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 inode *inode = mapping->host;
 513         struct page *page;
 514         int ret;
 515
 516         /* Hole page already exists? Return it...  */
 517         if (!radix_tree_exceptional_entry(*entry)) {
 518                 page = *entry;
 519                 goto finish_fault;
 520         }
 521
 522         /* This will replace locked radix tree entry with a hole page */
 523         page = find_or_create_page(mapping, vmf->pgoff,
 524                                    vmf->gfp_mask | __GFP_ZERO);
 525         if (!page) {
 526                 ret = VM_FAULT_OOM;
 527                 goto out;
 528         }
 529
 530 finish_fault:
 531         vmf->page = page;
 532         ret = finish_fault(vmf);
 533         vmf->page = NULL;
 534         *entry = page;
 535         if (!ret) {
 536                 /* Grab reference for PTE that is now referencing the page */
 537                 get_page(page);
 538                 ret = VM_FAULT_NOPAGE;
 539         }
 540 out:
 541         trace_dax_load_hole(inode, vmf, ret);
 542         return ret;
 543 }
 544
 545 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
 546                 sector_t sector, size_t size, struct page *to,
 547                 unsigned long vaddr)
 548 {
 549         void *vto, *kaddr;
 550         pgoff_t pgoff;
 551         pfn_t pfn;
 552         long rc;
 553         int id;
 554
 555         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 556         if (rc)
 557                 return rc;
 558
 559         id = dax_read_lock();
 560         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
 561         if (rc < 0) {
 562                 dax_read_unlock(id);
 563                 return rc;
 564         }
 565         vto = kmap_atomic(to);
 566         copy_user_page(vto, (void __force *)kaddr, vaddr, to);
 567         kunmap_atomic(vto);
 568         dax_read_unlock(id);
 569         return 0;
 570 }
 571
 572 /*
 573  * By this point grab_mapping_entry() has ensured that we have a locked entry
 574  * of the appropriate size so we don't have to worry about downgrading PMDs to
 575  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
 576  * already in the tree, we will skip the insertion and just dirty the PMD as
 577  * appropriate.
 578  */
 579 static void *dax_insert_mapping_entry(struct address_space *mapping,
 580                                       struct vm_fault *vmf,
 581                                       void *entry, sector_t sector,
 582                                       unsigned long flags)
 583 {
 584         struct radix_tree_root *page_tree = &mapping->page_tree;
 585         int error = 0;
 586         bool hole_fill = false;
 587         void *new_entry;
 588         pgoff_t index = vmf->pgoff;
 589
 590         if (vmf->flags & FAULT_FLAG_WRITE)
 591                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 592
 593         /* Replacing hole page with block mapping? */
 594         if (!radix_tree_exceptional_entry(entry)) {
 595                 hole_fill = true;
 596                 /*
 597                  * Unmap the page now before we remove it from page cache below.
 598                  * The page is locked so it cannot be faulted in again.
 599                  */
 600                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 601                                     PAGE_SIZE, 0);
 602                 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
 603                 if (error)
 604                         return ERR_PTR(error);
 605         } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) {
 606                 /* replacing huge zero page with PMD block mapping */
 607                 unmap_mapping_range(mapping,
 608                         (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
 609         }
 610
 611         spin_lock_irq(&mapping->tree_lock);
 612         new_entry = dax_radix_locked_entry(sector, flags);
 613
 614         if (hole_fill) {
 615                 __delete_from_page_cache(entry, NULL);
 616                 /* Drop pagecache reference */
 617                 put_page(entry);
 618                 error = __radix_tree_insert(page_tree, index,
 619                                 dax_radix_order(new_entry), new_entry);
 620                 if (error) {
 621                         new_entry = ERR_PTR(error);
 622                         goto unlock;
 623                 }
 624                 mapping->nrexceptional++;
 625         } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 626                 /*
 627                  * Only swap our new entry into the radix tree if the current
 628                  * entry is a zero page or an empty entry.  If a normal PTE or
 629                  * PMD entry is already in the tree, we leave it alone.  This
 630                  * means that if we are trying to insert a PTE and the
 631                  * existing entry is a PMD, we will just leave the PMD in the
 632                  * tree and dirty it if necessary.
 633                  */
 634                 struct radix_tree_node *node;
 635                 void **slot;
 636                 void *ret;
 637
 638                 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
 639                 WARN_ON_ONCE(ret != entry);
 640                 __radix_tree_replace(page_tree, node, slot,
 641                                      new_entry, NULL, NULL);
 642         }
 643         if (vmf->flags & FAULT_FLAG_WRITE)
 644                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 645  unlock:
 646         spin_unlock_irq(&mapping->tree_lock);
 647         if (hole_fill) {
 648                 radix_tree_preload_end();
 649                 /*
 650                  * We don't need hole page anymore, it has been replaced with
 651                  * locked radix tree entry now.
 652                  */
 653                 if (mapping->a_ops->freepage)
 654                         mapping->a_ops->freepage(entry);
 655                 unlock_page(entry);
 656                 put_page(entry);
 657         }
 658         return new_entry;
 659 }
 660
 661 static inline unsigned long
 662 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
 663 {
 664         unsigned long address;
 665
 666         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 667         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
 668         return address;
 669 }
 670
 671 /* Walk all mappings of a given index of a file and writeprotect them */
 672 static void dax_mapping_entry_mkclean(struct address_space *mapping,
 673                                       pgoff_t index, unsigned long pfn)
 674 {
 675         struct vm_area_struct *vma;
 676         pte_t pte, *ptep = NULL;
 677         pmd_t *pmdp = NULL;
 678         spinlock_t *ptl;
 679         bool changed;
 680
 681         i_mmap_lock_read(mapping);
 682         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
 683                 unsigned long address;
 684
 685                 cond_resched();
 686
 687                 if (!(vma->vm_flags & VM_SHARED))
 688                         continue;
 689
 690                 address = pgoff_address(index, vma);
 691                 changed = false;
 692                 if (follow_pte_pmd(vma->vm_mm, address, &ptep, &pmdp, &ptl))
 693                         continue;
 694
 695                 if (pmdp) {
 696 #ifdef CONFIG_FS_DAX_PMD
 697                         pmd_t pmd;
 698
 699                         if (pfn != pmd_pfn(*pmdp))
 700                                 goto unlock_pmd;
 701                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
 702                                 goto unlock_pmd;
 703
 704                         flush_cache_page(vma, address, pfn);
 705                         pmd = pmdp_huge_clear_flush(vma, address, pmdp);
 706                         pmd = pmd_wrprotect(pmd);
 707                         pmd = pmd_mkclean(pmd);
 708                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
 709                         changed = true;
 710 unlock_pmd:
 711                         spin_unlock(ptl);
 712 #endif
 713                 } else {
 714                         if (pfn != pte_pfn(*ptep))
 715                                 goto unlock_pte;
 716                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
 717                                 goto unlock_pte;
 718
 719                         flush_cache_page(vma, address, pfn);
 720                         pte = ptep_clear_flush(vma, address, ptep);
 721                         pte = pte_wrprotect(pte);
 722                         pte = pte_mkclean(pte);
 723                         set_pte_at(vma->vm_mm, address, ptep, pte);
 724                         changed = true;
 725 unlock_pte:
 726                         pte_unmap_unlock(ptep, ptl);
 727                 }
 728
 729                 if (changed)
 730                         mmu_notifier_invalidate_page(vma->vm_mm, address);
 731         }
 732         i_mmap_unlock_read(mapping);
 733 }
 734
 735 static int dax_writeback_one(struct block_device *bdev,
 736                 struct dax_device *dax_dev, struct address_space *mapping,
 737                 pgoff_t index, void *entry)
 738 {
 739         struct radix_tree_root *page_tree = &mapping->page_tree;
 740         void *entry2, **slot, *kaddr;
 741         long ret = 0, id;
 742         sector_t sector;
 743         pgoff_t pgoff;
 744         size_t size;
 745         pfn_t pfn;
 746
 747         /*
 748          * A page got tagged dirty in DAX mapping? Something is seriously
 749          * wrong.
 750          */
 751         if (WARN_ON(!radix_tree_exceptional_entry(entry)))
 752                 return -EIO;
 753
 754         spin_lock_irq(&mapping->tree_lock);
 755         entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
 756         /* Entry got punched out / reallocated? */
 757         if (!entry2 || !radix_tree_exceptional_entry(entry2))
 758                 goto put_unlocked;
 759         /*
 760          * Entry got reallocated elsewhere? No need to writeback. We have to
 761          * compare sectors as we must not bail out due to difference in lockbit
 762          * or entry type.
 763          */
 764         if (dax_radix_sector(entry2) != dax_radix_sector(entry))
 765                 goto put_unlocked;
 766         if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
 767                                 dax_is_zero_entry(entry))) {
 768                 ret = -EIO;
 769                 goto put_unlocked;
 770         }
 771
 772         /* Another fsync thread may have already written back this entry */
 773         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 774                 goto put_unlocked;
 775         /* Lock the entry to serialize with page faults */
 776         entry = lock_slot(mapping, slot);
 777         /*
 778          * We can clear the tag now but we have to be careful so that concurrent
 779          * dax_writeback_one() calls for the same index cannot finish before we
 780          * actually flush the caches. This is achieved as the calls will look
 781          * at the entry only under tree_lock and once they do that they will
 782          * see the entry locked and wait for it to unlock.
 783          */
 784         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 785         spin_unlock_irq(&mapping->tree_lock);
 786
 787         /*
 788          * Even if dax_writeback_mapping_range() was given a wbc->range_start
 789          * in the middle of a PMD, the 'index' we are given will be aligned to
 790          * the start index of the PMD, as will the sector we pull from
 791          * 'entry'.  This allows us to flush for PMD_SIZE and not have to
 792          * worry about partial PMD writebacks.
 793          */
 794         sector = dax_radix_sector(entry);
 795         size = PAGE_SIZE << dax_radix_order(entry);
 796
 797         id = dax_read_lock();
 798         ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 799         if (ret)
 800                 goto dax_unlock;
 801
 802         /*
 803          * dax_direct_access() may sleep, so cannot hold tree_lock over
 804          * its invocation.
 805          */
 806         ret = dax_direct_access(dax_dev, pgoff, size / PAGE_SIZE, &kaddr, &pfn);
 807         if (ret < 0)
 808                 goto dax_unlock;
 809
 810         if (WARN_ON_ONCE(ret < size / PAGE_SIZE)) {
 811                 ret = -EIO;
 812                 goto dax_unlock;
 813         }
 814
 815         dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(pfn));
 816         wb_cache_pmem(kaddr, size);
 817         /*
 818          * After we have flushed the cache, we can clear the dirty tag. There
 819          * cannot be new dirty data in the pfn after the flush has completed as
 820          * the pfn mappings are writeprotected and fault waits for mapping
 821          * entry lock.
 822          */
 823         spin_lock_irq(&mapping->tree_lock);
 824         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
 825         spin_unlock_irq(&mapping->tree_lock);
 826         trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT);
 827  dax_unlock:
 828         dax_read_unlock(id);
 829         put_locked_mapping_entry(mapping, index, entry);
 830         return ret;
 831
 832  put_unlocked:
 833         put_unlocked_mapping_entry(mapping, index, entry2);
 834         spin_unlock_irq(&mapping->tree_lock);
 835         return ret;
 836 }
 837
 838 /*
 839  * Flush the mapping to the persistent domain within the byte range of [start,
 840  * end]. This is required by data integrity operations to ensure file data is
 841  * on persistent storage prior to completion of the operation.
 842  */
 843 int dax_writeback_mapping_range(struct address_space *mapping,
 844                 struct block_device *bdev, struct writeback_control *wbc)
 845 {
 846         struct inode *inode = mapping->host;
 847         pgoff_t start_index, end_index;
 848         pgoff_t indices[PAGEVEC_SIZE];
 849         struct dax_device *dax_dev;
 850         struct pagevec pvec;
 851         bool done = false;
 852         int i, ret = 0;
 853
 854         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 855                 return -EIO;
 856
 857         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 858                 return 0;
 859
 860         dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
 861         if (!dax_dev)
 862                 return -EIO;
 863
 864         start_index = wbc->range_start >> PAGE_SHIFT;
 865         end_index = wbc->range_end >> PAGE_SHIFT;
 866
 867         trace_dax_writeback_range(inode, start_index, end_index);
 868
 869         tag_pages_for_writeback(mapping, start_index, end_index);
 870
 871         pagevec_init(&pvec, 0);
 872         while (!done) {
 873                 pvec.nr = find_get_entries_tag(mapping, start_index,
 874                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 875                                 pvec.pages, indices);
 876
 877                 if (pvec.nr == 0)
 878                         break;
 879
 880                 for (i = 0; i < pvec.nr; i++) {
 881                         if (indices[i] > end_index) {
 882                                 done = true;
 883                                 break;
 884                         }
 885
 886                         ret = dax_writeback_one(bdev, dax_dev, mapping,
 887                                         indices[i], pvec.pages[i]);
 888                         if (ret < 0)
 889                                 goto out;
 890                 }
 891         }
 892 out:
 893         put_dax(dax_dev);
 894         trace_dax_writeback_range_done(inode, start_index, end_index);
 895         return (ret < 0 ? ret : 0);
 896 }
 897 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 898
 899 static int dax_insert_mapping(struct address_space *mapping,
 900                 struct block_device *bdev, struct dax_device *dax_dev,
 901                 sector_t sector, size_t size, void **entryp,
 902                 struct vm_area_struct *vma, struct vm_fault *vmf)
 903 {
 904         unsigned long vaddr = vmf->address;
 905         void *entry = *entryp;
 906         void *ret, *kaddr;
 907         pgoff_t pgoff;
 908         int id, rc;
 909         pfn_t pfn;
 910
 911         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 912         if (rc)
 913                 return rc;
 914
 915         id = dax_read_lock();
 916         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
 917         if (rc < 0) {
 918                 dax_read_unlock(id);
 919                 return rc;
 920         }
 921         dax_read_unlock(id);
 922
 923         ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, 0);
 924         if (IS_ERR(ret))
 925                 return PTR_ERR(ret);
 926         *entryp = ret;
 927
 928         return vm_insert_mixed(vma, vaddr, pfn);
 929 }
 930
 931 /**
 932  * dax_pfn_mkwrite - handle first write to DAX page
 933  * @vmf: The description of the fault
 934  */
 935 int dax_pfn_mkwrite(struct vm_fault *vmf)
 936 {
 937         struct file *file = vmf->vma->vm_file;
 938         struct address_space *mapping = file->f_mapping;
 939         struct inode *inode = mapping->host;
 940         void *entry, **slot;
 941         pgoff_t index = vmf->pgoff;
 942
 943         spin_lock_irq(&mapping->tree_lock);
 944         entry = get_unlocked_mapping_entry(mapping, index, &slot);
 945         if (!entry || !radix_tree_exceptional_entry(entry)) {
 946                 if (entry)
 947                         put_unlocked_mapping_entry(mapping, index, entry);
 948                 spin_unlock_irq(&mapping->tree_lock);
 949                 trace_dax_pfn_mkwrite_no_entry(inode, vmf, VM_FAULT_NOPAGE);
 950                 return VM_FAULT_NOPAGE;
 951         }
 952         radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
 953         entry = lock_slot(mapping, slot);
 954         spin_unlock_irq(&mapping->tree_lock);
 955         /*
 956          * If we race with somebody updating the PTE and finish_mkwrite_fault()
 957          * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry
 958          * the fault in either case.
 959          */
 960         finish_mkwrite_fault(vmf);
 961         put_locked_mapping_entry(mapping, index, entry);
 962         trace_dax_pfn_mkwrite(inode, vmf, VM_FAULT_NOPAGE);
 963         return VM_FAULT_NOPAGE;
 964 }
 965 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
 966
 967 static bool dax_range_is_aligned(struct block_device *bdev,
 968                                  unsigned int offset, unsigned int length)
 969 {
 970         unsigned short sector_size = bdev_logical_block_size(bdev);
 971
 972         if (!IS_ALIGNED(offset, sector_size))
 973                 return false;
 974         if (!IS_ALIGNED(length, sector_size))
 975                 return false;
 976
 977         return true;
 978 }
 979
 980 int __dax_zero_page_range(struct block_device *bdev,
 981                 struct dax_device *dax_dev, sector_t sector,
 982                 unsigned int offset, unsigned int size)
 983 {
 984         if (dax_range_is_aligned(bdev, offset, size)) {
 985                 sector_t start_sector = sector + (offset >> 9);
 986
 987                 return blkdev_issue_zeroout(bdev, start_sector,
 988                                 size >> 9, GFP_NOFS, 0);
 989         } else {
 990                 pgoff_t pgoff;
 991                 long rc, id;
 992                 void *kaddr;
 993                 pfn_t pfn;
 994
 995                 rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 996                 if (rc)
 997                         return rc;
 998
 999                 id = dax_read_lock();
1000                 rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr,
1001                                 &pfn);
1002                 if (rc < 0) {
1003                         dax_read_unlock(id);
1004                         return rc;
1005                 }
1006                 clear_pmem(kaddr + offset, size);
1007                 dax_read_unlock(id);
1008         }
1009         return 0;
1010 }
1011 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1012
1013 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
1014 {
1015         return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
1016 }
1017
1018 static loff_t
1019 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1020                 struct iomap *iomap)
1021 {
1022         struct block_device *bdev = iomap->bdev;
1023         struct dax_device *dax_dev = iomap->dax_dev;
1024         struct iov_iter *iter = data;
1025         loff_t end = pos + length, done = 0;
1026         ssize_t ret = 0;
1027         int id;
1028
1029         if (iov_iter_rw(iter) == READ) {
1030                 end = min(end, i_size_read(inode));
1031                 if (pos >= end)
1032                         return 0;
1033
1034                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1035                         return iov_iter_zero(min(length, end - pos), iter);
1036         }
1037
1038         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1039                 return -EIO;
1040
1041         /*
1042          * Write can allocate block for an area which has a hole page mapped
1043          * into page tables. We have to tear down these mappings so that data
1044          * written by write(2) is visible in mmap.
1045          */
1046         if ((iomap->flags & IOMAP_F_NEW) && inode->i_mapping->nrpages) {
1047                 invalidate_inode_pages2_range(inode->i_mapping,
1048                                               pos >> PAGE_SHIFT,
1049                                               (end - 1) >> PAGE_SHIFT);
1050         }
1051
1052         id = dax_read_lock();
1053         while (pos < end) {
1054                 unsigned offset = pos & (PAGE_SIZE - 1);
1055                 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1056                 const sector_t sector = dax_iomap_sector(iomap, pos);
1057                 ssize_t map_len;
1058                 pgoff_t pgoff;
1059                 void *kaddr;
1060                 pfn_t pfn;
1061
1062                 if (fatal_signal_pending(current)) {
1063                         ret = -EINTR;
1064                         break;
1065                 }
1066
1067                 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1068                 if (ret)
1069                         break;
1070
1071                 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1072                                 &kaddr, &pfn);
1073                 if (map_len < 0) {
1074                         ret = map_len;
1075                         break;
1076                 }
1077
1078                 map_len = PFN_PHYS(map_len);
1079                 kaddr += offset;
1080                 map_len -= offset;
1081                 if (map_len > end - pos)
1082                         map_len = end - pos;
1083
1084                 if (iov_iter_rw(iter) == WRITE)
1085                         map_len = copy_from_iter_pmem(kaddr, map_len, iter);
1086                 else
1087                         map_len = copy_to_iter(kaddr, map_len, iter);
1088                 if (map_len <= 0) {
1089                         ret = map_len ? map_len : -EFAULT;
1090                         break;
1091                 }
1092
1093                 pos += map_len;
1094                 length -= map_len;
1095                 done += map_len;
1096         }
1097         dax_read_unlock(id);
1098
1099         return done ? done : ret;
1100 }
1101
1102 /**
1103  * dax_iomap_rw - Perform I/O to a DAX file
1104  * @iocb:       The control block for this I/O
1105  * @iter:       The addresses to do I/O from or to
1106  * @ops:        iomap ops passed from the file system
1107  *
1108  * This function performs read and write operations to directly mapped
1109  * persistent memory.  The callers needs to take care of read/write exclusion
1110  * and evicting any page cache pages in the region under I/O.
1111  */
1112 ssize_t
1113 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1114                 const struct iomap_ops *ops)
1115 {
1116         struct address_space *mapping = iocb->ki_filp->f_mapping;
1117         struct inode *inode = mapping->host;
1118         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1119         unsigned flags = 0;
1120
1121         if (iov_iter_rw(iter) == WRITE) {
1122                 lockdep_assert_held_exclusive(&inode->i_rwsem);
1123                 flags |= IOMAP_WRITE;
1124         } else {
1125                 lockdep_assert_held(&inode->i_rwsem);
1126         }
1127
1128         while (iov_iter_count(iter)) {
1129                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1130                                 iter, dax_iomap_actor);
1131                 if (ret <= 0)
1132                         break;
1133                 pos += ret;
1134                 done += ret;
1135         }
1136
1137         iocb->ki_pos += done;
1138         return done ? done : ret;
1139 }
1140 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1141
1142 static int dax_fault_return(int error)
1143 {
1144         if (error == 0)
1145                 return VM_FAULT_NOPAGE;
1146         if (error == -ENOMEM)
1147                 return VM_FAULT_OOM;
1148         return VM_FAULT_SIGBUS;
1149 }
1150
1151 static int dax_iomap_pte_fault(struct vm_fault *vmf,
1152                                const struct iomap_ops *ops)
1153 {
1154         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1155         struct inode *inode = mapping->host;
1156         unsigned long vaddr = vmf->address;
1157         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1158         sector_t sector;
1159         struct iomap iomap = { 0 };
1160         unsigned flags = IOMAP_FAULT;
1161         int error, major = 0;
1162         int vmf_ret = 0;
1163         void *entry;
1164
1165         trace_dax_pte_fault(inode, vmf, vmf_ret);
1166         /*
1167          * Check whether offset isn't beyond end of file now. Caller is supposed
1168          * to hold locks serializing us with truncate / punch hole so this is
1169          * a reliable test.
1170          */
1171         if (pos >= i_size_read(inode)) {
1172                 vmf_ret = VM_FAULT_SIGBUS;
1173                 goto out;
1174         }
1175
1176         if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1177                 flags |= IOMAP_WRITE;
1178
1179         /*
1180          * Note that we don't bother to use iomap_apply here: DAX required
1181          * the file system block size to be equal the page size, which means
1182          * that we never have to deal with more than a single extent here.
1183          */
1184         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1185         if (error) {
1186                 vmf_ret = dax_fault_return(error);
1187                 goto out;
1188         }
1189         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1190                 vmf_ret = dax_fault_return(-EIO);       /* fs corruption? */
1191                 goto finish_iomap;
1192         }
1193
1194         entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1195         if (IS_ERR(entry)) {
1196                 vmf_ret = dax_fault_return(PTR_ERR(entry));
1197                 goto finish_iomap;
1198         }
1199
1200         sector = dax_iomap_sector(&iomap, pos);
1201
1202         if (vmf->cow_page) {
1203                 switch (iomap.type) {
1204                 case IOMAP_HOLE:
1205                 case IOMAP_UNWRITTEN:
1206                         clear_user_highpage(vmf->cow_page, vaddr);
1207                         break;
1208                 case IOMAP_MAPPED:
1209                         error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1210                                         sector, PAGE_SIZE, vmf->cow_page, vaddr);
1211                         break;
1212                 default:
1213                         WARN_ON_ONCE(1);
1214                         error = -EIO;
1215                         break;
1216                 }
1217
1218                 if (error)
1219                         goto error_unlock_entry;
1220
1221                 __SetPageUptodate(vmf->cow_page);
1222                 vmf_ret = finish_fault(vmf);
1223                 if (!vmf_ret)
1224                         vmf_ret = VM_FAULT_DONE_COW;
1225                 goto unlock_entry;
1226         }
1227
1228         switch (iomap.type) {
1229         case IOMAP_MAPPED:
1230                 if (iomap.flags & IOMAP_F_NEW) {
1231                         count_vm_event(PGMAJFAULT);
1232                         mem_cgroup_count_vm_event(vmf->vma->vm_mm, PGMAJFAULT);
1233                         major = VM_FAULT_MAJOR;
1234                 }
1235                 error = dax_insert_mapping(mapping, iomap.bdev, iomap.dax_dev,
1236                                 sector, PAGE_SIZE, &entry, vmf->vma, vmf);
1237                 /* -EBUSY is fine, somebody else faulted on the same PTE */
1238                 if (error == -EBUSY)
1239                         error = 0;
1240                 break;
1241         case IOMAP_UNWRITTEN:
1242         case IOMAP_HOLE:
1243                 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
1244                         vmf_ret = dax_load_hole(mapping, &entry, vmf);
1245                         goto unlock_entry;
1246                 }
1247                 /*FALLTHRU*/
1248         default:
1249                 WARN_ON_ONCE(1);
1250                 error = -EIO;
1251                 break;
1252         }
1253
1254  error_unlock_entry:
1255         vmf_ret = dax_fault_return(error) | major;
1256  unlock_entry:
1257         put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1258  finish_iomap:
1259         if (ops->iomap_end) {
1260                 int copied = PAGE_SIZE;
1261
1262                 if (vmf_ret & VM_FAULT_ERROR)
1263                         copied = 0;
1264                 /*
1265                  * The fault is done by now and there's no way back (other
1266                  * thread may be already happily using PTE we have installed).
1267                  * Just ignore error from ->iomap_end since we cannot do much
1268                  * with it.
1269                  */
1270                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1271         }
1272 out:
1273         trace_dax_pte_fault_done(inode, vmf, vmf_ret);
1274         return vmf_ret;
1275 }
1276
1277 #ifdef CONFIG_FS_DAX_PMD
1278 /*
1279  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
1280  * more often than one might expect in the below functions.
1281  */
1282 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
1283
1284 static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap,
1285                 loff_t pos, void **entryp)
1286 {
1287         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1288         const sector_t sector = dax_iomap_sector(iomap, pos);
1289         struct dax_device *dax_dev = iomap->dax_dev;
1290         struct block_device *bdev = iomap->bdev;
1291         struct inode *inode = mapping->host;
1292         const size_t size = PMD_SIZE;
1293         void *ret = NULL, *kaddr;
1294         long length = 0;
1295         pgoff_t pgoff;
1296         pfn_t pfn;
1297         int id;
1298
1299         if (bdev_dax_pgoff(bdev, sector, size, &pgoff) != 0)
1300                 goto fallback;
1301
1302         id = dax_read_lock();
1303         length = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
1304         if (length < 0)
1305                 goto unlock_fallback;
1306         length = PFN_PHYS(length);
1307
1308         if (length < size)
1309                 goto unlock_fallback;
1310         if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR)
1311                 goto unlock_fallback;
1312         if (!pfn_t_devmap(pfn))
1313                 goto unlock_fallback;
1314         dax_read_unlock(id);
1315
1316         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, sector,
1317                         RADIX_DAX_PMD);
1318         if (IS_ERR(ret))
1319                 goto fallback;
1320         *entryp = ret;
1321
1322         trace_dax_pmd_insert_mapping(inode, vmf, length, pfn, ret);
1323         return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
1324                         pfn, vmf->flags & FAULT_FLAG_WRITE);
1325
1326 unlock_fallback:
1327         dax_read_unlock(id);
1328 fallback:
1329         trace_dax_pmd_insert_mapping_fallback(inode, vmf, length, pfn, ret);
1330         return VM_FAULT_FALLBACK;
1331 }
1332
1333 static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
1334                 void **entryp)
1335 {
1336         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1337         unsigned long pmd_addr = vmf->address & PMD_MASK;
1338         struct inode *inode = mapping->host;
1339         struct page *zero_page;
1340         void *ret = NULL;
1341         spinlock_t *ptl;
1342         pmd_t pmd_entry;
1343
1344         zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1345
1346         if (unlikely(!zero_page))
1347                 goto fallback;
1348
1349         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0,
1350                         RADIX_DAX_PMD | RADIX_DAX_HZP);
1351         if (IS_ERR(ret))
1352                 goto fallback;
1353         *entryp = ret;
1354
1355         ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1356         if (!pmd_none(*(vmf->pmd))) {
1357                 spin_unlock(ptl);
1358                 goto fallback;
1359         }
1360
1361         pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1362         pmd_entry = pmd_mkhuge(pmd_entry);
1363         set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1364         spin_unlock(ptl);
1365         trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
1366         return VM_FAULT_NOPAGE;
1367
1368 fallback:
1369         trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
1370         return VM_FAULT_FALLBACK;
1371 }
1372
1373 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1374                                const struct iomap_ops *ops)
1375 {
1376         struct vm_area_struct *vma = vmf->vma;
1377         struct address_space *mapping = vma->vm_file->f_mapping;
1378         unsigned long pmd_addr = vmf->address & PMD_MASK;
1379         bool write = vmf->flags & FAULT_FLAG_WRITE;
1380         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1381         struct inode *inode = mapping->host;
1382         int result = VM_FAULT_FALLBACK;
1383         struct iomap iomap = { 0 };
1384         pgoff_t max_pgoff, pgoff;
1385         void *entry;
1386         loff_t pos;
1387         int error;
1388
1389         /*
1390          * Check whether offset isn't beyond end of file now. Caller is
1391          * supposed to hold locks serializing us with truncate / punch hole so
1392          * this is a reliable test.
1393          */
1394         pgoff = linear_page_index(vma, pmd_addr);
1395         max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1396
1397         trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1398
1399         /* Fall back to PTEs if we're going to COW */
1400         if (write && !(vma->vm_flags & VM_SHARED))
1401                 goto fallback;
1402
1403         /* If the PMD would extend outside the VMA */
1404         if (pmd_addr < vma->vm_start)
1405                 goto fallback;
1406         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1407                 goto fallback;
1408
1409         if (pgoff > max_pgoff) {
1410                 result = VM_FAULT_SIGBUS;
1411                 goto out;
1412         }
1413
1414         /* If the PMD would extend beyond the file size */
1415         if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1416                 goto fallback;
1417
1418         /*
1419          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1420          * setting up a mapping, so really we're using iomap_begin() as a way
1421          * to look up our filesystem block.
1422          */
1423         pos = (loff_t)pgoff << PAGE_SHIFT;
1424         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1425         if (error)
1426                 goto fallback;
1427
1428         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1429                 goto finish_iomap;
1430
1431         /*
1432          * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1433          * PMD or a HZP entry.  If it can't (because a 4k page is already in
1434          * the tree, for instance), it will return -EEXIST and we just fall
1435          * back to 4k entries.
1436          */
1437         entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1438         if (IS_ERR(entry))
1439                 goto finish_iomap;
1440
1441         switch (iomap.type) {
1442         case IOMAP_MAPPED:
1443                 result = dax_pmd_insert_mapping(vmf, &iomap, pos, &entry);
1444                 break;
1445         case IOMAP_UNWRITTEN:
1446         case IOMAP_HOLE:
1447                 if (WARN_ON_ONCE(write))
1448                         goto unlock_entry;
1449                 result = dax_pmd_load_hole(vmf, &iomap, &entry);
1450                 break;
1451         default:
1452                 WARN_ON_ONCE(1);
1453                 break;
1454         }
1455
1456  unlock_entry:
1457         put_locked_mapping_entry(mapping, pgoff, entry);
1458  finish_iomap:
1459         if (ops->iomap_end) {
1460                 int copied = PMD_SIZE;
1461
1462                 if (result == VM_FAULT_FALLBACK)
1463                         copied = 0;
1464                 /*
1465                  * The fault is done by now and there's no way back (other
1466                  * thread may be already happily using PMD we have installed).
1467                  * Just ignore error from ->iomap_end since we cannot do much
1468                  * with it.
1469                  */
1470                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1471                                 &iomap);
1472         }
1473  fallback:
1474         if (result == VM_FAULT_FALLBACK) {
1475                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1476                 count_vm_event(THP_FAULT_FALLBACK);
1477         }
1478 out:
1479         trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1480         return result;
1481 }
1482 #else
1483 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1484                                const struct iomap_ops *ops)
1485 {
1486         return VM_FAULT_FALLBACK;
1487 }
1488 #endif /* CONFIG_FS_DAX_PMD */
1489
1490 /**
1491  * dax_iomap_fault - handle a page fault on a DAX file
1492  * @vmf: The description of the fault
1493  * @ops: iomap ops passed from the file system
1494  *
1495  * When a page fault occurs, filesystems may call this helper in
1496  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1497  * has done all the necessary locking for page fault to proceed
1498  * successfully.
1499  */
1500 int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1501                     const struct iomap_ops *ops)
1502 {
1503         switch (pe_size) {
1504         case PE_SIZE_PTE:
1505                 return dax_iomap_pte_fault(vmf, ops);
1506         case PE_SIZE_PMD:
1507                 return dax_iomap_pmd_fault(vmf, ops);
1508         default:
1509                 return VM_FAULT_FALLBACK;
1510         }
1511 }
1512 EXPORT_SYMBOL_GPL(dax_iomap_fault);