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 #include <linux/iomap.h>
  35 #include "internal.h"
  36
  37 /* We choose 4096 entries - same as per-zone page wait tables */
  38 #define DAX_WAIT_TABLE_BITS 12
  39 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
  40
  41 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
  42
  43 static int __init init_dax_wait_table(void)
  44 {
  45         int i;
  46
  47         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
  48                 init_waitqueue_head(wait_table + i);
  49         return 0;
  50 }
  51 fs_initcall(init_dax_wait_table);
  52
  53 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
  54 {
  55         struct request_queue *q = bdev->bd_queue;
  56         long rc = -EIO;
  57
  58         dax->addr = ERR_PTR(-EIO);
  59         if (blk_queue_enter(q, true) != 0)
  60                 return rc;
  61
  62         rc = bdev_direct_access(bdev, dax);
  63         if (rc < 0) {
  64                 dax->addr = ERR_PTR(rc);
  65                 blk_queue_exit(q);
  66                 return rc;
  67         }
  68         return rc;
  69 }
  70
  71 static void dax_unmap_atomic(struct block_device *bdev,
  72                 const struct blk_dax_ctl *dax)
  73 {
  74         if (IS_ERR(dax->addr))
  75                 return;
  76         blk_queue_exit(bdev->bd_queue);
  77 }
  78
  79 static int dax_is_pmd_entry(void *entry)
  80 {
  81         return (unsigned long)entry & RADIX_DAX_PMD;
  82 }
  83
  84 static int dax_is_pte_entry(void *entry)
  85 {
  86         return !((unsigned long)entry & RADIX_DAX_PMD);
  87 }
  88
  89 static int dax_is_zero_entry(void *entry)
  90 {
  91         return (unsigned long)entry & RADIX_DAX_HZP;
  92 }
  93
  94 static int dax_is_empty_entry(void *entry)
  95 {
  96         return (unsigned long)entry & RADIX_DAX_EMPTY;
  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 /*
 120  * DAX radix tree locking
 121  */
 122 struct exceptional_entry_key {
 123         struct address_space *mapping;
 124         pgoff_t entry_start;
 125 };
 126
 127 struct wait_exceptional_entry_queue {
 128         wait_queue_t wait;
 129         struct exceptional_entry_key key;
 130 };
 131
 132 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
 133                 pgoff_t index, void *entry, struct exceptional_entry_key *key)
 134 {
 135         unsigned long hash;
 136
 137         /*
 138          * If 'entry' is a PMD, align the 'index' that we use for the wait
 139          * queue to the start of that PMD.  This ensures that all offsets in
 140          * the range covered by the PMD map to the same bit lock.
 141          */
 142         if (dax_is_pmd_entry(entry))
 143                 index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
 144
 145         key->mapping = mapping;
 146         key->entry_start = index;
 147
 148         hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
 149         return wait_table + hash;
 150 }
 151
 152 static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
 153                                        int sync, void *keyp)
 154 {
 155         struct exceptional_entry_key *key = keyp;
 156         struct wait_exceptional_entry_queue *ewait =
 157                 container_of(wait, struct wait_exceptional_entry_queue, wait);
 158
 159         if (key->mapping != ewait->key.mapping ||
 160             key->entry_start != ewait->key.entry_start)
 161                 return 0;
 162         return autoremove_wake_function(wait, mode, sync, NULL);
 163 }
 164
 165 /*
 166  * Check whether the given slot is locked. The function must be called with
 167  * mapping->tree_lock held
 168  */
 169 static inline int slot_locked(struct address_space *mapping, void **slot)
 170 {
 171         unsigned long entry = (unsigned long)
 172                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 173         return entry & RADIX_DAX_ENTRY_LOCK;
 174 }
 175
 176 /*
 177  * Mark the given slot is locked. The function must be called with
 178  * mapping->tree_lock held
 179  */
 180 static inline void *lock_slot(struct address_space *mapping, void **slot)
 181 {
 182         unsigned long entry = (unsigned long)
 183                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 184
 185         entry |= RADIX_DAX_ENTRY_LOCK;
 186         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
 187         return (void *)entry;
 188 }
 189
 190 /*
 191  * Mark the given slot is unlocked. The function must be called with
 192  * mapping->tree_lock held
 193  */
 194 static inline void *unlock_slot(struct address_space *mapping, void **slot)
 195 {
 196         unsigned long entry = (unsigned long)
 197                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 198
 199         entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
 200         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
 201         return (void *)entry;
 202 }
 203
 204 /*
 205  * Lookup entry in radix tree, wait for it to become unlocked if it is
 206  * exceptional entry and return it. The caller must call
 207  * put_unlocked_mapping_entry() when he decided not to lock the entry or
 208  * put_locked_mapping_entry() when he locked the entry and now wants to
 209  * unlock it.
 210  *
 211  * The function must be called with mapping->tree_lock held.
 212  */
 213 static void *get_unlocked_mapping_entry(struct address_space *mapping,
 214                                         pgoff_t index, void ***slotp)
 215 {
 216         void *entry, **slot;
 217         struct wait_exceptional_entry_queue ewait;
 218         wait_queue_head_t *wq;
 219
 220         init_wait(&ewait.wait);
 221         ewait.wait.func = wake_exceptional_entry_func;
 222
 223         for (;;) {
 224                 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
 225                                           &slot);
 226                 if (!entry || !radix_tree_exceptional_entry(entry) ||
 227                     !slot_locked(mapping, slot)) {
 228                         if (slotp)
 229                                 *slotp = slot;
 230                         return entry;
 231                 }
 232
 233                 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
 234                 prepare_to_wait_exclusive(wq, &ewait.wait,
 235                                           TASK_UNINTERRUPTIBLE);
 236                 spin_unlock_irq(&mapping->tree_lock);
 237                 schedule();
 238                 finish_wait(wq, &ewait.wait);
 239                 spin_lock_irq(&mapping->tree_lock);
 240         }
 241 }
 242
 243 static void dax_unlock_mapping_entry(struct address_space *mapping,
 244                                      pgoff_t index)
 245 {
 246         void *entry, **slot;
 247
 248         spin_lock_irq(&mapping->tree_lock);
 249         entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
 250         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
 251                          !slot_locked(mapping, slot))) {
 252                 spin_unlock_irq(&mapping->tree_lock);
 253                 return;
 254         }
 255         unlock_slot(mapping, slot);
 256         spin_unlock_irq(&mapping->tree_lock);
 257         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 258 }
 259
 260 static void put_locked_mapping_entry(struct address_space *mapping,
 261                                      pgoff_t index, void *entry)
 262 {
 263         if (!radix_tree_exceptional_entry(entry)) {
 264                 unlock_page(entry);
 265                 put_page(entry);
 266         } else {
 267                 dax_unlock_mapping_entry(mapping, index);
 268         }
 269 }
 270
 271 /*
 272  * Called when we are done with radix tree entry we looked up via
 273  * get_unlocked_mapping_entry() and which we didn't lock in the end.
 274  */
 275 static void put_unlocked_mapping_entry(struct address_space *mapping,
 276                                        pgoff_t index, void *entry)
 277 {
 278         if (!radix_tree_exceptional_entry(entry))
 279                 return;
 280
 281         /* We have to wake up next waiter for the radix tree entry lock */
 282         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 283 }
 284
 285 /*
 286  * Find radix tree entry at given index. If it points to a page, return with
 287  * the page locked. If it points to the exceptional entry, return with the
 288  * radix tree entry locked. If the radix tree doesn't contain given index,
 289  * create empty exceptional entry for the index and return with it locked.
 290  *
 291  * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
 292  * either return that locked entry or will return an error.  This error will
 293  * happen if there are any 4k entries (either zero pages or DAX entries)
 294  * within the 2MiB range that we are requesting.
 295  *
 296  * We always favor 4k entries over 2MiB entries. There isn't a flow where we
 297  * evict 4k entries in order to 'upgrade' them to a 2MiB entry.  A 2MiB
 298  * insertion will fail if it finds any 4k entries already in the tree, and a
 299  * 4k insertion will cause an existing 2MiB entry to be unmapped and
 300  * downgraded to 4k entries.  This happens for both 2MiB huge zero pages as
 301  * well as 2MiB empty entries.
 302  *
 303  * The exception to this downgrade path is for 2MiB DAX PMD entries that have
 304  * real storage backing them.  We will leave these real 2MiB DAX entries in
 305  * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
 306  *
 307  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
 308  * persistent memory the benefit is doubtful. We can add that later if we can
 309  * show it helps.
 310  */
 311 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
 312                 unsigned long size_flag)
 313 {
 314         bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
 315         void *entry, **slot;
 316
 317 restart:
 318         spin_lock_irq(&mapping->tree_lock);
 319         entry = get_unlocked_mapping_entry(mapping, index, &slot);
 320
 321         if (entry) {
 322                 if (size_flag & RADIX_DAX_PMD) {
 323                         if (!radix_tree_exceptional_entry(entry) ||
 324                             dax_is_pte_entry(entry)) {
 325                                 put_unlocked_mapping_entry(mapping, index,
 326                                                 entry);
 327                                 entry = ERR_PTR(-EEXIST);
 328                                 goto out_unlock;
 329                         }
 330                 } else { /* trying to grab a PTE entry */
 331                         if (radix_tree_exceptional_entry(entry) &&
 332                             dax_is_pmd_entry(entry) &&
 333                             (dax_is_zero_entry(entry) ||
 334                              dax_is_empty_entry(entry))) {
 335                                 pmd_downgrade = true;
 336                         }
 337                 }
 338         }
 339
 340         /* No entry for given index? Make sure radix tree is big enough. */
 341         if (!entry || pmd_downgrade) {
 342                 int err;
 343
 344                 if (pmd_downgrade) {
 345                         /*
 346                          * Make sure 'entry' remains valid while we drop
 347                          * mapping->tree_lock.
 348                          */
 349                         entry = lock_slot(mapping, slot);
 350                 }
 351
 352                 spin_unlock_irq(&mapping->tree_lock);
 353                 /*
 354                  * Besides huge zero pages the only other thing that gets
 355                  * downgraded are empty entries which don't need to be
 356                  * unmapped.
 357                  */
 358                 if (pmd_downgrade && dax_is_zero_entry(entry))
 359                         unmap_mapping_range(mapping,
 360                                 (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
 361
 362                 err = radix_tree_preload(
 363                                 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
 364                 if (err) {
 365                         if (pmd_downgrade)
 366                                 put_locked_mapping_entry(mapping, index, entry);
 367                         return ERR_PTR(err);
 368                 }
 369                 spin_lock_irq(&mapping->tree_lock);
 370
 371                 if (pmd_downgrade) {
 372                         radix_tree_delete(&mapping->page_tree, index);
 373                         mapping->nrexceptional--;
 374                         dax_wake_mapping_entry_waiter(mapping, index, entry,
 375                                         true);
 376                 }
 377
 378                 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
 379
 380                 err = __radix_tree_insert(&mapping->page_tree, index,
 381                                 dax_radix_order(entry), entry);
 382                 radix_tree_preload_end();
 383                 if (err) {
 384                         spin_unlock_irq(&mapping->tree_lock);
 385                         /*
 386                          * Someone already created the entry?  This is a
 387                          * normal failure when inserting PMDs in a range
 388                          * that already contains PTEs.  In that case we want
 389                          * to return -EEXIST immediately.
 390                          */
 391                         if (err == -EEXIST && !(size_flag & RADIX_DAX_PMD))
 392                                 goto restart;
 393                         /*
 394                          * Our insertion of a DAX PMD entry failed, most
 395                          * likely because it collided with a PTE sized entry
 396                          * at a different index in the PMD range.  We haven't
 397                          * inserted anything into the radix tree and have no
 398                          * waiters to wake.
 399                          */
 400                         return ERR_PTR(err);
 401                 }
 402                 /* Good, we have inserted empty locked entry into the tree. */
 403                 mapping->nrexceptional++;
 404                 spin_unlock_irq(&mapping->tree_lock);
 405                 return entry;
 406         }
 407         /* Normal page in radix tree? */
 408         if (!radix_tree_exceptional_entry(entry)) {
 409                 struct page *page = entry;
 410
 411                 get_page(page);
 412                 spin_unlock_irq(&mapping->tree_lock);
 413                 lock_page(page);
 414                 /* Page got truncated? Retry... */
 415                 if (unlikely(page->mapping != mapping)) {
 416                         unlock_page(page);
 417                         put_page(page);
 418                         goto restart;
 419                 }
 420                 return page;
 421         }
 422         entry = lock_slot(mapping, slot);
 423  out_unlock:
 424         spin_unlock_irq(&mapping->tree_lock);
 425         return entry;
 426 }
 427
 428 /*
 429  * We do not necessarily hold the mapping->tree_lock when we call this
 430  * function so it is possible that 'entry' is no longer a valid item in the
 431  * radix tree.  This is okay because all we really need to do is to find the
 432  * correct waitqueue where tasks might be waiting for that old 'entry' and
 433  * wake them.
 434  */
 435 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
 436                 pgoff_t index, void *entry, bool wake_all)
 437 {
 438         struct exceptional_entry_key key;
 439         wait_queue_head_t *wq;
 440
 441         wq = dax_entry_waitqueue(mapping, index, entry, &key);
 442
 443         /*
 444          * Checking for locked entry and prepare_to_wait_exclusive() happens
 445          * under mapping->tree_lock, ditto for entry handling in our callers.
 446          * So at this point all tasks that could have seen our entry locked
 447          * must be in the waitqueue and the following check will see them.
 448          */
 449         if (waitqueue_active(wq))
 450                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
 451 }
 452
 453 /*
 454  * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
 455  * entry to get unlocked before deleting it.
 456  */
 457 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 458 {
 459         void *entry;
 460
 461         spin_lock_irq(&mapping->tree_lock);
 462         entry = get_unlocked_mapping_entry(mapping, index, NULL);
 463         /*
 464          * This gets called from truncate / punch_hole path. As such, the caller
 465          * must hold locks protecting against concurrent modifications of the
 466          * radix tree (usually fs-private i_mmap_sem for writing). Since the
 467          * caller has seen exceptional entry for this index, we better find it
 468          * at that index as well...
 469          */
 470         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry))) {
 471                 spin_unlock_irq(&mapping->tree_lock);
 472                 return 0;
 473         }
 474         radix_tree_delete(&mapping->page_tree, index);
 475         mapping->nrexceptional--;
 476         spin_unlock_irq(&mapping->tree_lock);
 477         dax_wake_mapping_entry_waiter(mapping, index, entry, true);
 478
 479         return 1;
 480 }
 481
 482 /*
 483  * The user has performed a load from a hole in the file.  Allocating
 484  * a new page in the file would cause excessive storage usage for
 485  * workloads with sparse files.  We allocate a page cache page instead.
 486  * We'll kick it out of the page cache if it's ever written to,
 487  * otherwise it will simply fall out of the page cache under memory
 488  * pressure without ever having been dirtied.
 489  */
 490 static int dax_load_hole(struct address_space *mapping, void *entry,
 491                          struct vm_fault *vmf)
 492 {
 493         struct page *page;
 494
 495         /* Hole page already exists? Return it...  */
 496         if (!radix_tree_exceptional_entry(entry)) {
 497                 vmf->page = entry;
 498                 return VM_FAULT_LOCKED;
 499         }
 500
 501         /* This will replace locked radix tree entry with a hole page */
 502         page = find_or_create_page(mapping, vmf->pgoff,
 503                                    vmf->gfp_mask | __GFP_ZERO);
 504         if (!page)
 505                 return VM_FAULT_OOM;
 506         vmf->page = page;
 507         return VM_FAULT_LOCKED;
 508 }
 509
 510 static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size,
 511                 struct page *to, unsigned long vaddr)
 512 {
 513         struct blk_dax_ctl dax = {
 514                 .sector = sector,
 515                 .size = size,
 516         };
 517         void *vto;
 518
 519         if (dax_map_atomic(bdev, &dax) < 0)
 520                 return PTR_ERR(dax.addr);
 521         vto = kmap_atomic(to);
 522         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 523         kunmap_atomic(vto);
 524         dax_unmap_atomic(bdev, &dax);
 525         return 0;
 526 }
 527
 528 /*
 529  * By this point grab_mapping_entry() has ensured that we have a locked entry
 530  * of the appropriate size so we don't have to worry about downgrading PMDs to
 531  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
 532  * already in the tree, we will skip the insertion and just dirty the PMD as
 533  * appropriate.
 534  */
 535 static void *dax_insert_mapping_entry(struct address_space *mapping,
 536                                       struct vm_fault *vmf,
 537                                       void *entry, sector_t sector,
 538                                       unsigned long flags)
 539 {
 540         struct radix_tree_root *page_tree = &mapping->page_tree;
 541         int error = 0;
 542         bool hole_fill = false;
 543         void *new_entry;
 544         pgoff_t index = vmf->pgoff;
 545
 546         if (vmf->flags & FAULT_FLAG_WRITE)
 547                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 548
 549         /* Replacing hole page with block mapping? */
 550         if (!radix_tree_exceptional_entry(entry)) {
 551                 hole_fill = true;
 552                 /*
 553                  * Unmap the page now before we remove it from page cache below.
 554                  * The page is locked so it cannot be faulted in again.
 555                  */
 556                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 557                                     PAGE_SIZE, 0);
 558                 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
 559                 if (error)
 560                         return ERR_PTR(error);
 561         } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) {
 562                 /* replacing huge zero page with PMD block mapping */
 563                 unmap_mapping_range(mapping,
 564                         (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
 565         }
 566
 567         spin_lock_irq(&mapping->tree_lock);
 568         new_entry = dax_radix_locked_entry(sector, flags);
 569
 570         if (hole_fill) {
 571                 __delete_from_page_cache(entry, NULL);
 572                 /* Drop pagecache reference */
 573                 put_page(entry);
 574                 error = __radix_tree_insert(page_tree, index,
 575                                 dax_radix_order(new_entry), new_entry);
 576                 if (error) {
 577                         new_entry = ERR_PTR(error);
 578                         goto unlock;
 579                 }
 580                 mapping->nrexceptional++;
 581         } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 582                 /*
 583                  * Only swap our new entry into the radix tree if the current
 584                  * entry is a zero page or an empty entry.  If a normal PTE or
 585                  * PMD entry is already in the tree, we leave it alone.  This
 586                  * means that if we are trying to insert a PTE and the
 587                  * existing entry is a PMD, we will just leave the PMD in the
 588                  * tree and dirty it if necessary.
 589                  */
 590                 struct radix_tree_node *node;
 591                 void **slot;
 592                 void *ret;
 593
 594                 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
 595                 WARN_ON_ONCE(ret != entry);
 596                 __radix_tree_replace(page_tree, node, slot,
 597                                      new_entry, NULL, NULL);
 598         }
 599         if (vmf->flags & FAULT_FLAG_WRITE)
 600                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 601  unlock:
 602         spin_unlock_irq(&mapping->tree_lock);
 603         if (hole_fill) {
 604                 radix_tree_preload_end();
 605                 /*
 606                  * We don't need hole page anymore, it has been replaced with
 607                  * locked radix tree entry now.
 608                  */
 609                 if (mapping->a_ops->freepage)
 610                         mapping->a_ops->freepage(entry);
 611                 unlock_page(entry);
 612                 put_page(entry);
 613         }
 614         return new_entry;
 615 }
 616
 617 static int dax_writeback_one(struct block_device *bdev,
 618                 struct address_space *mapping, pgoff_t index, void *entry)
 619 {
 620         struct radix_tree_root *page_tree = &mapping->page_tree;
 621         struct blk_dax_ctl dax;
 622         void *entry2, **slot;
 623         int ret = 0;
 624
 625         /*
 626          * A page got tagged dirty in DAX mapping? Something is seriously
 627          * wrong.
 628          */
 629         if (WARN_ON(!radix_tree_exceptional_entry(entry)))
 630                 return -EIO;
 631
 632         spin_lock_irq(&mapping->tree_lock);
 633         entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
 634         /* Entry got punched out / reallocated? */
 635         if (!entry2 || !radix_tree_exceptional_entry(entry2))
 636                 goto put_unlocked;
 637         /*
 638          * Entry got reallocated elsewhere? No need to writeback. We have to
 639          * compare sectors as we must not bail out due to difference in lockbit
 640          * or entry type.
 641          */
 642         if (dax_radix_sector(entry2) != dax_radix_sector(entry))
 643                 goto put_unlocked;
 644         if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
 645                                 dax_is_zero_entry(entry))) {
 646                 ret = -EIO;
 647                 goto put_unlocked;
 648         }
 649
 650         /* Another fsync thread may have already written back this entry */
 651         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 652                 goto put_unlocked;
 653         /* Lock the entry to serialize with page faults */
 654         entry = lock_slot(mapping, slot);
 655         /*
 656          * We can clear the tag now but we have to be careful so that concurrent
 657          * dax_writeback_one() calls for the same index cannot finish before we
 658          * actually flush the caches. This is achieved as the calls will look
 659          * at the entry only under tree_lock and once they do that they will
 660          * see the entry locked and wait for it to unlock.
 661          */
 662         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 663         spin_unlock_irq(&mapping->tree_lock);
 664
 665         /*
 666          * Even if dax_writeback_mapping_range() was given a wbc->range_start
 667          * in the middle of a PMD, the 'index' we are given will be aligned to
 668          * the start index of the PMD, as will the sector we pull from
 669          * 'entry'.  This allows us to flush for PMD_SIZE and not have to
 670          * worry about partial PMD writebacks.
 671          */
 672         dax.sector = dax_radix_sector(entry);
 673         dax.size = PAGE_SIZE << dax_radix_order(entry);
 674
 675         /*
 676          * We cannot hold tree_lock while calling dax_map_atomic() because it
 677          * eventually calls cond_resched().
 678          */
 679         ret = dax_map_atomic(bdev, &dax);
 680         if (ret < 0) {
 681                 put_locked_mapping_entry(mapping, index, entry);
 682                 return ret;
 683         }
 684
 685         if (WARN_ON_ONCE(ret < dax.size)) {
 686                 ret = -EIO;
 687                 goto unmap;
 688         }
 689
 690         wb_cache_pmem(dax.addr, dax.size);
 691  unmap:
 692         dax_unmap_atomic(bdev, &dax);
 693         put_locked_mapping_entry(mapping, index, entry);
 694         return ret;
 695
 696  put_unlocked:
 697         put_unlocked_mapping_entry(mapping, index, entry2);
 698         spin_unlock_irq(&mapping->tree_lock);
 699         return ret;
 700 }
 701
 702 /*
 703  * Flush the mapping to the persistent domain within the byte range of [start,
 704  * end]. This is required by data integrity operations to ensure file data is
 705  * on persistent storage prior to completion of the operation.
 706  */
 707 int dax_writeback_mapping_range(struct address_space *mapping,
 708                 struct block_device *bdev, struct writeback_control *wbc)
 709 {
 710         struct inode *inode = mapping->host;
 711         pgoff_t start_index, end_index;
 712         pgoff_t indices[PAGEVEC_SIZE];
 713         struct pagevec pvec;
 714         bool done = false;
 715         int i, ret = 0;
 716
 717         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 718                 return -EIO;
 719
 720         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 721                 return 0;
 722
 723         start_index = wbc->range_start >> PAGE_SHIFT;
 724         end_index = wbc->range_end >> PAGE_SHIFT;
 725
 726         tag_pages_for_writeback(mapping, start_index, end_index);
 727
 728         pagevec_init(&pvec, 0);
 729         while (!done) {
 730                 pvec.nr = find_get_entries_tag(mapping, start_index,
 731                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 732                                 pvec.pages, indices);
 733
 734                 if (pvec.nr == 0)
 735                         break;
 736
 737                 for (i = 0; i < pvec.nr; i++) {
 738                         if (indices[i] > end_index) {
 739                                 done = true;
 740                                 break;
 741                         }
 742
 743                         ret = dax_writeback_one(bdev, mapping, indices[i],
 744                                         pvec.pages[i]);
 745                         if (ret < 0)
 746                                 return ret;
 747                 }
 748         }
 749         return 0;
 750 }
 751 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 752
 753 static int dax_insert_mapping(struct address_space *mapping,
 754                 struct block_device *bdev, sector_t sector, size_t size,
 755                 void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf)
 756 {
 757         unsigned long vaddr = vmf->address;
 758         struct blk_dax_ctl dax = {
 759                 .sector = sector,
 760                 .size = size,
 761         };
 762         void *ret;
 763         void *entry = *entryp;
 764
 765         if (dax_map_atomic(bdev, &dax) < 0)
 766                 return PTR_ERR(dax.addr);
 767         dax_unmap_atomic(bdev, &dax);
 768
 769         ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector, 0);
 770         if (IS_ERR(ret))
 771                 return PTR_ERR(ret);
 772         *entryp = ret;
 773
 774         return vm_insert_mixed(vma, vaddr, dax.pfn);
 775 }
 776
 777 /**
 778  * dax_pfn_mkwrite - handle first write to DAX page
 779  * @vma: The virtual memory area where the fault occurred
 780  * @vmf: The description of the fault
 781  */
 782 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
 783 {
 784         struct file *file = vma->vm_file;
 785         struct address_space *mapping = file->f_mapping;
 786         void *entry, **slot;
 787         pgoff_t index = vmf->pgoff;
 788
 789         spin_lock_irq(&mapping->tree_lock);
 790         entry = get_unlocked_mapping_entry(mapping, index, &slot);
 791         if (!entry || !radix_tree_exceptional_entry(entry)) {
 792                 if (entry)
 793                         put_unlocked_mapping_entry(mapping, index, entry);
 794                 spin_unlock_irq(&mapping->tree_lock);
 795                 return VM_FAULT_NOPAGE;
 796         }
 797         radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
 798         entry = lock_slot(mapping, slot);
 799         spin_unlock_irq(&mapping->tree_lock);
 800         /*
 801          * If we race with somebody updating the PTE and finish_mkwrite_fault()
 802          * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry
 803          * the fault in either case.
 804          */
 805         finish_mkwrite_fault(vmf);
 806         put_locked_mapping_entry(mapping, index, entry);
 807         return VM_FAULT_NOPAGE;
 808 }
 809 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
 810
 811 static bool dax_range_is_aligned(struct block_device *bdev,
 812                                  unsigned int offset, unsigned int length)
 813 {
 814         unsigned short sector_size = bdev_logical_block_size(bdev);
 815
 816         if (!IS_ALIGNED(offset, sector_size))
 817                 return false;
 818         if (!IS_ALIGNED(length, sector_size))
 819                 return false;
 820
 821         return true;
 822 }
 823
 824 int __dax_zero_page_range(struct block_device *bdev, sector_t sector,
 825                 unsigned int offset, unsigned int length)
 826 {
 827         struct blk_dax_ctl dax = {
 828                 .sector         = sector,
 829                 .size           = PAGE_SIZE,
 830         };
 831
 832         if (dax_range_is_aligned(bdev, offset, length)) {
 833                 sector_t start_sector = dax.sector + (offset >> 9);
 834
 835                 return blkdev_issue_zeroout(bdev, start_sector,
 836                                 length >> 9, GFP_NOFS, true);
 837         } else {
 838                 if (dax_map_atomic(bdev, &dax) < 0)
 839                         return PTR_ERR(dax.addr);
 840                 clear_pmem(dax.addr + offset, length);
 841                 dax_unmap_atomic(bdev, &dax);
 842         }
 843         return 0;
 844 }
 845 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
 846
 847 #ifdef CONFIG_FS_IOMAP
 848 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
 849 {
 850         return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
 851 }
 852
 853 static loff_t
 854 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 855                 struct iomap *iomap)
 856 {
 857         struct iov_iter *iter = data;
 858         loff_t end = pos + length, done = 0;
 859         ssize_t ret = 0;
 860
 861         if (iov_iter_rw(iter) == READ) {
 862                 end = min(end, i_size_read(inode));
 863                 if (pos >= end)
 864                         return 0;
 865
 866                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
 867                         return iov_iter_zero(min(length, end - pos), iter);
 868         }
 869
 870         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
 871                 return -EIO;
 872
 873         while (pos < end) {
 874                 unsigned offset = pos & (PAGE_SIZE - 1);
 875                 struct blk_dax_ctl dax = { 0 };
 876                 ssize_t map_len;
 877
 878                 dax.sector = dax_iomap_sector(iomap, pos);
 879                 dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK;
 880                 map_len = dax_map_atomic(iomap->bdev, &dax);
 881                 if (map_len < 0) {
 882                         ret = map_len;
 883                         break;
 884                 }
 885
 886                 dax.addr += offset;
 887                 map_len -= offset;
 888                 if (map_len > end - pos)
 889                         map_len = end - pos;
 890
 891                 if (iov_iter_rw(iter) == WRITE)
 892                         map_len = copy_from_iter_pmem(dax.addr, map_len, iter);
 893                 else
 894                         map_len = copy_to_iter(dax.addr, map_len, iter);
 895                 dax_unmap_atomic(iomap->bdev, &dax);
 896                 if (map_len <= 0) {
 897                         ret = map_len ? map_len : -EFAULT;
 898                         break;
 899                 }
 900
 901                 pos += map_len;
 902                 length -= map_len;
 903                 done += map_len;
 904         }
 905
 906         return done ? done : ret;
 907 }
 908
 909 /**
 910  * dax_iomap_rw - Perform I/O to a DAX file
 911  * @iocb:       The control block for this I/O
 912  * @iter:       The addresses to do I/O from or to
 913  * @ops:        iomap ops passed from the file system
 914  *
 915  * This function performs read and write operations to directly mapped
 916  * persistent memory.  The callers needs to take care of read/write exclusion
 917  * and evicting any page cache pages in the region under I/O.
 918  */
 919 ssize_t
 920 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
 921                 struct iomap_ops *ops)
 922 {
 923         struct address_space *mapping = iocb->ki_filp->f_mapping;
 924         struct inode *inode = mapping->host;
 925         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
 926         unsigned flags = 0;
 927
 928         if (iov_iter_rw(iter) == WRITE)
 929                 flags |= IOMAP_WRITE;
 930
 931         /*
 932          * Yes, even DAX files can have page cache attached to them:  A zeroed
 933          * page is inserted into the pagecache when we have to serve a write
 934          * fault on a hole.  It should never be dirtied and can simply be
 935          * dropped from the pagecache once we get real data for the page.
 936          *
 937          * XXX: This is racy against mmap, and there's nothing we can do about
 938          * it. We'll eventually need to shift this down even further so that
 939          * we can check if we allocated blocks over a hole first.
 940          */
 941         if (mapping->nrpages) {
 942                 ret = invalidate_inode_pages2_range(mapping,
 943                                 pos >> PAGE_SHIFT,
 944                                 (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT);
 945                 WARN_ON_ONCE(ret);
 946         }
 947
 948         while (iov_iter_count(iter)) {
 949                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
 950                                 iter, dax_iomap_actor);
 951                 if (ret <= 0)
 952                         break;
 953                 pos += ret;
 954                 done += ret;
 955         }
 956
 957         iocb->ki_pos += done;
 958         return done ? done : ret;
 959 }
 960 EXPORT_SYMBOL_GPL(dax_iomap_rw);
 961
 962 /**
 963  * dax_iomap_fault - handle a page fault on a DAX file
 964  * @vma: The virtual memory area where the fault occurred
 965  * @vmf: The description of the fault
 966  * @ops: iomap ops passed from the file system
 967  *
 968  * When a page fault occurs, filesystems may call this helper in their fault
 969  * or mkwrite handler for DAX files. Assumes the caller has done all the
 970  * necessary locking for the page fault to proceed successfully.
 971  */
 972 int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 973                         struct iomap_ops *ops)
 974 {
 975         struct address_space *mapping = vma->vm_file->f_mapping;
 976         struct inode *inode = mapping->host;
 977         unsigned long vaddr = vmf->address;
 978         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
 979         sector_t sector;
 980         struct iomap iomap = { 0 };
 981         unsigned flags = IOMAP_FAULT;
 982         int error, major = 0;
 983         int vmf_ret = 0;
 984         void *entry;
 985
 986         /*
 987          * Check whether offset isn't beyond end of file now. Caller is supposed
 988          * to hold locks serializing us with truncate / punch hole so this is
 989          * a reliable test.
 990          */
 991         if (pos >= i_size_read(inode))
 992                 return VM_FAULT_SIGBUS;
 993
 994         entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
 995         if (IS_ERR(entry)) {
 996                 error = PTR_ERR(entry);
 997                 goto out;
 998         }
 999
1000         if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1001                 flags |= IOMAP_WRITE;
1002
1003         /*
1004          * Note that we don't bother to use iomap_apply here: DAX required
1005          * the file system block size to be equal the page size, which means
1006          * that we never have to deal with more than a single extent here.
1007          */
1008         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1009         if (error)
1010                 goto unlock_entry;
1011         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1012                 error = -EIO;           /* fs corruption? */
1013                 goto finish_iomap;
1014         }
1015
1016         sector = dax_iomap_sector(&iomap, pos);
1017
1018         if (vmf->cow_page) {
1019                 switch (iomap.type) {
1020                 case IOMAP_HOLE:
1021                 case IOMAP_UNWRITTEN:
1022                         clear_user_highpage(vmf->cow_page, vaddr);
1023                         break;
1024                 case IOMAP_MAPPED:
1025                         error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE,
1026                                         vmf->cow_page, vaddr);
1027                         break;
1028                 default:
1029                         WARN_ON_ONCE(1);
1030                         error = -EIO;
1031                         break;
1032                 }
1033
1034                 if (error)
1035                         goto finish_iomap;
1036
1037                 __SetPageUptodate(vmf->cow_page);
1038                 vmf_ret = finish_fault(vmf);
1039                 if (!vmf_ret)
1040                         vmf_ret = VM_FAULT_DONE_COW;
1041                 goto finish_iomap;
1042         }
1043
1044         switch (iomap.type) {
1045         case IOMAP_MAPPED:
1046                 if (iomap.flags & IOMAP_F_NEW) {
1047                         count_vm_event(PGMAJFAULT);
1048                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1049                         major = VM_FAULT_MAJOR;
1050                 }
1051                 error = dax_insert_mapping(mapping, iomap.bdev, sector,
1052                                 PAGE_SIZE, &entry, vma, vmf);
1053                 break;
1054         case IOMAP_UNWRITTEN:
1055         case IOMAP_HOLE:
1056                 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
1057                         vmf_ret = dax_load_hole(mapping, entry, vmf);
1058                         break;
1059                 }
1060                 /*FALLTHRU*/
1061         default:
1062                 WARN_ON_ONCE(1);
1063                 error = -EIO;
1064                 break;
1065         }
1066
1067  finish_iomap:
1068         if (ops->iomap_end) {
1069                 if (error || (vmf_ret & VM_FAULT_ERROR)) {
1070                         /* keep previous error */
1071                         ops->iomap_end(inode, pos, PAGE_SIZE, 0, flags,
1072                                         &iomap);
1073                 } else {
1074                         error = ops->iomap_end(inode, pos, PAGE_SIZE,
1075                                         PAGE_SIZE, flags, &iomap);
1076                 }
1077         }
1078  unlock_entry:
1079         if (vmf_ret != VM_FAULT_LOCKED || error)
1080                 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1081  out:
1082         if (error == -ENOMEM)
1083                 return VM_FAULT_OOM | major;
1084         /* -EBUSY is fine, somebody else faulted on the same PTE */
1085         if (error < 0 && error != -EBUSY)
1086                 return VM_FAULT_SIGBUS | major;
1087         if (vmf_ret) {
1088                 WARN_ON_ONCE(error); /* -EBUSY from ops->iomap_end? */
1089                 return vmf_ret;
1090         }
1091         return VM_FAULT_NOPAGE | major;
1092 }
1093 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1094
1095 #ifdef CONFIG_FS_DAX_PMD
1096 /*
1097  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
1098  * more often than one might expect in the below functions.
1099  */
1100 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
1101
1102 static int dax_pmd_insert_mapping(struct vm_area_struct *vma, pmd_t *pmd,
1103                 struct vm_fault *vmf, unsigned long address,
1104                 struct iomap *iomap, loff_t pos, bool write, void **entryp)
1105 {
1106         struct address_space *mapping = vma->vm_file->f_mapping;
1107         struct block_device *bdev = iomap->bdev;
1108         struct blk_dax_ctl dax = {
1109                 .sector = dax_iomap_sector(iomap, pos),
1110                 .size = PMD_SIZE,
1111         };
1112         long length = dax_map_atomic(bdev, &dax);
1113         void *ret;
1114
1115         if (length < 0) /* dax_map_atomic() failed */
1116                 return VM_FAULT_FALLBACK;
1117         if (length < PMD_SIZE)
1118                 goto unmap_fallback;
1119         if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR)
1120                 goto unmap_fallback;
1121         if (!pfn_t_devmap(dax.pfn))
1122                 goto unmap_fallback;
1123
1124         dax_unmap_atomic(bdev, &dax);
1125
1126         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, dax.sector,
1127                         RADIX_DAX_PMD);
1128         if (IS_ERR(ret))
1129                 return VM_FAULT_FALLBACK;
1130         *entryp = ret;
1131
1132         return vmf_insert_pfn_pmd(vma, address, pmd, dax.pfn, write);
1133
1134  unmap_fallback:
1135         dax_unmap_atomic(bdev, &dax);
1136         return VM_FAULT_FALLBACK;
1137 }
1138
1139 static int dax_pmd_load_hole(struct vm_area_struct *vma, pmd_t *pmd,
1140                 struct vm_fault *vmf, unsigned long address,
1141                 struct iomap *iomap, void **entryp)
1142 {
1143         struct address_space *mapping = vma->vm_file->f_mapping;
1144         unsigned long pmd_addr = address & PMD_MASK;
1145         struct page *zero_page;
1146         spinlock_t *ptl;
1147         pmd_t pmd_entry;
1148         void *ret;
1149
1150         zero_page = mm_get_huge_zero_page(vma->vm_mm);
1151
1152         if (unlikely(!zero_page))
1153                 return VM_FAULT_FALLBACK;
1154
1155         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0,
1156                         RADIX_DAX_PMD | RADIX_DAX_HZP);
1157         if (IS_ERR(ret))
1158                 return VM_FAULT_FALLBACK;
1159         *entryp = ret;
1160
1161         ptl = pmd_lock(vma->vm_mm, pmd);
1162         if (!pmd_none(*pmd)) {
1163                 spin_unlock(ptl);
1164                 return VM_FAULT_FALLBACK;
1165         }
1166
1167         pmd_entry = mk_pmd(zero_page, vma->vm_page_prot);
1168         pmd_entry = pmd_mkhuge(pmd_entry);
1169         set_pmd_at(vma->vm_mm, pmd_addr, pmd, pmd_entry);
1170         spin_unlock(ptl);
1171         return VM_FAULT_NOPAGE;
1172 }
1173
1174 int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1175                 pmd_t *pmd, unsigned int flags, struct iomap_ops *ops)
1176 {
1177         struct address_space *mapping = vma->vm_file->f_mapping;
1178         unsigned long pmd_addr = address & PMD_MASK;
1179         bool write = flags & FAULT_FLAG_WRITE;
1180         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1181         struct inode *inode = mapping->host;
1182         int result = VM_FAULT_FALLBACK;
1183         struct iomap iomap = { 0 };
1184         pgoff_t max_pgoff, pgoff;
1185         struct vm_fault vmf;
1186         void *entry;
1187         loff_t pos;
1188         int error;
1189
1190         /* Fall back to PTEs if we're going to COW */
1191         if (write && !(vma->vm_flags & VM_SHARED))
1192                 goto fallback;
1193
1194         /* If the PMD would extend outside the VMA */
1195         if (pmd_addr < vma->vm_start)
1196                 goto fallback;
1197         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1198                 goto fallback;
1199
1200         /*
1201          * Check whether offset isn't beyond end of file now. Caller is
1202          * supposed to hold locks serializing us with truncate / punch hole so
1203          * this is a reliable test.
1204          */
1205         pgoff = linear_page_index(vma, pmd_addr);
1206         max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1207
1208         if (pgoff > max_pgoff)
1209                 return VM_FAULT_SIGBUS;
1210
1211         /* If the PMD would extend beyond the file size */
1212         if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1213                 goto fallback;
1214
1215         /*
1216          * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1217          * PMD or a HZP entry.  If it can't (because a 4k page is already in
1218          * the tree, for instance), it will return -EEXIST and we just fall
1219          * back to 4k entries.
1220          */
1221         entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1222         if (IS_ERR(entry))
1223                 goto fallback;
1224
1225         /*
1226          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1227          * setting up a mapping, so really we're using iomap_begin() as a way
1228          * to look up our filesystem block.
1229          */
1230         pos = (loff_t)pgoff << PAGE_SHIFT;
1231         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1232         if (error)
1233                 goto unlock_entry;
1234         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1235                 goto finish_iomap;
1236
1237         vmf.pgoff = pgoff;
1238         vmf.flags = flags;
1239         vmf.gfp_mask = mapping_gfp_mask(mapping) | __GFP_IO;
1240
1241         switch (iomap.type) {
1242         case IOMAP_MAPPED:
1243                 result = dax_pmd_insert_mapping(vma, pmd, &vmf, address,
1244                                 &iomap, pos, write, &entry);
1245                 break;
1246         case IOMAP_UNWRITTEN:
1247         case IOMAP_HOLE:
1248                 if (WARN_ON_ONCE(write))
1249                         goto finish_iomap;
1250                 result = dax_pmd_load_hole(vma, pmd, &vmf, address, &iomap,
1251                                 &entry);
1252                 break;
1253         default:
1254                 WARN_ON_ONCE(1);
1255                 break;
1256         }
1257
1258  finish_iomap:
1259         if (ops->iomap_end) {
1260                 if (result == VM_FAULT_FALLBACK) {
1261                         ops->iomap_end(inode, pos, PMD_SIZE, 0, iomap_flags,
1262                                         &iomap);
1263                 } else {
1264                         error = ops->iomap_end(inode, pos, PMD_SIZE, PMD_SIZE,
1265                                         iomap_flags, &iomap);
1266                         if (error)
1267                                 result = VM_FAULT_FALLBACK;
1268                 }
1269         }
1270  unlock_entry:
1271         put_locked_mapping_entry(mapping, pgoff, entry);
1272  fallback:
1273         if (result == VM_FAULT_FALLBACK) {
1274                 split_huge_pmd(vma, pmd, address);
1275                 count_vm_event(THP_FAULT_FALLBACK);
1276         }
1277         return result;
1278 }
1279 EXPORT_SYMBOL_GPL(dax_iomap_pmd_fault);
1280 #endif /* CONFIG_FS_DAX_PMD */
1281 #endif /* CONFIG_FS_IOMAP */