fs/dax.c

   1 /*
   2  * fs/dax.c - Direct Access filesystem code
   3  * Copyright (c) 2013-2014 Intel Corporation
   4  * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
   5  * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms and conditions of the GNU General Public License,
   9  * version 2, as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  */
  16
  17 #include <linux/atomic.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/buffer_head.h>
  20 #include <linux/dax.h>
  21 #include <linux/fs.h>
  22 #include <linux/genhd.h>
  23 #include <linux/highmem.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm.h>
  26 #include <linux/mutex.h>
  27 #include <linux/pagevec.h>
  28 #include <linux/pmem.h>
  29 #include <linux/sched.h>
  30 #include <linux/sched/signal.h>
  31 #include <linux/uio.h>
  32 #include <linux/vmstat.h>
  33 #include <linux/pfn_t.h>
  34 #include <linux/sizes.h>
  35 #include <linux/mmu_notifier.h>
  36 #include <linux/iomap.h>
  37 #include "internal.h"
  38
  39 #define CREATE_TRACE_POINTS
  40 #include <trace/events/fs_dax.h>
  41
  42 /* We choose 4096 entries - same as per-zone page wait tables */
  43 #define DAX_WAIT_TABLE_BITS 12
  44 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
  45
  46 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
  47
  48 static int __init init_dax_wait_table(void)
  49 {
  50         int i;
  51
  52         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
  53                 init_waitqueue_head(wait_table + i);
  54         return 0;
  55 }
  56 fs_initcall(init_dax_wait_table);
  57
  58 static int dax_is_pmd_entry(void *entry)
  59 {
  60         return (unsigned long)entry & RADIX_DAX_PMD;
  61 }
  62
  63 static int dax_is_pte_entry(void *entry)
  64 {
  65         return !((unsigned long)entry & RADIX_DAX_PMD);
  66 }
  67
  68 static int dax_is_zero_entry(void *entry)
  69 {
  70         return (unsigned long)entry & RADIX_DAX_HZP;
  71 }
  72
  73 static int dax_is_empty_entry(void *entry)
  74 {
  75         return (unsigned long)entry & RADIX_DAX_EMPTY;
  76 }
  77
  78 /*
  79  * DAX radix tree locking
  80  */
  81 struct exceptional_entry_key {
  82         struct address_space *mapping;
  83         pgoff_t entry_start;
  84 };
  85
  86 struct wait_exceptional_entry_queue {
  87         wait_queue_entry_t wait;
  88         struct exceptional_entry_key key;
  89 };
  90
  91 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
  92                 pgoff_t index, void *entry, struct exceptional_entry_key *key)
  93 {
  94         unsigned long hash;
  95
  96         /*
  97          * If 'entry' is a PMD, align the 'index' that we use for the wait
  98          * queue to the start of that PMD.  This ensures that all offsets in
  99          * the range covered by the PMD map to the same bit lock.
 100          */
 101         if (dax_is_pmd_entry(entry))
 102                 index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
 103
 104         key->mapping = mapping;
 105         key->entry_start = index;
 106
 107         hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
 108         return wait_table + hash;
 109 }
 110
 111 static int wake_exceptional_entry_func(wait_queue_entry_t *wait, unsigned int mode,
 112                                        int sync, void *keyp)
 113 {
 114         struct exceptional_entry_key *key = keyp;
 115         struct wait_exceptional_entry_queue *ewait =
 116                 container_of(wait, struct wait_exceptional_entry_queue, wait);
 117
 118         if (key->mapping != ewait->key.mapping ||
 119             key->entry_start != ewait->key.entry_start)
 120                 return 0;
 121         return autoremove_wake_function(wait, mode, sync, NULL);
 122 }
 123
 124 /*
 125  * Check whether the given slot is locked. The function must be called with
 126  * mapping->tree_lock held
 127  */
 128 static inline int slot_locked(struct address_space *mapping, void **slot)
 129 {
 130         unsigned long entry = (unsigned long)
 131                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 132         return entry & RADIX_DAX_ENTRY_LOCK;
 133 }
 134
 135 /*
 136  * Mark the given slot is locked. The function must be called with
 137  * mapping->tree_lock held
 138  */
 139 static inline void *lock_slot(struct address_space *mapping, void **slot)
 140 {
 141         unsigned long entry = (unsigned long)
 142                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 143
 144         entry |= RADIX_DAX_ENTRY_LOCK;
 145         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
 146         return (void *)entry;
 147 }
 148
 149 /*
 150  * Mark the given slot is unlocked. The function must be called with
 151  * mapping->tree_lock held
 152  */
 153 static inline void *unlock_slot(struct address_space *mapping, void **slot)
 154 {
 155         unsigned long entry = (unsigned long)
 156                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 157
 158         entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
 159         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
 160         return (void *)entry;
 161 }
 162
 163 /*
 164  * Lookup entry in radix tree, wait for it to become unlocked if it is
 165  * exceptional entry and return it. The caller must call
 166  * put_unlocked_mapping_entry() when he decided not to lock the entry or
 167  * put_locked_mapping_entry() when he locked the entry and now wants to
 168  * unlock it.
 169  *
 170  * The function must be called with mapping->tree_lock held.
 171  */
 172 static void *get_unlocked_mapping_entry(struct address_space *mapping,
 173                                         pgoff_t index, void ***slotp)
 174 {
 175         void *entry, **slot;
 176         struct wait_exceptional_entry_queue ewait;
 177         wait_queue_head_t *wq;
 178
 179         init_wait(&ewait.wait);
 180         ewait.wait.func = wake_exceptional_entry_func;
 181
 182         for (;;) {
 183                 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
 184                                           &slot);
 185                 if (!entry || !radix_tree_exceptional_entry(entry) ||
 186                     !slot_locked(mapping, slot)) {
 187                         if (slotp)
 188                                 *slotp = slot;
 189                         return entry;
 190                 }
 191
 192                 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
 193                 prepare_to_wait_exclusive(wq, &ewait.wait,
 194                                           TASK_UNINTERRUPTIBLE);
 195                 spin_unlock_irq(&mapping->tree_lock);
 196                 schedule();
 197                 finish_wait(wq, &ewait.wait);
 198                 spin_lock_irq(&mapping->tree_lock);
 199         }
 200 }
 201
 202 static void dax_unlock_mapping_entry(struct address_space *mapping,
 203                                      pgoff_t index)
 204 {
 205         void *entry, **slot;
 206
 207         spin_lock_irq(&mapping->tree_lock);
 208         entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
 209         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
 210                          !slot_locked(mapping, slot))) {
 211                 spin_unlock_irq(&mapping->tree_lock);
 212                 return;
 213         }
 214         unlock_slot(mapping, slot);
 215         spin_unlock_irq(&mapping->tree_lock);
 216         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 217 }
 218
 219 static void put_locked_mapping_entry(struct address_space *mapping,
 220                                      pgoff_t index, void *entry)
 221 {
 222         if (!radix_tree_exceptional_entry(entry)) {
 223                 unlock_page(entry);
 224                 put_page(entry);
 225         } else {
 226                 dax_unlock_mapping_entry(mapping, index);
 227         }
 228 }
 229
 230 /*
 231  * Called when we are done with radix tree entry we looked up via
 232  * get_unlocked_mapping_entry() and which we didn't lock in the end.
 233  */
 234 static void put_unlocked_mapping_entry(struct address_space *mapping,
 235                                        pgoff_t index, void *entry)
 236 {
 237         if (!radix_tree_exceptional_entry(entry))
 238                 return;
 239
 240         /* We have to wake up next waiter for the radix tree entry lock */
 241         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 242 }
 243
 244 /*
 245  * Find radix tree entry at given index. If it points to a page, return with
 246  * the page locked. If it points to the exceptional entry, return with the
 247  * radix tree entry locked. If the radix tree doesn't contain given index,
 248  * create empty exceptional entry for the index and return with it locked.
 249  *
 250  * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
 251  * either return that locked entry or will return an error.  This error will
 252  * happen if there are any 4k entries (either zero pages or DAX entries)
 253  * within the 2MiB range that we are requesting.
 254  *
 255  * We always favor 4k entries over 2MiB entries. There isn't a flow where we
 256  * evict 4k entries in order to 'upgrade' them to a 2MiB entry.  A 2MiB
 257  * insertion will fail if it finds any 4k entries already in the tree, and a
 258  * 4k insertion will cause an existing 2MiB entry to be unmapped and
 259  * downgraded to 4k entries.  This happens for both 2MiB huge zero pages as
 260  * well as 2MiB empty entries.
 261  *
 262  * The exception to this downgrade path is for 2MiB DAX PMD entries that have
 263  * real storage backing them.  We will leave these real 2MiB DAX entries in
 264  * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
 265  *
 266  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
 267  * persistent memory the benefit is doubtful. We can add that later if we can
 268  * show it helps.
 269  */
 270 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
 271                 unsigned long size_flag)
 272 {
 273         bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
 274         void *entry, **slot;
 275
 276 restart:
 277         spin_lock_irq(&mapping->tree_lock);
 278         entry = get_unlocked_mapping_entry(mapping, index, &slot);
 279
 280         if (entry) {
 281                 if (size_flag & RADIX_DAX_PMD) {
 282                         if (!radix_tree_exceptional_entry(entry) ||
 283                             dax_is_pte_entry(entry)) {
 284                                 put_unlocked_mapping_entry(mapping, index,
 285                                                 entry);
 286                                 entry = ERR_PTR(-EEXIST);
 287                                 goto out_unlock;
 288                         }
 289                 } else { /* trying to grab a PTE entry */
 290                         if (radix_tree_exceptional_entry(entry) &&
 291                             dax_is_pmd_entry(entry) &&
 292                             (dax_is_zero_entry(entry) ||
 293                              dax_is_empty_entry(entry))) {
 294                                 pmd_downgrade = true;
 295                         }
 296                 }
 297         }
 298
 299         /* No entry for given index? Make sure radix tree is big enough. */
 300         if (!entry || pmd_downgrade) {
 301                 int err;
 302
 303                 if (pmd_downgrade) {
 304                         /*
 305                          * Make sure 'entry' remains valid while we drop
 306                          * mapping->tree_lock.
 307                          */
 308                         entry = lock_slot(mapping, slot);
 309                 }
 310
 311                 spin_unlock_irq(&mapping->tree_lock);
 312                 /*
 313                  * Besides huge zero pages the only other thing that gets
 314                  * downgraded are empty entries which don't need to be
 315                  * unmapped.
 316                  */
 317                 if (pmd_downgrade && dax_is_zero_entry(entry))
 318                         unmap_mapping_range(mapping,
 319                                 (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
 320
 321                 err = radix_tree_preload(
 322                                 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
 323                 if (err) {
 324                         if (pmd_downgrade)
 325                                 put_locked_mapping_entry(mapping, index, entry);
 326                         return ERR_PTR(err);
 327                 }
 328                 spin_lock_irq(&mapping->tree_lock);
 329
 330                 if (!entry) {
 331                         /*
 332                          * We needed to drop the page_tree lock while calling
 333                          * radix_tree_preload() and we didn't have an entry to
 334                          * lock.  See if another thread inserted an entry at
 335                          * our index during this time.
 336                          */
 337                         entry = __radix_tree_lookup(&mapping->page_tree, index,
 338                                         NULL, &slot);
 339                         if (entry) {
 340                                 radix_tree_preload_end();
 341                                 spin_unlock_irq(&mapping->tree_lock);
 342                                 goto restart;
 343                         }
 344                 }
 345
 346                 if (pmd_downgrade) {
 347                         radix_tree_delete(&mapping->page_tree, index);
 348                         mapping->nrexceptional--;
 349                         dax_wake_mapping_entry_waiter(mapping, index, entry,
 350                                         true);
 351                 }
 352
 353                 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
 354
 355                 err = __radix_tree_insert(&mapping->page_tree, index,
 356                                 dax_radix_order(entry), entry);
 357                 radix_tree_preload_end();
 358                 if (err) {
 359                         spin_unlock_irq(&mapping->tree_lock);
 360                         /*
 361                          * Our insertion of a DAX entry failed, most likely
 362                          * because we were inserting a PMD entry and it
 363                          * collided with a PTE sized entry at a different
 364                          * index in the PMD range.  We haven't inserted
 365                          * anything into the radix tree and have no waiters to
 366                          * wake.
 367                          */
 368                         return ERR_PTR(err);
 369                 }
 370                 /* Good, we have inserted empty locked entry into the tree. */
 371                 mapping->nrexceptional++;
 372                 spin_unlock_irq(&mapping->tree_lock);
 373                 return entry;
 374         }
 375         /* Normal page in radix tree? */
 376         if (!radix_tree_exceptional_entry(entry)) {
 377                 struct page *page = entry;
 378
 379                 get_page(page);
 380                 spin_unlock_irq(&mapping->tree_lock);
 381                 lock_page(page);
 382                 /* Page got truncated? Retry... */
 383                 if (unlikely(page->mapping != mapping)) {
 384                         unlock_page(page);
 385                         put_page(page);
 386                         goto restart;
 387                 }
 388                 return page;
 389         }
 390         entry = lock_slot(mapping, slot);
 391  out_unlock:
 392         spin_unlock_irq(&mapping->tree_lock);
 393         return entry;
 394 }
 395
 396 /*
 397  * We do not necessarily hold the mapping->tree_lock when we call this
 398  * function so it is possible that 'entry' is no longer a valid item in the
 399  * radix tree.  This is okay because all we really need to do is to find the
 400  * correct waitqueue where tasks might be waiting for that old 'entry' and
 401  * wake them.
 402  */
 403 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
 404                 pgoff_t index, void *entry, bool wake_all)
 405 {
 406         struct exceptional_entry_key key;
 407         wait_queue_head_t *wq;
 408
 409         wq = dax_entry_waitqueue(mapping, index, entry, &key);
 410
 411         /*
 412          * Checking for locked entry and prepare_to_wait_exclusive() happens
 413          * under mapping->tree_lock, ditto for entry handling in our callers.
 414          * So at this point all tasks that could have seen our entry locked
 415          * must be in the waitqueue and the following check will see them.
 416          */
 417         if (waitqueue_active(wq))
 418                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
 419 }
 420
 421 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
 422                                           pgoff_t index, bool trunc)
 423 {
 424         int ret = 0;
 425         void *entry;
 426         struct radix_tree_root *page_tree = &mapping->page_tree;
 427
 428         spin_lock_irq(&mapping->tree_lock);
 429         entry = get_unlocked_mapping_entry(mapping, index, NULL);
 430         if (!entry || !radix_tree_exceptional_entry(entry))
 431                 goto out;
 432         if (!trunc &&
 433             (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
 434              radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)))
 435                 goto out;
 436         radix_tree_delete(page_tree, index);
 437         mapping->nrexceptional--;
 438         ret = 1;
 439 out:
 440         put_unlocked_mapping_entry(mapping, index, entry);
 441         spin_unlock_irq(&mapping->tree_lock);
 442         return ret;
 443 }
 444 /*
 445  * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
 446  * entry to get unlocked before deleting it.
 447  */
 448 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 449 {
 450         int ret = __dax_invalidate_mapping_entry(mapping, index, true);
 451
 452         /*
 453          * This gets called from truncate / punch_hole path. As such, the caller
 454          * must hold locks protecting against concurrent modifications of the
 455          * radix tree (usually fs-private i_mmap_sem for writing). Since the
 456          * caller has seen exceptional entry for this index, we better find it
 457          * at that index as well...
 458          */
 459         WARN_ON_ONCE(!ret);
 460         return ret;
 461 }
 462
 463 /*
 464  * Invalidate exceptional DAX entry if it is clean.
 465  */
 466 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
 467                                       pgoff_t index)
 468 {
 469         return __dax_invalidate_mapping_entry(mapping, index, false);
 470 }
 471
 472 /*
 473  * The user has performed a load from a hole in the file.  Allocating
 474  * a new page in the file would cause excessive storage usage for
 475  * workloads with sparse files.  We allocate a page cache page instead.
 476  * We'll kick it out of the page cache if it's ever written to,
 477  * otherwise it will simply fall out of the page cache under memory
 478  * pressure without ever having been dirtied.
 479  */
 480 static int dax_load_hole(struct address_space *mapping, void **entry,
 481                          struct vm_fault *vmf)
 482 {
 483         struct inode *inode = mapping->host;
 484         struct page *page;
 485         int ret;
 486
 487         /* Hole page already exists? Return it...  */
 488         if (!radix_tree_exceptional_entry(*entry)) {
 489                 page = *entry;
 490                 goto finish_fault;
 491         }
 492
 493         /* This will replace locked radix tree entry with a hole page */
 494         page = find_or_create_page(mapping, vmf->pgoff,
 495                                    vmf->gfp_mask | __GFP_ZERO);
 496         if (!page) {
 497                 ret = VM_FAULT_OOM;
 498                 goto out;
 499         }
 500
 501 finish_fault:
 502         vmf->page = page;
 503         ret = finish_fault(vmf);
 504         vmf->page = NULL;
 505         *entry = page;
 506         if (!ret) {
 507                 /* Grab reference for PTE that is now referencing the page */
 508                 get_page(page);
 509                 ret = VM_FAULT_NOPAGE;
 510         }
 511 out:
 512         trace_dax_load_hole(inode, vmf, ret);
 513         return ret;
 514 }
 515
 516 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
 517                 sector_t sector, size_t size, struct page *to,
 518                 unsigned long vaddr)
 519 {
 520         void *vto, *kaddr;
 521         pgoff_t pgoff;
 522         pfn_t pfn;
 523         long rc;
 524         int id;
 525
 526         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 527         if (rc)
 528                 return rc;
 529
 530         id = dax_read_lock();
 531         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
 532         if (rc < 0) {
 533                 dax_read_unlock(id);
 534                 return rc;
 535         }
 536         vto = kmap_atomic(to);
 537         copy_user_page(vto, (void __force *)kaddr, vaddr, to);
 538         kunmap_atomic(vto);
 539         dax_read_unlock(id);
 540         return 0;
 541 }
 542
 543 /*
 544  * By this point grab_mapping_entry() has ensured that we have a locked entry
 545  * of the appropriate size so we don't have to worry about downgrading PMDs to
 546  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
 547  * already in the tree, we will skip the insertion and just dirty the PMD as
 548  * appropriate.
 549  */
 550 static void *dax_insert_mapping_entry(struct address_space *mapping,
 551                                       struct vm_fault *vmf,
 552                                       void *entry, sector_t sector,
 553                                       unsigned long flags)
 554 {
 555         struct radix_tree_root *page_tree = &mapping->page_tree;
 556         int error = 0;
 557         bool hole_fill = false;
 558         void *new_entry;
 559         pgoff_t index = vmf->pgoff;
 560
 561         if (vmf->flags & FAULT_FLAG_WRITE)
 562                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 563
 564         /* Replacing hole page with block mapping? */
 565         if (!radix_tree_exceptional_entry(entry)) {
 566                 hole_fill = true;
 567                 /*
 568                  * Unmap the page now before we remove it from page cache below.
 569                  * The page is locked so it cannot be faulted in again.
 570                  */
 571                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 572                                     PAGE_SIZE, 0);
 573                 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
 574                 if (error)
 575                         return ERR_PTR(error);
 576         } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) {
 577                 /* replacing huge zero page with PMD block mapping */
 578                 unmap_mapping_range(mapping,
 579                         (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
 580         }
 581
 582         spin_lock_irq(&mapping->tree_lock);
 583         new_entry = dax_radix_locked_entry(sector, flags);
 584
 585         if (hole_fill) {
 586                 __delete_from_page_cache(entry, NULL);
 587                 /* Drop pagecache reference */
 588                 put_page(entry);
 589                 error = __radix_tree_insert(page_tree, index,
 590                                 dax_radix_order(new_entry), new_entry);
 591                 if (error) {
 592                         new_entry = ERR_PTR(error);
 593                         goto unlock;
 594                 }
 595                 mapping->nrexceptional++;
 596         } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 597                 /*
 598                  * Only swap our new entry into the radix tree if the current
 599                  * entry is a zero page or an empty entry.  If a normal PTE or
 600                  * PMD entry is already in the tree, we leave it alone.  This
 601                  * means that if we are trying to insert a PTE and the
 602                  * existing entry is a PMD, we will just leave the PMD in the
 603                  * tree and dirty it if necessary.
 604                  */
 605                 struct radix_tree_node *node;
 606                 void **slot;
 607                 void *ret;
 608
 609                 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
 610                 WARN_ON_ONCE(ret != entry);
 611                 __radix_tree_replace(page_tree, node, slot,
 612                                      new_entry, NULL, NULL);
 613         }
 614         if (vmf->flags & FAULT_FLAG_WRITE)
 615                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 616  unlock:
 617         spin_unlock_irq(&mapping->tree_lock);
 618         if (hole_fill) {
 619                 radix_tree_preload_end();
 620                 /*
 621                  * We don't need hole page anymore, it has been replaced with
 622                  * locked radix tree entry now.
 623                  */
 624                 if (mapping->a_ops->freepage)
 625                         mapping->a_ops->freepage(entry);
 626                 unlock_page(entry);
 627                 put_page(entry);
 628         }
 629         return new_entry;
 630 }
 631
 632 static inline unsigned long
 633 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
 634 {
 635         unsigned long address;
 636
 637         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 638         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
 639         return address;
 640 }
 641
 642 /* Walk all mappings of a given index of a file and writeprotect them */
 643 static void dax_mapping_entry_mkclean(struct address_space *mapping,
 644                                       pgoff_t index, unsigned long pfn)
 645 {
 646         struct vm_area_struct *vma;
 647         pte_t pte, *ptep = NULL;
 648         pmd_t *pmdp = NULL;
 649         spinlock_t *ptl;
 650         bool changed;
 651
 652         i_mmap_lock_read(mapping);
 653         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
 654                 unsigned long address;
 655
 656                 cond_resched();
 657
 658                 if (!(vma->vm_flags & VM_SHARED))
 659                         continue;
 660
 661                 address = pgoff_address(index, vma);
 662                 changed = false;
 663                 if (follow_pte_pmd(vma->vm_mm, address, &ptep, &pmdp, &ptl))
 664                         continue;
 665
 666                 if (pmdp) {
 667 #ifdef CONFIG_FS_DAX_PMD
 668                         pmd_t pmd;
 669
 670                         if (pfn != pmd_pfn(*pmdp))
 671                                 goto unlock_pmd;
 672                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
 673                                 goto unlock_pmd;
 674
 675                         flush_cache_page(vma, address, pfn);
 676                         pmd = pmdp_huge_clear_flush(vma, address, pmdp);
 677                         pmd = pmd_wrprotect(pmd);
 678                         pmd = pmd_mkclean(pmd);
 679                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
 680                         changed = true;
 681 unlock_pmd:
 682                         spin_unlock(ptl);
 683 #endif
 684                 } else {
 685                         if (pfn != pte_pfn(*ptep))
 686                                 goto unlock_pte;
 687                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
 688                                 goto unlock_pte;
 689
 690                         flush_cache_page(vma, address, pfn);
 691                         pte = ptep_clear_flush(vma, address, ptep);
 692                         pte = pte_wrprotect(pte);
 693                         pte = pte_mkclean(pte);
 694                         set_pte_at(vma->vm_mm, address, ptep, pte);
 695                         changed = true;
 696 unlock_pte:
 697                         pte_unmap_unlock(ptep, ptl);
 698                 }
 699
 700                 if (changed)
 701                         mmu_notifier_invalidate_page(vma->vm_mm, address);
 702         }
 703         i_mmap_unlock_read(mapping);
 704 }
 705
 706 static int dax_writeback_one(struct block_device *bdev,
 707                 struct dax_device *dax_dev, struct address_space *mapping,
 708                 pgoff_t index, void *entry)
 709 {
 710         struct radix_tree_root *page_tree = &mapping->page_tree;
 711         void *entry2, **slot, *kaddr;
 712         long ret = 0, id;
 713         sector_t sector;
 714         pgoff_t pgoff;
 715         size_t size;
 716         pfn_t pfn;
 717
 718         /*
 719          * A page got tagged dirty in DAX mapping? Something is seriously
 720          * wrong.
 721          */
 722         if (WARN_ON(!radix_tree_exceptional_entry(entry)))
 723                 return -EIO;
 724
 725         spin_lock_irq(&mapping->tree_lock);
 726         entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
 727         /* Entry got punched out / reallocated? */
 728         if (!entry2 || !radix_tree_exceptional_entry(entry2))
 729                 goto put_unlocked;
 730         /*
 731          * Entry got reallocated elsewhere? No need to writeback. We have to
 732          * compare sectors as we must not bail out due to difference in lockbit
 733          * or entry type.
 734          */
 735         if (dax_radix_sector(entry2) != dax_radix_sector(entry))
 736                 goto put_unlocked;
 737         if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
 738                                 dax_is_zero_entry(entry))) {
 739                 ret = -EIO;
 740                 goto put_unlocked;
 741         }
 742
 743         /* Another fsync thread may have already written back this entry */
 744         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 745                 goto put_unlocked;
 746         /* Lock the entry to serialize with page faults */
 747         entry = lock_slot(mapping, slot);
 748         /*
 749          * We can clear the tag now but we have to be careful so that concurrent
 750          * dax_writeback_one() calls for the same index cannot finish before we
 751          * actually flush the caches. This is achieved as the calls will look
 752          * at the entry only under tree_lock and once they do that they will
 753          * see the entry locked and wait for it to unlock.
 754          */
 755         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 756         spin_unlock_irq(&mapping->tree_lock);
 757
 758         /*
 759          * Even if dax_writeback_mapping_range() was given a wbc->range_start
 760          * in the middle of a PMD, the 'index' we are given will be aligned to
 761          * the start index of the PMD, as will the sector we pull from
 762          * 'entry'.  This allows us to flush for PMD_SIZE and not have to
 763          * worry about partial PMD writebacks.
 764          */
 765         sector = dax_radix_sector(entry);
 766         size = PAGE_SIZE << dax_radix_order(entry);
 767
 768         id = dax_read_lock();
 769         ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 770         if (ret)
 771                 goto dax_unlock;
 772
 773         /*
 774          * dax_direct_access() may sleep, so cannot hold tree_lock over
 775          * its invocation.
 776          */
 777         ret = dax_direct_access(dax_dev, pgoff, size / PAGE_SIZE, &kaddr, &pfn);
 778         if (ret < 0)
 779                 goto dax_unlock;
 780
 781         if (WARN_ON_ONCE(ret < size / PAGE_SIZE)) {
 782                 ret = -EIO;
 783                 goto dax_unlock;
 784         }
 785
 786         dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(pfn));
 787         wb_cache_pmem(kaddr, size);
 788         /*
 789          * After we have flushed the cache, we can clear the dirty tag. There
 790          * cannot be new dirty data in the pfn after the flush has completed as
 791          * the pfn mappings are writeprotected and fault waits for mapping
 792          * entry lock.
 793          */
 794         spin_lock_irq(&mapping->tree_lock);
 795         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
 796         spin_unlock_irq(&mapping->tree_lock);
 797         trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT);
 798  dax_unlock:
 799         dax_read_unlock(id);
 800         put_locked_mapping_entry(mapping, index, entry);
 801         return ret;
 802
 803  put_unlocked:
 804         put_unlocked_mapping_entry(mapping, index, entry2);
 805         spin_unlock_irq(&mapping->tree_lock);
 806         return ret;
 807 }
 808
 809 /*
 810  * Flush the mapping to the persistent domain within the byte range of [start,
 811  * end]. This is required by data integrity operations to ensure file data is
 812  * on persistent storage prior to completion of the operation.
 813  */
 814 int dax_writeback_mapping_range(struct address_space *mapping,
 815                 struct block_device *bdev, struct writeback_control *wbc)
 816 {
 817         struct inode *inode = mapping->host;
 818         pgoff_t start_index, end_index;
 819         pgoff_t indices[PAGEVEC_SIZE];
 820         struct dax_device *dax_dev;
 821         struct pagevec pvec;
 822         bool done = false;
 823         int i, ret = 0;
 824
 825         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 826                 return -EIO;
 827
 828         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 829                 return 0;
 830
 831         dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
 832         if (!dax_dev)
 833                 return -EIO;
 834
 835         start_index = wbc->range_start >> PAGE_SHIFT;
 836         end_index = wbc->range_end >> PAGE_SHIFT;
 837
 838         trace_dax_writeback_range(inode, start_index, end_index);
 839
 840         tag_pages_for_writeback(mapping, start_index, end_index);
 841
 842         pagevec_init(&pvec, 0);
 843         while (!done) {
 844                 pvec.nr = find_get_entries_tag(mapping, start_index,
 845                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 846                                 pvec.pages, indices);
 847
 848                 if (pvec.nr == 0)
 849                         break;
 850
 851                 for (i = 0; i < pvec.nr; i++) {
 852                         if (indices[i] > end_index) {
 853                                 done = true;
 854                                 break;
 855                         }
 856
 857                         ret = dax_writeback_one(bdev, dax_dev, mapping,
 858                                         indices[i], pvec.pages[i]);
 859                         if (ret < 0)
 860                                 goto out;
 861                 }
 862                 start_index = indices[pvec.nr - 1] + 1;
 863         }
 864 out:
 865         put_dax(dax_dev);
 866         trace_dax_writeback_range_done(inode, start_index, end_index);
 867         return (ret < 0 ? ret : 0);
 868 }
 869 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 870
 871 static int dax_insert_mapping(struct address_space *mapping,
 872                 struct block_device *bdev, struct dax_device *dax_dev,
 873                 sector_t sector, size_t size, void **entryp,
 874                 struct vm_area_struct *vma, struct vm_fault *vmf)
 875 {
 876         unsigned long vaddr = vmf->address;
 877         void *entry = *entryp;
 878         void *ret, *kaddr;
 879         pgoff_t pgoff;
 880         int id, rc;
 881         pfn_t pfn;
 882
 883         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 884         if (rc)
 885                 return rc;
 886
 887         id = dax_read_lock();
 888         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
 889         if (rc < 0) {
 890                 dax_read_unlock(id);
 891                 return rc;
 892         }
 893         dax_read_unlock(id);
 894
 895         ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, 0);
 896         if (IS_ERR(ret))
 897                 return PTR_ERR(ret);
 898         *entryp = ret;
 899
 900         trace_dax_insert_mapping(mapping->host, vmf, ret);
 901         return vm_insert_mixed(vma, vaddr, pfn);
 902 }
 903
 904 /**
 905  * dax_pfn_mkwrite - handle first write to DAX page
 906  * @vmf: The description of the fault
 907  */
 908 int dax_pfn_mkwrite(struct vm_fault *vmf)
 909 {
 910         struct file *file = vmf->vma->vm_file;
 911         struct address_space *mapping = file->f_mapping;
 912         struct inode *inode = mapping->host;
 913         void *entry, **slot;
 914         pgoff_t index = vmf->pgoff;
 915
 916         spin_lock_irq(&mapping->tree_lock);
 917         entry = get_unlocked_mapping_entry(mapping, index, &slot);
 918         if (!entry || !radix_tree_exceptional_entry(entry)) {
 919                 if (entry)
 920                         put_unlocked_mapping_entry(mapping, index, entry);
 921                 spin_unlock_irq(&mapping->tree_lock);
 922                 trace_dax_pfn_mkwrite_no_entry(inode, vmf, VM_FAULT_NOPAGE);
 923                 return VM_FAULT_NOPAGE;
 924         }
 925         radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
 926         entry = lock_slot(mapping, slot);
 927         spin_unlock_irq(&mapping->tree_lock);
 928         /*
 929          * If we race with somebody updating the PTE and finish_mkwrite_fault()
 930          * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry
 931          * the fault in either case.
 932          */
 933         finish_mkwrite_fault(vmf);
 934         put_locked_mapping_entry(mapping, index, entry);
 935         trace_dax_pfn_mkwrite(inode, vmf, VM_FAULT_NOPAGE);
 936         return VM_FAULT_NOPAGE;
 937 }
 938 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
 939
 940 static bool dax_range_is_aligned(struct block_device *bdev,
 941                                  unsigned int offset, unsigned int length)
 942 {
 943         unsigned short sector_size = bdev_logical_block_size(bdev);
 944
 945         if (!IS_ALIGNED(offset, sector_size))
 946                 return false;
 947         if (!IS_ALIGNED(length, sector_size))
 948                 return false;
 949
 950         return true;
 951 }
 952
 953 int __dax_zero_page_range(struct block_device *bdev,
 954                 struct dax_device *dax_dev, sector_t sector,
 955                 unsigned int offset, unsigned int size)
 956 {
 957         if (dax_range_is_aligned(bdev, offset, size)) {
 958                 sector_t start_sector = sector + (offset >> 9);
 959
 960                 return blkdev_issue_zeroout(bdev, start_sector,
 961                                 size >> 9, GFP_NOFS, 0);
 962         } else {
 963                 pgoff_t pgoff;
 964                 long rc, id;
 965                 void *kaddr;
 966                 pfn_t pfn;
 967
 968                 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
 969                 if (rc)
 970                         return rc;
 971
 972                 id = dax_read_lock();
 973                 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr,
 974                                 &pfn);
 975                 if (rc < 0) {
 976                         dax_read_unlock(id);
 977                         return rc;
 978                 }
 979                 clear_pmem(kaddr + offset, size);
 980                 dax_read_unlock(id);
 981         }
 982         return 0;
 983 }
 984 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
 985
 986 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
 987 {
 988         return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
 989 }
 990
 991 static loff_t
 992 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 993                 struct iomap *iomap)
 994 {
 995         struct block_device *bdev = iomap->bdev;
 996         struct dax_device *dax_dev = iomap->dax_dev;
 997         struct iov_iter *iter = data;
 998         loff_t end = pos + length, done = 0;
 999         ssize_t ret = 0;
1000         int id;
1001
1002         if (iov_iter_rw(iter) == READ) {
1003                 end = min(end, i_size_read(inode));
1004                 if (pos >= end)
1005                         return 0;
1006
1007                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1008                         return iov_iter_zero(min(length, end - pos), iter);
1009         }
1010
1011         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1012                 return -EIO;
1013
1014         /*
1015          * Write can allocate block for an area which has a hole page mapped
1016          * into page tables. We have to tear down these mappings so that data
1017          * written by write(2) is visible in mmap.
1018          */
1019         if (iomap->flags & IOMAP_F_NEW) {
1020                 invalidate_inode_pages2_range(inode->i_mapping,
1021                                               pos >> PAGE_SHIFT,
1022                                               (end - 1) >> PAGE_SHIFT);
1023         }
1024
1025         id = dax_read_lock();
1026         while (pos < end) {
1027                 unsigned offset = pos & (PAGE_SIZE - 1);
1028                 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1029                 const sector_t sector = dax_iomap_sector(iomap, pos);
1030                 ssize_t map_len;
1031                 pgoff_t pgoff;
1032                 void *kaddr;
1033                 pfn_t pfn;
1034
1035                 if (fatal_signal_pending(current)) {
1036                         ret = -EINTR;
1037                         break;
1038                 }
1039
1040                 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1041                 if (ret)
1042                         break;
1043
1044                 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1045                                 &kaddr, &pfn);
1046                 if (map_len < 0) {
1047                         ret = map_len;
1048                         break;
1049                 }
1050
1051                 map_len = PFN_PHYS(map_len);
1052                 kaddr += offset;
1053                 map_len -= offset;
1054                 if (map_len > end - pos)
1055                         map_len = end - pos;
1056
1057                 if (iov_iter_rw(iter) == WRITE)
1058                         map_len = copy_from_iter_pmem(kaddr, map_len, iter);
1059                 else
1060                         map_len = copy_to_iter(kaddr, map_len, iter);
1061                 if (map_len <= 0) {
1062                         ret = map_len ? map_len : -EFAULT;
1063                         break;
1064                 }
1065
1066                 pos += map_len;
1067                 length -= map_len;
1068                 done += map_len;
1069         }
1070         dax_read_unlock(id);
1071
1072         return done ? done : ret;
1073 }
1074
1075 /**
1076  * dax_iomap_rw - Perform I/O to a DAX file
1077  * @iocb:       The control block for this I/O
1078  * @iter:       The addresses to do I/O from or to
1079  * @ops:        iomap ops passed from the file system
1080  *
1081  * This function performs read and write operations to directly mapped
1082  * persistent memory.  The callers needs to take care of read/write exclusion
1083  * and evicting any page cache pages in the region under I/O.
1084  */
1085 ssize_t
1086 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1087                 const struct iomap_ops *ops)
1088 {
1089         struct address_space *mapping = iocb->ki_filp->f_mapping;
1090         struct inode *inode = mapping->host;
1091         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1092         unsigned flags = 0;
1093
1094         if (iov_iter_rw(iter) == WRITE) {
1095                 lockdep_assert_held_exclusive(&inode->i_rwsem);
1096                 flags |= IOMAP_WRITE;
1097         } else {
1098                 lockdep_assert_held(&inode->i_rwsem);
1099         }
1100
1101         while (iov_iter_count(iter)) {
1102                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1103                                 iter, dax_iomap_actor);
1104                 if (ret <= 0)
1105                         break;
1106                 pos += ret;
1107                 done += ret;
1108         }
1109
1110         iocb->ki_pos += done;
1111         return done ? done : ret;
1112 }
1113 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1114
1115 static int dax_fault_return(int error)
1116 {
1117         if (error == 0)
1118                 return VM_FAULT_NOPAGE;
1119         if (error == -ENOMEM)
1120                 return VM_FAULT_OOM;
1121         return VM_FAULT_SIGBUS;
1122 }
1123
1124 static int dax_iomap_pte_fault(struct vm_fault *vmf,
1125                                const struct iomap_ops *ops)
1126 {
1127         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1128         struct inode *inode = mapping->host;
1129         unsigned long vaddr = vmf->address;
1130         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1131         sector_t sector;
1132         struct iomap iomap = { 0 };
1133         unsigned flags = IOMAP_FAULT;
1134         int error, major = 0;
1135         int vmf_ret = 0;
1136         void *entry;
1137
1138         trace_dax_pte_fault(inode, vmf, vmf_ret);
1139         /*
1140          * Check whether offset isn't beyond end of file now. Caller is supposed
1141          * to hold locks serializing us with truncate / punch hole so this is
1142          * a reliable test.
1143          */
1144         if (pos >= i_size_read(inode)) {
1145                 vmf_ret = VM_FAULT_SIGBUS;
1146                 goto out;
1147         }
1148
1149         if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1150                 flags |= IOMAP_WRITE;
1151
1152         entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1153         if (IS_ERR(entry)) {
1154                 vmf_ret = dax_fault_return(PTR_ERR(entry));
1155                 goto out;
1156         }
1157
1158         /*
1159          * It is possible, particularly with mixed reads & writes to private
1160          * mappings, that we have raced with a PMD fault that overlaps with
1161          * the PTE we need to set up.  If so just return and the fault will be
1162          * retried.
1163          */
1164         if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1165                 vmf_ret = VM_FAULT_NOPAGE;
1166                 goto unlock_entry;
1167         }
1168
1169         /*
1170          * Note that we don't bother to use iomap_apply here: DAX required
1171          * the file system block size to be equal the page size, which means
1172          * that we never have to deal with more than a single extent here.
1173          */
1174         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1175         if (error) {
1176                 vmf_ret = dax_fault_return(error);
1177                 goto unlock_entry;
1178         }
1179         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1180                 error = -EIO;   /* fs corruption? */
1181                 goto error_finish_iomap;
1182         }
1183
1184         sector = dax_iomap_sector(&iomap, pos);
1185
1186         if (vmf->cow_page) {
1187                 switch (iomap.type) {
1188                 case IOMAP_HOLE:
1189                 case IOMAP_UNWRITTEN:
1190                         clear_user_highpage(vmf->cow_page, vaddr);
1191                         break;
1192                 case IOMAP_MAPPED:
1193                         error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1194                                         sector, PAGE_SIZE, vmf->cow_page, vaddr);
1195                         break;
1196                 default:
1197                         WARN_ON_ONCE(1);
1198                         error = -EIO;
1199                         break;
1200                 }
1201
1202                 if (error)
1203                         goto error_finish_iomap;
1204
1205                 __SetPageUptodate(vmf->cow_page);
1206                 vmf_ret = finish_fault(vmf);
1207                 if (!vmf_ret)
1208                         vmf_ret = VM_FAULT_DONE_COW;
1209                 goto finish_iomap;
1210         }
1211
1212         switch (iomap.type) {
1213         case IOMAP_MAPPED:
1214                 if (iomap.flags & IOMAP_F_NEW) {
1215                         count_vm_event(PGMAJFAULT);
1216                         mem_cgroup_count_vm_event(vmf->vma->vm_mm, PGMAJFAULT);
1217                         major = VM_FAULT_MAJOR;
1218                 }
1219                 error = dax_insert_mapping(mapping, iomap.bdev, iomap.dax_dev,
1220                                 sector, PAGE_SIZE, &entry, vmf->vma, vmf);
1221                 /* -EBUSY is fine, somebody else faulted on the same PTE */
1222                 if (error == -EBUSY)
1223                         error = 0;
1224                 break;
1225         case IOMAP_UNWRITTEN:
1226         case IOMAP_HOLE:
1227                 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
1228                         vmf_ret = dax_load_hole(mapping, &entry, vmf);
1229                         goto finish_iomap;
1230                 }
1231                 /*FALLTHRU*/
1232         default:
1233                 WARN_ON_ONCE(1);
1234                 error = -EIO;
1235                 break;
1236         }
1237
1238  error_finish_iomap:
1239         vmf_ret = dax_fault_return(error) | major;
1240  finish_iomap:
1241         if (ops->iomap_end) {
1242                 int copied = PAGE_SIZE;
1243
1244                 if (vmf_ret & VM_FAULT_ERROR)
1245                         copied = 0;
1246                 /*
1247                  * The fault is done by now and there's no way back (other
1248                  * thread may be already happily using PTE we have installed).
1249                  * Just ignore error from ->iomap_end since we cannot do much
1250                  * with it.
1251                  */
1252                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1253         }
1254  unlock_entry:
1255         put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1256  out:
1257         trace_dax_pte_fault_done(inode, vmf, vmf_ret);
1258         return vmf_ret;
1259 }
1260
1261 #ifdef CONFIG_FS_DAX_PMD
1262 /*
1263  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
1264  * more often than one might expect in the below functions.
1265  */
1266 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
1267
1268 static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap,
1269                 loff_t pos, void **entryp)
1270 {
1271         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1272         const sector_t sector = dax_iomap_sector(iomap, pos);
1273         struct dax_device *dax_dev = iomap->dax_dev;
1274         struct block_device *bdev = iomap->bdev;
1275         struct inode *inode = mapping->host;
1276         const size_t size = PMD_SIZE;
1277         void *ret = NULL, *kaddr;
1278         long length = 0;
1279         pgoff_t pgoff;
1280         pfn_t pfn;
1281         int id;
1282
1283         if (bdev_dax_pgoff(bdev, sector, size, &pgoff) != 0)
1284                 goto fallback;
1285
1286         id = dax_read_lock();
1287         length = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
1288         if (length < 0)
1289                 goto unlock_fallback;
1290         length = PFN_PHYS(length);
1291
1292         if (length < size)
1293                 goto unlock_fallback;
1294         if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR)
1295                 goto unlock_fallback;
1296         if (!pfn_t_devmap(pfn))
1297                 goto unlock_fallback;
1298         dax_read_unlock(id);
1299
1300         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, sector,
1301                         RADIX_DAX_PMD);
1302         if (IS_ERR(ret))
1303                 goto fallback;
1304         *entryp = ret;
1305
1306         trace_dax_pmd_insert_mapping(inode, vmf, length, pfn, ret);
1307         return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
1308                         pfn, vmf->flags & FAULT_FLAG_WRITE);
1309
1310 unlock_fallback:
1311         dax_read_unlock(id);
1312 fallback:
1313         trace_dax_pmd_insert_mapping_fallback(inode, vmf, length, pfn, ret);
1314         return VM_FAULT_FALLBACK;
1315 }
1316
1317 static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
1318                 void **entryp)
1319 {
1320         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1321         unsigned long pmd_addr = vmf->address & PMD_MASK;
1322         struct inode *inode = mapping->host;
1323         struct page *zero_page;
1324         void *ret = NULL;
1325         spinlock_t *ptl;
1326         pmd_t pmd_entry;
1327
1328         zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1329
1330         if (unlikely(!zero_page))
1331                 goto fallback;
1332
1333         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0,
1334                         RADIX_DAX_PMD | RADIX_DAX_HZP);
1335         if (IS_ERR(ret))
1336                 goto fallback;
1337         *entryp = ret;
1338
1339         ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1340         if (!pmd_none(*(vmf->pmd))) {
1341                 spin_unlock(ptl);
1342                 goto fallback;
1343         }
1344
1345         pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1346         pmd_entry = pmd_mkhuge(pmd_entry);
1347         set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1348         spin_unlock(ptl);
1349         trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
1350         return VM_FAULT_NOPAGE;
1351
1352 fallback:
1353         trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
1354         return VM_FAULT_FALLBACK;
1355 }
1356
1357 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1358                                const struct iomap_ops *ops)
1359 {
1360         struct vm_area_struct *vma = vmf->vma;
1361         struct address_space *mapping = vma->vm_file->f_mapping;
1362         unsigned long pmd_addr = vmf->address & PMD_MASK;
1363         bool write = vmf->flags & FAULT_FLAG_WRITE;
1364         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1365         struct inode *inode = mapping->host;
1366         int result = VM_FAULT_FALLBACK;
1367         struct iomap iomap = { 0 };
1368         pgoff_t max_pgoff, pgoff;
1369         void *entry;
1370         loff_t pos;
1371         int error;
1372
1373         /*
1374          * Check whether offset isn't beyond end of file now. Caller is
1375          * supposed to hold locks serializing us with truncate / punch hole so
1376          * this is a reliable test.
1377          */
1378         pgoff = linear_page_index(vma, pmd_addr);
1379         max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1380
1381         trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1382
1383         /* Fall back to PTEs if we're going to COW */
1384         if (write && !(vma->vm_flags & VM_SHARED))
1385                 goto fallback;
1386
1387         /* If the PMD would extend outside the VMA */
1388         if (pmd_addr < vma->vm_start)
1389                 goto fallback;
1390         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1391                 goto fallback;
1392
1393         if (pgoff > max_pgoff) {
1394                 result = VM_FAULT_SIGBUS;
1395                 goto out;
1396         }
1397
1398         /* If the PMD would extend beyond the file size */
1399         if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1400                 goto fallback;
1401
1402         /*
1403          * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1404          * PMD or a HZP entry.  If it can't (because a 4k page is already in
1405          * the tree, for instance), it will return -EEXIST and we just fall
1406          * back to 4k entries.
1407          */
1408         entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1409         if (IS_ERR(entry))
1410                 goto fallback;
1411
1412         /*
1413          * It is possible, particularly with mixed reads & writes to private
1414          * mappings, that we have raced with a PTE fault that overlaps with
1415          * the PMD we need to set up.  If so just return and the fault will be
1416          * retried.
1417          */
1418         if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1419                         !pmd_devmap(*vmf->pmd)) {
1420                 result = 0;
1421                 goto unlock_entry;
1422         }
1423
1424         /*
1425          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1426          * setting up a mapping, so really we're using iomap_begin() as a way
1427          * to look up our filesystem block.
1428          */
1429         pos = (loff_t)pgoff << PAGE_SHIFT;
1430         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1431         if (error)
1432                 goto unlock_entry;
1433
1434         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1435                 goto finish_iomap;
1436
1437         switch (iomap.type) {
1438         case IOMAP_MAPPED:
1439                 result = dax_pmd_insert_mapping(vmf, &iomap, pos, &entry);
1440                 break;
1441         case IOMAP_UNWRITTEN:
1442         case IOMAP_HOLE:
1443                 if (WARN_ON_ONCE(write))
1444                         break;
1445                 result = dax_pmd_load_hole(vmf, &iomap, &entry);
1446                 break;
1447         default:
1448                 WARN_ON_ONCE(1);
1449                 break;
1450         }
1451
1452  finish_iomap:
1453         if (ops->iomap_end) {
1454                 int copied = PMD_SIZE;
1455
1456                 if (result == VM_FAULT_FALLBACK)
1457                         copied = 0;
1458                 /*
1459                  * The fault is done by now and there's no way back (other
1460                  * thread may be already happily using PMD we have installed).
1461                  * Just ignore error from ->iomap_end since we cannot do much
1462                  * with it.
1463                  */
1464                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1465                                 &iomap);
1466         }
1467  unlock_entry:
1468         put_locked_mapping_entry(mapping, pgoff, entry);
1469  fallback:
1470         if (result == VM_FAULT_FALLBACK) {
1471                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1472                 count_vm_event(THP_FAULT_FALLBACK);
1473         }
1474 out:
1475         trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1476         return result;
1477 }
1478 #else
1479 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1480                                const struct iomap_ops *ops)
1481 {
1482         return VM_FAULT_FALLBACK;
1483 }
1484 #endif /* CONFIG_FS_DAX_PMD */
1485
1486 /**
1487  * dax_iomap_fault - handle a page fault on a DAX file
1488  * @vmf: The description of the fault
1489  * @ops: iomap ops passed from the file system
1490  *
1491  * When a page fault occurs, filesystems may call this helper in
1492  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1493  * has done all the necessary locking for page fault to proceed
1494  * successfully.
1495  */
1496 int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1497                     const struct iomap_ops *ops)
1498 {
1499         switch (pe_size) {
1500         case PE_SIZE_PTE:
1501                 return dax_iomap_pte_fault(vmf, ops);
1502         case PE_SIZE_PMD:
1503                 return dax_iomap_pmd_fault(vmf, ops);
1504         default:
1505                 return VM_FAULT_FALLBACK;
1506         }
1507 }
1508 EXPORT_SYMBOL_GPL(dax_iomap_fault);