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/sched.h>
  29 #include <linux/sched/signal.h>
  30 #include <linux/uio.h>
  31 #include <linux/vmstat.h>
  32 #include <linux/pfn_t.h>
  33 #include <linux/sizes.h>
  34 #include <linux/mmu_notifier.h>
  35 #include <linux/iomap.h>
  36 #include "internal.h"
  37
  38 #define CREATE_TRACE_POINTS
  39 #include <trace/events/fs_dax.h>
  40
  41 /* We choose 4096 entries - same as per-zone page wait tables */
  42 #define DAX_WAIT_TABLE_BITS 12
  43 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
  44
  45 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
  46
  47 static int __init init_dax_wait_table(void)
  48 {
  49         int i;
  50
  51         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
  52                 init_waitqueue_head(wait_table + i);
  53         return 0;
  54 }
  55 fs_initcall(init_dax_wait_table);
  56
  57 static int dax_is_pmd_entry(void *entry)
  58 {
  59         return (unsigned long)entry & RADIX_DAX_PMD;
  60 }
  61
  62 static int dax_is_pte_entry(void *entry)
  63 {
  64         return !((unsigned long)entry & RADIX_DAX_PMD);
  65 }
  66
  67 static int dax_is_zero_entry(void *entry)
  68 {
  69         return (unsigned long)entry & RADIX_DAX_HZP;
  70 }
  71
  72 static int dax_is_empty_entry(void *entry)
  73 {
  74         return (unsigned long)entry & RADIX_DAX_EMPTY;
  75 }
  76
  77 /*
  78  * DAX radix tree locking
  79  */
  80 struct exceptional_entry_key {
  81         struct address_space *mapping;
  82         pgoff_t entry_start;
  83 };
  84
  85 struct wait_exceptional_entry_queue {
  86         wait_queue_entry_t wait;
  87         struct exceptional_entry_key key;
  88 };
  89
  90 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
  91                 pgoff_t index, void *entry, struct exceptional_entry_key *key)
  92 {
  93         unsigned long hash;
  94
  95         /*
  96          * If 'entry' is a PMD, align the 'index' that we use for the wait
  97          * queue to the start of that PMD.  This ensures that all offsets in
  98          * the range covered by the PMD map to the same bit lock.
  99          */
 100         if (dax_is_pmd_entry(entry))
 101                 index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
 102
 103         key->mapping = mapping;
 104         key->entry_start = index;
 105
 106         hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
 107         return wait_table + hash;
 108 }
 109
 110 static int wake_exceptional_entry_func(wait_queue_entry_t *wait, unsigned int mode,
 111                                        int sync, void *keyp)
 112 {
 113         struct exceptional_entry_key *key = keyp;
 114         struct wait_exceptional_entry_queue *ewait =
 115                 container_of(wait, struct wait_exceptional_entry_queue, wait);
 116
 117         if (key->mapping != ewait->key.mapping ||
 118             key->entry_start != ewait->key.entry_start)
 119                 return 0;
 120         return autoremove_wake_function(wait, mode, sync, NULL);
 121 }
 122
 123 /*
 124  * Check whether the given slot is locked. The function must be called with
 125  * mapping->tree_lock held
 126  */
 127 static inline int slot_locked(struct address_space *mapping, void **slot)
 128 {
 129         unsigned long entry = (unsigned long)
 130                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 131         return entry & RADIX_DAX_ENTRY_LOCK;
 132 }
 133
 134 /*
 135  * Mark the given slot is locked. The function must be called with
 136  * mapping->tree_lock held
 137  */
 138 static inline void *lock_slot(struct address_space *mapping, void **slot)
 139 {
 140         unsigned long entry = (unsigned long)
 141                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 142
 143         entry |= RADIX_DAX_ENTRY_LOCK;
 144         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
 145         return (void *)entry;
 146 }
 147
 148 /*
 149  * Mark the given slot is unlocked. The function must be called with
 150  * mapping->tree_lock held
 151  */
 152 static inline void *unlock_slot(struct address_space *mapping, void **slot)
 153 {
 154         unsigned long entry = (unsigned long)
 155                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 156
 157         entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
 158         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
 159         return (void *)entry;
 160 }
 161
 162 /*
 163  * Lookup entry in radix tree, wait for it to become unlocked if it is
 164  * exceptional entry and return it. The caller must call
 165  * put_unlocked_mapping_entry() when he decided not to lock the entry or
 166  * put_locked_mapping_entry() when he locked the entry and now wants to
 167  * unlock it.
 168  *
 169  * The function must be called with mapping->tree_lock held.
 170  */
 171 static void *get_unlocked_mapping_entry(struct address_space *mapping,
 172                                         pgoff_t index, void ***slotp)
 173 {
 174         void *entry, **slot;
 175         struct wait_exceptional_entry_queue ewait;
 176         wait_queue_head_t *wq;
 177
 178         init_wait(&ewait.wait);
 179         ewait.wait.func = wake_exceptional_entry_func;
 180
 181         for (;;) {
 182                 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
 183                                           &slot);
 184                 if (!entry || !radix_tree_exceptional_entry(entry) ||
 185                     !slot_locked(mapping, slot)) {
 186                         if (slotp)
 187                                 *slotp = slot;
 188                         return entry;
 189                 }
 190
 191                 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
 192                 prepare_to_wait_exclusive(wq, &ewait.wait,
 193                                           TASK_UNINTERRUPTIBLE);
 194                 spin_unlock_irq(&mapping->tree_lock);
 195                 schedule();
 196                 finish_wait(wq, &ewait.wait);
 197                 spin_lock_irq(&mapping->tree_lock);
 198         }
 199 }
 200
 201 static void dax_unlock_mapping_entry(struct address_space *mapping,
 202                                      pgoff_t index)
 203 {
 204         void *entry, **slot;
 205
 206         spin_lock_irq(&mapping->tree_lock);
 207         entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
 208         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
 209                          !slot_locked(mapping, slot))) {
 210                 spin_unlock_irq(&mapping->tree_lock);
 211                 return;
 212         }
 213         unlock_slot(mapping, slot);
 214         spin_unlock_irq(&mapping->tree_lock);
 215         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 216 }
 217
 218 static void put_locked_mapping_entry(struct address_space *mapping,
 219                                      pgoff_t index, void *entry)
 220 {
 221         if (!radix_tree_exceptional_entry(entry)) {
 222                 unlock_page(entry);
 223                 put_page(entry);
 224         } else {
 225                 dax_unlock_mapping_entry(mapping, index);
 226         }
 227 }
 228
 229 /*
 230  * Called when we are done with radix tree entry we looked up via
 231  * get_unlocked_mapping_entry() and which we didn't lock in the end.
 232  */
 233 static void put_unlocked_mapping_entry(struct address_space *mapping,
 234                                        pgoff_t index, void *entry)
 235 {
 236         if (!radix_tree_exceptional_entry(entry))
 237                 return;
 238
 239         /* We have to wake up next waiter for the radix tree entry lock */
 240         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 241 }
 242
 243 /*
 244  * Find radix tree entry at given index. If it points to a page, return with
 245  * the page locked. If it points to the exceptional entry, return with the
 246  * radix tree entry locked. If the radix tree doesn't contain given index,
 247  * create empty exceptional entry for the index and return with it locked.
 248  *
 249  * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
 250  * either return that locked entry or will return an error.  This error will
 251  * happen if there are any 4k entries (either zero pages or DAX entries)
 252  * within the 2MiB range that we are requesting.
 253  *
 254  * We always favor 4k entries over 2MiB entries. There isn't a flow where we
 255  * evict 4k entries in order to 'upgrade' them to a 2MiB entry.  A 2MiB
 256  * insertion will fail if it finds any 4k entries already in the tree, and a
 257  * 4k insertion will cause an existing 2MiB entry to be unmapped and
 258  * downgraded to 4k entries.  This happens for both 2MiB huge zero pages as
 259  * well as 2MiB empty entries.
 260  *
 261  * The exception to this downgrade path is for 2MiB DAX PMD entries that have
 262  * real storage backing them.  We will leave these real 2MiB DAX entries in
 263  * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
 264  *
 265  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
 266  * persistent memory the benefit is doubtful. We can add that later if we can
 267  * show it helps.
 268  */
 269 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
 270                 unsigned long size_flag)
 271 {
 272         bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
 273         void *entry, **slot;
 274
 275 restart:
 276         spin_lock_irq(&mapping->tree_lock);
 277         entry = get_unlocked_mapping_entry(mapping, index, &slot);
 278
 279         if (entry) {
 280                 if (size_flag & RADIX_DAX_PMD) {
 281                         if (!radix_tree_exceptional_entry(entry) ||
 282                             dax_is_pte_entry(entry)) {
 283                                 put_unlocked_mapping_entry(mapping, index,
 284                                                 entry);
 285                                 entry = ERR_PTR(-EEXIST);
 286                                 goto out_unlock;
 287                         }
 288                 } else { /* trying to grab a PTE entry */
 289                         if (radix_tree_exceptional_entry(entry) &&
 290                             dax_is_pmd_entry(entry) &&
 291                             (dax_is_zero_entry(entry) ||
 292                              dax_is_empty_entry(entry))) {
 293                                 pmd_downgrade = true;
 294                         }
 295                 }
 296         }
 297
 298         /* No entry for given index? Make sure radix tree is big enough. */
 299         if (!entry || pmd_downgrade) {
 300                 int err;
 301
 302                 if (pmd_downgrade) {
 303                         /*
 304                          * Make sure 'entry' remains valid while we drop
 305                          * mapping->tree_lock.
 306                          */
 307                         entry = lock_slot(mapping, slot);
 308                 }
 309
 310                 spin_unlock_irq(&mapping->tree_lock);
 311                 /*
 312                  * Besides huge zero pages the only other thing that gets
 313                  * downgraded are empty entries which don't need to be
 314                  * unmapped.
 315                  */
 316                 if (pmd_downgrade && dax_is_zero_entry(entry))
 317                         unmap_mapping_range(mapping,
 318                                 (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
 319
 320                 err = radix_tree_preload(
 321                                 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
 322                 if (err) {
 323                         if (pmd_downgrade)
 324                                 put_locked_mapping_entry(mapping, index, entry);
 325                         return ERR_PTR(err);
 326                 }
 327                 spin_lock_irq(&mapping->tree_lock);
 328
 329                 if (!entry) {
 330                         /*
 331                          * We needed to drop the page_tree lock while calling
 332                          * radix_tree_preload() and we didn't have an entry to
 333                          * lock.  See if another thread inserted an entry at
 334                          * our index during this time.
 335                          */
 336                         entry = __radix_tree_lookup(&mapping->page_tree, index,
 337                                         NULL, &slot);
 338                         if (entry) {
 339                                 radix_tree_preload_end();
 340                                 spin_unlock_irq(&mapping->tree_lock);
 341                                 goto restart;
 342                         }
 343                 }
 344
 345                 if (pmd_downgrade) {
 346                         radix_tree_delete(&mapping->page_tree, index);
 347                         mapping->nrexceptional--;
 348                         dax_wake_mapping_entry_waiter(mapping, index, entry,
 349                                         true);
 350                 }
 351
 352                 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
 353
 354                 err = __radix_tree_insert(&mapping->page_tree, index,
 355                                 dax_radix_order(entry), entry);
 356                 radix_tree_preload_end();
 357                 if (err) {
 358                         spin_unlock_irq(&mapping->tree_lock);
 359                         /*
 360                          * Our insertion of a DAX entry failed, most likely
 361                          * because we were inserting a PMD entry and it
 362                          * collided with a PTE sized entry at a different
 363                          * index in the PMD range.  We haven't inserted
 364                          * anything into the radix tree and have no waiters to
 365                          * wake.
 366                          */
 367                         return ERR_PTR(err);
 368                 }
 369                 /* Good, we have inserted empty locked entry into the tree. */
 370                 mapping->nrexceptional++;
 371                 spin_unlock_irq(&mapping->tree_lock);
 372                 return entry;
 373         }
 374         /* Normal page in radix tree? */
 375         if (!radix_tree_exceptional_entry(entry)) {
 376                 struct page *page = entry;
 377
 378                 get_page(page);
 379                 spin_unlock_irq(&mapping->tree_lock);
 380                 lock_page(page);
 381                 /* Page got truncated? Retry... */
 382                 if (unlikely(page->mapping != mapping)) {
 383                         unlock_page(page);
 384                         put_page(page);
 385                         goto restart;
 386                 }
 387                 return page;
 388         }
 389         entry = lock_slot(mapping, slot);
 390  out_unlock:
 391         spin_unlock_irq(&mapping->tree_lock);
 392         return entry;
 393 }
 394
 395 /*
 396  * We do not necessarily hold the mapping->tree_lock when we call this
 397  * function so it is possible that 'entry' is no longer a valid item in the
 398  * radix tree.  This is okay because all we really need to do is to find the
 399  * correct waitqueue where tasks might be waiting for that old 'entry' and
 400  * wake them.
 401  */
 402 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
 403                 pgoff_t index, void *entry, bool wake_all)
 404 {
 405         struct exceptional_entry_key key;
 406         wait_queue_head_t *wq;
 407
 408         wq = dax_entry_waitqueue(mapping, index, entry, &key);
 409
 410         /*
 411          * Checking for locked entry and prepare_to_wait_exclusive() happens
 412          * under mapping->tree_lock, ditto for entry handling in our callers.
 413          * So at this point all tasks that could have seen our entry locked
 414          * must be in the waitqueue and the following check will see them.
 415          */
 416         if (waitqueue_active(wq))
 417                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
 418 }
 419
 420 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
 421                                           pgoff_t index, bool trunc)
 422 {
 423         int ret = 0;
 424         void *entry;
 425         struct radix_tree_root *page_tree = &mapping->page_tree;
 426
 427         spin_lock_irq(&mapping->tree_lock);
 428         entry = get_unlocked_mapping_entry(mapping, index, NULL);
 429         if (!entry || !radix_tree_exceptional_entry(entry))
 430                 goto out;
 431         if (!trunc &&
 432             (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
 433              radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)))
 434                 goto out;
 435         radix_tree_delete(page_tree, index);
 436         mapping->nrexceptional--;
 437         ret = 1;
 438 out:
 439         put_unlocked_mapping_entry(mapping, index, entry);
 440         spin_unlock_irq(&mapping->tree_lock);
 441         return ret;
 442 }
 443 /*
 444  * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
 445  * entry to get unlocked before deleting it.
 446  */
 447 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 448 {
 449         int ret = __dax_invalidate_mapping_entry(mapping, index, true);
 450
 451         /*
 452          * This gets called from truncate / punch_hole path. As such, the caller
 453          * must hold locks protecting against concurrent modifications of the
 454          * radix tree (usually fs-private i_mmap_sem for writing). Since the
 455          * caller has seen exceptional entry for this index, we better find it
 456          * at that index as well...
 457          */
 458         WARN_ON_ONCE(!ret);
 459         return ret;
 460 }
 461
 462 /*
 463  * Invalidate exceptional DAX entry if it is clean.
 464  */
 465 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
 466                                       pgoff_t index)
 467 {
 468         return __dax_invalidate_mapping_entry(mapping, index, false);
 469 }
 470
 471 /*
 472  * The user has performed a load from a hole in the file.  Allocating
 473  * a new page in the file would cause excessive storage usage for
 474  * workloads with sparse files.  We allocate a page cache page instead.
 475  * We'll kick it out of the page cache if it's ever written to,
 476  * otherwise it will simply fall out of the page cache under memory
 477  * pressure without ever having been dirtied.
 478  */
 479 static int dax_load_hole(struct address_space *mapping, void **entry,
 480                          struct vm_fault *vmf)
 481 {
 482         struct inode *inode = mapping->host;
 483         struct page *page;
 484         int ret;
 485
 486         /* Hole page already exists? Return it...  */
 487         if (!radix_tree_exceptional_entry(*entry)) {
 488                 page = *entry;
 489                 goto finish_fault;
 490         }
 491
 492         /* This will replace locked radix tree entry with a hole page */
 493         page = find_or_create_page(mapping, vmf->pgoff,
 494                                    vmf->gfp_mask | __GFP_ZERO);
 495         if (!page) {
 496                 ret = VM_FAULT_OOM;
 497                 goto out;
 498         }
 499
 500 finish_fault:
 501         vmf->page = page;
 502         ret = finish_fault(vmf);
 503         vmf->page = NULL;
 504         *entry = page;
 505         if (!ret) {
 506                 /* Grab reference for PTE that is now referencing the page */
 507                 get_page(page);
 508                 ret = VM_FAULT_NOPAGE;
 509         }
 510 out:
 511         trace_dax_load_hole(inode, vmf, ret);
 512         return ret;
 513 }
 514
 515 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
 516                 sector_t sector, size_t size, struct page *to,
 517                 unsigned long vaddr)
 518 {
 519         void *vto, *kaddr;
 520         pgoff_t pgoff;
 521         pfn_t pfn;
 522         long rc;
 523         int id;
 524
 525         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 526         if (rc)
 527                 return rc;
 528
 529         id = dax_read_lock();
 530         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
 531         if (rc < 0) {
 532                 dax_read_unlock(id);
 533                 return rc;
 534         }
 535         vto = kmap_atomic(to);
 536         copy_user_page(vto, (void __force *)kaddr, vaddr, to);
 537         kunmap_atomic(vto);
 538         dax_read_unlock(id);
 539         return 0;
 540 }
 541
 542 /*
 543  * By this point grab_mapping_entry() has ensured that we have a locked entry
 544  * of the appropriate size so we don't have to worry about downgrading PMDs to
 545  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
 546  * already in the tree, we will skip the insertion and just dirty the PMD as
 547  * appropriate.
 548  */
 549 static void *dax_insert_mapping_entry(struct address_space *mapping,
 550                                       struct vm_fault *vmf,
 551                                       void *entry, sector_t sector,
 552                                       unsigned long flags)
 553 {
 554         struct radix_tree_root *page_tree = &mapping->page_tree;
 555         int error = 0;
 556         bool hole_fill = false;
 557         void *new_entry;
 558         pgoff_t index = vmf->pgoff;
 559
 560         if (vmf->flags & FAULT_FLAG_WRITE)
 561                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 562
 563         /* Replacing hole page with block mapping? */
 564         if (!radix_tree_exceptional_entry(entry)) {
 565                 hole_fill = true;
 566                 /*
 567                  * Unmap the page now before we remove it from page cache below.
 568                  * The page is locked so it cannot be faulted in again.
 569                  */
 570                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 571                                     PAGE_SIZE, 0);
 572                 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
 573                 if (error)
 574                         return ERR_PTR(error);
 575         } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) {
 576                 /* replacing huge zero page with PMD block mapping */
 577                 unmap_mapping_range(mapping,
 578                         (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
 579         }
 580
 581         spin_lock_irq(&mapping->tree_lock);
 582         new_entry = dax_radix_locked_entry(sector, flags);
 583
 584         if (hole_fill) {
 585                 __delete_from_page_cache(entry, NULL);
 586                 /* Drop pagecache reference */
 587                 put_page(entry);
 588                 error = __radix_tree_insert(page_tree, index,
 589                                 dax_radix_order(new_entry), new_entry);
 590                 if (error) {
 591                         new_entry = ERR_PTR(error);
 592                         goto unlock;
 593                 }
 594                 mapping->nrexceptional++;
 595         } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 596                 /*
 597                  * Only swap our new entry into the radix tree if the current
 598                  * entry is a zero page or an empty entry.  If a normal PTE or
 599                  * PMD entry is already in the tree, we leave it alone.  This
 600                  * means that if we are trying to insert a PTE and the
 601                  * existing entry is a PMD, we will just leave the PMD in the
 602                  * tree and dirty it if necessary.
 603                  */
 604                 struct radix_tree_node *node;
 605                 void **slot;
 606                 void *ret;
 607
 608                 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
 609                 WARN_ON_ONCE(ret != entry);
 610                 __radix_tree_replace(page_tree, node, slot,
 611                                      new_entry, NULL, NULL);
 612         }
 613         if (vmf->flags & FAULT_FLAG_WRITE)
 614                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 615  unlock:
 616         spin_unlock_irq(&mapping->tree_lock);
 617         if (hole_fill) {
 618                 radix_tree_preload_end();
 619                 /*
 620                  * We don't need hole page anymore, it has been replaced with
 621                  * locked radix tree entry now.
 622                  */
 623                 if (mapping->a_ops->freepage)
 624                         mapping->a_ops->freepage(entry);
 625                 unlock_page(entry);
 626                 put_page(entry);
 627         }
 628         return new_entry;
 629 }
 630
 631 static inline unsigned long
 632 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
 633 {
 634         unsigned long address;
 635
 636         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 637         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
 638         return address;
 639 }
 640
 641 /* Walk all mappings of a given index of a file and writeprotect them */
 642 static void dax_mapping_entry_mkclean(struct address_space *mapping,
 643                                       pgoff_t index, unsigned long pfn)
 644 {
 645         struct vm_area_struct *vma;
 646         pte_t pte, *ptep = NULL;
 647         pmd_t *pmdp = NULL;
 648         spinlock_t *ptl;
 649         bool changed;
 650
 651         i_mmap_lock_read(mapping);
 652         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
 653                 unsigned long address;
 654
 655                 cond_resched();
 656
 657                 if (!(vma->vm_flags & VM_SHARED))
 658                         continue;
 659
 660                 address = pgoff_address(index, vma);
 661                 changed = false;
 662                 if (follow_pte_pmd(vma->vm_mm, address, &ptep, &pmdp, &ptl))
 663                         continue;
 664
 665                 if (pmdp) {
 666 #ifdef CONFIG_FS_DAX_PMD
 667                         pmd_t pmd;
 668
 669                         if (pfn != pmd_pfn(*pmdp))
 670                                 goto unlock_pmd;
 671                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
 672                                 goto unlock_pmd;
 673
 674                         flush_cache_page(vma, address, pfn);
 675                         pmd = pmdp_huge_clear_flush(vma, address, pmdp);
 676                         pmd = pmd_wrprotect(pmd);
 677                         pmd = pmd_mkclean(pmd);
 678                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
 679                         changed = true;
 680 unlock_pmd:
 681                         spin_unlock(ptl);
 682 #endif
 683                 } else {
 684                         if (pfn != pte_pfn(*ptep))
 685                                 goto unlock_pte;
 686                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
 687                                 goto unlock_pte;
 688
 689                         flush_cache_page(vma, address, pfn);
 690                         pte = ptep_clear_flush(vma, address, ptep);
 691                         pte = pte_wrprotect(pte);
 692                         pte = pte_mkclean(pte);
 693                         set_pte_at(vma->vm_mm, address, ptep, pte);
 694                         changed = true;
 695 unlock_pte:
 696                         pte_unmap_unlock(ptep, ptl);
 697                 }
 698
 699                 if (changed)
 700                         mmu_notifier_invalidate_page(vma->vm_mm, address);
 701         }
 702         i_mmap_unlock_read(mapping);
 703 }
 704
 705 static int dax_writeback_one(struct block_device *bdev,
 706                 struct dax_device *dax_dev, struct address_space *mapping,
 707                 pgoff_t index, void *entry)
 708 {
 709         struct radix_tree_root *page_tree = &mapping->page_tree;
 710         void *entry2, **slot, *kaddr;
 711         long ret = 0, id;
 712         sector_t sector;
 713         pgoff_t pgoff;
 714         size_t size;
 715         pfn_t pfn;
 716
 717         /*
 718          * A page got tagged dirty in DAX mapping? Something is seriously
 719          * wrong.
 720          */
 721         if (WARN_ON(!radix_tree_exceptional_entry(entry)))
 722                 return -EIO;
 723
 724         spin_lock_irq(&mapping->tree_lock);
 725         entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
 726         /* Entry got punched out / reallocated? */
 727         if (!entry2 || !radix_tree_exceptional_entry(entry2))
 728                 goto put_unlocked;
 729         /*
 730          * Entry got reallocated elsewhere? No need to writeback. We have to
 731          * compare sectors as we must not bail out due to difference in lockbit
 732          * or entry type.
 733          */
 734         if (dax_radix_sector(entry2) != dax_radix_sector(entry))
 735                 goto put_unlocked;
 736         if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
 737                                 dax_is_zero_entry(entry))) {
 738                 ret = -EIO;
 739                 goto put_unlocked;
 740         }
 741
 742         /* Another fsync thread may have already written back this entry */
 743         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 744                 goto put_unlocked;
 745         /* Lock the entry to serialize with page faults */
 746         entry = lock_slot(mapping, slot);
 747         /*
 748          * We can clear the tag now but we have to be careful so that concurrent
 749          * dax_writeback_one() calls for the same index cannot finish before we
 750          * actually flush the caches. This is achieved as the calls will look
 751          * at the entry only under tree_lock and once they do that they will
 752          * see the entry locked and wait for it to unlock.
 753          */
 754         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 755         spin_unlock_irq(&mapping->tree_lock);
 756
 757         /*
 758          * Even if dax_writeback_mapping_range() was given a wbc->range_start
 759          * in the middle of a PMD, the 'index' we are given will be aligned to
 760          * the start index of the PMD, as will the sector we pull from
 761          * 'entry'.  This allows us to flush for PMD_SIZE and not have to
 762          * worry about partial PMD writebacks.
 763          */
 764         sector = dax_radix_sector(entry);
 765         size = PAGE_SIZE << dax_radix_order(entry);
 766
 767         id = dax_read_lock();
 768         ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 769         if (ret)
 770                 goto dax_unlock;
 771
 772         /*
 773          * dax_direct_access() may sleep, so cannot hold tree_lock over
 774          * its invocation.
 775          */
 776         ret = dax_direct_access(dax_dev, pgoff, size / PAGE_SIZE, &kaddr, &pfn);
 777         if (ret < 0)
 778                 goto dax_unlock;
 779
 780         if (WARN_ON_ONCE(ret < size / PAGE_SIZE)) {
 781                 ret = -EIO;
 782                 goto dax_unlock;
 783         }
 784
 785         dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(pfn));
 786         dax_flush(dax_dev, pgoff, kaddr, size);
 787         /*
 788          * After we have flushed the cache, we can clear the dirty tag. There
 789          * cannot be new dirty data in the pfn after the flush has completed as
 790          * the pfn mappings are writeprotected and fault waits for mapping
 791          * entry lock.
 792          */
 793         spin_lock_irq(&mapping->tree_lock);
 794         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
 795         spin_unlock_irq(&mapping->tree_lock);
 796         trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT);
 797  dax_unlock:
 798         dax_read_unlock(id);
 799         put_locked_mapping_entry(mapping, index, entry);
 800         return ret;
 801
 802  put_unlocked:
 803         put_unlocked_mapping_entry(mapping, index, entry2);
 804         spin_unlock_irq(&mapping->tree_lock);
 805         return ret;
 806 }
 807
 808 /*
 809  * Flush the mapping to the persistent domain within the byte range of [start,
 810  * end]. This is required by data integrity operations to ensure file data is
 811  * on persistent storage prior to completion of the operation.
 812  */
 813 int dax_writeback_mapping_range(struct address_space *mapping,
 814                 struct block_device *bdev, struct writeback_control *wbc)
 815 {
 816         struct inode *inode = mapping->host;
 817         pgoff_t start_index, end_index;
 818         pgoff_t indices[PAGEVEC_SIZE];
 819         struct dax_device *dax_dev;
 820         struct pagevec pvec;
 821         bool done = false;
 822         int i, ret = 0;
 823
 824         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 825                 return -EIO;
 826
 827         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 828                 return 0;
 829
 830         dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
 831         if (!dax_dev)
 832                 return -EIO;
 833
 834         start_index = wbc->range_start >> PAGE_SHIFT;
 835         end_index = wbc->range_end >> PAGE_SHIFT;
 836
 837         trace_dax_writeback_range(inode, start_index, end_index);
 838
 839         tag_pages_for_writeback(mapping, start_index, end_index);
 840
 841         pagevec_init(&pvec, 0);
 842         while (!done) {
 843                 pvec.nr = find_get_entries_tag(mapping, start_index,
 844                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 845                                 pvec.pages, indices);
 846
 847                 if (pvec.nr == 0)
 848                         break;
 849
 850                 for (i = 0; i < pvec.nr; i++) {
 851                         if (indices[i] > end_index) {
 852                                 done = true;
 853                                 break;
 854                         }
 855
 856                         ret = dax_writeback_one(bdev, dax_dev, mapping,
 857                                         indices[i], pvec.pages[i]);
 858                         if (ret < 0) {
 859                                 mapping_set_error(mapping, ret);
 860                                 goto out;
 861                         }
 862                 }
 863                 start_index = indices[pvec.nr - 1] + 1;
 864         }
 865 out:
 866         put_dax(dax_dev);
 867         trace_dax_writeback_range_done(inode, start_index, end_index);
 868         return (ret < 0 ? ret : 0);
 869 }
 870 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 871
 872 static int dax_insert_mapping(struct address_space *mapping,
 873                 struct block_device *bdev, struct dax_device *dax_dev,
 874                 sector_t sector, size_t size, void **entryp,
 875                 struct vm_area_struct *vma, struct vm_fault *vmf)
 876 {
 877         unsigned long vaddr = vmf->address;
 878         void *entry = *entryp;
 879         void *ret, *kaddr;
 880         pgoff_t pgoff;
 881         int id, rc;
 882         pfn_t pfn;
 883
 884         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
 885         if (rc)
 886                 return rc;
 887
 888         id = dax_read_lock();
 889         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
 890         if (rc < 0) {
 891                 dax_read_unlock(id);
 892                 return rc;
 893         }
 894         dax_read_unlock(id);
 895
 896         ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, 0);
 897         if (IS_ERR(ret))
 898                 return PTR_ERR(ret);
 899         *entryp = ret;
 900
 901         trace_dax_insert_mapping(mapping->host, vmf, ret);
 902         return vm_insert_mixed(vma, vaddr, pfn);
 903 }
 904
 905 /**
 906  * dax_pfn_mkwrite - handle first write to DAX page
 907  * @vmf: The description of the fault
 908  */
 909 int dax_pfn_mkwrite(struct vm_fault *vmf)
 910 {
 911         struct file *file = vmf->vma->vm_file;
 912         struct address_space *mapping = file->f_mapping;
 913         struct inode *inode = mapping->host;
 914         void *entry, **slot;
 915         pgoff_t index = vmf->pgoff;
 916
 917         spin_lock_irq(&mapping->tree_lock);
 918         entry = get_unlocked_mapping_entry(mapping, index, &slot);
 919         if (!entry || !radix_tree_exceptional_entry(entry)) {
 920                 if (entry)
 921                         put_unlocked_mapping_entry(mapping, index, entry);
 922                 spin_unlock_irq(&mapping->tree_lock);
 923                 trace_dax_pfn_mkwrite_no_entry(inode, vmf, VM_FAULT_NOPAGE);
 924                 return VM_FAULT_NOPAGE;
 925         }
 926         radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
 927         entry = lock_slot(mapping, slot);
 928         spin_unlock_irq(&mapping->tree_lock);
 929         /*
 930          * If we race with somebody updating the PTE and finish_mkwrite_fault()
 931          * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry
 932          * the fault in either case.
 933          */
 934         finish_mkwrite_fault(vmf);
 935         put_locked_mapping_entry(mapping, index, entry);
 936         trace_dax_pfn_mkwrite(inode, vmf, VM_FAULT_NOPAGE);
 937         return VM_FAULT_NOPAGE;
 938 }
 939 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
 940
 941 static bool dax_range_is_aligned(struct block_device *bdev,
 942                                  unsigned int offset, unsigned int length)
 943 {
 944         unsigned short sector_size = bdev_logical_block_size(bdev);
 945
 946         if (!IS_ALIGNED(offset, sector_size))
 947                 return false;
 948         if (!IS_ALIGNED(length, sector_size))
 949                 return false;
 950
 951         return true;
 952 }
 953
 954 int __dax_zero_page_range(struct block_device *bdev,
 955                 struct dax_device *dax_dev, sector_t sector,
 956                 unsigned int offset, unsigned int size)
 957 {
 958         if (dax_range_is_aligned(bdev, offset, size)) {
 959                 sector_t start_sector = sector + (offset >> 9);
 960
 961                 return blkdev_issue_zeroout(bdev, start_sector,
 962                                 size >> 9, GFP_NOFS, 0);
 963         } else {
 964                 pgoff_t pgoff;
 965                 long rc, id;
 966                 void *kaddr;
 967                 pfn_t pfn;
 968
 969                 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
 970                 if (rc)
 971                         return rc;
 972
 973                 id = dax_read_lock();
 974                 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr,
 975                                 &pfn);
 976                 if (rc < 0) {
 977                         dax_read_unlock(id);
 978                         return rc;
 979                 }
 980                 memset(kaddr + offset, 0, size);
 981                 dax_flush(dax_dev, pgoff, kaddr + offset, size);
 982                 dax_read_unlock(id);
 983         }
 984         return 0;
 985 }
 986 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
 987
 988 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
 989 {
 990         return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
 991 }
 992
 993 static loff_t
 994 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 995                 struct iomap *iomap)
 996 {
 997         struct block_device *bdev = iomap->bdev;
 998         struct dax_device *dax_dev = iomap->dax_dev;
 999         struct iov_iter *iter = data;
1000         loff_t end = pos + length, done = 0;
1001         ssize_t ret = 0;
1002         int id;
1003
1004         if (iov_iter_rw(iter) == READ) {
1005                 end = min(end, i_size_read(inode));
1006                 if (pos >= end)
1007                         return 0;
1008
1009                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1010                         return iov_iter_zero(min(length, end - pos), iter);
1011         }
1012
1013         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1014                 return -EIO;
1015
1016         /*
1017          * Write can allocate block for an area which has a hole page mapped
1018          * into page tables. We have to tear down these mappings so that data
1019          * written by write(2) is visible in mmap.
1020          */
1021         if (iomap->flags & IOMAP_F_NEW) {
1022                 invalidate_inode_pages2_range(inode->i_mapping,
1023                                               pos >> PAGE_SHIFT,
1024                                               (end - 1) >> PAGE_SHIFT);
1025         }
1026
1027         id = dax_read_lock();
1028         while (pos < end) {
1029                 unsigned offset = pos & (PAGE_SIZE - 1);
1030                 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1031                 const sector_t sector = dax_iomap_sector(iomap, pos);
1032                 ssize_t map_len;
1033                 pgoff_t pgoff;
1034                 void *kaddr;
1035                 pfn_t pfn;
1036
1037                 if (fatal_signal_pending(current)) {
1038                         ret = -EINTR;
1039                         break;
1040                 }
1041
1042                 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1043                 if (ret)
1044                         break;
1045
1046                 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1047                                 &kaddr, &pfn);
1048                 if (map_len < 0) {
1049                         ret = map_len;
1050                         break;
1051                 }
1052
1053                 map_len = PFN_PHYS(map_len);
1054                 kaddr += offset;
1055                 map_len -= offset;
1056                 if (map_len > end - pos)
1057                         map_len = end - pos;
1058
1059                 if (iov_iter_rw(iter) == WRITE)
1060                         map_len = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1061                                         map_len, iter);
1062                 else
1063                         map_len = copy_to_iter(kaddr, map_len, iter);
1064                 if (map_len <= 0) {
1065                         ret = map_len ? map_len : -EFAULT;
1066                         break;
1067                 }
1068
1069                 pos += map_len;
1070                 length -= map_len;
1071                 done += map_len;
1072         }
1073         dax_read_unlock(id);
1074
1075         return done ? done : ret;
1076 }
1077
1078 /**
1079  * dax_iomap_rw - Perform I/O to a DAX file
1080  * @iocb:       The control block for this I/O
1081  * @iter:       The addresses to do I/O from or to
1082  * @ops:        iomap ops passed from the file system
1083  *
1084  * This function performs read and write operations to directly mapped
1085  * persistent memory.  The callers needs to take care of read/write exclusion
1086  * and evicting any page cache pages in the region under I/O.
1087  */
1088 ssize_t
1089 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1090                 const struct iomap_ops *ops)
1091 {
1092         struct address_space *mapping = iocb->ki_filp->f_mapping;
1093         struct inode *inode = mapping->host;
1094         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1095         unsigned flags = 0;
1096
1097         if (iov_iter_rw(iter) == WRITE) {
1098                 lockdep_assert_held_exclusive(&inode->i_rwsem);
1099                 flags |= IOMAP_WRITE;
1100         } else {
1101                 lockdep_assert_held(&inode->i_rwsem);
1102         }
1103
1104         while (iov_iter_count(iter)) {
1105                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1106                                 iter, dax_iomap_actor);
1107                 if (ret <= 0)
1108                         break;
1109                 pos += ret;
1110                 done += ret;
1111         }
1112
1113         iocb->ki_pos += done;
1114         return done ? done : ret;
1115 }
1116 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1117
1118 static int dax_fault_return(int error)
1119 {
1120         if (error == 0)
1121                 return VM_FAULT_NOPAGE;
1122         if (error == -ENOMEM)
1123                 return VM_FAULT_OOM;
1124         return VM_FAULT_SIGBUS;
1125 }
1126
1127 static int dax_iomap_pte_fault(struct vm_fault *vmf,
1128                                const struct iomap_ops *ops)
1129 {
1130         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1131         struct inode *inode = mapping->host;
1132         unsigned long vaddr = vmf->address;
1133         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1134         sector_t sector;
1135         struct iomap iomap = { 0 };
1136         unsigned flags = IOMAP_FAULT;
1137         int error, major = 0;
1138         int vmf_ret = 0;
1139         void *entry;
1140
1141         trace_dax_pte_fault(inode, vmf, vmf_ret);
1142         /*
1143          * Check whether offset isn't beyond end of file now. Caller is supposed
1144          * to hold locks serializing us with truncate / punch hole so this is
1145          * a reliable test.
1146          */
1147         if (pos >= i_size_read(inode)) {
1148                 vmf_ret = VM_FAULT_SIGBUS;
1149                 goto out;
1150         }
1151
1152         if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1153                 flags |= IOMAP_WRITE;
1154
1155         entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1156         if (IS_ERR(entry)) {
1157                 vmf_ret = dax_fault_return(PTR_ERR(entry));
1158                 goto out;
1159         }
1160
1161         /*
1162          * It is possible, particularly with mixed reads & writes to private
1163          * mappings, that we have raced with a PMD fault that overlaps with
1164          * the PTE we need to set up.  If so just return and the fault will be
1165          * retried.
1166          */
1167         if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1168                 vmf_ret = VM_FAULT_NOPAGE;
1169                 goto unlock_entry;
1170         }
1171
1172         /*
1173          * Note that we don't bother to use iomap_apply here: DAX required
1174          * the file system block size to be equal the page size, which means
1175          * that we never have to deal with more than a single extent here.
1176          */
1177         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1178         if (error) {
1179                 vmf_ret = dax_fault_return(error);
1180                 goto unlock_entry;
1181         }
1182         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1183                 error = -EIO;   /* fs corruption? */
1184                 goto error_finish_iomap;
1185         }
1186
1187         sector = dax_iomap_sector(&iomap, pos);
1188
1189         if (vmf->cow_page) {
1190                 switch (iomap.type) {
1191                 case IOMAP_HOLE:
1192                 case IOMAP_UNWRITTEN:
1193                         clear_user_highpage(vmf->cow_page, vaddr);
1194                         break;
1195                 case IOMAP_MAPPED:
1196                         error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1197                                         sector, PAGE_SIZE, vmf->cow_page, vaddr);
1198                         break;
1199                 default:
1200                         WARN_ON_ONCE(1);
1201                         error = -EIO;
1202                         break;
1203                 }
1204
1205                 if (error)
1206                         goto error_finish_iomap;
1207
1208                 __SetPageUptodate(vmf->cow_page);
1209                 vmf_ret = finish_fault(vmf);
1210                 if (!vmf_ret)
1211                         vmf_ret = VM_FAULT_DONE_COW;
1212                 goto finish_iomap;
1213         }
1214
1215         switch (iomap.type) {
1216         case IOMAP_MAPPED:
1217                 if (iomap.flags & IOMAP_F_NEW) {
1218                         count_vm_event(PGMAJFAULT);
1219                         count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
1220                         major = VM_FAULT_MAJOR;
1221                 }
1222                 error = dax_insert_mapping(mapping, iomap.bdev, iomap.dax_dev,
1223                                 sector, PAGE_SIZE, &entry, vmf->vma, vmf);
1224                 /* -EBUSY is fine, somebody else faulted on the same PTE */
1225                 if (error == -EBUSY)
1226                         error = 0;
1227                 break;
1228         case IOMAP_UNWRITTEN:
1229         case IOMAP_HOLE:
1230                 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
1231                         vmf_ret = dax_load_hole(mapping, &entry, vmf);
1232                         goto finish_iomap;
1233                 }
1234                 /*FALLTHRU*/
1235         default:
1236                 WARN_ON_ONCE(1);
1237                 error = -EIO;
1238                 break;
1239         }
1240
1241  error_finish_iomap:
1242         vmf_ret = dax_fault_return(error) | major;
1243  finish_iomap:
1244         if (ops->iomap_end) {
1245                 int copied = PAGE_SIZE;
1246
1247                 if (vmf_ret & VM_FAULT_ERROR)
1248                         copied = 0;
1249                 /*
1250                  * The fault is done by now and there's no way back (other
1251                  * thread may be already happily using PTE we have installed).
1252                  * Just ignore error from ->iomap_end since we cannot do much
1253                  * with it.
1254                  */
1255                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1256         }
1257  unlock_entry:
1258         put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1259  out:
1260         trace_dax_pte_fault_done(inode, vmf, vmf_ret);
1261         return vmf_ret;
1262 }
1263
1264 #ifdef CONFIG_FS_DAX_PMD
1265 /*
1266  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
1267  * more often than one might expect in the below functions.
1268  */
1269 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
1270
1271 static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap,
1272                 loff_t pos, void **entryp)
1273 {
1274         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1275         const sector_t sector = dax_iomap_sector(iomap, pos);
1276         struct dax_device *dax_dev = iomap->dax_dev;
1277         struct block_device *bdev = iomap->bdev;
1278         struct inode *inode = mapping->host;
1279         const size_t size = PMD_SIZE;
1280         void *ret = NULL, *kaddr;
1281         long length = 0;
1282         pgoff_t pgoff;
1283         pfn_t pfn;
1284         int id;
1285
1286         if (bdev_dax_pgoff(bdev, sector, size, &pgoff) != 0)
1287                 goto fallback;
1288
1289         id = dax_read_lock();
1290         length = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
1291         if (length < 0)
1292                 goto unlock_fallback;
1293         length = PFN_PHYS(length);
1294
1295         if (length < size)
1296                 goto unlock_fallback;
1297         if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR)
1298                 goto unlock_fallback;
1299         if (!pfn_t_devmap(pfn))
1300                 goto unlock_fallback;
1301         dax_read_unlock(id);
1302
1303         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, sector,
1304                         RADIX_DAX_PMD);
1305         if (IS_ERR(ret))
1306                 goto fallback;
1307         *entryp = ret;
1308
1309         trace_dax_pmd_insert_mapping(inode, vmf, length, pfn, ret);
1310         return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
1311                         pfn, vmf->flags & FAULT_FLAG_WRITE);
1312
1313 unlock_fallback:
1314         dax_read_unlock(id);
1315 fallback:
1316         trace_dax_pmd_insert_mapping_fallback(inode, vmf, length, pfn, ret);
1317         return VM_FAULT_FALLBACK;
1318 }
1319
1320 static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
1321                 void **entryp)
1322 {
1323         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1324         unsigned long pmd_addr = vmf->address & PMD_MASK;
1325         struct inode *inode = mapping->host;
1326         struct page *zero_page;
1327         void *ret = NULL;
1328         spinlock_t *ptl;
1329         pmd_t pmd_entry;
1330
1331         zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1332
1333         if (unlikely(!zero_page))
1334                 goto fallback;
1335
1336         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0,
1337                         RADIX_DAX_PMD | RADIX_DAX_HZP);
1338         if (IS_ERR(ret))
1339                 goto fallback;
1340         *entryp = ret;
1341
1342         ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1343         if (!pmd_none(*(vmf->pmd))) {
1344                 spin_unlock(ptl);
1345                 goto fallback;
1346         }
1347
1348         pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1349         pmd_entry = pmd_mkhuge(pmd_entry);
1350         set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1351         spin_unlock(ptl);
1352         trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
1353         return VM_FAULT_NOPAGE;
1354
1355 fallback:
1356         trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
1357         return VM_FAULT_FALLBACK;
1358 }
1359
1360 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1361                                const struct iomap_ops *ops)
1362 {
1363         struct vm_area_struct *vma = vmf->vma;
1364         struct address_space *mapping = vma->vm_file->f_mapping;
1365         unsigned long pmd_addr = vmf->address & PMD_MASK;
1366         bool write = vmf->flags & FAULT_FLAG_WRITE;
1367         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1368         struct inode *inode = mapping->host;
1369         int result = VM_FAULT_FALLBACK;
1370         struct iomap iomap = { 0 };
1371         pgoff_t max_pgoff, pgoff;
1372         void *entry;
1373         loff_t pos;
1374         int error;
1375
1376         /*
1377          * Check whether offset isn't beyond end of file now. Caller is
1378          * supposed to hold locks serializing us with truncate / punch hole so
1379          * this is a reliable test.
1380          */
1381         pgoff = linear_page_index(vma, pmd_addr);
1382         max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1383
1384         trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1385
1386         /* Fall back to PTEs if we're going to COW */
1387         if (write && !(vma->vm_flags & VM_SHARED))
1388                 goto fallback;
1389
1390         /* If the PMD would extend outside the VMA */
1391         if (pmd_addr < vma->vm_start)
1392                 goto fallback;
1393         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1394                 goto fallback;
1395
1396         if (pgoff > max_pgoff) {
1397                 result = VM_FAULT_SIGBUS;
1398                 goto out;
1399         }
1400
1401         /* If the PMD would extend beyond the file size */
1402         if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1403                 goto fallback;
1404
1405         /*
1406          * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1407          * PMD or a HZP entry.  If it can't (because a 4k page is already in
1408          * the tree, for instance), it will return -EEXIST and we just fall
1409          * back to 4k entries.
1410          */
1411         entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1412         if (IS_ERR(entry))
1413                 goto fallback;
1414
1415         /*
1416          * It is possible, particularly with mixed reads & writes to private
1417          * mappings, that we have raced with a PTE fault that overlaps with
1418          * the PMD we need to set up.  If so just return and the fault will be
1419          * retried.
1420          */
1421         if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1422                         !pmd_devmap(*vmf->pmd)) {
1423                 result = 0;
1424                 goto unlock_entry;
1425         }
1426
1427         /*
1428          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1429          * setting up a mapping, so really we're using iomap_begin() as a way
1430          * to look up our filesystem block.
1431          */
1432         pos = (loff_t)pgoff << PAGE_SHIFT;
1433         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1434         if (error)
1435                 goto unlock_entry;
1436
1437         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1438                 goto finish_iomap;
1439
1440         switch (iomap.type) {
1441         case IOMAP_MAPPED:
1442                 result = dax_pmd_insert_mapping(vmf, &iomap, pos, &entry);
1443                 break;
1444         case IOMAP_UNWRITTEN:
1445         case IOMAP_HOLE:
1446                 if (WARN_ON_ONCE(write))
1447                         break;
1448                 result = dax_pmd_load_hole(vmf, &iomap, &entry);
1449                 break;
1450         default:
1451                 WARN_ON_ONCE(1);
1452                 break;
1453         }
1454
1455  finish_iomap:
1456         if (ops->iomap_end) {
1457                 int copied = PMD_SIZE;
1458
1459                 if (result == VM_FAULT_FALLBACK)
1460                         copied = 0;
1461                 /*
1462                  * The fault is done by now and there's no way back (other
1463                  * thread may be already happily using PMD we have installed).
1464                  * Just ignore error from ->iomap_end since we cannot do much
1465                  * with it.
1466                  */
1467                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1468                                 &iomap);
1469         }
1470  unlock_entry:
1471         put_locked_mapping_entry(mapping, pgoff, entry);
1472  fallback:
1473         if (result == VM_FAULT_FALLBACK) {
1474                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1475                 count_vm_event(THP_FAULT_FALLBACK);
1476         }
1477 out:
1478         trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1479         return result;
1480 }
1481 #else
1482 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1483                                const struct iomap_ops *ops)
1484 {
1485         return VM_FAULT_FALLBACK;
1486 }
1487 #endif /* CONFIG_FS_DAX_PMD */
1488
1489 /**
1490  * dax_iomap_fault - handle a page fault on a DAX file
1491  * @vmf: The description of the fault
1492  * @ops: iomap ops passed from the file system
1493  *
1494  * When a page fault occurs, filesystems may call this helper in
1495  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1496  * has done all the necessary locking for page fault to proceed
1497  * successfully.
1498  */
1499 int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1500                     const struct iomap_ops *ops)
1501 {
1502         switch (pe_size) {
1503         case PE_SIZE_PTE:
1504                 return dax_iomap_pte_fault(vmf, ops);
1505         case PE_SIZE_PMD:
1506                 return dax_iomap_pmd_fault(vmf, ops);
1507         default:
1508                 return VM_FAULT_FALLBACK;
1509         }
1510 }
1511 EXPORT_SYMBOL_GPL(dax_iomap_fault);