root/fs/dax.c

/* [<][>][^][v][top][bottom][index][help] */

DEFINITIONS

This source file includes following definitions.
  1. pe_order
  2. init_dax_wait_table
  3. dax_to_pfn
  4. dax_make_entry
  5. dax_is_locked
  6. dax_entry_order
  7. dax_is_pmd_entry
  8. dax_is_pte_entry
  9. dax_is_zero_entry
  10. dax_is_empty_entry
  11. dax_is_conflict
  12. dax_entry_waitqueue
  13. wake_exceptional_entry_func
  14. dax_wake_entry
  15. get_unlocked_entry
  16. wait_entry_unlocked
  17. put_unlocked_entry
  18. dax_unlock_entry
  19. dax_lock_entry
  20. dax_entry_size
  21. dax_end_pfn
  22. dax_associate_entry
  23. dax_disassociate_entry
  24. dax_busy_page
  25. dax_lock_page
  26. dax_unlock_page
  27. grab_mapping_entry
  28. dax_layout_busy_page
  29. __dax_invalidate_entry
  30. dax_delete_mapping_entry
  31. dax_invalidate_mapping_entry_sync
  32. copy_user_dax
  33. dax_insert_entry
  34. pgoff_address
  35. dax_entry_mkclean
  36. dax_writeback_one
  37. dax_writeback_mapping_range
  38. dax_iomap_sector
  39. dax_iomap_pfn
  40. dax_load_hole
  41. dax_range_is_aligned
  42. __dax_zero_page_range
  43. dax_iomap_actor
  44. dax_iomap_rw
  45. dax_fault_return
  46. dax_fault_is_synchronous
  47. dax_iomap_pte_fault
  48. dax_pmd_load_hole
  49. dax_iomap_pmd_fault
  50. dax_iomap_pmd_fault
  51. dax_iomap_fault
  52. dax_insert_pfn_mkwrite
  53. dax_finish_sync_fault

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * fs/dax.c - Direct Access filesystem code
   4  * Copyright (c) 2013-2014 Intel Corporation
   5  * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
   6  * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
   7  */
   8 
   9 #include <linux/atomic.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/buffer_head.h>
  12 #include <linux/dax.h>
  13 #include <linux/fs.h>
  14 #include <linux/genhd.h>
  15 #include <linux/highmem.h>
  16 #include <linux/memcontrol.h>
  17 #include <linux/mm.h>
  18 #include <linux/mutex.h>
  19 #include <linux/pagevec.h>
  20 #include <linux/sched.h>
  21 #include <linux/sched/signal.h>
  22 #include <linux/uio.h>
  23 #include <linux/vmstat.h>
  24 #include <linux/pfn_t.h>
  25 #include <linux/sizes.h>
  26 #include <linux/mmu_notifier.h>
  27 #include <linux/iomap.h>
  28 #include <asm/pgalloc.h>
  29 
  30 #define CREATE_TRACE_POINTS
  31 #include <trace/events/fs_dax.h>
  32 
  33 static inline unsigned int pe_order(enum page_entry_size pe_size)
  34 {
  35         if (pe_size == PE_SIZE_PTE)
  36                 return PAGE_SHIFT - PAGE_SHIFT;
  37         if (pe_size == PE_SIZE_PMD)
  38                 return PMD_SHIFT - PAGE_SHIFT;
  39         if (pe_size == PE_SIZE_PUD)
  40                 return PUD_SHIFT - PAGE_SHIFT;
  41         return ~0;
  42 }
  43 
  44 /* We choose 4096 entries - same as per-zone page wait tables */
  45 #define DAX_WAIT_TABLE_BITS 12
  46 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
  47 
  48 /* The 'colour' (ie low bits) within a PMD of a page offset.  */
  49 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
  50 #define PG_PMD_NR       (PMD_SIZE >> PAGE_SHIFT)
  51 
  52 /* The order of a PMD entry */
  53 #define PMD_ORDER       (PMD_SHIFT - PAGE_SHIFT)
  54 
  55 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
  56 
  57 static int __init init_dax_wait_table(void)
  58 {
  59         int i;
  60 
  61         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
  62                 init_waitqueue_head(wait_table + i);
  63         return 0;
  64 }
  65 fs_initcall(init_dax_wait_table);
  66 
  67 /*
  68  * DAX pagecache entries use XArray value entries so they can't be mistaken
  69  * for pages.  We use one bit for locking, one bit for the entry size (PMD)
  70  * and two more to tell us if the entry is a zero page or an empty entry that
  71  * is just used for locking.  In total four special bits.
  72  *
  73  * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
  74  * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
  75  * block allocation.
  76  */
  77 #define DAX_SHIFT       (4)
  78 #define DAX_LOCKED      (1UL << 0)
  79 #define DAX_PMD         (1UL << 1)
  80 #define DAX_ZERO_PAGE   (1UL << 2)
  81 #define DAX_EMPTY       (1UL << 3)
  82 
  83 static unsigned long dax_to_pfn(void *entry)
  84 {
  85         return xa_to_value(entry) >> DAX_SHIFT;
  86 }
  87 
  88 static void *dax_make_entry(pfn_t pfn, unsigned long flags)
  89 {
  90         return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
  91 }
  92 
  93 static bool dax_is_locked(void *entry)
  94 {
  95         return xa_to_value(entry) & DAX_LOCKED;
  96 }
  97 
  98 static unsigned int dax_entry_order(void *entry)
  99 {
 100         if (xa_to_value(entry) & DAX_PMD)
 101                 return PMD_ORDER;
 102         return 0;
 103 }
 104 
 105 static unsigned long dax_is_pmd_entry(void *entry)
 106 {
 107         return xa_to_value(entry) & DAX_PMD;
 108 }
 109 
 110 static bool dax_is_pte_entry(void *entry)
 111 {
 112         return !(xa_to_value(entry) & DAX_PMD);
 113 }
 114 
 115 static int dax_is_zero_entry(void *entry)
 116 {
 117         return xa_to_value(entry) & DAX_ZERO_PAGE;
 118 }
 119 
 120 static int dax_is_empty_entry(void *entry)
 121 {
 122         return xa_to_value(entry) & DAX_EMPTY;
 123 }
 124 
 125 /*
 126  * true if the entry that was found is of a smaller order than the entry
 127  * we were looking for
 128  */
 129 static bool dax_is_conflict(void *entry)
 130 {
 131         return entry == XA_RETRY_ENTRY;
 132 }
 133 
 134 /*
 135  * DAX page cache entry locking
 136  */
 137 struct exceptional_entry_key {
 138         struct xarray *xa;
 139         pgoff_t entry_start;
 140 };
 141 
 142 struct wait_exceptional_entry_queue {
 143         wait_queue_entry_t wait;
 144         struct exceptional_entry_key key;
 145 };
 146 
 147 static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
 148                 void *entry, struct exceptional_entry_key *key)
 149 {
 150         unsigned long hash;
 151         unsigned long index = xas->xa_index;
 152 
 153         /*
 154          * If 'entry' is a PMD, align the 'index' that we use for the wait
 155          * queue to the start of that PMD.  This ensures that all offsets in
 156          * the range covered by the PMD map to the same bit lock.
 157          */
 158         if (dax_is_pmd_entry(entry))
 159                 index &= ~PG_PMD_COLOUR;
 160         key->xa = xas->xa;
 161         key->entry_start = index;
 162 
 163         hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
 164         return wait_table + hash;
 165 }
 166 
 167 static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
 168                 unsigned int mode, int sync, void *keyp)
 169 {
 170         struct exceptional_entry_key *key = keyp;
 171         struct wait_exceptional_entry_queue *ewait =
 172                 container_of(wait, struct wait_exceptional_entry_queue, wait);
 173 
 174         if (key->xa != ewait->key.xa ||
 175             key->entry_start != ewait->key.entry_start)
 176                 return 0;
 177         return autoremove_wake_function(wait, mode, sync, NULL);
 178 }
 179 
 180 /*
 181  * @entry may no longer be the entry at the index in the mapping.
 182  * The important information it's conveying is whether the entry at
 183  * this index used to be a PMD entry.
 184  */
 185 static void dax_wake_entry(struct xa_state *xas, void *entry, bool wake_all)
 186 {
 187         struct exceptional_entry_key key;
 188         wait_queue_head_t *wq;
 189 
 190         wq = dax_entry_waitqueue(xas, entry, &key);
 191 
 192         /*
 193          * Checking for locked entry and prepare_to_wait_exclusive() happens
 194          * under the i_pages lock, ditto for entry handling in our callers.
 195          * So at this point all tasks that could have seen our entry locked
 196          * must be in the waitqueue and the following check will see them.
 197          */
 198         if (waitqueue_active(wq))
 199                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
 200 }
 201 
 202 /*
 203  * Look up entry in page cache, wait for it to become unlocked if it
 204  * is a DAX entry and return it.  The caller must subsequently call
 205  * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
 206  * if it did.  The entry returned may have a larger order than @order.
 207  * If @order is larger than the order of the entry found in i_pages, this
 208  * function returns a dax_is_conflict entry.
 209  *
 210  * Must be called with the i_pages lock held.
 211  */
 212 static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
 213 {
 214         void *entry;
 215         struct wait_exceptional_entry_queue ewait;
 216         wait_queue_head_t *wq;
 217 
 218         init_wait(&ewait.wait);
 219         ewait.wait.func = wake_exceptional_entry_func;
 220 
 221         for (;;) {
 222                 entry = xas_find_conflict(xas);
 223                 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
 224                         return entry;
 225                 if (dax_entry_order(entry) < order)
 226                         return XA_RETRY_ENTRY;
 227                 if (!dax_is_locked(entry))
 228                         return entry;
 229 
 230                 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
 231                 prepare_to_wait_exclusive(wq, &ewait.wait,
 232                                           TASK_UNINTERRUPTIBLE);
 233                 xas_unlock_irq(xas);
 234                 xas_reset(xas);
 235                 schedule();
 236                 finish_wait(wq, &ewait.wait);
 237                 xas_lock_irq(xas);
 238         }
 239 }
 240 
 241 /*
 242  * The only thing keeping the address space around is the i_pages lock
 243  * (it's cycled in clear_inode() after removing the entries from i_pages)
 244  * After we call xas_unlock_irq(), we cannot touch xas->xa.
 245  */
 246 static void wait_entry_unlocked(struct xa_state *xas, void *entry)
 247 {
 248         struct wait_exceptional_entry_queue ewait;
 249         wait_queue_head_t *wq;
 250 
 251         init_wait(&ewait.wait);
 252         ewait.wait.func = wake_exceptional_entry_func;
 253 
 254         wq = dax_entry_waitqueue(xas, entry, &ewait.key);
 255         /*
 256          * Unlike get_unlocked_entry() there is no guarantee that this
 257          * path ever successfully retrieves an unlocked entry before an
 258          * inode dies. Perform a non-exclusive wait in case this path
 259          * never successfully performs its own wake up.
 260          */
 261         prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
 262         xas_unlock_irq(xas);
 263         schedule();
 264         finish_wait(wq, &ewait.wait);
 265 }
 266 
 267 static void put_unlocked_entry(struct xa_state *xas, void *entry)
 268 {
 269         /* If we were the only waiter woken, wake the next one */
 270         if (entry && !dax_is_conflict(entry))
 271                 dax_wake_entry(xas, entry, false);
 272 }
 273 
 274 /*
 275  * We used the xa_state to get the entry, but then we locked the entry and
 276  * dropped the xa_lock, so we know the xa_state is stale and must be reset
 277  * before use.
 278  */
 279 static void dax_unlock_entry(struct xa_state *xas, void *entry)
 280 {
 281         void *old;
 282 
 283         BUG_ON(dax_is_locked(entry));
 284         xas_reset(xas);
 285         xas_lock_irq(xas);
 286         old = xas_store(xas, entry);
 287         xas_unlock_irq(xas);
 288         BUG_ON(!dax_is_locked(old));
 289         dax_wake_entry(xas, entry, false);
 290 }
 291 
 292 /*
 293  * Return: The entry stored at this location before it was locked.
 294  */
 295 static void *dax_lock_entry(struct xa_state *xas, void *entry)
 296 {
 297         unsigned long v = xa_to_value(entry);
 298         return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
 299 }
 300 
 301 static unsigned long dax_entry_size(void *entry)
 302 {
 303         if (dax_is_zero_entry(entry))
 304                 return 0;
 305         else if (dax_is_empty_entry(entry))
 306                 return 0;
 307         else if (dax_is_pmd_entry(entry))
 308                 return PMD_SIZE;
 309         else
 310                 return PAGE_SIZE;
 311 }
 312 
 313 static unsigned long dax_end_pfn(void *entry)
 314 {
 315         return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
 316 }
 317 
 318 /*
 319  * Iterate through all mapped pfns represented by an entry, i.e. skip
 320  * 'empty' and 'zero' entries.
 321  */
 322 #define for_each_mapped_pfn(entry, pfn) \
 323         for (pfn = dax_to_pfn(entry); \
 324                         pfn < dax_end_pfn(entry); pfn++)
 325 
 326 /*
 327  * TODO: for reflink+dax we need a way to associate a single page with
 328  * multiple address_space instances at different linear_page_index()
 329  * offsets.
 330  */
 331 static void dax_associate_entry(void *entry, struct address_space *mapping,
 332                 struct vm_area_struct *vma, unsigned long address)
 333 {
 334         unsigned long size = dax_entry_size(entry), pfn, index;
 335         int i = 0;
 336 
 337         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
 338                 return;
 339 
 340         index = linear_page_index(vma, address & ~(size - 1));
 341         for_each_mapped_pfn(entry, pfn) {
 342                 struct page *page = pfn_to_page(pfn);
 343 
 344                 WARN_ON_ONCE(page->mapping);
 345                 page->mapping = mapping;
 346                 page->index = index + i++;
 347         }
 348 }
 349 
 350 static void dax_disassociate_entry(void *entry, struct address_space *mapping,
 351                 bool trunc)
 352 {
 353         unsigned long pfn;
 354 
 355         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
 356                 return;
 357 
 358         for_each_mapped_pfn(entry, pfn) {
 359                 struct page *page = pfn_to_page(pfn);
 360 
 361                 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
 362                 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
 363                 page->mapping = NULL;
 364                 page->index = 0;
 365         }
 366 }
 367 
 368 static struct page *dax_busy_page(void *entry)
 369 {
 370         unsigned long pfn;
 371 
 372         for_each_mapped_pfn(entry, pfn) {
 373                 struct page *page = pfn_to_page(pfn);
 374 
 375                 if (page_ref_count(page) > 1)
 376                         return page;
 377         }
 378         return NULL;
 379 }
 380 
 381 /*
 382  * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page
 383  * @page: The page whose entry we want to lock
 384  *
 385  * Context: Process context.
 386  * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
 387  * not be locked.
 388  */
 389 dax_entry_t dax_lock_page(struct page *page)
 390 {
 391         XA_STATE(xas, NULL, 0);
 392         void *entry;
 393 
 394         /* Ensure page->mapping isn't freed while we look at it */
 395         rcu_read_lock();
 396         for (;;) {
 397                 struct address_space *mapping = READ_ONCE(page->mapping);
 398 
 399                 entry = NULL;
 400                 if (!mapping || !dax_mapping(mapping))
 401                         break;
 402 
 403                 /*
 404                  * In the device-dax case there's no need to lock, a
 405                  * struct dev_pagemap pin is sufficient to keep the
 406                  * inode alive, and we assume we have dev_pagemap pin
 407                  * otherwise we would not have a valid pfn_to_page()
 408                  * translation.
 409                  */
 410                 entry = (void *)~0UL;
 411                 if (S_ISCHR(mapping->host->i_mode))
 412                         break;
 413 
 414                 xas.xa = &mapping->i_pages;
 415                 xas_lock_irq(&xas);
 416                 if (mapping != page->mapping) {
 417                         xas_unlock_irq(&xas);
 418                         continue;
 419                 }
 420                 xas_set(&xas, page->index);
 421                 entry = xas_load(&xas);
 422                 if (dax_is_locked(entry)) {
 423                         rcu_read_unlock();
 424                         wait_entry_unlocked(&xas, entry);
 425                         rcu_read_lock();
 426                         continue;
 427                 }
 428                 dax_lock_entry(&xas, entry);
 429                 xas_unlock_irq(&xas);
 430                 break;
 431         }
 432         rcu_read_unlock();
 433         return (dax_entry_t)entry;
 434 }
 435 
 436 void dax_unlock_page(struct page *page, dax_entry_t cookie)
 437 {
 438         struct address_space *mapping = page->mapping;
 439         XA_STATE(xas, &mapping->i_pages, page->index);
 440 
 441         if (S_ISCHR(mapping->host->i_mode))
 442                 return;
 443 
 444         dax_unlock_entry(&xas, (void *)cookie);
 445 }
 446 
 447 /*
 448  * Find page cache entry at given index. If it is a DAX entry, return it
 449  * with the entry locked. If the page cache doesn't contain an entry at
 450  * that index, add a locked empty entry.
 451  *
 452  * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
 453  * either return that locked entry or will return VM_FAULT_FALLBACK.
 454  * This will happen if there are any PTE entries within the PMD range
 455  * that we are requesting.
 456  *
 457  * We always favor PTE entries over PMD entries. There isn't a flow where we
 458  * evict PTE entries in order to 'upgrade' them to a PMD entry.  A PMD
 459  * insertion will fail if it finds any PTE entries already in the tree, and a
 460  * PTE insertion will cause an existing PMD entry to be unmapped and
 461  * downgraded to PTE entries.  This happens for both PMD zero pages as
 462  * well as PMD empty entries.
 463  *
 464  * The exception to this downgrade path is for PMD entries that have
 465  * real storage backing them.  We will leave these real PMD entries in
 466  * the tree, and PTE writes will simply dirty the entire PMD entry.
 467  *
 468  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
 469  * persistent memory the benefit is doubtful. We can add that later if we can
 470  * show it helps.
 471  *
 472  * On error, this function does not return an ERR_PTR.  Instead it returns
 473  * a VM_FAULT code, encoded as an xarray internal entry.  The ERR_PTR values
 474  * overlap with xarray value entries.
 475  */
 476 static void *grab_mapping_entry(struct xa_state *xas,
 477                 struct address_space *mapping, unsigned int order)
 478 {
 479         unsigned long index = xas->xa_index;
 480         bool pmd_downgrade = false; /* splitting PMD entry into PTE entries? */
 481         void *entry;
 482 
 483 retry:
 484         xas_lock_irq(xas);
 485         entry = get_unlocked_entry(xas, order);
 486 
 487         if (entry) {
 488                 if (dax_is_conflict(entry))
 489                         goto fallback;
 490                 if (!xa_is_value(entry)) {
 491                         xas_set_err(xas, EIO);
 492                         goto out_unlock;
 493                 }
 494 
 495                 if (order == 0) {
 496                         if (dax_is_pmd_entry(entry) &&
 497                             (dax_is_zero_entry(entry) ||
 498                              dax_is_empty_entry(entry))) {
 499                                 pmd_downgrade = true;
 500                         }
 501                 }
 502         }
 503 
 504         if (pmd_downgrade) {
 505                 /*
 506                  * Make sure 'entry' remains valid while we drop
 507                  * the i_pages lock.
 508                  */
 509                 dax_lock_entry(xas, entry);
 510 
 511                 /*
 512                  * Besides huge zero pages the only other thing that gets
 513                  * downgraded are empty entries which don't need to be
 514                  * unmapped.
 515                  */
 516                 if (dax_is_zero_entry(entry)) {
 517                         xas_unlock_irq(xas);
 518                         unmap_mapping_pages(mapping,
 519                                         xas->xa_index & ~PG_PMD_COLOUR,
 520                                         PG_PMD_NR, false);
 521                         xas_reset(xas);
 522                         xas_lock_irq(xas);
 523                 }
 524 
 525                 dax_disassociate_entry(entry, mapping, false);
 526                 xas_store(xas, NULL);   /* undo the PMD join */
 527                 dax_wake_entry(xas, entry, true);
 528                 mapping->nrexceptional--;
 529                 entry = NULL;
 530                 xas_set(xas, index);
 531         }
 532 
 533         if (entry) {
 534                 dax_lock_entry(xas, entry);
 535         } else {
 536                 unsigned long flags = DAX_EMPTY;
 537 
 538                 if (order > 0)
 539                         flags |= DAX_PMD;
 540                 entry = dax_make_entry(pfn_to_pfn_t(0), flags);
 541                 dax_lock_entry(xas, entry);
 542                 if (xas_error(xas))
 543                         goto out_unlock;
 544                 mapping->nrexceptional++;
 545         }
 546 
 547 out_unlock:
 548         xas_unlock_irq(xas);
 549         if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
 550                 goto retry;
 551         if (xas->xa_node == XA_ERROR(-ENOMEM))
 552                 return xa_mk_internal(VM_FAULT_OOM);
 553         if (xas_error(xas))
 554                 return xa_mk_internal(VM_FAULT_SIGBUS);
 555         return entry;
 556 fallback:
 557         xas_unlock_irq(xas);
 558         return xa_mk_internal(VM_FAULT_FALLBACK);
 559 }
 560 
 561 /**
 562  * dax_layout_busy_page - find first pinned page in @mapping
 563  * @mapping: address space to scan for a page with ref count > 1
 564  *
 565  * DAX requires ZONE_DEVICE mapped pages. These pages are never
 566  * 'onlined' to the page allocator so they are considered idle when
 567  * page->count == 1. A filesystem uses this interface to determine if
 568  * any page in the mapping is busy, i.e. for DMA, or other
 569  * get_user_pages() usages.
 570  *
 571  * It is expected that the filesystem is holding locks to block the
 572  * establishment of new mappings in this address_space. I.e. it expects
 573  * to be able to run unmap_mapping_range() and subsequently not race
 574  * mapping_mapped() becoming true.
 575  */
 576 struct page *dax_layout_busy_page(struct address_space *mapping)
 577 {
 578         XA_STATE(xas, &mapping->i_pages, 0);
 579         void *entry;
 580         unsigned int scanned = 0;
 581         struct page *page = NULL;
 582 
 583         /*
 584          * In the 'limited' case get_user_pages() for dax is disabled.
 585          */
 586         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
 587                 return NULL;
 588 
 589         if (!dax_mapping(mapping) || !mapping_mapped(mapping))
 590                 return NULL;
 591 
 592         /*
 593          * If we race get_user_pages_fast() here either we'll see the
 594          * elevated page count in the iteration and wait, or
 595          * get_user_pages_fast() will see that the page it took a reference
 596          * against is no longer mapped in the page tables and bail to the
 597          * get_user_pages() slow path.  The slow path is protected by
 598          * pte_lock() and pmd_lock(). New references are not taken without
 599          * holding those locks, and unmap_mapping_range() will not zero the
 600          * pte or pmd without holding the respective lock, so we are
 601          * guaranteed to either see new references or prevent new
 602          * references from being established.
 603          */
 604         unmap_mapping_range(mapping, 0, 0, 0);
 605 
 606         xas_lock_irq(&xas);
 607         xas_for_each(&xas, entry, ULONG_MAX) {
 608                 if (WARN_ON_ONCE(!xa_is_value(entry)))
 609                         continue;
 610                 if (unlikely(dax_is_locked(entry)))
 611                         entry = get_unlocked_entry(&xas, 0);
 612                 if (entry)
 613                         page = dax_busy_page(entry);
 614                 put_unlocked_entry(&xas, entry);
 615                 if (page)
 616                         break;
 617                 if (++scanned % XA_CHECK_SCHED)
 618                         continue;
 619 
 620                 xas_pause(&xas);
 621                 xas_unlock_irq(&xas);
 622                 cond_resched();
 623                 xas_lock_irq(&xas);
 624         }
 625         xas_unlock_irq(&xas);
 626         return page;
 627 }
 628 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
 629 
 630 static int __dax_invalidate_entry(struct address_space *mapping,
 631                                           pgoff_t index, bool trunc)
 632 {
 633         XA_STATE(xas, &mapping->i_pages, index);
 634         int ret = 0;
 635         void *entry;
 636 
 637         xas_lock_irq(&xas);
 638         entry = get_unlocked_entry(&xas, 0);
 639         if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
 640                 goto out;
 641         if (!trunc &&
 642             (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
 643              xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
 644                 goto out;
 645         dax_disassociate_entry(entry, mapping, trunc);
 646         xas_store(&xas, NULL);
 647         mapping->nrexceptional--;
 648         ret = 1;
 649 out:
 650         put_unlocked_entry(&xas, entry);
 651         xas_unlock_irq(&xas);
 652         return ret;
 653 }
 654 
 655 /*
 656  * Delete DAX entry at @index from @mapping.  Wait for it
 657  * to be unlocked before deleting it.
 658  */
 659 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 660 {
 661         int ret = __dax_invalidate_entry(mapping, index, true);
 662 
 663         /*
 664          * This gets called from truncate / punch_hole path. As such, the caller
 665          * must hold locks protecting against concurrent modifications of the
 666          * page cache (usually fs-private i_mmap_sem for writing). Since the
 667          * caller has seen a DAX entry for this index, we better find it
 668          * at that index as well...
 669          */
 670         WARN_ON_ONCE(!ret);
 671         return ret;
 672 }
 673 
 674 /*
 675  * Invalidate DAX entry if it is clean.
 676  */
 677 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
 678                                       pgoff_t index)
 679 {
 680         return __dax_invalidate_entry(mapping, index, false);
 681 }
 682 
 683 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
 684                 sector_t sector, size_t size, struct page *to,
 685                 unsigned long vaddr)
 686 {
 687         void *vto, *kaddr;
 688         pgoff_t pgoff;
 689         long rc;
 690         int id;
 691 
 692         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 693         if (rc)
 694                 return rc;
 695 
 696         id = dax_read_lock();
 697         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, NULL);
 698         if (rc < 0) {
 699                 dax_read_unlock(id);
 700                 return rc;
 701         }
 702         vto = kmap_atomic(to);
 703         copy_user_page(vto, (void __force *)kaddr, vaddr, to);
 704         kunmap_atomic(vto);
 705         dax_read_unlock(id);
 706         return 0;
 707 }
 708 
 709 /*
 710  * By this point grab_mapping_entry() has ensured that we have a locked entry
 711  * of the appropriate size so we don't have to worry about downgrading PMDs to
 712  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
 713  * already in the tree, we will skip the insertion and just dirty the PMD as
 714  * appropriate.
 715  */
 716 static void *dax_insert_entry(struct xa_state *xas,
 717                 struct address_space *mapping, struct vm_fault *vmf,
 718                 void *entry, pfn_t pfn, unsigned long flags, bool dirty)
 719 {
 720         void *new_entry = dax_make_entry(pfn, flags);
 721 
 722         if (dirty)
 723                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 724 
 725         if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) {
 726                 unsigned long index = xas->xa_index;
 727                 /* we are replacing a zero page with block mapping */
 728                 if (dax_is_pmd_entry(entry))
 729                         unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
 730                                         PG_PMD_NR, false);
 731                 else /* pte entry */
 732                         unmap_mapping_pages(mapping, index, 1, false);
 733         }
 734 
 735         xas_reset(xas);
 736         xas_lock_irq(xas);
 737         if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 738                 void *old;
 739 
 740                 dax_disassociate_entry(entry, mapping, false);
 741                 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address);
 742                 /*
 743                  * Only swap our new entry into the page cache if the current
 744                  * entry is a zero page or an empty entry.  If a normal PTE or
 745                  * PMD entry is already in the cache, we leave it alone.  This
 746                  * means that if we are trying to insert a PTE and the
 747                  * existing entry is a PMD, we will just leave the PMD in the
 748                  * tree and dirty it if necessary.
 749                  */
 750                 old = dax_lock_entry(xas, new_entry);
 751                 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
 752                                         DAX_LOCKED));
 753                 entry = new_entry;
 754         } else {
 755                 xas_load(xas);  /* Walk the xa_state */
 756         }
 757 
 758         if (dirty)
 759                 xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
 760 
 761         xas_unlock_irq(xas);
 762         return entry;
 763 }
 764 
 765 static inline
 766 unsigned long pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
 767 {
 768         unsigned long address;
 769 
 770         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 771         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
 772         return address;
 773 }
 774 
 775 /* Walk all mappings of a given index of a file and writeprotect them */
 776 static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
 777                 unsigned long pfn)
 778 {
 779         struct vm_area_struct *vma;
 780         pte_t pte, *ptep = NULL;
 781         pmd_t *pmdp = NULL;
 782         spinlock_t *ptl;
 783 
 784         i_mmap_lock_read(mapping);
 785         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
 786                 struct mmu_notifier_range range;
 787                 unsigned long address;
 788 
 789                 cond_resched();
 790 
 791                 if (!(vma->vm_flags & VM_SHARED))
 792                         continue;
 793 
 794                 address = pgoff_address(index, vma);
 795 
 796                 /*
 797                  * Note because we provide range to follow_pte_pmd it will
 798                  * call mmu_notifier_invalidate_range_start() on our behalf
 799                  * before taking any lock.
 800                  */
 801                 if (follow_pte_pmd(vma->vm_mm, address, &range,
 802                                    &ptep, &pmdp, &ptl))
 803                         continue;
 804 
 805                 /*
 806                  * No need to call mmu_notifier_invalidate_range() as we are
 807                  * downgrading page table protection not changing it to point
 808                  * to a new page.
 809                  *
 810                  * See Documentation/vm/mmu_notifier.rst
 811                  */
 812                 if (pmdp) {
 813 #ifdef CONFIG_FS_DAX_PMD
 814                         pmd_t pmd;
 815 
 816                         if (pfn != pmd_pfn(*pmdp))
 817                                 goto unlock_pmd;
 818                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
 819                                 goto unlock_pmd;
 820 
 821                         flush_cache_page(vma, address, pfn);
 822                         pmd = pmdp_invalidate(vma, address, pmdp);
 823                         pmd = pmd_wrprotect(pmd);
 824                         pmd = pmd_mkclean(pmd);
 825                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
 826 unlock_pmd:
 827 #endif
 828                         spin_unlock(ptl);
 829                 } else {
 830                         if (pfn != pte_pfn(*ptep))
 831                                 goto unlock_pte;
 832                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
 833                                 goto unlock_pte;
 834 
 835                         flush_cache_page(vma, address, pfn);
 836                         pte = ptep_clear_flush(vma, address, ptep);
 837                         pte = pte_wrprotect(pte);
 838                         pte = pte_mkclean(pte);
 839                         set_pte_at(vma->vm_mm, address, ptep, pte);
 840 unlock_pte:
 841                         pte_unmap_unlock(ptep, ptl);
 842                 }
 843 
 844                 mmu_notifier_invalidate_range_end(&range);
 845         }
 846         i_mmap_unlock_read(mapping);
 847 }
 848 
 849 static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
 850                 struct address_space *mapping, void *entry)
 851 {
 852         unsigned long pfn, index, count;
 853         long ret = 0;
 854 
 855         /*
 856          * A page got tagged dirty in DAX mapping? Something is seriously
 857          * wrong.
 858          */
 859         if (WARN_ON(!xa_is_value(entry)))
 860                 return -EIO;
 861 
 862         if (unlikely(dax_is_locked(entry))) {
 863                 void *old_entry = entry;
 864 
 865                 entry = get_unlocked_entry(xas, 0);
 866 
 867                 /* Entry got punched out / reallocated? */
 868                 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
 869                         goto put_unlocked;
 870                 /*
 871                  * Entry got reallocated elsewhere? No need to writeback.
 872                  * We have to compare pfns as we must not bail out due to
 873                  * difference in lockbit or entry type.
 874                  */
 875                 if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
 876                         goto put_unlocked;
 877                 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
 878                                         dax_is_zero_entry(entry))) {
 879                         ret = -EIO;
 880                         goto put_unlocked;
 881                 }
 882 
 883                 /* Another fsync thread may have already done this entry */
 884                 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
 885                         goto put_unlocked;
 886         }
 887 
 888         /* Lock the entry to serialize with page faults */
 889         dax_lock_entry(xas, entry);
 890 
 891         /*
 892          * We can clear the tag now but we have to be careful so that concurrent
 893          * dax_writeback_one() calls for the same index cannot finish before we
 894          * actually flush the caches. This is achieved as the calls will look
 895          * at the entry only under the i_pages lock and once they do that
 896          * they will see the entry locked and wait for it to unlock.
 897          */
 898         xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
 899         xas_unlock_irq(xas);
 900 
 901         /*
 902          * If dax_writeback_mapping_range() was given a wbc->range_start
 903          * in the middle of a PMD, the 'index' we use needs to be
 904          * aligned to the start of the PMD.
 905          * This allows us to flush for PMD_SIZE and not have to worry about
 906          * partial PMD writebacks.
 907          */
 908         pfn = dax_to_pfn(entry);
 909         count = 1UL << dax_entry_order(entry);
 910         index = xas->xa_index & ~(count - 1);
 911 
 912         dax_entry_mkclean(mapping, index, pfn);
 913         dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
 914         /*
 915          * After we have flushed the cache, we can clear the dirty tag. There
 916          * cannot be new dirty data in the pfn after the flush has completed as
 917          * the pfn mappings are writeprotected and fault waits for mapping
 918          * entry lock.
 919          */
 920         xas_reset(xas);
 921         xas_lock_irq(xas);
 922         xas_store(xas, entry);
 923         xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
 924         dax_wake_entry(xas, entry, false);
 925 
 926         trace_dax_writeback_one(mapping->host, index, count);
 927         return ret;
 928 
 929  put_unlocked:
 930         put_unlocked_entry(xas, entry);
 931         return ret;
 932 }
 933 
 934 /*
 935  * Flush the mapping to the persistent domain within the byte range of [start,
 936  * end]. This is required by data integrity operations to ensure file data is
 937  * on persistent storage prior to completion of the operation.
 938  */
 939 int dax_writeback_mapping_range(struct address_space *mapping,
 940                 struct block_device *bdev, struct writeback_control *wbc)
 941 {
 942         XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
 943         struct inode *inode = mapping->host;
 944         pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
 945         struct dax_device *dax_dev;
 946         void *entry;
 947         int ret = 0;
 948         unsigned int scanned = 0;
 949 
 950         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 951                 return -EIO;
 952 
 953         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 954                 return 0;
 955 
 956         dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
 957         if (!dax_dev)
 958                 return -EIO;
 959 
 960         trace_dax_writeback_range(inode, xas.xa_index, end_index);
 961 
 962         tag_pages_for_writeback(mapping, xas.xa_index, end_index);
 963 
 964         xas_lock_irq(&xas);
 965         xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
 966                 ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
 967                 if (ret < 0) {
 968                         mapping_set_error(mapping, ret);
 969                         break;
 970                 }
 971                 if (++scanned % XA_CHECK_SCHED)
 972                         continue;
 973 
 974                 xas_pause(&xas);
 975                 xas_unlock_irq(&xas);
 976                 cond_resched();
 977                 xas_lock_irq(&xas);
 978         }
 979         xas_unlock_irq(&xas);
 980         put_dax(dax_dev);
 981         trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
 982         return ret;
 983 }
 984 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 985 
 986 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
 987 {
 988         return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
 989 }
 990 
 991 static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
 992                          pfn_t *pfnp)
 993 {
 994         const sector_t sector = dax_iomap_sector(iomap, pos);
 995         pgoff_t pgoff;
 996         int id, rc;
 997         long length;
 998 
 999         rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
1000         if (rc)
1001                 return rc;
1002         id = dax_read_lock();
1003         length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1004                                    NULL, pfnp);
1005         if (length < 0) {
1006                 rc = length;
1007                 goto out;
1008         }
1009         rc = -EINVAL;
1010         if (PFN_PHYS(length) < size)
1011                 goto out;
1012         if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1013                 goto out;
1014         /* For larger pages we need devmap */
1015         if (length > 1 && !pfn_t_devmap(*pfnp))
1016                 goto out;
1017         rc = 0;
1018 out:
1019         dax_read_unlock(id);
1020         return rc;
1021 }
1022 
1023 /*
1024  * The user has performed a load from a hole in the file.  Allocating a new
1025  * page in the file would cause excessive storage usage for workloads with
1026  * sparse files.  Instead we insert a read-only mapping of the 4k zero page.
1027  * If this page is ever written to we will re-fault and change the mapping to
1028  * point to real DAX storage instead.
1029  */
1030 static vm_fault_t dax_load_hole(struct xa_state *xas,
1031                 struct address_space *mapping, void **entry,
1032                 struct vm_fault *vmf)
1033 {
1034         struct inode *inode = mapping->host;
1035         unsigned long vaddr = vmf->address;
1036         pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1037         vm_fault_t ret;
1038 
1039         *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1040                         DAX_ZERO_PAGE, false);
1041 
1042         ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1043         trace_dax_load_hole(inode, vmf, ret);
1044         return ret;
1045 }
1046 
1047 static bool dax_range_is_aligned(struct block_device *bdev,
1048                                  unsigned int offset, unsigned int length)
1049 {
1050         unsigned short sector_size = bdev_logical_block_size(bdev);
1051 
1052         if (!IS_ALIGNED(offset, sector_size))
1053                 return false;
1054         if (!IS_ALIGNED(length, sector_size))
1055                 return false;
1056 
1057         return true;
1058 }
1059 
1060 int __dax_zero_page_range(struct block_device *bdev,
1061                 struct dax_device *dax_dev, sector_t sector,
1062                 unsigned int offset, unsigned int size)
1063 {
1064         if (dax_range_is_aligned(bdev, offset, size)) {
1065                 sector_t start_sector = sector + (offset >> 9);
1066 
1067                 return blkdev_issue_zeroout(bdev, start_sector,
1068                                 size >> 9, GFP_NOFS, 0);
1069         } else {
1070                 pgoff_t pgoff;
1071                 long rc, id;
1072                 void *kaddr;
1073 
1074                 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
1075                 if (rc)
1076                         return rc;
1077 
1078                 id = dax_read_lock();
1079                 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, NULL);
1080                 if (rc < 0) {
1081                         dax_read_unlock(id);
1082                         return rc;
1083                 }
1084                 memset(kaddr + offset, 0, size);
1085                 dax_flush(dax_dev, kaddr + offset, size);
1086                 dax_read_unlock(id);
1087         }
1088         return 0;
1089 }
1090 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1091 
1092 static loff_t
1093 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1094                 struct iomap *iomap)
1095 {
1096         struct block_device *bdev = iomap->bdev;
1097         struct dax_device *dax_dev = iomap->dax_dev;
1098         struct iov_iter *iter = data;
1099         loff_t end = pos + length, done = 0;
1100         ssize_t ret = 0;
1101         size_t xfer;
1102         int id;
1103 
1104         if (iov_iter_rw(iter) == READ) {
1105                 end = min(end, i_size_read(inode));
1106                 if (pos >= end)
1107                         return 0;
1108 
1109                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1110                         return iov_iter_zero(min(length, end - pos), iter);
1111         }
1112 
1113         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1114                 return -EIO;
1115 
1116         /*
1117          * Write can allocate block for an area which has a hole page mapped
1118          * into page tables. We have to tear down these mappings so that data
1119          * written by write(2) is visible in mmap.
1120          */
1121         if (iomap->flags & IOMAP_F_NEW) {
1122                 invalidate_inode_pages2_range(inode->i_mapping,
1123                                               pos >> PAGE_SHIFT,
1124                                               (end - 1) >> PAGE_SHIFT);
1125         }
1126 
1127         id = dax_read_lock();
1128         while (pos < end) {
1129                 unsigned offset = pos & (PAGE_SIZE - 1);
1130                 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1131                 const sector_t sector = dax_iomap_sector(iomap, pos);
1132                 ssize_t map_len;
1133                 pgoff_t pgoff;
1134                 void *kaddr;
1135 
1136                 if (fatal_signal_pending(current)) {
1137                         ret = -EINTR;
1138                         break;
1139                 }
1140 
1141                 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1142                 if (ret)
1143                         break;
1144 
1145                 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1146                                 &kaddr, NULL);
1147                 if (map_len < 0) {
1148                         ret = map_len;
1149                         break;
1150                 }
1151 
1152                 map_len = PFN_PHYS(map_len);
1153                 kaddr += offset;
1154                 map_len -= offset;
1155                 if (map_len > end - pos)
1156                         map_len = end - pos;
1157 
1158                 /*
1159                  * The userspace address for the memory copy has already been
1160                  * validated via access_ok() in either vfs_read() or
1161                  * vfs_write(), depending on which operation we are doing.
1162                  */
1163                 if (iov_iter_rw(iter) == WRITE)
1164                         xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1165                                         map_len, iter);
1166                 else
1167                         xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1168                                         map_len, iter);
1169 
1170                 pos += xfer;
1171                 length -= xfer;
1172                 done += xfer;
1173 
1174                 if (xfer == 0)
1175                         ret = -EFAULT;
1176                 if (xfer < map_len)
1177                         break;
1178         }
1179         dax_read_unlock(id);
1180 
1181         return done ? done : ret;
1182 }
1183 
1184 /**
1185  * dax_iomap_rw - Perform I/O to a DAX file
1186  * @iocb:       The control block for this I/O
1187  * @iter:       The addresses to do I/O from or to
1188  * @ops:        iomap ops passed from the file system
1189  *
1190  * This function performs read and write operations to directly mapped
1191  * persistent memory.  The callers needs to take care of read/write exclusion
1192  * and evicting any page cache pages in the region under I/O.
1193  */
1194 ssize_t
1195 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1196                 const struct iomap_ops *ops)
1197 {
1198         struct address_space *mapping = iocb->ki_filp->f_mapping;
1199         struct inode *inode = mapping->host;
1200         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1201         unsigned flags = 0;
1202 
1203         if (iov_iter_rw(iter) == WRITE) {
1204                 lockdep_assert_held_write(&inode->i_rwsem);
1205                 flags |= IOMAP_WRITE;
1206         } else {
1207                 lockdep_assert_held(&inode->i_rwsem);
1208         }
1209 
1210         if (iocb->ki_flags & IOCB_NOWAIT)
1211                 flags |= IOMAP_NOWAIT;
1212 
1213         while (iov_iter_count(iter)) {
1214                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1215                                 iter, dax_iomap_actor);
1216                 if (ret <= 0)
1217                         break;
1218                 pos += ret;
1219                 done += ret;
1220         }
1221 
1222         iocb->ki_pos += done;
1223         return done ? done : ret;
1224 }
1225 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1226 
1227 static vm_fault_t dax_fault_return(int error)
1228 {
1229         if (error == 0)
1230                 return VM_FAULT_NOPAGE;
1231         return vmf_error(error);
1232 }
1233 
1234 /*
1235  * MAP_SYNC on a dax mapping guarantees dirty metadata is
1236  * flushed on write-faults (non-cow), but not read-faults.
1237  */
1238 static bool dax_fault_is_synchronous(unsigned long flags,
1239                 struct vm_area_struct *vma, struct iomap *iomap)
1240 {
1241         return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1242                 && (iomap->flags & IOMAP_F_DIRTY);
1243 }
1244 
1245 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1246                                int *iomap_errp, const struct iomap_ops *ops)
1247 {
1248         struct vm_area_struct *vma = vmf->vma;
1249         struct address_space *mapping = vma->vm_file->f_mapping;
1250         XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1251         struct inode *inode = mapping->host;
1252         unsigned long vaddr = vmf->address;
1253         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1254         struct iomap iomap = { 0 };
1255         unsigned flags = IOMAP_FAULT;
1256         int error, major = 0;
1257         bool write = vmf->flags & FAULT_FLAG_WRITE;
1258         bool sync;
1259         vm_fault_t ret = 0;
1260         void *entry;
1261         pfn_t pfn;
1262 
1263         trace_dax_pte_fault(inode, vmf, ret);
1264         /*
1265          * Check whether offset isn't beyond end of file now. Caller is supposed
1266          * to hold locks serializing us with truncate / punch hole so this is
1267          * a reliable test.
1268          */
1269         if (pos >= i_size_read(inode)) {
1270                 ret = VM_FAULT_SIGBUS;
1271                 goto out;
1272         }
1273 
1274         if (write && !vmf->cow_page)
1275                 flags |= IOMAP_WRITE;
1276 
1277         entry = grab_mapping_entry(&xas, mapping, 0);
1278         if (xa_is_internal(entry)) {
1279                 ret = xa_to_internal(entry);
1280                 goto out;
1281         }
1282 
1283         /*
1284          * It is possible, particularly with mixed reads & writes to private
1285          * mappings, that we have raced with a PMD fault that overlaps with
1286          * the PTE we need to set up.  If so just return and the fault will be
1287          * retried.
1288          */
1289         if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1290                 ret = VM_FAULT_NOPAGE;
1291                 goto unlock_entry;
1292         }
1293 
1294         /*
1295          * Note that we don't bother to use iomap_apply here: DAX required
1296          * the file system block size to be equal the page size, which means
1297          * that we never have to deal with more than a single extent here.
1298          */
1299         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1300         if (iomap_errp)
1301                 *iomap_errp = error;
1302         if (error) {
1303                 ret = dax_fault_return(error);
1304                 goto unlock_entry;
1305         }
1306         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1307                 error = -EIO;   /* fs corruption? */
1308                 goto error_finish_iomap;
1309         }
1310 
1311         if (vmf->cow_page) {
1312                 sector_t sector = dax_iomap_sector(&iomap, pos);
1313 
1314                 switch (iomap.type) {
1315                 case IOMAP_HOLE:
1316                 case IOMAP_UNWRITTEN:
1317                         clear_user_highpage(vmf->cow_page, vaddr);
1318                         break;
1319                 case IOMAP_MAPPED:
1320                         error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1321                                         sector, PAGE_SIZE, vmf->cow_page, vaddr);
1322                         break;
1323                 default:
1324                         WARN_ON_ONCE(1);
1325                         error = -EIO;
1326                         break;
1327                 }
1328 
1329                 if (error)
1330                         goto error_finish_iomap;
1331 
1332                 __SetPageUptodate(vmf->cow_page);
1333                 ret = finish_fault(vmf);
1334                 if (!ret)
1335                         ret = VM_FAULT_DONE_COW;
1336                 goto finish_iomap;
1337         }
1338 
1339         sync = dax_fault_is_synchronous(flags, vma, &iomap);
1340 
1341         switch (iomap.type) {
1342         case IOMAP_MAPPED:
1343                 if (iomap.flags & IOMAP_F_NEW) {
1344                         count_vm_event(PGMAJFAULT);
1345                         count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1346                         major = VM_FAULT_MAJOR;
1347                 }
1348                 error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1349                 if (error < 0)
1350                         goto error_finish_iomap;
1351 
1352                 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1353                                                  0, write && !sync);
1354 
1355                 /*
1356                  * If we are doing synchronous page fault and inode needs fsync,
1357                  * we can insert PTE into page tables only after that happens.
1358                  * Skip insertion for now and return the pfn so that caller can
1359                  * insert it after fsync is done.
1360                  */
1361                 if (sync) {
1362                         if (WARN_ON_ONCE(!pfnp)) {
1363                                 error = -EIO;
1364                                 goto error_finish_iomap;
1365                         }
1366                         *pfnp = pfn;
1367                         ret = VM_FAULT_NEEDDSYNC | major;
1368                         goto finish_iomap;
1369                 }
1370                 trace_dax_insert_mapping(inode, vmf, entry);
1371                 if (write)
1372                         ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn);
1373                 else
1374                         ret = vmf_insert_mixed(vma, vaddr, pfn);
1375 
1376                 goto finish_iomap;
1377         case IOMAP_UNWRITTEN:
1378         case IOMAP_HOLE:
1379                 if (!write) {
1380                         ret = dax_load_hole(&xas, mapping, &entry, vmf);
1381                         goto finish_iomap;
1382                 }
1383                 /*FALLTHRU*/
1384         default:
1385                 WARN_ON_ONCE(1);
1386                 error = -EIO;
1387                 break;
1388         }
1389 
1390  error_finish_iomap:
1391         ret = dax_fault_return(error);
1392  finish_iomap:
1393         if (ops->iomap_end) {
1394                 int copied = PAGE_SIZE;
1395 
1396                 if (ret & VM_FAULT_ERROR)
1397                         copied = 0;
1398                 /*
1399                  * The fault is done by now and there's no way back (other
1400                  * thread may be already happily using PTE we have installed).
1401                  * Just ignore error from ->iomap_end since we cannot do much
1402                  * with it.
1403                  */
1404                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1405         }
1406  unlock_entry:
1407         dax_unlock_entry(&xas, entry);
1408  out:
1409         trace_dax_pte_fault_done(inode, vmf, ret);
1410         return ret | major;
1411 }
1412 
1413 #ifdef CONFIG_FS_DAX_PMD
1414 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1415                 struct iomap *iomap, void **entry)
1416 {
1417         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1418         unsigned long pmd_addr = vmf->address & PMD_MASK;
1419         struct vm_area_struct *vma = vmf->vma;
1420         struct inode *inode = mapping->host;
1421         pgtable_t pgtable = NULL;
1422         struct page *zero_page;
1423         spinlock_t *ptl;
1424         pmd_t pmd_entry;
1425         pfn_t pfn;
1426 
1427         zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1428 
1429         if (unlikely(!zero_page))
1430                 goto fallback;
1431 
1432         pfn = page_to_pfn_t(zero_page);
1433         *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1434                         DAX_PMD | DAX_ZERO_PAGE, false);
1435 
1436         if (arch_needs_pgtable_deposit()) {
1437                 pgtable = pte_alloc_one(vma->vm_mm);
1438                 if (!pgtable)
1439                         return VM_FAULT_OOM;
1440         }
1441 
1442         ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1443         if (!pmd_none(*(vmf->pmd))) {
1444                 spin_unlock(ptl);
1445                 goto fallback;
1446         }
1447 
1448         if (pgtable) {
1449                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1450                 mm_inc_nr_ptes(vma->vm_mm);
1451         }
1452         pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1453         pmd_entry = pmd_mkhuge(pmd_entry);
1454         set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1455         spin_unlock(ptl);
1456         trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1457         return VM_FAULT_NOPAGE;
1458 
1459 fallback:
1460         if (pgtable)
1461                 pte_free(vma->vm_mm, pgtable);
1462         trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1463         return VM_FAULT_FALLBACK;
1464 }
1465 
1466 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1467                                const struct iomap_ops *ops)
1468 {
1469         struct vm_area_struct *vma = vmf->vma;
1470         struct address_space *mapping = vma->vm_file->f_mapping;
1471         XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1472         unsigned long pmd_addr = vmf->address & PMD_MASK;
1473         bool write = vmf->flags & FAULT_FLAG_WRITE;
1474         bool sync;
1475         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1476         struct inode *inode = mapping->host;
1477         vm_fault_t result = VM_FAULT_FALLBACK;
1478         struct iomap iomap = { 0 };
1479         pgoff_t max_pgoff;
1480         void *entry;
1481         loff_t pos;
1482         int error;
1483         pfn_t pfn;
1484 
1485         /*
1486          * Check whether offset isn't beyond end of file now. Caller is
1487          * supposed to hold locks serializing us with truncate / punch hole so
1488          * this is a reliable test.
1489          */
1490         max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1491 
1492         trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1493 
1494         /*
1495          * Make sure that the faulting address's PMD offset (color) matches
1496          * the PMD offset from the start of the file.  This is necessary so
1497          * that a PMD range in the page table overlaps exactly with a PMD
1498          * range in the page cache.
1499          */
1500         if ((vmf->pgoff & PG_PMD_COLOUR) !=
1501             ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1502                 goto fallback;
1503 
1504         /* Fall back to PTEs if we're going to COW */
1505         if (write && !(vma->vm_flags & VM_SHARED))
1506                 goto fallback;
1507 
1508         /* If the PMD would extend outside the VMA */
1509         if (pmd_addr < vma->vm_start)
1510                 goto fallback;
1511         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1512                 goto fallback;
1513 
1514         if (xas.xa_index >= max_pgoff) {
1515                 result = VM_FAULT_SIGBUS;
1516                 goto out;
1517         }
1518 
1519         /* If the PMD would extend beyond the file size */
1520         if ((xas.xa_index | PG_PMD_COLOUR) >= max_pgoff)
1521                 goto fallback;
1522 
1523         /*
1524          * grab_mapping_entry() will make sure we get an empty PMD entry,
1525          * a zero PMD entry or a DAX PMD.  If it can't (because a PTE
1526          * entry is already in the array, for instance), it will return
1527          * VM_FAULT_FALLBACK.
1528          */
1529         entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1530         if (xa_is_internal(entry)) {
1531                 result = xa_to_internal(entry);
1532                 goto fallback;
1533         }
1534 
1535         /*
1536          * It is possible, particularly with mixed reads & writes to private
1537          * mappings, that we have raced with a PTE fault that overlaps with
1538          * the PMD we need to set up.  If so just return and the fault will be
1539          * retried.
1540          */
1541         if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1542                         !pmd_devmap(*vmf->pmd)) {
1543                 result = 0;
1544                 goto unlock_entry;
1545         }
1546 
1547         /*
1548          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1549          * setting up a mapping, so really we're using iomap_begin() as a way
1550          * to look up our filesystem block.
1551          */
1552         pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1553         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1554         if (error)
1555                 goto unlock_entry;
1556 
1557         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1558                 goto finish_iomap;
1559 
1560         sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1561 
1562         switch (iomap.type) {
1563         case IOMAP_MAPPED:
1564                 error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1565                 if (error < 0)
1566                         goto finish_iomap;
1567 
1568                 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1569                                                 DAX_PMD, write && !sync);
1570 
1571                 /*
1572                  * If we are doing synchronous page fault and inode needs fsync,
1573                  * we can insert PMD into page tables only after that happens.
1574                  * Skip insertion for now and return the pfn so that caller can
1575                  * insert it after fsync is done.
1576                  */
1577                 if (sync) {
1578                         if (WARN_ON_ONCE(!pfnp))
1579                                 goto finish_iomap;
1580                         *pfnp = pfn;
1581                         result = VM_FAULT_NEEDDSYNC;
1582                         goto finish_iomap;
1583                 }
1584 
1585                 trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
1586                 result = vmf_insert_pfn_pmd(vmf, pfn, write);
1587                 break;
1588         case IOMAP_UNWRITTEN:
1589         case IOMAP_HOLE:
1590                 if (WARN_ON_ONCE(write))
1591                         break;
1592                 result = dax_pmd_load_hole(&xas, vmf, &iomap, &entry);
1593                 break;
1594         default:
1595                 WARN_ON_ONCE(1);
1596                 break;
1597         }
1598 
1599  finish_iomap:
1600         if (ops->iomap_end) {
1601                 int copied = PMD_SIZE;
1602 
1603                 if (result == VM_FAULT_FALLBACK)
1604                         copied = 0;
1605                 /*
1606                  * The fault is done by now and there's no way back (other
1607                  * thread may be already happily using PMD we have installed).
1608                  * Just ignore error from ->iomap_end since we cannot do much
1609                  * with it.
1610                  */
1611                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1612                                 &iomap);
1613         }
1614  unlock_entry:
1615         dax_unlock_entry(&xas, entry);
1616  fallback:
1617         if (result == VM_FAULT_FALLBACK) {
1618                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1619                 count_vm_event(THP_FAULT_FALLBACK);
1620         }
1621 out:
1622         trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1623         return result;
1624 }
1625 #else
1626 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1627                                const struct iomap_ops *ops)
1628 {
1629         return VM_FAULT_FALLBACK;
1630 }
1631 #endif /* CONFIG_FS_DAX_PMD */
1632 
1633 /**
1634  * dax_iomap_fault - handle a page fault on a DAX file
1635  * @vmf: The description of the fault
1636  * @pe_size: Size of the page to fault in
1637  * @pfnp: PFN to insert for synchronous faults if fsync is required
1638  * @iomap_errp: Storage for detailed error code in case of error
1639  * @ops: Iomap ops passed from the file system
1640  *
1641  * When a page fault occurs, filesystems may call this helper in
1642  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1643  * has done all the necessary locking for page fault to proceed
1644  * successfully.
1645  */
1646 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1647                     pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1648 {
1649         switch (pe_size) {
1650         case PE_SIZE_PTE:
1651                 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1652         case PE_SIZE_PMD:
1653                 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1654         default:
1655                 return VM_FAULT_FALLBACK;
1656         }
1657 }
1658 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1659 
1660 /*
1661  * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1662  * @vmf: The description of the fault
1663  * @pfn: PFN to insert
1664  * @order: Order of entry to insert.
1665  *
1666  * This function inserts a writeable PTE or PMD entry into the page tables
1667  * for an mmaped DAX file.  It also marks the page cache entry as dirty.
1668  */
1669 static vm_fault_t
1670 dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1671 {
1672         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1673         XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1674         void *entry;
1675         vm_fault_t ret;
1676 
1677         xas_lock_irq(&xas);
1678         entry = get_unlocked_entry(&xas, order);
1679         /* Did we race with someone splitting entry or so? */
1680         if (!entry || dax_is_conflict(entry) ||
1681             (order == 0 && !dax_is_pte_entry(entry))) {
1682                 put_unlocked_entry(&xas, entry);
1683                 xas_unlock_irq(&xas);
1684                 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1685                                                       VM_FAULT_NOPAGE);
1686                 return VM_FAULT_NOPAGE;
1687         }
1688         xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1689         dax_lock_entry(&xas, entry);
1690         xas_unlock_irq(&xas);
1691         if (order == 0)
1692                 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1693 #ifdef CONFIG_FS_DAX_PMD
1694         else if (order == PMD_ORDER)
1695                 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1696 #endif
1697         else
1698                 ret = VM_FAULT_FALLBACK;
1699         dax_unlock_entry(&xas, entry);
1700         trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1701         return ret;
1702 }
1703 
1704 /**
1705  * dax_finish_sync_fault - finish synchronous page fault
1706  * @vmf: The description of the fault
1707  * @pe_size: Size of entry to be inserted
1708  * @pfn: PFN to insert
1709  *
1710  * This function ensures that the file range touched by the page fault is
1711  * stored persistently on the media and handles inserting of appropriate page
1712  * table entry.
1713  */
1714 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1715                 enum page_entry_size pe_size, pfn_t pfn)
1716 {
1717         int err;
1718         loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1719         unsigned int order = pe_order(pe_size);
1720         size_t len = PAGE_SIZE << order;
1721 
1722         err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1723         if (err)
1724                 return VM_FAULT_SIGBUS;
1725         return dax_insert_pfn_mkwrite(vmf, pfn, order);
1726 }
1727 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);

/* [<][>][^][v][top][bottom][index][help] */