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