
 326 static inline int slot_locked(struct address_space *mapping, void **slot)
 327 {
 328         unsigned long entry = (unsigned long)
 329                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 330         return entry & RADIX_DAX_ENTRY_LOCK;
 331 }
 332
 333 /*
 334  * Mark the given slot is locked. The function must be called with
 335  * mapping->tree_lock held
 336  */
 337 static inline void *lock_slot(struct address_space *mapping, void **slot)
 338 {
 339         unsigned long entry = (unsigned long)
 340                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 341
 342         entry |= RADIX_DAX_ENTRY_LOCK;
 343         radix_tree_replace_slot(slot, (void *)entry);
 344         return (void *)entry;
 345 }
 346
 347 /*
 348  * Mark the given slot is unlocked. The function must be called with
 349  * mapping->tree_lock held
 350  */
 351 static inline void *unlock_slot(struct address_space *mapping, void **slot)
 352 {
 353         unsigned long entry = (unsigned long)
 354                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 355
 356         entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
 357         radix_tree_replace_slot(slot, (void *)entry);
 358         return (void *)entry;
 359 }
 360
 361 /*
 362  * Lookup entry in radix tree, wait for it to become unlocked if it is
 363  * exceptional entry and return it. The caller must call
 364  * put_unlocked_mapping_entry() when he decided not to lock the entry or
 365  * put_locked_mapping_entry() when he locked the entry and now wants to
 366  * unlock it.
 367  *
 368  * The function must be called with mapping->tree_lock held.
 369  */
 370 static void *get_unlocked_mapping_entry(struct address_space *mapping,
 371                                         pgoff_t index, void ***slotp)
 372 {
 373         void *ret, **slot;
 374         struct wait_exceptional_entry_queue ewait;
 375         wait_queue_head_t *wq = dax_entry_waitqueue(mapping, index);
 376
 377         init_wait(&ewait.wait);
 378         ewait.wait.func = wake_exceptional_entry_func;
 379         ewait.key.mapping = mapping;
 380         ewait.key.index = index;
 381
 382         for (;;) {
 383                 ret = __radix_tree_lookup(&mapping->page_tree, index, NULL,
 384                                           &slot);
 385                 if (!ret || !radix_tree_exceptional_entry(ret) ||
 386                     !slot_locked(mapping, slot)) {
 387                         if (slotp)
 388                                 *slotp = slot;
 389                         return ret;
 390                 }
 391                 prepare_to_wait_exclusive(wq, &ewait.wait,
 392                                           TASK_UNINTERRUPTIBLE);
 393                 spin_unlock_irq(&mapping->tree_lock);
 394                 schedule();
 395                 finish_wait(wq, &ewait.wait);
 396                 spin_lock_irq(&mapping->tree_lock);
 397         }
 398 }
 399
 400 /*
 401  * Find radix tree entry at given index. If it points to a page, return with
 402  * the page locked. If it points to the exceptional entry, return with the
 403  * radix tree entry locked. If the radix tree doesn't contain given index,
 404  * create empty exceptional entry for the index and return with it locked.
 405  *
 406  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
 407  * persistent memory the benefit is doubtful. We can add that later if we can
 408  * show it helps.
 409  */
 410 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index)
 411 {
 412         void *ret, **slot;
 413
 414 restart:
 415         spin_lock_irq(&mapping->tree_lock);
 416         ret = get_unlocked_mapping_entry(mapping, index, &slot);
 417         /* No entry for given index? Make sure radix tree is big enough. */
 418         if (!ret) {
 419                 int err;
 420
 421                 spin_unlock_irq(&mapping->tree_lock);
 422                 err = radix_tree_preload(
 423                                 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
 424                 if (err)
 425                         return ERR_PTR(err);
 426                 ret = (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY |
 427                                RADIX_DAX_ENTRY_LOCK);
 428                 spin_lock_irq(&mapping->tree_lock);
 429                 err = radix_tree_insert(&mapping->page_tree, index, ret);
 430                 radix_tree_preload_end();
 431                 if (err) {
 432                         spin_unlock_irq(&mapping->tree_lock);
 433                         /* Someone already created the entry? */
 434                         if (err == -EEXIST)
 435                                 goto restart;
 436                         return ERR_PTR(err);
 437                 }
 438                 /* Good, we have inserted empty locked entry into the tree. */
 439                 mapping->nrexceptional++;
 440                 spin_unlock_irq(&mapping->tree_lock);
 441                 return ret;
 442         }
 443         /* Normal page in radix tree? */
 444         if (!radix_tree_exceptional_entry(ret)) {
 445                 struct page *page = ret;
 446
 447                 get_page(page);
 448                 spin_unlock_irq(&mapping->tree_lock);
 449                 lock_page(page);
 450                 /* Page got truncated? Retry... */
 451                 if (unlikely(page->mapping != mapping)) {
 452                         unlock_page(page);
 453                         put_page(page);
 454                         goto restart;
 455                 }
 456                 return page;
 457         }
 458         ret = lock_slot(mapping, slot);
 459         spin_unlock_irq(&mapping->tree_lock);
 460         return ret;
 461 }
 462
 463 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
 464                                    pgoff_t index, bool wake_all)
 465 {
 466         wait_queue_head_t *wq = dax_entry_waitqueue(mapping, index);
 467
 468         /*
 469          * Checking for locked entry and prepare_to_wait_exclusive() happens
 470          * under mapping->tree_lock, ditto for entry handling in our callers.
 471          * So at this point all tasks that could have seen our entry locked
 472          * must be in the waitqueue and the following check will see them.
 473          */
 474         if (waitqueue_active(wq)) {
 475                 struct exceptional_entry_key key;
 476
 477                 key.mapping = mapping;
 478                 key.index = index;
 479                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
 480         }
 481 }
 482
 483 void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index)
 484 {
 485         void *ret, **slot;
 486
 487         spin_lock_irq(&mapping->tree_lock);
 488         ret = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
 489         if (WARN_ON_ONCE(!ret || !radix_tree_exceptional_entry(ret) ||
 490                          !slot_locked(mapping, slot))) {
 491                 spin_unlock_irq(&mapping->tree_lock);
 492                 return;
 493         }
 494         unlock_slot(mapping, slot);
 495         spin_unlock_irq(&mapping->tree_lock);
 496         dax_wake_mapping_entry_waiter(mapping, index, false);
 497 }
 498
 499 static void put_locked_mapping_entry(struct address_space *mapping,
 500                                      pgoff_t index, void *entry)
 501 {
 502         if (!radix_tree_exceptional_entry(entry)) {
 503                 unlock_page(entry);
 504                 put_page(entry);
 505         } else {
 506                 dax_unlock_mapping_entry(mapping, index);
 507         }
 508 }
 509
 510 /*
 511  * Called when we are done with radix tree entry we looked up via
 512  * get_unlocked_mapping_entry() and which we didn't lock in the end.
 513  */
 514 static void put_unlocked_mapping_entry(struct address_space *mapping,
 515                                        pgoff_t index, void *entry)
 516 {
 517         if (!radix_tree_exceptional_entry(entry))
 518                 return;
 519
 520         /* We have to wake up next waiter for the radix tree entry lock */
 521         dax_wake_mapping_entry_waiter(mapping, index, false);
 522 }
 523
 524 /*
 525  * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
 526  * entry to get unlocked before deleting it.
 527  */
 528 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 529 {
 530         void *entry;
 531
 532         spin_lock_irq(&mapping->tree_lock);
 533         entry = get_unlocked_mapping_entry(mapping, index, NULL);
 534         /*
 535          * This gets called from truncate / punch_hole path. As such, the caller
 536          * must hold locks protecting against concurrent modifications of the
 537          * radix tree (usually fs-private i_mmap_sem for writing). Since the
 538          * caller has seen exceptional entry for this index, we better find it
 539          * at that index as well...
 540          */
 541         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry))) {
 542                 spin_unlock_irq(&mapping->tree_lock);
 543                 return 0;
 544         }
 545         radix_tree_delete(&mapping->page_tree, index);
 546         mapping->nrexceptional--;
 547         spin_unlock_irq(&mapping->tree_lock);
 548         dax_wake_mapping_entry_waiter(mapping, index, true);
 549
 550         return 1;
 551 }
 552
 553 /*
 554  * The user has performed a load from a hole in the file.  Allocating
 555  * a new page in the file would cause excessive storage usage for
 556  * workloads with sparse files.  We allocate a page cache page instead.
 557  * We'll kick it out of the page cache if it's ever written to,
 558  * otherwise it will simply fall out of the page cache under memory
 559  * pressure without ever having been dirtied.
 560  */
 561 static int dax_load_hole(struct address_space *mapping, void *entry,
 562                          struct vm_fault *vmf)
 563 {
 564         struct page *page;
 565
 566         /* Hole page already exists? Return it...  */
 567         if (!radix_tree_exceptional_entry(entry)) {
 568                 vmf->page = entry;
 569                 return VM_FAULT_LOCKED;
 570         }
 571
 572         /* This will replace locked radix tree entry with a hole page */
 573         page = find_or_create_page(mapping, vmf->pgoff,
 574                                    vmf->gfp_mask | __GFP_ZERO);
 575         if (!page) {
 576                 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
 577                 return VM_FAULT_OOM;
 578         }
 579         vmf->page = page;
 580         return VM_FAULT_LOCKED;
 581 }
 582
 583 static int copy_user_bh(struct page *to, struct inode *inode,
 584                 struct buffer_head *bh, unsigned long vaddr)
 585 {
 586         struct blk_dax_ctl dax = {
 587                 .sector = to_sector(bh, inode),
 588                 .size = bh->b_size,
 589         };
 590         struct block_device *bdev = bh->b_bdev;
 591         void *vto;
 592
 593         if (dax_map_atomic(bdev, &dax) < 0)
 594                 return PTR_ERR(dax.addr);
 595         vto = kmap_atomic(to);
 596         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 597         kunmap_atomic(vto);
 598         dax_unmap_atomic(bdev, &dax);
 599         return 0;
 600 }
 601
 602 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
 603
 604 static void *dax_insert_mapping_entry(struct address_space *mapping,
 605                                       struct vm_fault *vmf,
 606                                       void *entry, sector_t sector)
 607 {
 608         struct radix_tree_root *page_tree = &mapping->page_tree;
 609         int error = 0;
 610         bool hole_fill = false;
 611         void *new_entry;
 612         pgoff_t index = vmf->pgoff;
 613
 614         if (vmf->flags & FAULT_FLAG_WRITE)
 615                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 616
 617         /* Replacing hole page with block mapping? */
 618         if (!radix_tree_exceptional_entry(entry)) {
 619                 hole_fill = true;
 620                 /*
 621                  * Unmap the page now before we remove it from page cache below.
 622                  * The page is locked so it cannot be faulted in again.
 623                  */
 624                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 625                                     PAGE_SIZE, 0);
 626                 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
 627                 if (error)
 628                         return ERR_PTR(error);
 629         }
 630
 631         spin_lock_irq(&mapping->tree_lock);
 632         new_entry = (void *)((unsigned long)RADIX_DAX_ENTRY(sector, false) |
 633                        RADIX_DAX_ENTRY_LOCK);
 634         if (hole_fill) {
 635                 __delete_from_page_cache(entry, NULL);
 636                 /* Drop pagecache reference */
 637                 put_page(entry);
 638                 error = radix_tree_insert(page_tree, index, new_entry);
 639                 if (error) {
 640                         new_entry = ERR_PTR(error);
 641                         goto unlock;
 642                 }
 643                 mapping->nrexceptional++;
 644         } else {
 645                 void **slot;
 646                 void *ret;
 647
 648                 ret = __radix_tree_lookup(page_tree, index, NULL, &slot);
 649                 WARN_ON_ONCE(ret != entry);
 650                 radix_tree_replace_slot(slot, new_entry);
 651         }
 652         if (vmf->flags & FAULT_FLAG_WRITE)
 653                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 654  unlock:
 655         spin_unlock_irq(&mapping->tree_lock);
 656         if (hole_fill) {
 657                 radix_tree_preload_end();
 658                 /*
 659                  * We don't need hole page anymore, it has been replaced with
 660                  * locked radix tree entry now.
 661                  */
 662                 if (mapping->a_ops->freepage)
 663                         mapping->a_ops->freepage(entry);
 664                 unlock_page(entry);
 665                 put_page(entry);
 666         }
 667         return new_entry;
 668 }
 669
 670 static int dax_writeback_one(struct block_device *bdev,
 671                 struct address_space *mapping, pgoff_t index, void *entry)
 672 {
 673         struct radix_tree_root *page_tree = &mapping->page_tree;
 674         int type = RADIX_DAX_TYPE(entry);
 675         struct radix_tree_node *node;
 676         struct blk_dax_ctl dax;
 677         void **slot;
 678         int ret = 0;
 679
 680         spin_lock_irq(&mapping->tree_lock);
 681         /*
 682          * Regular page slots are stabilized by the page lock even
 683          * without the tree itself locked.  These unlocked entries
 684          * need verification under the tree lock.
 685          */
 686         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 687                 goto unlock;
 688         if (*slot != entry)
 689                 goto unlock;
 690
 691         /* another fsync thread may have already written back this entry */
 692         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 693                 goto unlock;
 694
 695         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 696                 ret = -EIO;
 697                 goto unlock;
 698         }
 699
 700         dax.sector = RADIX_DAX_SECTOR(entry);
 701         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 702         spin_unlock_irq(&mapping->tree_lock);
 703
 704         /*
 705          * We cannot hold tree_lock while calling dax_map_atomic() because it
 706          * eventually calls cond_resched().
 707          */
 708         ret = dax_map_atomic(bdev, &dax);
 709         if (ret < 0)
 710                 return ret;
 711
 712         if (WARN_ON_ONCE(ret < dax.size)) {
 713                 ret = -EIO;
 714                 goto unmap;
 715         }
 716
 717         wb_cache_pmem(dax.addr, dax.size);
 718
 719         spin_lock_irq(&mapping->tree_lock);
 720         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 721         spin_unlock_irq(&mapping->tree_lock);
 722  unmap:
 723         dax_unmap_atomic(bdev, &dax);
 724         return ret;
 725
 726  unlock:
 727         spin_unlock_irq(&mapping->tree_lock);
 728         return ret;
 729 }
 730
 731 /*
 732  * Flush the mapping to the persistent domain within the byte range of [start,
 733  * end]. This is required by data integrity operations to ensure file data is
 734  * on persistent storage prior to completion of the operation.
 735  */
 736 int dax_writeback_mapping_range(struct address_space *mapping,
 737                 struct block_device *bdev, struct writeback_control *wbc)
 738 {
 739         struct inode *inode = mapping->host;
 740         pgoff_t start_index, end_index, pmd_index;
 741         pgoff_t indices[PAGEVEC_SIZE];
 742         struct pagevec pvec;
 743         bool done = false;
 744         int i, ret = 0;
 745         void *entry;
 746
 747         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 748                 return -EIO;
 749
 750         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 751                 return 0;
 752
 753         start_index = wbc->range_start >> PAGE_SHIFT;
 754         end_index = wbc->range_end >> PAGE_SHIFT;
 755         pmd_index = DAX_PMD_INDEX(start_index);
 756
 757         rcu_read_lock();
 758         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 759         rcu_read_unlock();
 760
 761         /* see if the start of our range is covered by a PMD entry */
 762         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 763                 start_index = pmd_index;
 764
 765         tag_pages_for_writeback(mapping, start_index, end_index);
 766
 767         pagevec_init(&pvec, 0);
 768         while (!done) {
 769                 pvec.nr = find_get_entries_tag(mapping, start_index,
 770                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 771                                 pvec.pages, indices);
 772
 773                 if (pvec.nr == 0)
 774                         break;
 775
 776                 for (i = 0; i < pvec.nr; i++) {
 777                         if (indices[i] > end_index) {
 778                                 done = true;
 779                                 break;
 780                         }
 781
 782                         ret = dax_writeback_one(bdev, mapping, indices[i],
 783                                         pvec.pages[i]);
 784                         if (ret < 0)
 785                                 return ret;
 786                 }
 787         }
 788         return 0;
 789 }
 790 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 791
 792 static int dax_insert_mapping(struct address_space *mapping,
 793                         struct buffer_head *bh, void **entryp,
 794                         struct vm_area_struct *vma, struct vm_fault *vmf)
 795 {
 796         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 797         struct block_device *bdev = bh->b_bdev;
 798         struct blk_dax_ctl dax = {
 799                 .sector = to_sector(bh, mapping->host),
 800                 .size = bh->b_size,
 801         };
 802         void *ret;
 803         void *entry = *entryp;
 804
 805         if (dax_map_atomic(bdev, &dax) < 0)
 806                 return PTR_ERR(dax.addr);
 807         dax_unmap_atomic(bdev, &dax);
 808
 809         ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector);
 810         if (IS_ERR(ret))
 811                 return PTR_ERR(ret);
 812         *entryp = ret;
 813
 814         return vm_insert_mixed(vma, vaddr, dax.pfn);
 815 }
 816
 817 /**
 818  * dax_fault - handle a page fault on a DAX file
 819  * @vma: The virtual memory area where the fault occurred
 820  * @vmf: The description of the fault
 821  * @get_block: The filesystem method used to translate file offsets to blocks
 822  *
 823  * When a page fault occurs, filesystems may call this helper in their
 824  * fault handler for DAX files. dax_fault() assumes the caller has done all
 825  * the necessary locking for the page fault to proceed successfully.
 826  */
 827 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 828                         get_block_t get_block)
 829 {
 830         struct file *file = vma->vm_file;
 831         struct address_space *mapping = file->f_mapping;
 832         struct inode *inode = mapping->host;
 833         void *entry;
 834         struct buffer_head bh;
 835         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 836         unsigned blkbits = inode->i_blkbits;
 837         sector_t block;
 838         pgoff_t size;
 839         int error;
 840         int major = 0;
 841
 842         /*
 843          * Check whether offset isn't beyond end of file now. Caller is supposed
 844          * to hold locks serializing us with truncate / punch hole so this is
 845          * a reliable test.
 846          */
 847         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 848         if (vmf->pgoff >= size)
 849                 return VM_FAULT_SIGBUS;
 850
 851         memset(&bh, 0, sizeof(bh));
 852         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 853         bh.b_bdev = inode->i_sb->s_bdev;
 854         bh.b_size = PAGE_SIZE;
 855
 856         entry = grab_mapping_entry(mapping, vmf->pgoff);
 857         if (IS_ERR(entry)) {
 858                 error = PTR_ERR(entry);
 859                 goto out;
 860         }
 861
 862         error = get_block(inode, block, &bh, 0);
 863         if (!error && (bh.b_size < PAGE_SIZE))
 864                 error = -EIO;           /* fs corruption? */
 865         if (error)
 866                 goto unlock_entry;
 867
 868         if (vmf->cow_page) {
 869                 struct page *new_page = vmf->cow_page;
 870                 if (buffer_written(&bh))
 871                         error = copy_user_bh(new_page, inode, &bh, vaddr);
 872                 else
 873                         clear_user_highpage(new_page, vaddr);
 874                 if (error)
 875                         goto unlock_entry;
 876                 if (!radix_tree_exceptional_entry(entry)) {
 877                         vmf->page = entry;
 878                         return VM_FAULT_LOCKED;
 879                 }
 880                 vmf->entry = entry;
 881                 return VM_FAULT_DAX_LOCKED;
 882         }
 883
 884         if (!buffer_mapped(&bh)) {
 885                 if (vmf->flags & FAULT_FLAG_WRITE) {
 886                         error = get_block(inode, block, &bh, 1);
 887                         count_vm_event(PGMAJFAULT);
 888                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 889                         major = VM_FAULT_MAJOR;
 890                         if (!error && (bh.b_size < PAGE_SIZE))
 891                                 error = -EIO;
 892                         if (error)
 893                                 goto unlock_entry;
 894                 } else {
 895                         return dax_load_hole(mapping, entry, vmf);
 896                 }
 897         }
 898
 899         /* Filesystem should not return unwritten buffers to us! */
 900         WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
 901         error = dax_insert_mapping(mapping, &bh, &entry, vma, vmf);
 902  unlock_entry:
 903         put_locked_mapping_entry(mapping, vmf->pgoff, entry);
 904  out:
 905         if (error == -ENOMEM)
 906                 return VM_FAULT_OOM | major;
 907         /* -EBUSY is fine, somebody else faulted on the same PTE */
 908         if ((error < 0) && (error != -EBUSY))
 909                 return VM_FAULT_SIGBUS | major;
 910         return VM_FAULT_NOPAGE | major;
 911 }
 912 EXPORT_SYMBOL_GPL(dax_fault);
 913
 914 #if defined(CONFIG_TRANSPARENT_HUGEPAGE)
 915 /*
 916  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 917  * more often than one might expect in the below function.
 918  */
 919 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 920
 921 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 922                 const char *reason, const char *fn)
 923 {
 924         if (bh) {
 925                 char bname[BDEVNAME_SIZE];
 926                 bdevname(bh->b_bdev, bname);
 927                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 928                         "length %zd fallback: %s\n", fn, current->comm,
 929                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 930                         bh->b_size, reason);
 931         } else {
 932                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 933                         current->comm, address, reason);
 934         }
 935 }
 936
 937 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 938
 939 /**
 940  * dax_pmd_fault - handle a PMD fault on a DAX file
 941  * @vma: The virtual memory area where the fault occurred
 942  * @vmf: The description of the fault
 943  * @get_block: The filesystem method used to translate file offsets to blocks
 944  *
 945  * When a page fault occurs, filesystems may call this helper in their
 946  * pmd_fault handler for DAX files.
 947  */
 948 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 949                 pmd_t *pmd, unsigned int flags, get_block_t get_block)
 950 {
 951         struct file *file = vma->vm_file;
 952         struct address_space *mapping = file->f_mapping;
 953         struct inode *inode = mapping->host;
 954         struct buffer_head bh;
 955         unsigned blkbits = inode->i_blkbits;
 956         unsigned long pmd_addr = address & PMD_MASK;
 957         bool write = flags & FAULT_FLAG_WRITE;
 958         struct block_device *bdev;
 959         pgoff_t size, pgoff;
 960         sector_t block;
 961         int result = 0;
 962         bool alloc = false;
 963
 964         /* dax pmd mappings require pfn_t_devmap() */
 965         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 966                 return VM_FAULT_FALLBACK;
 967
 968         /* Fall back to PTEs if we're going to COW */
 969         if (write && !(vma->vm_flags & VM_SHARED)) {
 970                 split_huge_pmd(vma, pmd, address);
 971                 dax_pmd_dbg(NULL, address, "cow write");
 972                 return VM_FAULT_FALLBACK;
 973         }
 974         /* If the PMD would extend outside the VMA */
 975         if (pmd_addr < vma->vm_start) {
 976                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 977                 return VM_FAULT_FALLBACK;
 978         }
 979         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 980                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 981                 return VM_FAULT_FALLBACK;
 982         }
 983
 984         pgoff = linear_page_index(vma, pmd_addr);
 985         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 986         if (pgoff >= size)
 987                 return VM_FAULT_SIGBUS;
 988         /* If the PMD would cover blocks out of the file */
 989         if ((pgoff | PG_PMD_COLOUR) >= size) {
 990                 dax_pmd_dbg(NULL, address,
 991                                 "offset + huge page size > file size");
 992                 return VM_FAULT_FALLBACK;
 993         }
 994
 995         memset(&bh, 0, sizeof(bh));
 996         bh.b_bdev = inode->i_sb->s_bdev;
 997         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 998
 999         bh.b_size = PMD_SIZE;
1000
1001         if (get_block(inode, block, &bh, 0) != 0)
1002                 return VM_FAULT_SIGBUS;
1003
1004         if (!buffer_mapped(&bh) && write) {
1005                 if (get_block(inode, block, &bh, 1) != 0)
1006                         return VM_FAULT_SIGBUS;
1007                 alloc = true;
1008                 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
1009         }
1010
1011         bdev = bh.b_bdev;
1012
1013         /*
1014          * If the filesystem isn't willing to tell us the length of a hole,
1015          * just fall back to PTEs.  Calling get_block 512 times in a loop
1016          * would be silly.
1017          */
1018         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
1019                 dax_pmd_dbg(&bh, address, "allocated block too small");
1020                 return VM_FAULT_FALLBACK;
1021         }
1022
1023         /*
1024          * If we allocated new storage, make sure no process has any
1025          * zero pages covering this hole
1026          */
1027         if (alloc) {
1028                 loff_t lstart = pgoff << PAGE_SHIFT;
1029                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
1030
1031                 truncate_pagecache_range(inode, lstart, lend);
1032         }
1033
1034         if (!write && !buffer_mapped(&bh)) {
1035                 spinlock_t *ptl;
1036                 pmd_t entry;
1037                 struct page *zero_page = get_huge_zero_page();
1038
1039                 if (unlikely(!zero_page)) {
1040                         dax_pmd_dbg(&bh, address, "no zero page");
1041                         goto fallback;
1042                 }
1043
1044                 ptl = pmd_lock(vma->vm_mm, pmd);
1045                 if (!pmd_none(*pmd)) {
1046                         spin_unlock(ptl);
1047                         dax_pmd_dbg(&bh, address, "pmd already present");
1048                         goto fallback;
1049                 }
1050
1051                 dev_dbg(part_to_dev(bdev->bd_part),
1052                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
1053                                 __func__, current->comm, address,
1054                                 (unsigned long long) to_sector(&bh, inode));
1055
1056                 entry = mk_pmd(zero_page, vma->vm_page_prot);
1057                 entry = pmd_mkhuge(entry);
1058                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
1059                 result = VM_FAULT_NOPAGE;
1060                 spin_unlock(ptl);
1061         } else {
1062                 struct blk_dax_ctl dax = {
1063                         .sector = to_sector(&bh, inode),
1064                         .size = PMD_SIZE,
1065                 };
1066                 long length = dax_map_atomic(bdev, &dax);
1067
1068                 if (length < 0) {
1069                         dax_pmd_dbg(&bh, address, "dax-error fallback");
1070                         goto fallback;
1071                 }
1072                 if (length < PMD_SIZE) {
1073                         dax_pmd_dbg(&bh, address, "dax-length too small");
1074                         dax_unmap_atomic(bdev, &dax);
1075                         goto fallback;
1076                 }
1077                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
1078                         dax_pmd_dbg(&bh, address, "pfn unaligned");
1079                         dax_unmap_atomic(bdev, &dax);
1080                         goto fallback;
1081                 }
1082
1083                 if (!pfn_t_devmap(dax.pfn)) {
1084                         dax_unmap_atomic(bdev, &dax);
1085                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
1086                         goto fallback;
1087                 }
1088                 dax_unmap_atomic(bdev, &dax);
1089
1090                 /*
1091                  * For PTE faults we insert a radix tree entry for reads, and
1092                  * leave it clean.  Then on the first write we dirty the radix
1093                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
1094                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
1095                  * call into get_block() to translate the pgoff to a sector in
1096                  * order to be able to create a new radix tree entry.
1097                  *
1098                  * The PMD path doesn't have an equivalent to
1099                  * dax_pfn_mkwrite(), though, so for a read followed by a
1100                  * write we traverse all the way through dax_pmd_fault()
1101                  * twice.  This means we can just skip inserting a radix tree
1102                  * entry completely on the initial read and just wait until
1103                  * the write to insert a dirty entry.
1104                  */
1105                 if (write) {
1106                         /*
1107                          * We should insert radix-tree entry and dirty it here.
1108                          * For now this is broken...
1109                          */
1110                 }
1111
1112                 dev_dbg(part_to_dev(bdev->bd_part),
1113                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
1114                                 __func__, current->comm, address,
1115                                 pfn_t_to_pfn(dax.pfn),
1116                                 (unsigned long long) dax.sector);
1117                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
1118                                 dax.pfn, write);
1119         }
1120
1121  out:
1122         return result;
1123
1124  fallback:
1125         count_vm_event(THP_FAULT_FALLBACK);
1126         result = VM_FAULT_FALLBACK;
1127         goto out;
1128 }
1129 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1130 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1131
1132 /**
1133  * dax_pfn_mkwrite - handle first write to DAX page
1134  * @vma: The virtual memory area where the fault occurred
1135  * @vmf: The description of the fault
1136  */
1137 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1138 {
1139         struct file *file = vma->vm_file;
1140         struct address_space *mapping = file->f_mapping;
1141         void *entry;
1142         pgoff_t index = vmf->pgoff;
1143
1144         spin_lock_irq(&mapping->tree_lock);
1145         entry = get_unlocked_mapping_entry(mapping, index, NULL);
1146         if (!entry || !radix_tree_exceptional_entry(entry))
1147                 goto out;
1148         radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
1149         put_unlocked_mapping_entry(mapping, index, entry);
1150 out:
1151         spin_unlock_irq(&mapping->tree_lock);
1152         return VM_FAULT_NOPAGE;
1153 }
1154 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1155
1156 static bool dax_range_is_aligned(struct block_device *bdev,
1157                                  unsigned int offset, unsigned int length)
1158 {
1159         unsigned short sector_size = bdev_logical_block_size(bdev);
1160
1161         if (!IS_ALIGNED(offset, sector_size))
1162                 return false;
1163         if (!IS_ALIGNED(length, sector_size))
1164                 return false;
1165
1166         return true;
1167 }
1168
1169 int __dax_zero_page_range(struct block_device *bdev, sector_t sector,
1170                 unsigned int offset, unsigned int length)
1171 {
1172         struct blk_dax_ctl dax = {
1173                 .sector         = sector,
1174                 .size           = PAGE_SIZE,
1175         };
1176
1177         if (dax_range_is_aligned(bdev, offset, length)) {
1178                 sector_t start_sector = dax.sector + (offset >> 9);
1179
1180                 return blkdev_issue_zeroout(bdev, start_sector,
1181                                 length >> 9, GFP_NOFS, true);
1182         } else {
1183                 if (dax_map_atomic(bdev, &dax) < 0)
1184                         return PTR_ERR(dax.addr);
1185                 clear_pmem(dax.addr + offset, length);
1186                 dax_unmap_atomic(bdev, &dax);
1187         }
1188         return 0;
1189 }
1190 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1191
1192 /**
1193  * dax_zero_page_range - zero a range within a page of a DAX file
1194  * @inode: The file being truncated
1195  * @from: The file offset that is being truncated to
1196  * @length: The number of bytes to zero
1197  * @get_block: The filesystem method used to translate file offsets to blocks
1198  *
1199  * This function can be called by a filesystem when it is zeroing part of a
1200  * page in a DAX file.  This is intended for hole-punch operations.  If
1201  * you are truncating a file, the helper function dax_truncate_page() may be
1202  * more convenient.
1203  */
1204 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1205                                                         get_block_t get_block)
1206 {
1207         struct buffer_head bh;
1208         pgoff_t index = from >> PAGE_SHIFT;
1209         unsigned offset = from & (PAGE_SIZE-1);
1210         int err;
1211
1212         /* Block boundary? Nothing to do */
1213         if (!length)
1214                 return 0;
1215         BUG_ON((offset + length) > PAGE_SIZE);
1216
1217         memset(&bh, 0, sizeof(bh));
1218         bh.b_bdev = inode->i_sb->s_bdev;
1219         bh.b_size = PAGE_SIZE;
1220         err = get_block(inode, index, &bh, 0);
1221         if (err < 0 || !buffer_written(&bh))
1222                 return err;
1223
1224         return __dax_zero_page_range(bh.b_bdev, to_sector(&bh, inode),
1225                         offset, length);
1226 }
1227 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1228
1229 /**
1230  * dax_truncate_page - handle a partial page being truncated in a DAX file
1231  * @inode: The file being truncated
1232  * @from: The file offset that is being truncated to
1233  * @get_block: The filesystem method used to translate file offsets to blocks
1234  *
1235  * Similar to block_truncate_page(), this function can be called by a
1236  * filesystem when it is truncating a DAX file to handle the partial page.
1237  */
1238 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1239 {
1240         unsigned length = PAGE_ALIGN(from) - from;
1241         return dax_zero_page_range(inode, from, length, get_block);
1242 }
1243 EXPORT_SYMBOL_GPL(dax_truncate_page);