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/sched.h>
  29 #include <linux/sched/signal.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/mmu_notifier.h>
  35 #include <linux/iomap.h>
  36 #include "internal.h"
  37
  38 #define CREATE_TRACE_POINTS
  39 #include <trace/events/fs_dax.h>
  40
  41 /* We choose 4096 entries - same as per-zone page wait tables */
  42 #define DAX_WAIT_TABLE_BITS 12
  43 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
  44
  45 /* The 'colour' (ie low bits) within a PMD of a page offset.  */
  46 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
  47
  48 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
  49
  50 static int __init init_dax_wait_table(void)
  51 {
  52         int i;
  53
  54         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
  55                 init_waitqueue_head(wait_table + i);
  56         return 0;
  57 }
  58 fs_initcall(init_dax_wait_table);
  59
  60 /*
  61  * We use lowest available bit in exceptional entry for locking, one bit for
  62  * the entry size (PMD) and two more to tell us if the entry is a zero page or
  63  * an empty entry that is just used for locking.  In total four special bits.
  64  *
  65  * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
  66  * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
  67  * block allocation.
  68  */
  69 #define RADIX_DAX_SHIFT         (RADIX_TREE_EXCEPTIONAL_SHIFT + 4)
  70 #define RADIX_DAX_ENTRY_LOCK    (1 << RADIX_TREE_EXCEPTIONAL_SHIFT)
  71 #define RADIX_DAX_PMD           (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1))
  72 #define RADIX_DAX_ZERO_PAGE     (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2))
  73 #define RADIX_DAX_EMPTY         (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 3))
  74
  75 static unsigned long dax_radix_sector(void *entry)
  76 {
  77         return (unsigned long)entry >> RADIX_DAX_SHIFT;
  78 }
  79
  80 static void *dax_radix_locked_entry(sector_t sector, unsigned long flags)
  81 {
  82         return (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY | flags |
  83                         ((unsigned long)sector << RADIX_DAX_SHIFT) |
  84                         RADIX_DAX_ENTRY_LOCK);
  85 }
  86
  87 static unsigned int dax_radix_order(void *entry)
  88 {
  89         if ((unsigned long)entry & RADIX_DAX_PMD)
  90                 return PMD_SHIFT - PAGE_SHIFT;
  91         return 0;
  92 }
  93
  94 static int dax_is_pmd_entry(void *entry)
  95 {
  96         return (unsigned long)entry & RADIX_DAX_PMD;
  97 }
  98
  99 static int dax_is_pte_entry(void *entry)
 100 {
 101         return !((unsigned long)entry & RADIX_DAX_PMD);
 102 }
 103
 104 static int dax_is_zero_entry(void *entry)
 105 {
 106         return (unsigned long)entry & RADIX_DAX_ZERO_PAGE;
 107 }
 108
 109 static int dax_is_empty_entry(void *entry)
 110 {
 111         return (unsigned long)entry & RADIX_DAX_EMPTY;
 112 }
 113
 114 /*
 115  * DAX radix tree locking
 116  */
 117 struct exceptional_entry_key {
 118         struct address_space *mapping;
 119         pgoff_t entry_start;
 120 };
 121
 122 struct wait_exceptional_entry_queue {
 123         wait_queue_entry_t wait;
 124         struct exceptional_entry_key key;
 125 };
 126
 127 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
 128                 pgoff_t index, void *entry, struct exceptional_entry_key *key)
 129 {
 130         unsigned long hash;
 131
 132         /*
 133          * If 'entry' is a PMD, align the 'index' that we use for the wait
 134          * queue to the start of that PMD.  This ensures that all offsets in
 135          * the range covered by the PMD map to the same bit lock.
 136          */
 137         if (dax_is_pmd_entry(entry))
 138                 index &= ~PG_PMD_COLOUR;
 139
 140         key->mapping = mapping;
 141         key->entry_start = index;
 142
 143         hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
 144         return wait_table + hash;
 145 }
 146
 147 static int wake_exceptional_entry_func(wait_queue_entry_t *wait, unsigned int mode,
 148                                        int sync, void *keyp)
 149 {
 150         struct exceptional_entry_key *key = keyp;
 151         struct wait_exceptional_entry_queue *ewait =
 152                 container_of(wait, struct wait_exceptional_entry_queue, wait);
 153
 154         if (key->mapping != ewait->key.mapping ||
 155             key->entry_start != ewait->key.entry_start)
 156                 return 0;
 157         return autoremove_wake_function(wait, mode, sync, NULL);
 158 }
 159
 160 /*
 161  * We do not necessarily hold the mapping->tree_lock when we call this
 162  * function so it is possible that 'entry' is no longer a valid item in the
 163  * radix tree.  This is okay because all we really need to do is to find the
 164  * correct waitqueue where tasks might be waiting for that old 'entry' and
 165  * wake them.
 166  */
 167 static void dax_wake_mapping_entry_waiter(struct address_space *mapping,
 168                 pgoff_t index, void *entry, bool wake_all)
 169 {
 170         struct exceptional_entry_key key;
 171         wait_queue_head_t *wq;
 172
 173         wq = dax_entry_waitqueue(mapping, index, entry, &key);
 174
 175         /*
 176          * Checking for locked entry and prepare_to_wait_exclusive() happens
 177          * under mapping->tree_lock, ditto for entry handling in our callers.
 178          * So at this point all tasks that could have seen our entry locked
 179          * must be in the waitqueue and the following check will see them.
 180          */
 181         if (waitqueue_active(wq))
 182                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
 183 }
 184
 185 /*
 186  * Check whether the given slot is locked. The function must be called with
 187  * mapping->tree_lock held
 188  */
 189 static inline int slot_locked(struct address_space *mapping, void **slot)
 190 {
 191         unsigned long entry = (unsigned long)
 192                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 193         return entry & RADIX_DAX_ENTRY_LOCK;
 194 }
 195
 196 /*
 197  * Mark the given slot is locked. The function must be called with
 198  * mapping->tree_lock held
 199  */
 200 static inline void *lock_slot(struct address_space *mapping, void **slot)
 201 {
 202         unsigned long entry = (unsigned long)
 203                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 204
 205         entry |= RADIX_DAX_ENTRY_LOCK;
 206         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
 207         return (void *)entry;
 208 }
 209
 210 /*
 211  * Mark the given slot is unlocked. The function must be called with
 212  * mapping->tree_lock held
 213  */
 214 static inline void *unlock_slot(struct address_space *mapping, void **slot)
 215 {
 216         unsigned long entry = (unsigned long)
 217                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 218
 219         entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
 220         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
 221         return (void *)entry;
 222 }
 223
 224 /*
 225  * Lookup entry in radix tree, wait for it to become unlocked if it is
 226  * exceptional entry and return it. The caller must call
 227  * put_unlocked_mapping_entry() when he decided not to lock the entry or
 228  * put_locked_mapping_entry() when he locked the entry and now wants to
 229  * unlock it.
 230  *
 231  * The function must be called with mapping->tree_lock held.
 232  */
 233 static void *get_unlocked_mapping_entry(struct address_space *mapping,
 234                                         pgoff_t index, void ***slotp)
 235 {
 236         void *entry, **slot;
 237         struct wait_exceptional_entry_queue ewait;
 238         wait_queue_head_t *wq;
 239
 240         init_wait(&ewait.wait);
 241         ewait.wait.func = wake_exceptional_entry_func;
 242
 243         for (;;) {
 244                 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
 245                                           &slot);
 246                 if (!entry ||
 247                     WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)) ||
 248                     !slot_locked(mapping, slot)) {
 249                         if (slotp)
 250                                 *slotp = slot;
 251                         return entry;
 252                 }
 253
 254                 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
 255                 prepare_to_wait_exclusive(wq, &ewait.wait,
 256                                           TASK_UNINTERRUPTIBLE);
 257                 spin_unlock_irq(&mapping->tree_lock);
 258                 schedule();
 259                 finish_wait(wq, &ewait.wait);
 260                 spin_lock_irq(&mapping->tree_lock);
 261         }
 262 }
 263
 264 static void dax_unlock_mapping_entry(struct address_space *mapping,
 265                                      pgoff_t index)
 266 {
 267         void *entry, **slot;
 268
 269         spin_lock_irq(&mapping->tree_lock);
 270         entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
 271         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
 272                          !slot_locked(mapping, slot))) {
 273                 spin_unlock_irq(&mapping->tree_lock);
 274                 return;
 275         }
 276         unlock_slot(mapping, slot);
 277         spin_unlock_irq(&mapping->tree_lock);
 278         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 279 }
 280
 281 static void put_locked_mapping_entry(struct address_space *mapping,
 282                 pgoff_t index)
 283 {
 284         dax_unlock_mapping_entry(mapping, index);
 285 }
 286
 287 /*
 288  * Called when we are done with radix tree entry we looked up via
 289  * get_unlocked_mapping_entry() and which we didn't lock in the end.
 290  */
 291 static void put_unlocked_mapping_entry(struct address_space *mapping,
 292                                        pgoff_t index, void *entry)
 293 {
 294         if (!entry)
 295                 return;
 296
 297         /* We have to wake up next waiter for the radix tree entry lock */
 298         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 299 }
 300
 301 /*
 302  * Find radix tree entry at given index. If it points to an exceptional entry,
 303  * return it with the radix tree entry locked. If the radix tree doesn't
 304  * contain given index, create an empty exceptional entry for the index and
 305  * return with it locked.
 306  *
 307  * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
 308  * either return that locked entry or will return an error.  This error will
 309  * happen if there are any 4k entries within the 2MiB range that we are
 310  * requesting.
 311  *
 312  * We always favor 4k entries over 2MiB entries. There isn't a flow where we
 313  * evict 4k entries in order to 'upgrade' them to a 2MiB entry.  A 2MiB
 314  * insertion will fail if it finds any 4k entries already in the tree, and a
 315  * 4k insertion will cause an existing 2MiB entry to be unmapped and
 316  * downgraded to 4k entries.  This happens for both 2MiB huge zero pages as
 317  * well as 2MiB empty entries.
 318  *
 319  * The exception to this downgrade path is for 2MiB DAX PMD entries that have
 320  * real storage backing them.  We will leave these real 2MiB DAX entries in
 321  * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
 322  *
 323  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
 324  * persistent memory the benefit is doubtful. We can add that later if we can
 325  * show it helps.
 326  */
 327 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
 328                 unsigned long size_flag)
 329 {
 330         bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
 331         void *entry, **slot;
 332
 333 restart:
 334         spin_lock_irq(&mapping->tree_lock);
 335         entry = get_unlocked_mapping_entry(mapping, index, &slot);
 336
 337         if (WARN_ON_ONCE(entry && !radix_tree_exceptional_entry(entry))) {
 338                 entry = ERR_PTR(-EIO);
 339                 goto out_unlock;
 340         }
 341
 342         if (entry) {
 343                 if (size_flag & RADIX_DAX_PMD) {
 344                         if (dax_is_pte_entry(entry)) {
 345                                 put_unlocked_mapping_entry(mapping, index,
 346                                                 entry);
 347                                 entry = ERR_PTR(-EEXIST);
 348                                 goto out_unlock;
 349                         }
 350                 } else { /* trying to grab a PTE entry */
 351                         if (dax_is_pmd_entry(entry) &&
 352                             (dax_is_zero_entry(entry) ||
 353                              dax_is_empty_entry(entry))) {
 354                                 pmd_downgrade = true;
 355                         }
 356                 }
 357         }
 358
 359         /* No entry for given index? Make sure radix tree is big enough. */
 360         if (!entry || pmd_downgrade) {
 361                 int err;
 362
 363                 if (pmd_downgrade) {
 364                         /*
 365                          * Make sure 'entry' remains valid while we drop
 366                          * mapping->tree_lock.
 367                          */
 368                         entry = lock_slot(mapping, slot);
 369                 }
 370
 371                 spin_unlock_irq(&mapping->tree_lock);
 372                 /*
 373                  * Besides huge zero pages the only other thing that gets
 374                  * downgraded are empty entries which don't need to be
 375                  * unmapped.
 376                  */
 377                 if (pmd_downgrade && dax_is_zero_entry(entry))
 378                         unmap_mapping_range(mapping,
 379                                 (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
 380
 381                 err = radix_tree_preload(
 382                                 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
 383                 if (err) {
 384                         if (pmd_downgrade)
 385                                 put_locked_mapping_entry(mapping, index);
 386                         return ERR_PTR(err);
 387                 }
 388                 spin_lock_irq(&mapping->tree_lock);
 389
 390                 if (!entry) {
 391                         /*
 392                          * We needed to drop the page_tree lock while calling
 393                          * radix_tree_preload() and we didn't have an entry to
 394                          * lock.  See if another thread inserted an entry at
 395                          * our index during this time.
 396                          */
 397                         entry = __radix_tree_lookup(&mapping->page_tree, index,
 398                                         NULL, &slot);
 399                         if (entry) {
 400                                 radix_tree_preload_end();
 401                                 spin_unlock_irq(&mapping->tree_lock);
 402                                 goto restart;
 403                         }
 404                 }
 405
 406                 if (pmd_downgrade) {
 407                         radix_tree_delete(&mapping->page_tree, index);
 408                         mapping->nrexceptional--;
 409                         dax_wake_mapping_entry_waiter(mapping, index, entry,
 410                                         true);
 411                 }
 412
 413                 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
 414
 415                 err = __radix_tree_insert(&mapping->page_tree, index,
 416                                 dax_radix_order(entry), entry);
 417                 radix_tree_preload_end();
 418                 if (err) {
 419                         spin_unlock_irq(&mapping->tree_lock);
 420                         /*
 421                          * Our insertion of a DAX entry failed, most likely
 422                          * because we were inserting a PMD entry and it
 423                          * collided with a PTE sized entry at a different
 424                          * index in the PMD range.  We haven't inserted
 425                          * anything into the radix tree and have no waiters to
 426                          * wake.
 427                          */
 428                         return ERR_PTR(err);
 429                 }
 430                 /* Good, we have inserted empty locked entry into the tree. */
 431                 mapping->nrexceptional++;
 432                 spin_unlock_irq(&mapping->tree_lock);
 433                 return entry;
 434         }
 435         entry = lock_slot(mapping, slot);
 436  out_unlock:
 437         spin_unlock_irq(&mapping->tree_lock);
 438         return entry;
 439 }
 440
 441 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
 442                                           pgoff_t index, bool trunc)
 443 {
 444         int ret = 0;
 445         void *entry;
 446         struct radix_tree_root *page_tree = &mapping->page_tree;
 447
 448         spin_lock_irq(&mapping->tree_lock);
 449         entry = get_unlocked_mapping_entry(mapping, index, NULL);
 450         if (!entry || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)))
 451                 goto out;
 452         if (!trunc &&
 453             (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
 454              radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)))
 455                 goto out;
 456         radix_tree_delete(page_tree, index);
 457         mapping->nrexceptional--;
 458         ret = 1;
 459 out:
 460         put_unlocked_mapping_entry(mapping, index, entry);
 461         spin_unlock_irq(&mapping->tree_lock);
 462         return ret;
 463 }
 464 /*
 465  * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
 466  * entry to get unlocked before deleting it.
 467  */
 468 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 469 {
 470         int ret = __dax_invalidate_mapping_entry(mapping, index, true);
 471
 472         /*
 473          * This gets called from truncate / punch_hole path. As such, the caller
 474          * must hold locks protecting against concurrent modifications of the
 475          * radix tree (usually fs-private i_mmap_sem for writing). Since the
 476          * caller has seen exceptional entry for this index, we better find it
 477          * at that index as well...
 478          */
 479         WARN_ON_ONCE(!ret);
 480         return ret;
 481 }
 482
 483 /*
 484  * Invalidate exceptional DAX entry if it is clean.
 485  */
 486 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
 487                                       pgoff_t index)
 488 {
 489         return __dax_invalidate_mapping_entry(mapping, index, false);
 490 }
 491
 492 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
 493                 sector_t sector, size_t size, struct page *to,
 494                 unsigned long vaddr)
 495 {
 496         void *vto, *kaddr;
 497         pgoff_t pgoff;
 498         pfn_t pfn;
 499         long rc;
 500         int id;
 501
 502         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 503         if (rc)
 504                 return rc;
 505
 506         id = dax_read_lock();
 507         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
 508         if (rc < 0) {
 509                 dax_read_unlock(id);
 510                 return rc;
 511         }
 512         vto = kmap_atomic(to);
 513         copy_user_page(vto, (void __force *)kaddr, vaddr, to);
 514         kunmap_atomic(vto);
 515         dax_read_unlock(id);
 516         return 0;
 517 }
 518
 519 /*
 520  * By this point grab_mapping_entry() has ensured that we have a locked entry
 521  * of the appropriate size so we don't have to worry about downgrading PMDs to
 522  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
 523  * already in the tree, we will skip the insertion and just dirty the PMD as
 524  * appropriate.
 525  */
 526 static void *dax_insert_mapping_entry(struct address_space *mapping,
 527                                       struct vm_fault *vmf,
 528                                       void *entry, sector_t sector,
 529                                       unsigned long flags)
 530 {
 531         struct radix_tree_root *page_tree = &mapping->page_tree;
 532         void *new_entry;
 533         pgoff_t index = vmf->pgoff;
 534
 535         if (vmf->flags & FAULT_FLAG_WRITE)
 536                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 537
 538         if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_ZERO_PAGE)) {
 539                 /* we are replacing a zero page with block mapping */
 540                 if (dax_is_pmd_entry(entry))
 541                         unmap_mapping_range(mapping,
 542                                         (vmf->pgoff << PAGE_SHIFT) & PMD_MASK,
 543                                         PMD_SIZE, 0);
 544                 else /* pte entry */
 545                         unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 546                                         PAGE_SIZE, 0);
 547         }
 548
 549         spin_lock_irq(&mapping->tree_lock);
 550         new_entry = dax_radix_locked_entry(sector, flags);
 551
 552         if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 553                 /*
 554                  * Only swap our new entry into the radix tree if the current
 555                  * entry is a zero page or an empty entry.  If a normal PTE or
 556                  * PMD entry is already in the tree, we leave it alone.  This
 557                  * means that if we are trying to insert a PTE and the
 558                  * existing entry is a PMD, we will just leave the PMD in the
 559                  * tree and dirty it if necessary.
 560                  */
 561                 struct radix_tree_node *node;
 562                 void **slot;
 563                 void *ret;
 564
 565                 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
 566                 WARN_ON_ONCE(ret != entry);
 567                 __radix_tree_replace(page_tree, node, slot,
 568                                      new_entry, NULL, NULL);
 569                 entry = new_entry;
 570         }
 571
 572         if (vmf->flags & FAULT_FLAG_WRITE)
 573                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 574
 575         spin_unlock_irq(&mapping->tree_lock);
 576         return entry;
 577 }
 578
 579 static inline unsigned long
 580 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
 581 {
 582         unsigned long address;
 583
 584         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 585         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
 586         return address;
 587 }
 588
 589 /* Walk all mappings of a given index of a file and writeprotect them */
 590 static void dax_mapping_entry_mkclean(struct address_space *mapping,
 591                                       pgoff_t index, unsigned long pfn)
 592 {
 593         struct vm_area_struct *vma;
 594         pte_t pte, *ptep = NULL;
 595         pmd_t *pmdp = NULL;
 596         spinlock_t *ptl;
 597
 598         i_mmap_lock_read(mapping);
 599         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
 600                 unsigned long address, start, end;
 601
 602                 cond_resched();
 603
 604                 if (!(vma->vm_flags & VM_SHARED))
 605                         continue;
 606
 607                 address = pgoff_address(index, vma);
 608
 609                 /*
 610                  * Note because we provide start/end to follow_pte_pmd it will
 611                  * call mmu_notifier_invalidate_range_start() on our behalf
 612                  * before taking any lock.
 613                  */
 614                 if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl))
 615                         continue;
 616
 617                 if (pmdp) {
 618 #ifdef CONFIG_FS_DAX_PMD
 619                         pmd_t pmd;
 620
 621                         if (pfn != pmd_pfn(*pmdp))
 622                                 goto unlock_pmd;
 623                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
 624                                 goto unlock_pmd;
 625
 626                         flush_cache_page(vma, address, pfn);
 627                         pmd = pmdp_huge_clear_flush(vma, address, pmdp);
 628                         pmd = pmd_wrprotect(pmd);
 629                         pmd = pmd_mkclean(pmd);
 630                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
 631                         mmu_notifier_invalidate_range(vma->vm_mm, start, end);
 632 unlock_pmd:
 633                         spin_unlock(ptl);
 634 #endif
 635                 } else {
 636                         if (pfn != pte_pfn(*ptep))
 637                                 goto unlock_pte;
 638                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
 639                                 goto unlock_pte;
 640
 641                         flush_cache_page(vma, address, pfn);
 642                         pte = ptep_clear_flush(vma, address, ptep);
 643                         pte = pte_wrprotect(pte);
 644                         pte = pte_mkclean(pte);
 645                         set_pte_at(vma->vm_mm, address, ptep, pte);
 646                         mmu_notifier_invalidate_range(vma->vm_mm, start, end);
 647 unlock_pte:
 648                         pte_unmap_unlock(ptep, ptl);
 649                 }
 650
 651                 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
 652         }
 653         i_mmap_unlock_read(mapping);
 654 }
 655
 656 static int dax_writeback_one(struct block_device *bdev,
 657                 struct dax_device *dax_dev, struct address_space *mapping,
 658                 pgoff_t index, void *entry)
 659 {
 660         struct radix_tree_root *page_tree = &mapping->page_tree;
 661         void *entry2, **slot, *kaddr;
 662         long ret = 0, id;
 663         sector_t sector;
 664         pgoff_t pgoff;
 665         size_t size;
 666         pfn_t pfn;
 667
 668         /*
 669          * A page got tagged dirty in DAX mapping? Something is seriously
 670          * wrong.
 671          */
 672         if (WARN_ON(!radix_tree_exceptional_entry(entry)))
 673                 return -EIO;
 674
 675         spin_lock_irq(&mapping->tree_lock);
 676         entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
 677         /* Entry got punched out / reallocated? */
 678         if (!entry2 || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry2)))
 679                 goto put_unlocked;
 680         /*
 681          * Entry got reallocated elsewhere? No need to writeback. We have to
 682          * compare sectors as we must not bail out due to difference in lockbit
 683          * or entry type.
 684          */
 685         if (dax_radix_sector(entry2) != dax_radix_sector(entry))
 686                 goto put_unlocked;
 687         if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
 688                                 dax_is_zero_entry(entry))) {
 689                 ret = -EIO;
 690                 goto put_unlocked;
 691         }
 692
 693         /* Another fsync thread may have already written back this entry */
 694         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 695                 goto put_unlocked;
 696         /* Lock the entry to serialize with page faults */
 697         entry = lock_slot(mapping, slot);
 698         /*
 699          * We can clear the tag now but we have to be careful so that concurrent
 700          * dax_writeback_one() calls for the same index cannot finish before we
 701          * actually flush the caches. This is achieved as the calls will look
 702          * at the entry only under tree_lock and once they do that they will
 703          * see the entry locked and wait for it to unlock.
 704          */
 705         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 706         spin_unlock_irq(&mapping->tree_lock);
 707
 708         /*
 709          * Even if dax_writeback_mapping_range() was given a wbc->range_start
 710          * in the middle of a PMD, the 'index' we are given will be aligned to
 711          * the start index of the PMD, as will the sector we pull from
 712          * 'entry'.  This allows us to flush for PMD_SIZE and not have to
 713          * worry about partial PMD writebacks.
 714          */
 715         sector = dax_radix_sector(entry);
 716         size = PAGE_SIZE << dax_radix_order(entry);
 717
 718         id = dax_read_lock();
 719         ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 720         if (ret)
 721                 goto dax_unlock;
 722
 723         /*
 724          * dax_direct_access() may sleep, so cannot hold tree_lock over
 725          * its invocation.
 726          */
 727         ret = dax_direct_access(dax_dev, pgoff, size / PAGE_SIZE, &kaddr, &pfn);
 728         if (ret < 0)
 729                 goto dax_unlock;
 730
 731         if (WARN_ON_ONCE(ret < size / PAGE_SIZE)) {
 732                 ret = -EIO;
 733                 goto dax_unlock;
 734         }
 735
 736         dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(pfn));
 737         dax_flush(dax_dev, kaddr, size);
 738         /*
 739          * After we have flushed the cache, we can clear the dirty tag. There
 740          * cannot be new dirty data in the pfn after the flush has completed as
 741          * the pfn mappings are writeprotected and fault waits for mapping
 742          * entry lock.
 743          */
 744         spin_lock_irq(&mapping->tree_lock);
 745         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
 746         spin_unlock_irq(&mapping->tree_lock);
 747         trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT);
 748  dax_unlock:
 749         dax_read_unlock(id);
 750         put_locked_mapping_entry(mapping, index);
 751         return ret;
 752
 753  put_unlocked:
 754         put_unlocked_mapping_entry(mapping, index, entry2);
 755         spin_unlock_irq(&mapping->tree_lock);
 756         return ret;
 757 }
 758
 759 /*
 760  * Flush the mapping to the persistent domain within the byte range of [start,
 761  * end]. This is required by data integrity operations to ensure file data is
 762  * on persistent storage prior to completion of the operation.
 763  */
 764 int dax_writeback_mapping_range(struct address_space *mapping,
 765                 struct block_device *bdev, struct writeback_control *wbc)
 766 {
 767         struct inode *inode = mapping->host;
 768         pgoff_t start_index, end_index;
 769         pgoff_t indices[PAGEVEC_SIZE];
 770         struct dax_device *dax_dev;
 771         struct pagevec pvec;
 772         bool done = false;
 773         int i, ret = 0;
 774
 775         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 776                 return -EIO;
 777
 778         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 779                 return 0;
 780
 781         dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
 782         if (!dax_dev)
 783                 return -EIO;
 784
 785         start_index = wbc->range_start >> PAGE_SHIFT;
 786         end_index = wbc->range_end >> PAGE_SHIFT;
 787
 788         trace_dax_writeback_range(inode, start_index, end_index);
 789
 790         tag_pages_for_writeback(mapping, start_index, end_index);
 791
 792         pagevec_init(&pvec, 0);
 793         while (!done) {
 794                 pvec.nr = find_get_entries_tag(mapping, start_index,
 795                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 796                                 pvec.pages, indices);
 797
 798                 if (pvec.nr == 0)
 799                         break;
 800
 801                 for (i = 0; i < pvec.nr; i++) {
 802                         if (indices[i] > end_index) {
 803                                 done = true;
 804                                 break;
 805                         }
 806
 807                         ret = dax_writeback_one(bdev, dax_dev, mapping,
 808                                         indices[i], pvec.pages[i]);
 809                         if (ret < 0) {
 810                                 mapping_set_error(mapping, ret);
 811                                 goto out;
 812                         }
 813                 }
 814                 start_index = indices[pvec.nr - 1] + 1;
 815         }
 816 out:
 817         put_dax(dax_dev);
 818         trace_dax_writeback_range_done(inode, start_index, end_index);
 819         return (ret < 0 ? ret : 0);
 820 }
 821 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 822
 823 static int dax_insert_mapping(struct address_space *mapping,
 824                 struct block_device *bdev, struct dax_device *dax_dev,
 825                 sector_t sector, size_t size, void *entry,
 826                 struct vm_area_struct *vma, struct vm_fault *vmf)
 827 {
 828         unsigned long vaddr = vmf->address;
 829         void *ret, *kaddr;
 830         pgoff_t pgoff;
 831         int id, rc;
 832         pfn_t pfn;
 833
 834         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 835         if (rc)
 836                 return rc;
 837
 838         id = dax_read_lock();
 839         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
 840         if (rc < 0) {
 841                 dax_read_unlock(id);
 842                 return rc;
 843         }
 844         dax_read_unlock(id);
 845
 846         ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, 0);
 847         if (IS_ERR(ret))
 848                 return PTR_ERR(ret);
 849
 850         trace_dax_insert_mapping(mapping->host, vmf, ret);
 851         if (vmf->flags & FAULT_FLAG_WRITE)
 852                 return vm_insert_mixed_mkwrite(vma, vaddr, pfn);
 853         else
 854                 return vm_insert_mixed(vma, vaddr, pfn);
 855 }
 856
 857 /*
 858  * The user has performed a load from a hole in the file.  Allocating a new
 859  * page in the file would cause excessive storage usage for workloads with
 860  * sparse files.  Instead we insert a read-only mapping of the 4k zero page.
 861  * If this page is ever written to we will re-fault and change the mapping to
 862  * point to real DAX storage instead.
 863  */
 864 static int dax_load_hole(struct address_space *mapping, void *entry,
 865                          struct vm_fault *vmf)
 866 {
 867         struct inode *inode = mapping->host;
 868         unsigned long vaddr = vmf->address;
 869         int ret = VM_FAULT_NOPAGE;
 870         struct page *zero_page;
 871         void *entry2;
 872
 873         zero_page = ZERO_PAGE(0);
 874         if (unlikely(!zero_page)) {
 875                 ret = VM_FAULT_OOM;
 876                 goto out;
 877         }
 878
 879         entry2 = dax_insert_mapping_entry(mapping, vmf, entry, 0,
 880                         RADIX_DAX_ZERO_PAGE);
 881         if (IS_ERR(entry2)) {
 882                 ret = VM_FAULT_SIGBUS;
 883                 goto out;
 884         }
 885
 886         vm_insert_mixed(vmf->vma, vaddr, page_to_pfn_t(zero_page));
 887 out:
 888         trace_dax_load_hole(inode, vmf, ret);
 889         return ret;
 890 }
 891
 892 static bool dax_range_is_aligned(struct block_device *bdev,
 893                                  unsigned int offset, unsigned int length)
 894 {
 895         unsigned short sector_size = bdev_logical_block_size(bdev);
 896
 897         if (!IS_ALIGNED(offset, sector_size))
 898                 return false;
 899         if (!IS_ALIGNED(length, sector_size))
 900                 return false;
 901
 902         return true;
 903 }
 904
 905 int __dax_zero_page_range(struct block_device *bdev,
 906                 struct dax_device *dax_dev, sector_t sector,
 907                 unsigned int offset, unsigned int size)
 908 {
 909         if (dax_range_is_aligned(bdev, offset, size)) {
 910                 sector_t start_sector = sector + (offset >> 9);
 911
 912                 return blkdev_issue_zeroout(bdev, start_sector,
 913                                 size >> 9, GFP_NOFS, 0);
 914         } else {
 915                 pgoff_t pgoff;
 916                 long rc, id;
 917                 void *kaddr;
 918                 pfn_t pfn;
 919
 920                 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
 921                 if (rc)
 922                         return rc;
 923
 924                 id = dax_read_lock();
 925                 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr,
 926                                 &pfn);
 927                 if (rc < 0) {
 928                         dax_read_unlock(id);
 929                         return rc;
 930                 }
 931                 memset(kaddr + offset, 0, size);
 932                 dax_flush(dax_dev, kaddr + offset, size);
 933                 dax_read_unlock(id);
 934         }
 935         return 0;
 936 }
 937 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
 938
 939 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
 940 {
 941         return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
 942 }
 943
 944 static loff_t
 945 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 946                 struct iomap *iomap)
 947 {
 948         struct block_device *bdev = iomap->bdev;
 949         struct dax_device *dax_dev = iomap->dax_dev;
 950         struct iov_iter *iter = data;
 951         loff_t end = pos + length, done = 0;
 952         ssize_t ret = 0;
 953         int id;
 954
 955         if (iov_iter_rw(iter) == READ) {
 956                 end = min(end, i_size_read(inode));
 957                 if (pos >= end)
 958                         return 0;
 959
 960                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
 961                         return iov_iter_zero(min(length, end - pos), iter);
 962         }
 963
 964         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
 965                 return -EIO;
 966
 967         /*
 968          * Write can allocate block for an area which has a hole page mapped
 969          * into page tables. We have to tear down these mappings so that data
 970          * written by write(2) is visible in mmap.
 971          */
 972         if (iomap->flags & IOMAP_F_NEW) {
 973                 invalidate_inode_pages2_range(inode->i_mapping,
 974                                               pos >> PAGE_SHIFT,
 975                                               (end - 1) >> PAGE_SHIFT);
 976         }
 977
 978         id = dax_read_lock();
 979         while (pos < end) {
 980                 unsigned offset = pos & (PAGE_SIZE - 1);
 981                 const size_t size = ALIGN(length + offset, PAGE_SIZE);
 982                 const sector_t sector = dax_iomap_sector(iomap, pos);
 983                 ssize_t map_len;
 984                 pgoff_t pgoff;
 985                 void *kaddr;
 986                 pfn_t pfn;
 987
 988                 if (fatal_signal_pending(current)) {
 989                         ret = -EINTR;
 990                         break;
 991                 }
 992
 993                 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 994                 if (ret)
 995                         break;
 996
 997                 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
 998                                 &kaddr, &pfn);
 999                 if (map_len < 0) {
1000                         ret = map_len;
1001                         break;
1002                 }
1003
1004                 map_len = PFN_PHYS(map_len);
1005                 kaddr += offset;
1006                 map_len -= offset;
1007                 if (map_len > end - pos)
1008                         map_len = end - pos;
1009
1010                 /*
1011                  * The userspace address for the memory copy has already been
1012                  * validated via access_ok() in either vfs_read() or
1013                  * vfs_write(), depending on which operation we are doing.
1014                  */
1015                 if (iov_iter_rw(iter) == WRITE)
1016                         map_len = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1017                                         map_len, iter);
1018                 else
1019                         map_len = copy_to_iter(kaddr, map_len, iter);
1020                 if (map_len <= 0) {
1021                         ret = map_len ? map_len : -EFAULT;
1022                         break;
1023                 }
1024
1025                 pos += map_len;
1026                 length -= map_len;
1027                 done += map_len;
1028         }
1029         dax_read_unlock(id);
1030
1031         return done ? done : ret;
1032 }
1033
1034 /**
1035  * dax_iomap_rw - Perform I/O to a DAX file
1036  * @iocb:       The control block for this I/O
1037  * @iter:       The addresses to do I/O from or to
1038  * @ops:        iomap ops passed from the file system
1039  *
1040  * This function performs read and write operations to directly mapped
1041  * persistent memory.  The callers needs to take care of read/write exclusion
1042  * and evicting any page cache pages in the region under I/O.
1043  */
1044 ssize_t
1045 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1046                 const struct iomap_ops *ops)
1047 {
1048         struct address_space *mapping = iocb->ki_filp->f_mapping;
1049         struct inode *inode = mapping->host;
1050         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1051         unsigned flags = 0;
1052
1053         if (iov_iter_rw(iter) == WRITE) {
1054                 lockdep_assert_held_exclusive(&inode->i_rwsem);
1055                 flags |= IOMAP_WRITE;
1056         } else {
1057                 lockdep_assert_held(&inode->i_rwsem);
1058         }
1059
1060         while (iov_iter_count(iter)) {
1061                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1062                                 iter, dax_iomap_actor);
1063                 if (ret <= 0)
1064                         break;
1065                 pos += ret;
1066                 done += ret;
1067         }
1068
1069         iocb->ki_pos += done;
1070         return done ? done : ret;
1071 }
1072 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1073
1074 static int dax_fault_return(int error)
1075 {
1076         if (error == 0)
1077                 return VM_FAULT_NOPAGE;
1078         if (error == -ENOMEM)
1079                 return VM_FAULT_OOM;
1080         return VM_FAULT_SIGBUS;
1081 }
1082
1083 static int dax_iomap_pte_fault(struct vm_fault *vmf,
1084                                const struct iomap_ops *ops)
1085 {
1086         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1087         struct inode *inode = mapping->host;
1088         unsigned long vaddr = vmf->address;
1089         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1090         sector_t sector;
1091         struct iomap iomap = { 0 };
1092         unsigned flags = IOMAP_FAULT;
1093         int error, major = 0;
1094         int vmf_ret = 0;
1095         void *entry;
1096
1097         trace_dax_pte_fault(inode, vmf, vmf_ret);
1098         /*
1099          * Check whether offset isn't beyond end of file now. Caller is supposed
1100          * to hold locks serializing us with truncate / punch hole so this is
1101          * a reliable test.
1102          */
1103         if (pos >= i_size_read(inode)) {
1104                 vmf_ret = VM_FAULT_SIGBUS;
1105                 goto out;
1106         }
1107
1108         if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1109                 flags |= IOMAP_WRITE;
1110
1111         entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1112         if (IS_ERR(entry)) {
1113                 vmf_ret = dax_fault_return(PTR_ERR(entry));
1114                 goto out;
1115         }
1116
1117         /*
1118          * It is possible, particularly with mixed reads & writes to private
1119          * mappings, that we have raced with a PMD fault that overlaps with
1120          * the PTE we need to set up.  If so just return and the fault will be
1121          * retried.
1122          */
1123         if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1124                 vmf_ret = VM_FAULT_NOPAGE;
1125                 goto unlock_entry;
1126         }
1127
1128         /*
1129          * Note that we don't bother to use iomap_apply here: DAX required
1130          * the file system block size to be equal the page size, which means
1131          * that we never have to deal with more than a single extent here.
1132          */
1133         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1134         if (error) {
1135                 vmf_ret = dax_fault_return(error);
1136                 goto unlock_entry;
1137         }
1138         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1139                 error = -EIO;   /* fs corruption? */
1140                 goto error_finish_iomap;
1141         }
1142
1143         sector = dax_iomap_sector(&iomap, pos);
1144
1145         if (vmf->cow_page) {
1146                 switch (iomap.type) {
1147                 case IOMAP_HOLE:
1148                 case IOMAP_UNWRITTEN:
1149                         clear_user_highpage(vmf->cow_page, vaddr);
1150                         break;
1151                 case IOMAP_MAPPED:
1152                         error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1153                                         sector, PAGE_SIZE, vmf->cow_page, vaddr);
1154                         break;
1155                 default:
1156                         WARN_ON_ONCE(1);
1157                         error = -EIO;
1158                         break;
1159                 }
1160
1161                 if (error)
1162                         goto error_finish_iomap;
1163
1164                 __SetPageUptodate(vmf->cow_page);
1165                 vmf_ret = finish_fault(vmf);
1166                 if (!vmf_ret)
1167                         vmf_ret = VM_FAULT_DONE_COW;
1168                 goto finish_iomap;
1169         }
1170
1171         switch (iomap.type) {
1172         case IOMAP_MAPPED:
1173                 if (iomap.flags & IOMAP_F_NEW) {
1174                         count_vm_event(PGMAJFAULT);
1175                         count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
1176                         major = VM_FAULT_MAJOR;
1177                 }
1178                 error = dax_insert_mapping(mapping, iomap.bdev, iomap.dax_dev,
1179                                 sector, PAGE_SIZE, entry, vmf->vma, vmf);
1180                 /* -EBUSY is fine, somebody else faulted on the same PTE */
1181                 if (error == -EBUSY)
1182                         error = 0;
1183                 break;
1184         case IOMAP_UNWRITTEN:
1185         case IOMAP_HOLE:
1186                 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
1187                         vmf_ret = dax_load_hole(mapping, entry, vmf);
1188                         goto finish_iomap;
1189                 }
1190                 /*FALLTHRU*/
1191         default:
1192                 WARN_ON_ONCE(1);
1193                 error = -EIO;
1194                 break;
1195         }
1196
1197  error_finish_iomap:
1198         vmf_ret = dax_fault_return(error) | major;
1199  finish_iomap:
1200         if (ops->iomap_end) {
1201                 int copied = PAGE_SIZE;
1202
1203                 if (vmf_ret & VM_FAULT_ERROR)
1204                         copied = 0;
1205                 /*
1206                  * The fault is done by now and there's no way back (other
1207                  * thread may be already happily using PTE we have installed).
1208                  * Just ignore error from ->iomap_end since we cannot do much
1209                  * with it.
1210                  */
1211                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1212         }
1213  unlock_entry:
1214         put_locked_mapping_entry(mapping, vmf->pgoff);
1215  out:
1216         trace_dax_pte_fault_done(inode, vmf, vmf_ret);
1217         return vmf_ret;
1218 }
1219
1220 #ifdef CONFIG_FS_DAX_PMD
1221 static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap,
1222                 loff_t pos, void *entry)
1223 {
1224         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1225         const sector_t sector = dax_iomap_sector(iomap, pos);
1226         struct dax_device *dax_dev = iomap->dax_dev;
1227         struct block_device *bdev = iomap->bdev;
1228         struct inode *inode = mapping->host;
1229         const size_t size = PMD_SIZE;
1230         void *ret = NULL, *kaddr;
1231         long length = 0;
1232         pgoff_t pgoff;
1233         pfn_t pfn = {};
1234         int id;
1235
1236         if (bdev_dax_pgoff(bdev, sector, size, &pgoff) != 0)
1237                 goto fallback;
1238
1239         id = dax_read_lock();
1240         length = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
1241         if (length < 0)
1242                 goto unlock_fallback;
1243         length = PFN_PHYS(length);
1244
1245         if (length < size)
1246                 goto unlock_fallback;
1247         if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR)
1248                 goto unlock_fallback;
1249         if (!pfn_t_devmap(pfn))
1250                 goto unlock_fallback;
1251         dax_read_unlock(id);
1252
1253         ret = dax_insert_mapping_entry(mapping, vmf, entry, sector,
1254                         RADIX_DAX_PMD);
1255         if (IS_ERR(ret))
1256                 goto fallback;
1257
1258         trace_dax_pmd_insert_mapping(inode, vmf, length, pfn, ret);
1259         return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
1260                         pfn, vmf->flags & FAULT_FLAG_WRITE);
1261
1262 unlock_fallback:
1263         dax_read_unlock(id);
1264 fallback:
1265         trace_dax_pmd_insert_mapping_fallback(inode, vmf, length, pfn, ret);
1266         return VM_FAULT_FALLBACK;
1267 }
1268
1269 static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
1270                 void *entry)
1271 {
1272         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1273         unsigned long pmd_addr = vmf->address & PMD_MASK;
1274         struct inode *inode = mapping->host;
1275         struct page *zero_page;
1276         void *ret = NULL;
1277         spinlock_t *ptl;
1278         pmd_t pmd_entry;
1279
1280         zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1281
1282         if (unlikely(!zero_page))
1283                 goto fallback;
1284
1285         ret = dax_insert_mapping_entry(mapping, vmf, entry, 0,
1286                         RADIX_DAX_PMD | RADIX_DAX_ZERO_PAGE);
1287         if (IS_ERR(ret))
1288                 goto fallback;
1289
1290         ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1291         if (!pmd_none(*(vmf->pmd))) {
1292                 spin_unlock(ptl);
1293                 goto fallback;
1294         }
1295
1296         pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1297         pmd_entry = pmd_mkhuge(pmd_entry);
1298         set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1299         spin_unlock(ptl);
1300         trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
1301         return VM_FAULT_NOPAGE;
1302
1303 fallback:
1304         trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
1305         return VM_FAULT_FALLBACK;
1306 }
1307
1308 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1309                                const struct iomap_ops *ops)
1310 {
1311         struct vm_area_struct *vma = vmf->vma;
1312         struct address_space *mapping = vma->vm_file->f_mapping;
1313         unsigned long pmd_addr = vmf->address & PMD_MASK;
1314         bool write = vmf->flags & FAULT_FLAG_WRITE;
1315         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1316         struct inode *inode = mapping->host;
1317         int result = VM_FAULT_FALLBACK;
1318         struct iomap iomap = { 0 };
1319         pgoff_t max_pgoff, pgoff;
1320         void *entry;
1321         loff_t pos;
1322         int error;
1323
1324         /*
1325          * Check whether offset isn't beyond end of file now. Caller is
1326          * supposed to hold locks serializing us with truncate / punch hole so
1327          * this is a reliable test.
1328          */
1329         pgoff = linear_page_index(vma, pmd_addr);
1330         max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1331
1332         trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1333
1334         /*
1335          * Make sure that the faulting address's PMD offset (color) matches
1336          * the PMD offset from the start of the file.  This is necessary so
1337          * that a PMD range in the page table overlaps exactly with a PMD
1338          * range in the radix tree.
1339          */
1340         if ((vmf->pgoff & PG_PMD_COLOUR) !=
1341             ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1342                 goto fallback;
1343
1344         /* Fall back to PTEs if we're going to COW */
1345         if (write && !(vma->vm_flags & VM_SHARED))
1346                 goto fallback;
1347
1348         /* If the PMD would extend outside the VMA */
1349         if (pmd_addr < vma->vm_start)
1350                 goto fallback;
1351         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1352                 goto fallback;
1353
1354         if (pgoff > max_pgoff) {
1355                 result = VM_FAULT_SIGBUS;
1356                 goto out;
1357         }
1358
1359         /* If the PMD would extend beyond the file size */
1360         if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1361                 goto fallback;
1362
1363         /*
1364          * grab_mapping_entry() will make sure we get a 2MiB empty entry, a
1365          * 2MiB zero page entry or a DAX PMD.  If it can't (because a 4k page
1366          * is already in the tree, for instance), it will return -EEXIST and
1367          * we just fall back to 4k entries.
1368          */
1369         entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1370         if (IS_ERR(entry))
1371                 goto fallback;
1372
1373         /*
1374          * It is possible, particularly with mixed reads & writes to private
1375          * mappings, that we have raced with a PTE fault that overlaps with
1376          * the PMD we need to set up.  If so just return and the fault will be
1377          * retried.
1378          */
1379         if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1380                         !pmd_devmap(*vmf->pmd)) {
1381                 result = 0;
1382                 goto unlock_entry;
1383         }
1384
1385         /*
1386          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1387          * setting up a mapping, so really we're using iomap_begin() as a way
1388          * to look up our filesystem block.
1389          */
1390         pos = (loff_t)pgoff << PAGE_SHIFT;
1391         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1392         if (error)
1393                 goto unlock_entry;
1394
1395         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1396                 goto finish_iomap;
1397
1398         switch (iomap.type) {
1399         case IOMAP_MAPPED:
1400                 result = dax_pmd_insert_mapping(vmf, &iomap, pos, entry);
1401                 break;
1402         case IOMAP_UNWRITTEN:
1403         case IOMAP_HOLE:
1404                 if (WARN_ON_ONCE(write))
1405                         break;
1406                 result = dax_pmd_load_hole(vmf, &iomap, entry);
1407                 break;
1408         default:
1409                 WARN_ON_ONCE(1);
1410                 break;
1411         }
1412
1413  finish_iomap:
1414         if (ops->iomap_end) {
1415                 int copied = PMD_SIZE;
1416
1417                 if (result == VM_FAULT_FALLBACK)
1418                         copied = 0;
1419                 /*
1420                  * The fault is done by now and there's no way back (other
1421                  * thread may be already happily using PMD we have installed).
1422                  * Just ignore error from ->iomap_end since we cannot do much
1423                  * with it.
1424                  */
1425                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1426                                 &iomap);
1427         }
1428  unlock_entry:
1429         put_locked_mapping_entry(mapping, pgoff);
1430  fallback:
1431         if (result == VM_FAULT_FALLBACK) {
1432                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1433                 count_vm_event(THP_FAULT_FALLBACK);
1434         }
1435 out:
1436         trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1437         return result;
1438 }
1439 #else
1440 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1441                                const struct iomap_ops *ops)
1442 {
1443         return VM_FAULT_FALLBACK;
1444 }
1445 #endif /* CONFIG_FS_DAX_PMD */
1446
1447 /**
1448  * dax_iomap_fault - handle a page fault on a DAX file
1449  * @vmf: The description of the fault
1450  * @ops: iomap ops passed from the file system
1451  *
1452  * When a page fault occurs, filesystems may call this helper in
1453  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1454  * has done all the necessary locking for page fault to proceed
1455  * successfully.
1456  */
1457 int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1458                     const struct iomap_ops *ops)
1459 {
1460         switch (pe_size) {
1461         case PE_SIZE_PTE:
1462                 return dax_iomap_pte_fault(vmf, ops);
1463         case PE_SIZE_PMD:
1464                 return dax_iomap_pmd_fault(vmf, ops);
1465         default:
1466                 return VM_FAULT_FALLBACK;
1467         }
1468 }
1469 EXPORT_SYMBOL_GPL(dax_iomap_fault);