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