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