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/sched/signal.h>
  31 #include <linux/uio.h>
  32 #include <linux/vmstat.h>
  33 #include <linux/pfn_t.h>
  34 #include <linux/sizes.h>
  35 #include <linux/mmu_notifier.h>
  36 #include <linux/iomap.h>
  37 #include "internal.h"
  38
  39 #define CREATE_TRACE_POINTS
  40 #include <trace/events/fs_dax.h>
  41
  42 /* We choose 4096 entries - same as per-zone page wait tables */
  43 #define DAX_WAIT_TABLE_BITS 12
  44 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
  45
  46 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
  47
  48 static int __init init_dax_wait_table(void)
  49 {
  50         int i;
  51
  52         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
  53                 init_waitqueue_head(wait_table + i);
  54         return 0;
  55 }
  56 fs_initcall(init_dax_wait_table);
  57
  58 static int dax_is_pmd_entry(void *entry)
  59 {
  60         return (unsigned long)entry & RADIX_DAX_PMD;
  61 }
  62
  63 static int dax_is_pte_entry(void *entry)
  64 {
  65         return !((unsigned long)entry & RADIX_DAX_PMD);
  66 }
  67
  68 static int dax_is_zero_entry(void *entry)
  69 {
  70         return (unsigned long)entry & RADIX_DAX_HZP;
  71 }
  72
  73 static int dax_is_empty_entry(void *entry)
  74 {
  75         return (unsigned long)entry & RADIX_DAX_EMPTY;
  76 }
  77
  78 /*
  79  * DAX radix tree locking
  80  */
  81 struct exceptional_entry_key {
  82         struct address_space *mapping;
  83         pgoff_t entry_start;
  84 };
  85
  86 struct wait_exceptional_entry_queue {
  87         wait_queue_t wait;
  88         struct exceptional_entry_key key;
  89 };
  90
  91 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
  92                 pgoff_t index, void *entry, struct exceptional_entry_key *key)
  93 {
  94         unsigned long hash;
  95
  96         /*
  97          * If 'entry' is a PMD, align the 'index' that we use for the wait
  98          * queue to the start of that PMD.  This ensures that all offsets in
  99          * the range covered by the PMD map to the same bit lock.
 100          */
 101         if (dax_is_pmd_entry(entry))
 102                 index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
 103
 104         key->mapping = mapping;
 105         key->entry_start = index;
 106
 107         hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
 108         return wait_table + hash;
 109 }
 110
 111 static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
 112                                        int sync, void *keyp)
 113 {
 114         struct exceptional_entry_key *key = keyp;
 115         struct wait_exceptional_entry_queue *ewait =
 116                 container_of(wait, struct wait_exceptional_entry_queue, wait);
 117
 118         if (key->mapping != ewait->key.mapping ||
 119             key->entry_start != ewait->key.entry_start)
 120                 return 0;
 121         return autoremove_wake_function(wait, mode, sync, NULL);
 122 }
 123
 124 /*
 125  * Check whether the given slot is locked. The function must be called with
 126  * mapping->tree_lock held
 127  */
 128 static inline int slot_locked(struct address_space *mapping, void **slot)
 129 {
 130         unsigned long entry = (unsigned long)
 131                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 132         return entry & RADIX_DAX_ENTRY_LOCK;
 133 }
 134
 135 /*
 136  * Mark the given slot is locked. The function must be called with
 137  * mapping->tree_lock held
 138  */
 139 static inline void *lock_slot(struct address_space *mapping, void **slot)
 140 {
 141         unsigned long entry = (unsigned long)
 142                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 143
 144         entry |= RADIX_DAX_ENTRY_LOCK;
 145         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
 146         return (void *)entry;
 147 }
 148
 149 /*
 150  * Mark the given slot is unlocked. The function must be called with
 151  * mapping->tree_lock held
 152  */
 153 static inline void *unlock_slot(struct address_space *mapping, void **slot)
 154 {
 155         unsigned long entry = (unsigned long)
 156                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 157
 158         entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
 159         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
 160         return (void *)entry;
 161 }
 162
 163 /*
 164  * Lookup entry in radix tree, wait for it to become unlocked if it is
 165  * exceptional entry and return it. The caller must call
 166  * put_unlocked_mapping_entry() when he decided not to lock the entry or
 167  * put_locked_mapping_entry() when he locked the entry and now wants to
 168  * unlock it.
 169  *
 170  * The function must be called with mapping->tree_lock held.
 171  */
 172 static void *get_unlocked_mapping_entry(struct address_space *mapping,
 173                                         pgoff_t index, void ***slotp)
 174 {
 175         void *entry, **slot;
 176         struct wait_exceptional_entry_queue ewait;
 177         wait_queue_head_t *wq;
 178
 179         init_wait(&ewait.wait);
 180         ewait.wait.func = wake_exceptional_entry_func;
 181
 182         for (;;) {
 183                 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
 184                                           &slot);
 185                 if (!entry || !radix_tree_exceptional_entry(entry) ||
 186                     !slot_locked(mapping, slot)) {
 187                         if (slotp)
 188                                 *slotp = slot;
 189                         return entry;
 190                 }
 191
 192                 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
 193                 prepare_to_wait_exclusive(wq, &ewait.wait,
 194                                           TASK_UNINTERRUPTIBLE);
 195                 spin_unlock_irq(&mapping->tree_lock);
 196                 schedule();
 197                 finish_wait(wq, &ewait.wait);
 198                 spin_lock_irq(&mapping->tree_lock);
 199         }
 200 }
 201
 202 static void dax_unlock_mapping_entry(struct address_space *mapping,
 203                                      pgoff_t index)
 204 {
 205         void *entry, **slot;
 206
 207         spin_lock_irq(&mapping->tree_lock);
 208         entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
 209         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
 210                          !slot_locked(mapping, slot))) {
 211                 spin_unlock_irq(&mapping->tree_lock);
 212                 return;
 213         }
 214         unlock_slot(mapping, slot);
 215         spin_unlock_irq(&mapping->tree_lock);
 216         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 217 }
 218
 219 static void put_locked_mapping_entry(struct address_space *mapping,
 220                                      pgoff_t index, void *entry)
 221 {
 222         if (!radix_tree_exceptional_entry(entry)) {
 223                 unlock_page(entry);
 224                 put_page(entry);
 225         } else {
 226                 dax_unlock_mapping_entry(mapping, index);
 227         }
 228 }
 229
 230 /*
 231  * Called when we are done with radix tree entry we looked up via
 232  * get_unlocked_mapping_entry() and which we didn't lock in the end.
 233  */
 234 static void put_unlocked_mapping_entry(struct address_space *mapping,
 235                                        pgoff_t index, void *entry)
 236 {
 237         if (!radix_tree_exceptional_entry(entry))
 238                 return;
 239
 240         /* We have to wake up next waiter for the radix tree entry lock */
 241         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 242 }
 243
 244 /*
 245  * Find radix tree entry at given index. If it points to a page, return with
 246  * the page locked. If it points to the exceptional entry, return with the
 247  * radix tree entry locked. If the radix tree doesn't contain given index,
 248  * create empty exceptional entry for the index and return with it locked.
 249  *
 250  * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
 251  * either return that locked entry or will return an error.  This error will
 252  * happen if there are any 4k entries (either zero pages or DAX entries)
 253  * within the 2MiB range that we are requesting.
 254  *
 255  * We always favor 4k entries over 2MiB entries. There isn't a flow where we
 256  * evict 4k entries in order to 'upgrade' them to a 2MiB entry.  A 2MiB
 257  * insertion will fail if it finds any 4k entries already in the tree, and a
 258  * 4k insertion will cause an existing 2MiB entry to be unmapped and
 259  * downgraded to 4k entries.  This happens for both 2MiB huge zero pages as
 260  * well as 2MiB empty entries.
 261  *
 262  * The exception to this downgrade path is for 2MiB DAX PMD entries that have
 263  * real storage backing them.  We will leave these real 2MiB DAX entries in
 264  * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
 265  *
 266  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
 267  * persistent memory the benefit is doubtful. We can add that later if we can
 268  * show it helps.
 269  */
 270 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
 271                 unsigned long size_flag)
 272 {
 273         bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
 274         void *entry, **slot;
 275
 276 restart:
 277         spin_lock_irq(&mapping->tree_lock);
 278         entry = get_unlocked_mapping_entry(mapping, index, &slot);
 279
 280         if (entry) {
 281                 if (size_flag & RADIX_DAX_PMD) {
 282                         if (!radix_tree_exceptional_entry(entry) ||
 283                             dax_is_pte_entry(entry)) {
 284                                 put_unlocked_mapping_entry(mapping, index,
 285                                                 entry);
 286                                 entry = ERR_PTR(-EEXIST);
 287                                 goto out_unlock;
 288                         }
 289                 } else { /* trying to grab a PTE entry */
 290                         if (radix_tree_exceptional_entry(entry) &&
 291                             dax_is_pmd_entry(entry) &&
 292                             (dax_is_zero_entry(entry) ||
 293                              dax_is_empty_entry(entry))) {
 294                                 pmd_downgrade = true;
 295                         }
 296                 }
 297         }
 298
 299         /* No entry for given index? Make sure radix tree is big enough. */
 300         if (!entry || pmd_downgrade) {
 301                 int err;
 302
 303                 if (pmd_downgrade) {
 304                         /*
 305                          * Make sure 'entry' remains valid while we drop
 306                          * mapping->tree_lock.
 307                          */
 308                         entry = lock_slot(mapping, slot);
 309                 }
 310
 311                 spin_unlock_irq(&mapping->tree_lock);
 312                 /*
 313                  * Besides huge zero pages the only other thing that gets
 314                  * downgraded are empty entries which don't need to be
 315                  * unmapped.
 316                  */
 317                 if (pmd_downgrade && dax_is_zero_entry(entry))
 318                         unmap_mapping_range(mapping,
 319                                 (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
 320
 321                 err = radix_tree_preload(
 322                                 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
 323                 if (err) {
 324                         if (pmd_downgrade)
 325                                 put_locked_mapping_entry(mapping, index, entry);
 326                         return ERR_PTR(err);
 327                 }
 328                 spin_lock_irq(&mapping->tree_lock);
 329
 330                 if (!entry) {
 331                         /*
 332                          * We needed to drop the page_tree lock while calling
 333                          * radix_tree_preload() and we didn't have an entry to
 334                          * lock.  See if another thread inserted an entry at
 335                          * our index during this time.
 336                          */
 337                         entry = __radix_tree_lookup(&mapping->page_tree, index,
 338                                         NULL, &slot);
 339                         if (entry) {
 340                                 radix_tree_preload_end();
 341                                 spin_unlock_irq(&mapping->tree_lock);
 342                                 goto restart;
 343                         }
 344                 }
 345
 346                 if (pmd_downgrade) {
 347                         radix_tree_delete(&mapping->page_tree, index);
 348                         mapping->nrexceptional--;
 349                         dax_wake_mapping_entry_waiter(mapping, index, entry,
 350                                         true);
 351                 }
 352
 353                 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
 354
 355                 err = __radix_tree_insert(&mapping->page_tree, index,
 356                                 dax_radix_order(entry), entry);
 357                 radix_tree_preload_end();
 358                 if (err) {
 359                         spin_unlock_irq(&mapping->tree_lock);
 360                         /*
 361                          * Our insertion of a DAX entry failed, most likely
 362                          * because we were inserting a PMD entry and it
 363                          * collided with a PTE sized entry at a different
 364                          * index in the PMD range.  We haven't inserted
 365                          * anything into the radix tree and have no waiters to
 366                          * wake.
 367                          */
 368                         return ERR_PTR(err);
 369                 }
 370                 /* Good, we have inserted empty locked entry into the tree. */
 371                 mapping->nrexceptional++;
 372                 spin_unlock_irq(&mapping->tree_lock);
 373                 return entry;
 374         }
 375         /* Normal page in radix tree? */
 376         if (!radix_tree_exceptional_entry(entry)) {
 377                 struct page *page = entry;
 378
 379                 get_page(page);
 380                 spin_unlock_irq(&mapping->tree_lock);
 381                 lock_page(page);
 382                 /* Page got truncated? Retry... */
 383                 if (unlikely(page->mapping != mapping)) {
 384                         unlock_page(page);
 385                         put_page(page);
 386                         goto restart;
 387                 }
 388                 return page;
 389         }
 390         entry = lock_slot(mapping, slot);
 391  out_unlock:
 392         spin_unlock_irq(&mapping->tree_lock);
 393         return entry;
 394 }
 395
 396 /*
 397  * We do not necessarily hold the mapping->tree_lock when we call this
 398  * function so it is possible that 'entry' is no longer a valid item in the
 399  * radix tree.  This is okay because all we really need to do is to find the
 400  * correct waitqueue where tasks might be waiting for that old 'entry' and
 401  * wake them.
 402  */
 403 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
 404                 pgoff_t index, void *entry, bool wake_all)
 405 {
 406         struct exceptional_entry_key key;
 407         wait_queue_head_t *wq;
 408
 409         wq = dax_entry_waitqueue(mapping, index, entry, &key);
 410
 411         /*
 412          * Checking for locked entry and prepare_to_wait_exclusive() happens
 413          * under mapping->tree_lock, ditto for entry handling in our callers.
 414          * So at this point all tasks that could have seen our entry locked
 415          * must be in the waitqueue and the following check will see them.
 416          */
 417         if (waitqueue_active(wq))
 418                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
 419 }
 420
 421 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
 422                                           pgoff_t index, bool trunc)
 423 {
 424         int ret = 0;
 425         void *entry;
 426         struct radix_tree_root *page_tree = &mapping->page_tree;
 427
 428         spin_lock_irq(&mapping->tree_lock);
 429         entry = get_unlocked_mapping_entry(mapping, index, NULL);
 430         if (!entry || !radix_tree_exceptional_entry(entry))
 431                 goto out;
 432         if (!trunc &&
 433             (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
 434              radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)))
 435                 goto out;
 436         radix_tree_delete(page_tree, index);
 437         mapping->nrexceptional--;
 438         ret = 1;
 439 out:
 440         put_unlocked_mapping_entry(mapping, index, entry);
 441         spin_unlock_irq(&mapping->tree_lock);
 442         return ret;
 443 }
 444 /*
 445  * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
 446  * entry to get unlocked before deleting it.
 447  */
 448 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 449 {
 450         int ret = __dax_invalidate_mapping_entry(mapping, index, true);
 451
 452         /*
 453          * This gets called from truncate / punch_hole path. As such, the caller
 454          * must hold locks protecting against concurrent modifications of the
 455          * radix tree (usually fs-private i_mmap_sem for writing). Since the
 456          * caller has seen exceptional entry for this index, we better find it
 457          * at that index as well...
 458          */
 459         WARN_ON_ONCE(!ret);
 460         return ret;
 461 }
 462
 463 /*
 464  * Invalidate exceptional DAX entry if easily possible. This handles DAX
 465  * entries for invalidate_inode_pages() so we evict the entry only if we can
 466  * do so without blocking.
 467  */
 468 int dax_invalidate_mapping_entry(struct address_space *mapping, pgoff_t index)
 469 {
 470         int ret = 0;
 471         void *entry, **slot;
 472         struct radix_tree_root *page_tree = &mapping->page_tree;
 473
 474         spin_lock_irq(&mapping->tree_lock);
 475         entry = __radix_tree_lookup(page_tree, index, NULL, &slot);
 476         if (!entry || !radix_tree_exceptional_entry(entry) ||
 477             slot_locked(mapping, slot))
 478                 goto out;
 479         if (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
 480             radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 481                 goto out;
 482         radix_tree_delete(page_tree, index);
 483         mapping->nrexceptional--;
 484         ret = 1;
 485 out:
 486         spin_unlock_irq(&mapping->tree_lock);
 487         if (ret)
 488                 dax_wake_mapping_entry_waiter(mapping, index, entry, true);
 489         return ret;
 490 }
 491
 492 /*
 493  * Invalidate exceptional DAX entry if it is clean.
 494  */
 495 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
 496                                       pgoff_t index)
 497 {
 498         return __dax_invalidate_mapping_entry(mapping, index, false);
 499 }
 500
 501 /*
 502  * The user has performed a load from a hole in the file.  Allocating
 503  * a new page in the file would cause excessive storage usage for
 504  * workloads with sparse files.  We allocate a page cache page instead.
 505  * We'll kick it out of the page cache if it's ever written to,
 506  * otherwise it will simply fall out of the page cache under memory
 507  * pressure without ever having been dirtied.
 508  */
 509 static int dax_load_hole(struct address_space *mapping, void **entry,
 510                          struct vm_fault *vmf)
 511 {
 512         struct page *page;
 513         int ret;
 514
 515         /* Hole page already exists? Return it...  */
 516         if (!radix_tree_exceptional_entry(*entry)) {
 517                 page = *entry;
 518                 goto out;
 519         }
 520
 521         /* This will replace locked radix tree entry with a hole page */
 522         page = find_or_create_page(mapping, vmf->pgoff,
 523                                    vmf->gfp_mask | __GFP_ZERO);
 524         if (!page)
 525                 return VM_FAULT_OOM;
 526  out:
 527         vmf->page = page;
 528         ret = finish_fault(vmf);
 529         vmf->page = NULL;
 530         *entry = page;
 531         if (!ret) {
 532                 /* Grab reference for PTE that is now referencing the page */
 533                 get_page(page);
 534                 return VM_FAULT_NOPAGE;
 535         }
 536         return ret;
 537 }
 538
 539 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
 540                 sector_t sector, size_t size, struct page *to,
 541                 unsigned long vaddr)
 542 {
 543         void *vto, *kaddr;
 544         pgoff_t pgoff;
 545         pfn_t pfn;
 546         long rc;
 547         int id;
 548
 549         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 550         if (rc)
 551                 return rc;
 552
 553         id = dax_read_lock();
 554         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
 555         if (rc < 0) {
 556                 dax_read_unlock(id);
 557                 return rc;
 558         }
 559         vto = kmap_atomic(to);
 560         copy_user_page(vto, (void __force *)kaddr, vaddr, to);
 561         kunmap_atomic(vto);
 562         dax_read_unlock(id);
 563         return 0;
 564 }
 565
 566 /*
 567  * By this point grab_mapping_entry() has ensured that we have a locked entry
 568  * of the appropriate size so we don't have to worry about downgrading PMDs to
 569  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
 570  * already in the tree, we will skip the insertion and just dirty the PMD as
 571  * appropriate.
 572  */
 573 static void *dax_insert_mapping_entry(struct address_space *mapping,
 574                                       struct vm_fault *vmf,
 575                                       void *entry, sector_t sector,
 576                                       unsigned long flags)
 577 {
 578         struct radix_tree_root *page_tree = &mapping->page_tree;
 579         int error = 0;
 580         bool hole_fill = false;
 581         void *new_entry;
 582         pgoff_t index = vmf->pgoff;
 583
 584         if (vmf->flags & FAULT_FLAG_WRITE)
 585                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 586
 587         /* Replacing hole page with block mapping? */
 588         if (!radix_tree_exceptional_entry(entry)) {
 589                 hole_fill = true;
 590                 /*
 591                  * Unmap the page now before we remove it from page cache below.
 592                  * The page is locked so it cannot be faulted in again.
 593                  */
 594                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 595                                     PAGE_SIZE, 0);
 596                 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
 597                 if (error)
 598                         return ERR_PTR(error);
 599         } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) {
 600                 /* replacing huge zero page with PMD block mapping */
 601                 unmap_mapping_range(mapping,
 602                         (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
 603         }
 604
 605         spin_lock_irq(&mapping->tree_lock);
 606         new_entry = dax_radix_locked_entry(sector, flags);
 607
 608         if (hole_fill) {
 609                 __delete_from_page_cache(entry, NULL);
 610                 /* Drop pagecache reference */
 611                 put_page(entry);
 612                 error = __radix_tree_insert(page_tree, index,
 613                                 dax_radix_order(new_entry), new_entry);
 614                 if (error) {
 615                         new_entry = ERR_PTR(error);
 616                         goto unlock;
 617                 }
 618                 mapping->nrexceptional++;
 619         } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 620                 /*
 621                  * Only swap our new entry into the radix tree if the current
 622                  * entry is a zero page or an empty entry.  If a normal PTE or
 623                  * PMD entry is already in the tree, we leave it alone.  This
 624                  * means that if we are trying to insert a PTE and the
 625                  * existing entry is a PMD, we will just leave the PMD in the
 626                  * tree and dirty it if necessary.
 627                  */
 628                 struct radix_tree_node *node;
 629                 void **slot;
 630                 void *ret;
 631
 632                 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
 633                 WARN_ON_ONCE(ret != entry);
 634                 __radix_tree_replace(page_tree, node, slot,
 635                                      new_entry, NULL, NULL);
 636         }
 637         if (vmf->flags & FAULT_FLAG_WRITE)
 638                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 639  unlock:
 640         spin_unlock_irq(&mapping->tree_lock);
 641         if (hole_fill) {
 642                 radix_tree_preload_end();
 643                 /*
 644                  * We don't need hole page anymore, it has been replaced with
 645                  * locked radix tree entry now.
 646                  */
 647                 if (mapping->a_ops->freepage)
 648                         mapping->a_ops->freepage(entry);
 649                 unlock_page(entry);
 650                 put_page(entry);
 651         }
 652         return new_entry;
 653 }
 654
 655 static inline unsigned long
 656 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
 657 {
 658         unsigned long address;
 659
 660         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 661         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
 662         return address;
 663 }
 664
 665 /* Walk all mappings of a given index of a file and writeprotect them */
 666 static void dax_mapping_entry_mkclean(struct address_space *mapping,
 667                                       pgoff_t index, unsigned long pfn)
 668 {
 669         struct vm_area_struct *vma;
 670         pte_t pte, *ptep = NULL;
 671         pmd_t *pmdp = NULL;
 672         spinlock_t *ptl;
 673         bool changed;
 674
 675         i_mmap_lock_read(mapping);
 676         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
 677                 unsigned long address;
 678
 679                 cond_resched();
 680
 681                 if (!(vma->vm_flags & VM_SHARED))
 682                         continue;
 683
 684                 address = pgoff_address(index, vma);
 685                 changed = false;
 686                 if (follow_pte_pmd(vma->vm_mm, address, &ptep, &pmdp, &ptl))
 687                         continue;
 688
 689                 if (pmdp) {
 690 #ifdef CONFIG_FS_DAX_PMD
 691                         pmd_t pmd;
 692
 693                         if (pfn != pmd_pfn(*pmdp))
 694                                 goto unlock_pmd;
 695                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
 696                                 goto unlock_pmd;
 697
 698                         flush_cache_page(vma, address, pfn);
 699                         pmd = pmdp_huge_clear_flush(vma, address, pmdp);
 700                         pmd = pmd_wrprotect(pmd);
 701                         pmd = pmd_mkclean(pmd);
 702                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
 703                         changed = true;
 704 unlock_pmd:
 705                         spin_unlock(ptl);
 706 #endif
 707                 } else {
 708                         if (pfn != pte_pfn(*ptep))
 709                                 goto unlock_pte;
 710                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
 711                                 goto unlock_pte;
 712
 713                         flush_cache_page(vma, address, pfn);
 714                         pte = ptep_clear_flush(vma, address, ptep);
 715                         pte = pte_wrprotect(pte);
 716                         pte = pte_mkclean(pte);
 717                         set_pte_at(vma->vm_mm, address, ptep, pte);
 718                         changed = true;
 719 unlock_pte:
 720                         pte_unmap_unlock(ptep, ptl);
 721                 }
 722
 723                 if (changed)
 724                         mmu_notifier_invalidate_page(vma->vm_mm, address);
 725         }
 726         i_mmap_unlock_read(mapping);
 727 }
 728
 729 static int dax_writeback_one(struct block_device *bdev,
 730                 struct dax_device *dax_dev, struct address_space *mapping,
 731                 pgoff_t index, void *entry)
 732 {
 733         struct radix_tree_root *page_tree = &mapping->page_tree;
 734         void *entry2, **slot, *kaddr;
 735         long ret = 0, id;
 736         sector_t sector;
 737         pgoff_t pgoff;
 738         size_t size;
 739         pfn_t pfn;
 740
 741         /*
 742          * A page got tagged dirty in DAX mapping? Something is seriously
 743          * wrong.
 744          */
 745         if (WARN_ON(!radix_tree_exceptional_entry(entry)))
 746                 return -EIO;
 747
 748         spin_lock_irq(&mapping->tree_lock);
 749         entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
 750         /* Entry got punched out / reallocated? */
 751         if (!entry2 || !radix_tree_exceptional_entry(entry2))
 752                 goto put_unlocked;
 753         /*
 754          * Entry got reallocated elsewhere? No need to writeback. We have to
 755          * compare sectors as we must not bail out due to difference in lockbit
 756          * or entry type.
 757          */
 758         if (dax_radix_sector(entry2) != dax_radix_sector(entry))
 759                 goto put_unlocked;
 760         if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
 761                                 dax_is_zero_entry(entry))) {
 762                 ret = -EIO;
 763                 goto put_unlocked;
 764         }
 765
 766         /* Another fsync thread may have already written back this entry */
 767         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 768                 goto put_unlocked;
 769         /* Lock the entry to serialize with page faults */
 770         entry = lock_slot(mapping, slot);
 771         /*
 772          * We can clear the tag now but we have to be careful so that concurrent
 773          * dax_writeback_one() calls for the same index cannot finish before we
 774          * actually flush the caches. This is achieved as the calls will look
 775          * at the entry only under tree_lock and once they do that they will
 776          * see the entry locked and wait for it to unlock.
 777          */
 778         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 779         spin_unlock_irq(&mapping->tree_lock);
 780
 781         /*
 782          * Even if dax_writeback_mapping_range() was given a wbc->range_start
 783          * in the middle of a PMD, the 'index' we are given will be aligned to
 784          * the start index of the PMD, as will the sector we pull from
 785          * 'entry'.  This allows us to flush for PMD_SIZE and not have to
 786          * worry about partial PMD writebacks.
 787          */
 788         sector = dax_radix_sector(entry);
 789         size = PAGE_SIZE << dax_radix_order(entry);
 790
 791         id = dax_read_lock();
 792         ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 793         if (ret)
 794                 goto dax_unlock;
 795
 796         /*
 797          * dax_direct_access() may sleep, so cannot hold tree_lock over
 798          * its invocation.
 799          */
 800         ret = dax_direct_access(dax_dev, pgoff, size / PAGE_SIZE, &kaddr, &pfn);
 801         if (ret < 0)
 802                 goto dax_unlock;
 803
 804         if (WARN_ON_ONCE(ret < size / PAGE_SIZE)) {
 805                 ret = -EIO;
 806                 goto dax_unlock;
 807         }
 808
 809         dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(pfn));
 810         wb_cache_pmem(kaddr, size);
 811         /*
 812          * After we have flushed the cache, we can clear the dirty tag. There
 813          * cannot be new dirty data in the pfn after the flush has completed as
 814          * the pfn mappings are writeprotected and fault waits for mapping
 815          * entry lock.
 816          */
 817         spin_lock_irq(&mapping->tree_lock);
 818         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
 819         spin_unlock_irq(&mapping->tree_lock);
 820  dax_unlock:
 821         dax_read_unlock(id);
 822         put_locked_mapping_entry(mapping, index, entry);
 823         return ret;
 824
 825  put_unlocked:
 826         put_unlocked_mapping_entry(mapping, index, entry2);
 827         spin_unlock_irq(&mapping->tree_lock);
 828         return ret;
 829 }
 830
 831 /*
 832  * Flush the mapping to the persistent domain within the byte range of [start,
 833  * end]. This is required by data integrity operations to ensure file data is
 834  * on persistent storage prior to completion of the operation.
 835  */
 836 int dax_writeback_mapping_range(struct address_space *mapping,
 837                 struct block_device *bdev, struct writeback_control *wbc)
 838 {
 839         struct inode *inode = mapping->host;
 840         pgoff_t start_index, end_index;
 841         pgoff_t indices[PAGEVEC_SIZE];
 842         struct dax_device *dax_dev;
 843         struct pagevec pvec;
 844         bool done = false;
 845         int i, ret = 0;
 846
 847         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 848                 return -EIO;
 849
 850         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 851                 return 0;
 852
 853         dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
 854         if (!dax_dev)
 855                 return -EIO;
 856
 857         start_index = wbc->range_start >> PAGE_SHIFT;
 858         end_index = wbc->range_end >> PAGE_SHIFT;
 859
 860         tag_pages_for_writeback(mapping, start_index, end_index);
 861
 862         pagevec_init(&pvec, 0);
 863         while (!done) {
 864                 pvec.nr = find_get_entries_tag(mapping, start_index,
 865                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 866                                 pvec.pages, indices);
 867
 868                 if (pvec.nr == 0)
 869                         break;
 870
 871                 for (i = 0; i < pvec.nr; i++) {
 872                         if (indices[i] > end_index) {
 873                                 done = true;
 874                                 break;
 875                         }
 876
 877                         ret = dax_writeback_one(bdev, dax_dev, mapping,
 878                                         indices[i], pvec.pages[i]);
 879                         if (ret < 0) {
 880                                 put_dax(dax_dev);
 881                                 return ret;
 882                         }
 883                 }
 884         }
 885         put_dax(dax_dev);
 886         return 0;
 887 }
 888 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 889
 890 static int dax_insert_mapping(struct address_space *mapping,
 891                 struct block_device *bdev, struct dax_device *dax_dev,
 892                 sector_t sector, size_t size, void **entryp,
 893                 struct vm_area_struct *vma, struct vm_fault *vmf)
 894 {
 895         unsigned long vaddr = vmf->address;
 896         void *entry = *entryp;
 897         void *ret, *kaddr;
 898         pgoff_t pgoff;
 899         int id, rc;
 900         pfn_t pfn;
 901
 902         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 903         if (rc)
 904                 return rc;
 905
 906         id = dax_read_lock();
 907         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
 908         if (rc < 0) {
 909                 dax_read_unlock(id);
 910                 return rc;
 911         }
 912         dax_read_unlock(id);
 913
 914         ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, 0);
 915         if (IS_ERR(ret))
 916                 return PTR_ERR(ret);
 917         *entryp = ret;
 918
 919         return vm_insert_mixed(vma, vaddr, pfn);
 920 }
 921
 922 /**
 923  * dax_pfn_mkwrite - handle first write to DAX page
 924  * @vmf: The description of the fault
 925  */
 926 int dax_pfn_mkwrite(struct vm_fault *vmf)
 927 {
 928         struct file *file = vmf->vma->vm_file;
 929         struct address_space *mapping = file->f_mapping;
 930         void *entry, **slot;
 931         pgoff_t index = vmf->pgoff;
 932
 933         spin_lock_irq(&mapping->tree_lock);
 934         entry = get_unlocked_mapping_entry(mapping, index, &slot);
 935         if (!entry || !radix_tree_exceptional_entry(entry)) {
 936                 if (entry)
 937                         put_unlocked_mapping_entry(mapping, index, entry);
 938                 spin_unlock_irq(&mapping->tree_lock);
 939                 return VM_FAULT_NOPAGE;
 940         }
 941         radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
 942         entry = lock_slot(mapping, slot);
 943         spin_unlock_irq(&mapping->tree_lock);
 944         /*
 945          * If we race with somebody updating the PTE and finish_mkwrite_fault()
 946          * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry
 947          * the fault in either case.
 948          */
 949         finish_mkwrite_fault(vmf);
 950         put_locked_mapping_entry(mapping, index, entry);
 951         return VM_FAULT_NOPAGE;
 952 }
 953 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
 954
 955 static bool dax_range_is_aligned(struct block_device *bdev,
 956                                  unsigned int offset, unsigned int length)
 957 {
 958         unsigned short sector_size = bdev_logical_block_size(bdev);
 959
 960         if (!IS_ALIGNED(offset, sector_size))
 961                 return false;
 962         if (!IS_ALIGNED(length, sector_size))
 963                 return false;
 964
 965         return true;
 966 }
 967
 968 int __dax_zero_page_range(struct block_device *bdev,
 969                 struct dax_device *dax_dev, sector_t sector,
 970                 unsigned int offset, unsigned int size)
 971 {
 972         if (dax_range_is_aligned(bdev, offset, size)) {
 973                 sector_t start_sector = sector + (offset >> 9);
 974
 975                 return blkdev_issue_zeroout(bdev, start_sector,
 976                                 size >> 9, GFP_NOFS, 0);
 977         } else {
 978                 pgoff_t pgoff;
 979                 long rc, id;
 980                 void *kaddr;
 981                 pfn_t pfn;
 982
 983                 rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 984                 if (rc)
 985                         return rc;
 986
 987                 id = dax_read_lock();
 988                 rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr,
 989                                 &pfn);
 990                 if (rc < 0) {
 991                         dax_read_unlock(id);
 992                         return rc;
 993                 }
 994                 clear_pmem(kaddr + offset, size);
 995                 dax_read_unlock(id);
 996         }
 997         return 0;
 998 }
 999 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1000
1001 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
1002 {
1003         return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
1004 }
1005
1006 static loff_t
1007 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1008                 struct iomap *iomap)
1009 {
1010         struct block_device *bdev = iomap->bdev;
1011         struct dax_device *dax_dev = iomap->dax_dev;
1012         struct iov_iter *iter = data;
1013         loff_t end = pos + length, done = 0;
1014         ssize_t ret = 0;
1015         int id;
1016
1017         if (iov_iter_rw(iter) == READ) {
1018                 end = min(end, i_size_read(inode));
1019                 if (pos >= end)
1020                         return 0;
1021
1022                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1023                         return iov_iter_zero(min(length, end - pos), iter);
1024         }
1025
1026         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1027                 return -EIO;
1028
1029         /*
1030          * Write can allocate block for an area which has a hole page mapped
1031          * into page tables. We have to tear down these mappings so that data
1032          * written by write(2) is visible in mmap.
1033          */
1034         if ((iomap->flags & IOMAP_F_NEW) && inode->i_mapping->nrpages) {
1035                 invalidate_inode_pages2_range(inode->i_mapping,
1036                                               pos >> PAGE_SHIFT,
1037                                               (end - 1) >> PAGE_SHIFT);
1038         }
1039
1040         id = dax_read_lock();
1041         while (pos < end) {
1042                 unsigned offset = pos & (PAGE_SIZE - 1);
1043                 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1044                 const sector_t sector = dax_iomap_sector(iomap, pos);
1045                 ssize_t map_len;
1046                 pgoff_t pgoff;
1047                 void *kaddr;
1048                 pfn_t pfn;
1049
1050                 if (fatal_signal_pending(current)) {
1051                         ret = -EINTR;
1052                         break;
1053                 }
1054
1055                 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1056                 if (ret)
1057                         break;
1058
1059                 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1060                                 &kaddr, &pfn);
1061                 if (map_len < 0) {
1062                         ret = map_len;
1063                         break;
1064                 }
1065
1066                 map_len = PFN_PHYS(map_len);
1067                 kaddr += offset;
1068                 map_len -= offset;
1069                 if (map_len > end - pos)
1070                         map_len = end - pos;
1071
1072                 if (iov_iter_rw(iter) == WRITE)
1073                         map_len = copy_from_iter_pmem(kaddr, map_len, iter);
1074                 else
1075                         map_len = copy_to_iter(kaddr, map_len, iter);
1076                 if (map_len <= 0) {
1077                         ret = map_len ? map_len : -EFAULT;
1078                         break;
1079                 }
1080
1081                 pos += map_len;
1082                 length -= map_len;
1083                 done += map_len;
1084         }
1085         dax_read_unlock(id);
1086
1087         return done ? done : ret;
1088 }
1089
1090 /**
1091  * dax_iomap_rw - Perform I/O to a DAX file
1092  * @iocb:       The control block for this I/O
1093  * @iter:       The addresses to do I/O from or to
1094  * @ops:        iomap ops passed from the file system
1095  *
1096  * This function performs read and write operations to directly mapped
1097  * persistent memory.  The callers needs to take care of read/write exclusion
1098  * and evicting any page cache pages in the region under I/O.
1099  */
1100 ssize_t
1101 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1102                 const struct iomap_ops *ops)
1103 {
1104         struct address_space *mapping = iocb->ki_filp->f_mapping;
1105         struct inode *inode = mapping->host;
1106         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1107         unsigned flags = 0;
1108
1109         if (iov_iter_rw(iter) == WRITE) {
1110                 lockdep_assert_held_exclusive(&inode->i_rwsem);
1111                 flags |= IOMAP_WRITE;
1112         } else {
1113                 lockdep_assert_held(&inode->i_rwsem);
1114         }
1115
1116         while (iov_iter_count(iter)) {
1117                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1118                                 iter, dax_iomap_actor);
1119                 if (ret <= 0)
1120                         break;
1121                 pos += ret;
1122                 done += ret;
1123         }
1124
1125         iocb->ki_pos += done;
1126         return done ? done : ret;
1127 }
1128 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1129
1130 static int dax_fault_return(int error)
1131 {
1132         if (error == 0)
1133                 return VM_FAULT_NOPAGE;
1134         if (error == -ENOMEM)
1135                 return VM_FAULT_OOM;
1136         return VM_FAULT_SIGBUS;
1137 }
1138
1139 static int dax_iomap_pte_fault(struct vm_fault *vmf,
1140                                const struct iomap_ops *ops)
1141 {
1142         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1143         struct inode *inode = mapping->host;
1144         unsigned long vaddr = vmf->address;
1145         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1146         sector_t sector;
1147         struct iomap iomap = { 0 };
1148         unsigned flags = IOMAP_FAULT;
1149         int error, major = 0;
1150         int vmf_ret = 0;
1151         void *entry;
1152
1153         /*
1154          * Check whether offset isn't beyond end of file now. Caller is supposed
1155          * to hold locks serializing us with truncate / punch hole so this is
1156          * a reliable test.
1157          */
1158         if (pos >= i_size_read(inode))
1159                 return VM_FAULT_SIGBUS;
1160
1161         if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1162                 flags |= IOMAP_WRITE;
1163
1164         /*
1165          * Note that we don't bother to use iomap_apply here: DAX required
1166          * the file system block size to be equal the page size, which means
1167          * that we never have to deal with more than a single extent here.
1168          */
1169         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1170         if (error)
1171                 return dax_fault_return(error);
1172         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1173                 vmf_ret = dax_fault_return(-EIO);       /* fs corruption? */
1174                 goto finish_iomap;
1175         }
1176
1177         entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1178         if (IS_ERR(entry)) {
1179                 vmf_ret = dax_fault_return(PTR_ERR(entry));
1180                 goto finish_iomap;
1181         }
1182
1183         sector = dax_iomap_sector(&iomap, pos);
1184
1185         if (vmf->cow_page) {
1186                 switch (iomap.type) {
1187                 case IOMAP_HOLE:
1188                 case IOMAP_UNWRITTEN:
1189                         clear_user_highpage(vmf->cow_page, vaddr);
1190                         break;
1191                 case IOMAP_MAPPED:
1192                         error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1193                                         sector, PAGE_SIZE, vmf->cow_page, vaddr);
1194                         break;
1195                 default:
1196                         WARN_ON_ONCE(1);
1197                         error = -EIO;
1198                         break;
1199                 }
1200
1201                 if (error)
1202                         goto error_unlock_entry;
1203
1204                 __SetPageUptodate(vmf->cow_page);
1205                 vmf_ret = finish_fault(vmf);
1206                 if (!vmf_ret)
1207                         vmf_ret = VM_FAULT_DONE_COW;
1208                 goto unlock_entry;
1209         }
1210
1211         switch (iomap.type) {
1212         case IOMAP_MAPPED:
1213                 if (iomap.flags & IOMAP_F_NEW) {
1214                         count_vm_event(PGMAJFAULT);
1215                         mem_cgroup_count_vm_event(vmf->vma->vm_mm, PGMAJFAULT);
1216                         major = VM_FAULT_MAJOR;
1217                 }
1218                 error = dax_insert_mapping(mapping, iomap.bdev, iomap.dax_dev,
1219                                 sector, PAGE_SIZE, &entry, vmf->vma, vmf);
1220                 /* -EBUSY is fine, somebody else faulted on the same PTE */
1221                 if (error == -EBUSY)
1222                         error = 0;
1223                 break;
1224         case IOMAP_UNWRITTEN:
1225         case IOMAP_HOLE:
1226                 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
1227                         vmf_ret = dax_load_hole(mapping, &entry, vmf);
1228                         goto unlock_entry;
1229                 }
1230                 /*FALLTHRU*/
1231         default:
1232                 WARN_ON_ONCE(1);
1233                 error = -EIO;
1234                 break;
1235         }
1236
1237  error_unlock_entry:
1238         vmf_ret = dax_fault_return(error) | major;
1239  unlock_entry:
1240         put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1241  finish_iomap:
1242         if (ops->iomap_end) {
1243                 int copied = PAGE_SIZE;
1244
1245                 if (vmf_ret & VM_FAULT_ERROR)
1246                         copied = 0;
1247                 /*
1248                  * The fault is done by now and there's no way back (other
1249                  * thread may be already happily using PTE we have installed).
1250                  * Just ignore error from ->iomap_end since we cannot do much
1251                  * with it.
1252                  */
1253                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1254         }
1255         return vmf_ret;
1256 }
1257
1258 #ifdef CONFIG_FS_DAX_PMD
1259 /*
1260  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
1261  * more often than one might expect in the below functions.
1262  */
1263 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
1264
1265 static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap,
1266                 loff_t pos, void **entryp)
1267 {
1268         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1269         const sector_t sector = dax_iomap_sector(iomap, pos);
1270         struct dax_device *dax_dev = iomap->dax_dev;
1271         struct block_device *bdev = iomap->bdev;
1272         struct inode *inode = mapping->host;
1273         const size_t size = PMD_SIZE;
1274         void *ret = NULL, *kaddr;
1275         long length = 0;
1276         pgoff_t pgoff;
1277         pfn_t pfn;
1278         int id;
1279
1280         if (bdev_dax_pgoff(bdev, sector, size, &pgoff) != 0)
1281                 goto fallback;
1282
1283         id = dax_read_lock();
1284         length = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
1285         if (length < 0)
1286                 goto unlock_fallback;
1287         length = PFN_PHYS(length);
1288
1289         if (length < size)
1290                 goto unlock_fallback;
1291         if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR)
1292                 goto unlock_fallback;
1293         if (!pfn_t_devmap(pfn))
1294                 goto unlock_fallback;
1295         dax_read_unlock(id);
1296
1297         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, sector,
1298                         RADIX_DAX_PMD);
1299         if (IS_ERR(ret))
1300                 goto fallback;
1301         *entryp = ret;
1302
1303         trace_dax_pmd_insert_mapping(inode, vmf, length, pfn, ret);
1304         return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
1305                         pfn, vmf->flags & FAULT_FLAG_WRITE);
1306
1307 unlock_fallback:
1308         dax_read_unlock(id);
1309 fallback:
1310         trace_dax_pmd_insert_mapping_fallback(inode, vmf, length, pfn, ret);
1311         return VM_FAULT_FALLBACK;
1312 }
1313
1314 static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
1315                 void **entryp)
1316 {
1317         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1318         unsigned long pmd_addr = vmf->address & PMD_MASK;
1319         struct inode *inode = mapping->host;
1320         struct page *zero_page;
1321         void *ret = NULL;
1322         spinlock_t *ptl;
1323         pmd_t pmd_entry;
1324
1325         zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1326
1327         if (unlikely(!zero_page))
1328                 goto fallback;
1329
1330         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0,
1331                         RADIX_DAX_PMD | RADIX_DAX_HZP);
1332         if (IS_ERR(ret))
1333                 goto fallback;
1334         *entryp = ret;
1335
1336         ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1337         if (!pmd_none(*(vmf->pmd))) {
1338                 spin_unlock(ptl);
1339                 goto fallback;
1340         }
1341
1342         pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1343         pmd_entry = pmd_mkhuge(pmd_entry);
1344         set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1345         spin_unlock(ptl);
1346         trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
1347         return VM_FAULT_NOPAGE;
1348
1349 fallback:
1350         trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
1351         return VM_FAULT_FALLBACK;
1352 }
1353
1354 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1355                                const struct iomap_ops *ops)
1356 {
1357         struct vm_area_struct *vma = vmf->vma;
1358         struct address_space *mapping = vma->vm_file->f_mapping;
1359         unsigned long pmd_addr = vmf->address & PMD_MASK;
1360         bool write = vmf->flags & FAULT_FLAG_WRITE;
1361         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1362         struct inode *inode = mapping->host;
1363         int result = VM_FAULT_FALLBACK;
1364         struct iomap iomap = { 0 };
1365         pgoff_t max_pgoff, pgoff;
1366         void *entry;
1367         loff_t pos;
1368         int error;
1369
1370         /*
1371          * Check whether offset isn't beyond end of file now. Caller is
1372          * supposed to hold locks serializing us with truncate / punch hole so
1373          * this is a reliable test.
1374          */
1375         pgoff = linear_page_index(vma, pmd_addr);
1376         max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1377
1378         trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1379
1380         /* Fall back to PTEs if we're going to COW */
1381         if (write && !(vma->vm_flags & VM_SHARED))
1382                 goto fallback;
1383
1384         /* If the PMD would extend outside the VMA */
1385         if (pmd_addr < vma->vm_start)
1386                 goto fallback;
1387         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1388                 goto fallback;
1389
1390         if (pgoff > max_pgoff) {
1391                 result = VM_FAULT_SIGBUS;
1392                 goto out;
1393         }
1394
1395         /* If the PMD would extend beyond the file size */
1396         if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1397                 goto fallback;
1398
1399         /*
1400          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1401          * setting up a mapping, so really we're using iomap_begin() as a way
1402          * to look up our filesystem block.
1403          */
1404         pos = (loff_t)pgoff << PAGE_SHIFT;
1405         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1406         if (error)
1407                 goto fallback;
1408
1409         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1410                 goto finish_iomap;
1411
1412         /*
1413          * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1414          * PMD or a HZP entry.  If it can't (because a 4k page is already in
1415          * the tree, for instance), it will return -EEXIST and we just fall
1416          * back to 4k entries.
1417          */
1418         entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1419         if (IS_ERR(entry))
1420                 goto finish_iomap;
1421
1422         switch (iomap.type) {
1423         case IOMAP_MAPPED:
1424                 result = dax_pmd_insert_mapping(vmf, &iomap, pos, &entry);
1425                 break;
1426         case IOMAP_UNWRITTEN:
1427         case IOMAP_HOLE:
1428                 if (WARN_ON_ONCE(write))
1429                         goto unlock_entry;
1430                 result = dax_pmd_load_hole(vmf, &iomap, &entry);
1431                 break;
1432         default:
1433                 WARN_ON_ONCE(1);
1434                 break;
1435         }
1436
1437  unlock_entry:
1438         put_locked_mapping_entry(mapping, pgoff, entry);
1439  finish_iomap:
1440         if (ops->iomap_end) {
1441                 int copied = PMD_SIZE;
1442
1443                 if (result == VM_FAULT_FALLBACK)
1444                         copied = 0;
1445                 /*
1446                  * The fault is done by now and there's no way back (other
1447                  * thread may be already happily using PMD we have installed).
1448                  * Just ignore error from ->iomap_end since we cannot do much
1449                  * with it.
1450                  */
1451                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1452                                 &iomap);
1453         }
1454  fallback:
1455         if (result == VM_FAULT_FALLBACK) {
1456                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1457                 count_vm_event(THP_FAULT_FALLBACK);
1458         }
1459 out:
1460         trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1461         return result;
1462 }
1463 #else
1464 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1465                                const struct iomap_ops *ops)
1466 {
1467         return VM_FAULT_FALLBACK;
1468 }
1469 #endif /* CONFIG_FS_DAX_PMD */
1470
1471 /**
1472  * dax_iomap_fault - handle a page fault on a DAX file
1473  * @vmf: The description of the fault
1474  * @ops: iomap ops passed from the file system
1475  *
1476  * When a page fault occurs, filesystems may call this helper in
1477  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1478  * has done all the necessary locking for page fault to proceed
1479  * successfully.
1480  */
1481 int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1482                     const struct iomap_ops *ops)
1483 {
1484         switch (pe_size) {
1485         case PE_SIZE_PTE:
1486                 return dax_iomap_pte_fault(vmf, ops);
1487         case PE_SIZE_PMD:
1488                 return dax_iomap_pmd_fault(vmf, ops);
1489         default:
1490                 return VM_FAULT_FALLBACK;
1491         }
1492 }
1493 EXPORT_SYMBOL_GPL(dax_iomap_fault);