1 // SPDX-License-Identifier: GPL-2.0-only
3 * fs/dax.c - Direct Access filesystem code
4 * Copyright (c) 2013-2014 Intel Corporation
5 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
6 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
9 #include <linux/atomic.h>
10 #include <linux/blkdev.h>
11 #include <linux/buffer_head.h>
12 #include <linux/dax.h>
14 #include <linux/genhd.h>
15 #include <linux/highmem.h>
16 #include <linux/memcontrol.h>
18 #include <linux/mutex.h>
19 #include <linux/pagevec.h>
20 #include <linux/sched.h>
21 #include <linux/sched/signal.h>
22 #include <linux/uio.h>
23 #include <linux/vmstat.h>
24 #include <linux/pfn_t.h>
25 #include <linux/sizes.h>
26 #include <linux/mmu_notifier.h>
27 #include <linux/iomap.h>
28 #include <asm/pgalloc.h>
30 #define CREATE_TRACE_POINTS
31 #include <trace/events/fs_dax.h>
33 static inline unsigned int pe_order(enum page_entry_size pe_size)
35 if (pe_size == PE_SIZE_PTE)
36 return PAGE_SHIFT - PAGE_SHIFT;
37 if (pe_size == PE_SIZE_PMD)
38 return PMD_SHIFT - PAGE_SHIFT;
39 if (pe_size == PE_SIZE_PUD)
40 return PUD_SHIFT - PAGE_SHIFT;
44 /* We choose 4096 entries - same as per-zone page wait tables */
45 #define DAX_WAIT_TABLE_BITS 12
46 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
48 /* The 'colour' (ie low bits) within a PMD of a page offset. */
49 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
50 #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
52 /* The order of a PMD entry */
53 #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT)
55 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
57 static int __init init_dax_wait_table(void)
61 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
62 init_waitqueue_head(wait_table + i);
65 fs_initcall(init_dax_wait_table);
68 * DAX pagecache entries use XArray value entries so they can't be mistaken
69 * for pages. We use one bit for locking, one bit for the entry size (PMD)
70 * and two more to tell us if the entry is a zero page or an empty entry that
71 * is just used for locking. In total four special bits.
73 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
74 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
78 #define DAX_LOCKED (1UL << 0)
79 #define DAX_PMD (1UL << 1)
80 #define DAX_ZERO_PAGE (1UL << 2)
81 #define DAX_EMPTY (1UL << 3)
83 static unsigned long dax_to_pfn(void *entry)
85 return xa_to_value(entry) >> DAX_SHIFT;
88 static void *dax_make_entry(pfn_t pfn, unsigned long flags)
90 return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
93 static bool dax_is_locked(void *entry)
95 return xa_to_value(entry) & DAX_LOCKED;
98 static unsigned int dax_entry_order(void *entry)
100 if (xa_to_value(entry) & DAX_PMD)
105 static unsigned long dax_is_pmd_entry(void *entry)
107 return xa_to_value(entry) & DAX_PMD;
110 static bool dax_is_pte_entry(void *entry)
112 return !(xa_to_value(entry) & DAX_PMD);
115 static int dax_is_zero_entry(void *entry)
117 return xa_to_value(entry) & DAX_ZERO_PAGE;
120 static int dax_is_empty_entry(void *entry)
122 return xa_to_value(entry) & DAX_EMPTY;
126 * true if the entry that was found is of a smaller order than the entry
127 * we were looking for
129 static bool dax_is_conflict(void *entry)
131 return entry == XA_RETRY_ENTRY;
135 * DAX page cache entry locking
137 struct exceptional_entry_key {
142 struct wait_exceptional_entry_queue {
143 wait_queue_entry_t wait;
144 struct exceptional_entry_key key;
147 static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
148 void *entry, struct exceptional_entry_key *key)
151 unsigned long index = xas->xa_index;
154 * If 'entry' is a PMD, align the 'index' that we use for the wait
155 * queue to the start of that PMD. This ensures that all offsets in
156 * the range covered by the PMD map to the same bit lock.
158 if (dax_is_pmd_entry(entry))
159 index &= ~PG_PMD_COLOUR;
161 key->entry_start = index;
163 hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
164 return wait_table + hash;
167 static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
168 unsigned int mode, int sync, void *keyp)
170 struct exceptional_entry_key *key = keyp;
171 struct wait_exceptional_entry_queue *ewait =
172 container_of(wait, struct wait_exceptional_entry_queue, wait);
174 if (key->xa != ewait->key.xa ||
175 key->entry_start != ewait->key.entry_start)
177 return autoremove_wake_function(wait, mode, sync, NULL);
181 * @entry may no longer be the entry at the index in the mapping.
182 * The important information it's conveying is whether the entry at
183 * this index used to be a PMD entry.
185 static void dax_wake_entry(struct xa_state *xas, void *entry, bool wake_all)
187 struct exceptional_entry_key key;
188 wait_queue_head_t *wq;
190 wq = dax_entry_waitqueue(xas, entry, &key);
193 * Checking for locked entry and prepare_to_wait_exclusive() happens
194 * under the i_pages lock, ditto for entry handling in our callers.
195 * So at this point all tasks that could have seen our entry locked
196 * must be in the waitqueue and the following check will see them.
198 if (waitqueue_active(wq))
199 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
203 * Look up entry in page cache, wait for it to become unlocked if it
204 * is a DAX entry and return it. The caller must subsequently call
205 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
206 * if it did. The entry returned may have a larger order than @order.
207 * If @order is larger than the order of the entry found in i_pages, this
208 * function returns a dax_is_conflict entry.
210 * Must be called with the i_pages lock held.
212 static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
215 struct wait_exceptional_entry_queue ewait;
216 wait_queue_head_t *wq;
218 init_wait(&ewait.wait);
219 ewait.wait.func = wake_exceptional_entry_func;
222 entry = xas_find_conflict(xas);
223 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
225 if (dax_entry_order(entry) < order)
226 return XA_RETRY_ENTRY;
227 if (!dax_is_locked(entry))
230 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
231 prepare_to_wait_exclusive(wq, &ewait.wait,
232 TASK_UNINTERRUPTIBLE);
236 finish_wait(wq, &ewait.wait);
242 * The only thing keeping the address space around is the i_pages lock
243 * (it's cycled in clear_inode() after removing the entries from i_pages)
244 * After we call xas_unlock_irq(), we cannot touch xas->xa.
246 static void wait_entry_unlocked(struct xa_state *xas, void *entry)
248 struct wait_exceptional_entry_queue ewait;
249 wait_queue_head_t *wq;
251 init_wait(&ewait.wait);
252 ewait.wait.func = wake_exceptional_entry_func;
254 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
256 * Unlike get_unlocked_entry() there is no guarantee that this
257 * path ever successfully retrieves an unlocked entry before an
258 * inode dies. Perform a non-exclusive wait in case this path
259 * never successfully performs its own wake up.
261 prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
264 finish_wait(wq, &ewait.wait);
267 static void put_unlocked_entry(struct xa_state *xas, void *entry)
269 /* If we were the only waiter woken, wake the next one */
270 if (entry && !dax_is_conflict(entry))
271 dax_wake_entry(xas, entry, false);
275 * We used the xa_state to get the entry, but then we locked the entry and
276 * dropped the xa_lock, so we know the xa_state is stale and must be reset
279 static void dax_unlock_entry(struct xa_state *xas, void *entry)
283 BUG_ON(dax_is_locked(entry));
286 old = xas_store(xas, entry);
288 BUG_ON(!dax_is_locked(old));
289 dax_wake_entry(xas, entry, false);
293 * Return: The entry stored at this location before it was locked.
295 static void *dax_lock_entry(struct xa_state *xas, void *entry)
297 unsigned long v = xa_to_value(entry);
298 return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
301 static unsigned long dax_entry_size(void *entry)
303 if (dax_is_zero_entry(entry))
305 else if (dax_is_empty_entry(entry))
307 else if (dax_is_pmd_entry(entry))
313 static unsigned long dax_end_pfn(void *entry)
315 return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
319 * Iterate through all mapped pfns represented by an entry, i.e. skip
320 * 'empty' and 'zero' entries.
322 #define for_each_mapped_pfn(entry, pfn) \
323 for (pfn = dax_to_pfn(entry); \
324 pfn < dax_end_pfn(entry); pfn++)
327 * TODO: for reflink+dax we need a way to associate a single page with
328 * multiple address_space instances at different linear_page_index()
331 static void dax_associate_entry(void *entry, struct address_space *mapping,
332 struct vm_area_struct *vma, unsigned long address)
334 unsigned long size = dax_entry_size(entry), pfn, index;
337 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
340 index = linear_page_index(vma, address & ~(size - 1));
341 for_each_mapped_pfn(entry, pfn) {
342 struct page *page = pfn_to_page(pfn);
344 WARN_ON_ONCE(page->mapping);
345 page->mapping = mapping;
346 page->index = index + i++;
350 static void dax_disassociate_entry(void *entry, struct address_space *mapping,
355 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
358 for_each_mapped_pfn(entry, pfn) {
359 struct page *page = pfn_to_page(pfn);
361 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
362 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
363 page->mapping = NULL;
368 static struct page *dax_busy_page(void *entry)
372 for_each_mapped_pfn(entry, pfn) {
373 struct page *page = pfn_to_page(pfn);
375 if (page_ref_count(page) > 1)
382 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page
383 * @page: The page whose entry we want to lock
385 * Context: Process context.
386 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
389 dax_entry_t dax_lock_page(struct page *page)
391 XA_STATE(xas, NULL, 0);
394 /* Ensure page->mapping isn't freed while we look at it */
397 struct address_space *mapping = READ_ONCE(page->mapping);
400 if (!mapping || !dax_mapping(mapping))
404 * In the device-dax case there's no need to lock, a
405 * struct dev_pagemap pin is sufficient to keep the
406 * inode alive, and we assume we have dev_pagemap pin
407 * otherwise we would not have a valid pfn_to_page()
410 entry = (void *)~0UL;
411 if (S_ISCHR(mapping->host->i_mode))
414 xas.xa = &mapping->i_pages;
416 if (mapping != page->mapping) {
417 xas_unlock_irq(&xas);
420 xas_set(&xas, page->index);
421 entry = xas_load(&xas);
422 if (dax_is_locked(entry)) {
424 wait_entry_unlocked(&xas, entry);
428 dax_lock_entry(&xas, entry);
429 xas_unlock_irq(&xas);
433 return (dax_entry_t)entry;
436 void dax_unlock_page(struct page *page, dax_entry_t cookie)
438 struct address_space *mapping = page->mapping;
439 XA_STATE(xas, &mapping->i_pages, page->index);
441 if (S_ISCHR(mapping->host->i_mode))
444 dax_unlock_entry(&xas, (void *)cookie);
448 * Find page cache entry at given index. If it is a DAX entry, return it
449 * with the entry locked. If the page cache doesn't contain an entry at
450 * that index, add a locked empty entry.
452 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
453 * either return that locked entry or will return VM_FAULT_FALLBACK.
454 * This will happen if there are any PTE entries within the PMD range
455 * that we are requesting.
457 * We always favor PTE entries over PMD entries. There isn't a flow where we
458 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
459 * insertion will fail if it finds any PTE entries already in the tree, and a
460 * PTE insertion will cause an existing PMD entry to be unmapped and
461 * downgraded to PTE entries. This happens for both PMD zero pages as
462 * well as PMD empty entries.
464 * The exception to this downgrade path is for PMD entries that have
465 * real storage backing them. We will leave these real PMD entries in
466 * the tree, and PTE writes will simply dirty the entire PMD entry.
468 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
469 * persistent memory the benefit is doubtful. We can add that later if we can
472 * On error, this function does not return an ERR_PTR. Instead it returns
473 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
474 * overlap with xarray value entries.
476 static void *grab_mapping_entry(struct xa_state *xas,
477 struct address_space *mapping, unsigned int order)
479 unsigned long index = xas->xa_index;
480 bool pmd_downgrade; /* splitting PMD entry into PTE entries? */
484 pmd_downgrade = false;
486 entry = get_unlocked_entry(xas, order);
489 if (dax_is_conflict(entry))
491 if (!xa_is_value(entry)) {
492 xas_set_err(xas, EIO);
497 if (dax_is_pmd_entry(entry) &&
498 (dax_is_zero_entry(entry) ||
499 dax_is_empty_entry(entry))) {
500 pmd_downgrade = true;
507 * Make sure 'entry' remains valid while we drop
510 dax_lock_entry(xas, entry);
513 * Besides huge zero pages the only other thing that gets
514 * downgraded are empty entries which don't need to be
517 if (dax_is_zero_entry(entry)) {
519 unmap_mapping_pages(mapping,
520 xas->xa_index & ~PG_PMD_COLOUR,
526 dax_disassociate_entry(entry, mapping, false);
527 xas_store(xas, NULL); /* undo the PMD join */
528 dax_wake_entry(xas, entry, true);
529 mapping->nrexceptional--;
535 dax_lock_entry(xas, entry);
537 unsigned long flags = DAX_EMPTY;
541 entry = dax_make_entry(pfn_to_pfn_t(0), flags);
542 dax_lock_entry(xas, entry);
545 mapping->nrexceptional++;
550 if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
552 if (xas->xa_node == XA_ERROR(-ENOMEM))
553 return xa_mk_internal(VM_FAULT_OOM);
555 return xa_mk_internal(VM_FAULT_SIGBUS);
559 return xa_mk_internal(VM_FAULT_FALLBACK);
563 * dax_layout_busy_page - find first pinned page in @mapping
564 * @mapping: address space to scan for a page with ref count > 1
566 * DAX requires ZONE_DEVICE mapped pages. These pages are never
567 * 'onlined' to the page allocator so they are considered idle when
568 * page->count == 1. A filesystem uses this interface to determine if
569 * any page in the mapping is busy, i.e. for DMA, or other
570 * get_user_pages() usages.
572 * It is expected that the filesystem is holding locks to block the
573 * establishment of new mappings in this address_space. I.e. it expects
574 * to be able to run unmap_mapping_range() and subsequently not race
575 * mapping_mapped() becoming true.
577 struct page *dax_layout_busy_page(struct address_space *mapping)
579 XA_STATE(xas, &mapping->i_pages, 0);
581 unsigned int scanned = 0;
582 struct page *page = NULL;
585 * In the 'limited' case get_user_pages() for dax is disabled.
587 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
590 if (!dax_mapping(mapping) || !mapping_mapped(mapping))
594 * If we race get_user_pages_fast() here either we'll see the
595 * elevated page count in the iteration and wait, or
596 * get_user_pages_fast() will see that the page it took a reference
597 * against is no longer mapped in the page tables and bail to the
598 * get_user_pages() slow path. The slow path is protected by
599 * pte_lock() and pmd_lock(). New references are not taken without
600 * holding those locks, and unmap_mapping_range() will not zero the
601 * pte or pmd without holding the respective lock, so we are
602 * guaranteed to either see new references or prevent new
603 * references from being established.
605 unmap_mapping_range(mapping, 0, 0, 0);
608 xas_for_each(&xas, entry, ULONG_MAX) {
609 if (WARN_ON_ONCE(!xa_is_value(entry)))
611 if (unlikely(dax_is_locked(entry)))
612 entry = get_unlocked_entry(&xas, 0);
614 page = dax_busy_page(entry);
615 put_unlocked_entry(&xas, entry);
618 if (++scanned % XA_CHECK_SCHED)
622 xas_unlock_irq(&xas);
626 xas_unlock_irq(&xas);
629 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
631 static int __dax_invalidate_entry(struct address_space *mapping,
632 pgoff_t index, bool trunc)
634 XA_STATE(xas, &mapping->i_pages, index);
639 entry = get_unlocked_entry(&xas, 0);
640 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
643 (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
644 xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
646 dax_disassociate_entry(entry, mapping, trunc);
647 xas_store(&xas, NULL);
648 mapping->nrexceptional--;
651 put_unlocked_entry(&xas, entry);
652 xas_unlock_irq(&xas);
657 * Delete DAX entry at @index from @mapping. Wait for it
658 * to be unlocked before deleting it.
660 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
662 int ret = __dax_invalidate_entry(mapping, index, true);
665 * This gets called from truncate / punch_hole path. As such, the caller
666 * must hold locks protecting against concurrent modifications of the
667 * page cache (usually fs-private i_mmap_sem for writing). Since the
668 * caller has seen a DAX entry for this index, we better find it
669 * at that index as well...
676 * Invalidate DAX entry if it is clean.
678 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
681 return __dax_invalidate_entry(mapping, index, false);
684 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
685 sector_t sector, size_t size, struct page *to,
693 rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
697 id = dax_read_lock();
698 rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, NULL);
703 vto = kmap_atomic(to);
704 copy_user_page(vto, (void __force *)kaddr, vaddr, to);
711 * By this point grab_mapping_entry() has ensured that we have a locked entry
712 * of the appropriate size so we don't have to worry about downgrading PMDs to
713 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
714 * already in the tree, we will skip the insertion and just dirty the PMD as
717 static void *dax_insert_entry(struct xa_state *xas,
718 struct address_space *mapping, struct vm_fault *vmf,
719 void *entry, pfn_t pfn, unsigned long flags, bool dirty)
721 void *new_entry = dax_make_entry(pfn, flags);
724 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
726 if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) {
727 unsigned long index = xas->xa_index;
728 /* we are replacing a zero page with block mapping */
729 if (dax_is_pmd_entry(entry))
730 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
733 unmap_mapping_pages(mapping, index, 1, false);
738 if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
741 dax_disassociate_entry(entry, mapping, false);
742 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address);
744 * Only swap our new entry into the page cache if the current
745 * entry is a zero page or an empty entry. If a normal PTE or
746 * PMD entry is already in the cache, we leave it alone. This
747 * means that if we are trying to insert a PTE and the
748 * existing entry is a PMD, we will just leave the PMD in the
749 * tree and dirty it if necessary.
751 old = dax_lock_entry(xas, new_entry);
752 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
756 xas_load(xas); /* Walk the xa_state */
760 xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
767 unsigned long pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
769 unsigned long address;
771 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
772 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
776 /* Walk all mappings of a given index of a file and writeprotect them */
777 static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
780 struct vm_area_struct *vma;
781 pte_t pte, *ptep = NULL;
785 i_mmap_lock_read(mapping);
786 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
787 struct mmu_notifier_range range;
788 unsigned long address;
792 if (!(vma->vm_flags & VM_SHARED))
795 address = pgoff_address(index, vma);
798 * follow_invalidate_pte() will use the range to call
799 * mmu_notifier_invalidate_range_start() on our behalf before
802 if (follow_invalidate_pte(vma->vm_mm, address, &range, &ptep,
807 * No need to call mmu_notifier_invalidate_range() as we are
808 * downgrading page table protection not changing it to point
811 * See Documentation/vm/mmu_notifier.rst
814 #ifdef CONFIG_FS_DAX_PMD
817 if (pfn != pmd_pfn(*pmdp))
819 if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
822 flush_cache_range(vma, address,
823 address + HPAGE_PMD_SIZE);
824 pmd = pmdp_invalidate(vma, address, pmdp);
825 pmd = pmd_wrprotect(pmd);
826 pmd = pmd_mkclean(pmd);
827 set_pmd_at(vma->vm_mm, address, pmdp, pmd);
832 if (pfn != pte_pfn(*ptep))
834 if (!pte_dirty(*ptep) && !pte_write(*ptep))
837 flush_cache_page(vma, address, pfn);
838 pte = ptep_clear_flush(vma, address, ptep);
839 pte = pte_wrprotect(pte);
840 pte = pte_mkclean(pte);
841 set_pte_at(vma->vm_mm, address, ptep, pte);
843 pte_unmap_unlock(ptep, ptl);
846 mmu_notifier_invalidate_range_end(&range);
848 i_mmap_unlock_read(mapping);
851 static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
852 struct address_space *mapping, void *entry)
854 unsigned long pfn, index, count;
858 * A page got tagged dirty in DAX mapping? Something is seriously
861 if (WARN_ON(!xa_is_value(entry)))
864 if (unlikely(dax_is_locked(entry))) {
865 void *old_entry = entry;
867 entry = get_unlocked_entry(xas, 0);
869 /* Entry got punched out / reallocated? */
870 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
873 * Entry got reallocated elsewhere? No need to writeback.
874 * We have to compare pfns as we must not bail out due to
875 * difference in lockbit or entry type.
877 if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
879 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
880 dax_is_zero_entry(entry))) {
885 /* Another fsync thread may have already done this entry */
886 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
890 /* Lock the entry to serialize with page faults */
891 dax_lock_entry(xas, entry);
894 * We can clear the tag now but we have to be careful so that concurrent
895 * dax_writeback_one() calls for the same index cannot finish before we
896 * actually flush the caches. This is achieved as the calls will look
897 * at the entry only under the i_pages lock and once they do that
898 * they will see the entry locked and wait for it to unlock.
900 xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
904 * If dax_writeback_mapping_range() was given a wbc->range_start
905 * in the middle of a PMD, the 'index' we use needs to be
906 * aligned to the start of the PMD.
907 * This allows us to flush for PMD_SIZE and not have to worry about
908 * partial PMD writebacks.
910 pfn = dax_to_pfn(entry);
911 count = 1UL << dax_entry_order(entry);
912 index = xas->xa_index & ~(count - 1);
914 dax_entry_mkclean(mapping, index, pfn);
915 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
917 * After we have flushed the cache, we can clear the dirty tag. There
918 * cannot be new dirty data in the pfn after the flush has completed as
919 * the pfn mappings are writeprotected and fault waits for mapping
924 xas_store(xas, entry);
925 xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
926 dax_wake_entry(xas, entry, false);
928 trace_dax_writeback_one(mapping->host, index, count);
932 put_unlocked_entry(xas, entry);
937 * Flush the mapping to the persistent domain within the byte range of [start,
938 * end]. This is required by data integrity operations to ensure file data is
939 * on persistent storage prior to completion of the operation.
941 int dax_writeback_mapping_range(struct address_space *mapping,
942 struct block_device *bdev, struct writeback_control *wbc)
944 XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
945 struct inode *inode = mapping->host;
946 pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
947 struct dax_device *dax_dev;
950 unsigned int scanned = 0;
952 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
955 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
958 dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
962 trace_dax_writeback_range(inode, xas.xa_index, end_index);
964 tag_pages_for_writeback(mapping, xas.xa_index, end_index);
967 xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
968 ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
970 mapping_set_error(mapping, ret);
973 if (++scanned % XA_CHECK_SCHED)
977 xas_unlock_irq(&xas);
981 xas_unlock_irq(&xas);
983 trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
986 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
988 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
990 return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
993 static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
996 const sector_t sector = dax_iomap_sector(iomap, pos);
1001 rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
1004 id = dax_read_lock();
1005 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1012 if (PFN_PHYS(length) < size)
1014 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1016 /* For larger pages we need devmap */
1017 if (length > 1 && !pfn_t_devmap(*pfnp))
1021 dax_read_unlock(id);
1026 * The user has performed a load from a hole in the file. Allocating a new
1027 * page in the file would cause excessive storage usage for workloads with
1028 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1029 * If this page is ever written to we will re-fault and change the mapping to
1030 * point to real DAX storage instead.
1032 static vm_fault_t dax_load_hole(struct xa_state *xas,
1033 struct address_space *mapping, void **entry,
1034 struct vm_fault *vmf)
1036 struct inode *inode = mapping->host;
1037 unsigned long vaddr = vmf->address;
1038 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1041 *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1042 DAX_ZERO_PAGE, false);
1044 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1045 trace_dax_load_hole(inode, vmf, ret);
1049 static bool dax_range_is_aligned(struct block_device *bdev,
1050 unsigned int offset, unsigned int length)
1052 unsigned short sector_size = bdev_logical_block_size(bdev);
1054 if (!IS_ALIGNED(offset, sector_size))
1056 if (!IS_ALIGNED(length, sector_size))
1062 int __dax_zero_page_range(struct block_device *bdev,
1063 struct dax_device *dax_dev, sector_t sector,
1064 unsigned int offset, unsigned int size)
1066 if (dax_range_is_aligned(bdev, offset, size)) {
1067 sector_t start_sector = sector + (offset >> 9);
1069 return blkdev_issue_zeroout(bdev, start_sector,
1070 size >> 9, GFP_NOFS, 0);
1076 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
1080 id = dax_read_lock();
1081 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, NULL);
1083 dax_read_unlock(id);
1086 memset(kaddr + offset, 0, size);
1087 dax_flush(dax_dev, kaddr + offset, size);
1088 dax_read_unlock(id);
1092 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1095 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1096 struct iomap *iomap)
1098 struct block_device *bdev = iomap->bdev;
1099 struct dax_device *dax_dev = iomap->dax_dev;
1100 struct iov_iter *iter = data;
1101 loff_t end = pos + length, done = 0;
1106 if (iov_iter_rw(iter) == READ) {
1107 end = min(end, i_size_read(inode));
1111 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1112 return iov_iter_zero(min(length, end - pos), iter);
1115 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1119 * Write can allocate block for an area which has a hole page mapped
1120 * into page tables. We have to tear down these mappings so that data
1121 * written by write(2) is visible in mmap.
1123 if (iomap->flags & IOMAP_F_NEW) {
1124 invalidate_inode_pages2_range(inode->i_mapping,
1126 (end - 1) >> PAGE_SHIFT);
1129 id = dax_read_lock();
1131 unsigned offset = pos & (PAGE_SIZE - 1);
1132 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1133 const sector_t sector = dax_iomap_sector(iomap, pos);
1138 if (fatal_signal_pending(current)) {
1143 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1147 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1154 map_len = PFN_PHYS(map_len);
1157 if (map_len > end - pos)
1158 map_len = end - pos;
1161 * The userspace address for the memory copy has already been
1162 * validated via access_ok() in either vfs_read() or
1163 * vfs_write(), depending on which operation we are doing.
1165 if (iov_iter_rw(iter) == WRITE)
1166 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1169 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1181 dax_read_unlock(id);
1183 return done ? done : ret;
1187 * dax_iomap_rw - Perform I/O to a DAX file
1188 * @iocb: The control block for this I/O
1189 * @iter: The addresses to do I/O from or to
1190 * @ops: iomap ops passed from the file system
1192 * This function performs read and write operations to directly mapped
1193 * persistent memory. The callers needs to take care of read/write exclusion
1194 * and evicting any page cache pages in the region under I/O.
1197 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1198 const struct iomap_ops *ops)
1200 struct address_space *mapping = iocb->ki_filp->f_mapping;
1201 struct inode *inode = mapping->host;
1202 loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1205 if (iov_iter_rw(iter) == WRITE) {
1206 lockdep_assert_held_write(&inode->i_rwsem);
1207 flags |= IOMAP_WRITE;
1209 lockdep_assert_held(&inode->i_rwsem);
1212 if (iocb->ki_flags & IOCB_NOWAIT)
1213 flags |= IOMAP_NOWAIT;
1215 while (iov_iter_count(iter)) {
1216 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1217 iter, dax_iomap_actor);
1224 iocb->ki_pos += done;
1225 return done ? done : ret;
1227 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1229 static vm_fault_t dax_fault_return(int error)
1232 return VM_FAULT_NOPAGE;
1233 return vmf_error(error);
1237 * MAP_SYNC on a dax mapping guarantees dirty metadata is
1238 * flushed on write-faults (non-cow), but not read-faults.
1240 static bool dax_fault_is_synchronous(unsigned long flags,
1241 struct vm_area_struct *vma, struct iomap *iomap)
1243 return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1244 && (iomap->flags & IOMAP_F_DIRTY);
1247 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1248 int *iomap_errp, const struct iomap_ops *ops)
1250 struct vm_area_struct *vma = vmf->vma;
1251 struct address_space *mapping = vma->vm_file->f_mapping;
1252 XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1253 struct inode *inode = mapping->host;
1254 unsigned long vaddr = vmf->address;
1255 loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1256 struct iomap iomap = { 0 };
1257 unsigned flags = IOMAP_FAULT;
1258 int error, major = 0;
1259 bool write = vmf->flags & FAULT_FLAG_WRITE;
1265 trace_dax_pte_fault(inode, vmf, ret);
1267 * Check whether offset isn't beyond end of file now. Caller is supposed
1268 * to hold locks serializing us with truncate / punch hole so this is
1271 if (pos >= i_size_read(inode)) {
1272 ret = VM_FAULT_SIGBUS;
1276 if (write && !vmf->cow_page)
1277 flags |= IOMAP_WRITE;
1279 entry = grab_mapping_entry(&xas, mapping, 0);
1280 if (xa_is_internal(entry)) {
1281 ret = xa_to_internal(entry);
1286 * It is possible, particularly with mixed reads & writes to private
1287 * mappings, that we have raced with a PMD fault that overlaps with
1288 * the PTE we need to set up. If so just return and the fault will be
1291 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1292 ret = VM_FAULT_NOPAGE;
1297 * Note that we don't bother to use iomap_apply here: DAX required
1298 * the file system block size to be equal the page size, which means
1299 * that we never have to deal with more than a single extent here.
1301 error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1303 *iomap_errp = error;
1305 ret = dax_fault_return(error);
1308 if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1309 error = -EIO; /* fs corruption? */
1310 goto error_finish_iomap;
1313 if (vmf->cow_page) {
1314 sector_t sector = dax_iomap_sector(&iomap, pos);
1316 switch (iomap.type) {
1318 case IOMAP_UNWRITTEN:
1319 clear_user_highpage(vmf->cow_page, vaddr);
1322 error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1323 sector, PAGE_SIZE, vmf->cow_page, vaddr);
1332 goto error_finish_iomap;
1334 __SetPageUptodate(vmf->cow_page);
1335 ret = finish_fault(vmf);
1337 ret = VM_FAULT_DONE_COW;
1341 sync = dax_fault_is_synchronous(flags, vma, &iomap);
1343 switch (iomap.type) {
1345 if (iomap.flags & IOMAP_F_NEW) {
1346 count_vm_event(PGMAJFAULT);
1347 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1348 major = VM_FAULT_MAJOR;
1350 error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1352 goto error_finish_iomap;
1354 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1358 * If we are doing synchronous page fault and inode needs fsync,
1359 * we can insert PTE into page tables only after that happens.
1360 * Skip insertion for now and return the pfn so that caller can
1361 * insert it after fsync is done.
1364 if (WARN_ON_ONCE(!pfnp)) {
1366 goto error_finish_iomap;
1369 ret = VM_FAULT_NEEDDSYNC | major;
1372 trace_dax_insert_mapping(inode, vmf, entry);
1374 ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn);
1376 ret = vmf_insert_mixed(vma, vaddr, pfn);
1379 case IOMAP_UNWRITTEN:
1382 ret = dax_load_hole(&xas, mapping, &entry, vmf);
1393 ret = dax_fault_return(error);
1395 if (ops->iomap_end) {
1396 int copied = PAGE_SIZE;
1398 if (ret & VM_FAULT_ERROR)
1401 * The fault is done by now and there's no way back (other
1402 * thread may be already happily using PTE we have installed).
1403 * Just ignore error from ->iomap_end since we cannot do much
1406 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1409 dax_unlock_entry(&xas, entry);
1411 trace_dax_pte_fault_done(inode, vmf, ret);
1415 #ifdef CONFIG_FS_DAX_PMD
1416 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1417 struct iomap *iomap, void **entry)
1419 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1420 unsigned long pmd_addr = vmf->address & PMD_MASK;
1421 struct vm_area_struct *vma = vmf->vma;
1422 struct inode *inode = mapping->host;
1423 pgtable_t pgtable = NULL;
1424 struct page *zero_page;
1429 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1431 if (unlikely(!zero_page))
1434 pfn = page_to_pfn_t(zero_page);
1435 *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1436 DAX_PMD | DAX_ZERO_PAGE, false);
1438 if (arch_needs_pgtable_deposit()) {
1439 pgtable = pte_alloc_one(vma->vm_mm);
1441 return VM_FAULT_OOM;
1444 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1445 if (!pmd_none(*(vmf->pmd))) {
1451 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1452 mm_inc_nr_ptes(vma->vm_mm);
1454 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1455 pmd_entry = pmd_mkhuge(pmd_entry);
1456 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1458 trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1459 return VM_FAULT_NOPAGE;
1463 pte_free(vma->vm_mm, pgtable);
1464 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1465 return VM_FAULT_FALLBACK;
1468 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1469 const struct iomap_ops *ops)
1471 struct vm_area_struct *vma = vmf->vma;
1472 struct address_space *mapping = vma->vm_file->f_mapping;
1473 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1474 unsigned long pmd_addr = vmf->address & PMD_MASK;
1475 bool write = vmf->flags & FAULT_FLAG_WRITE;
1477 unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1478 struct inode *inode = mapping->host;
1479 vm_fault_t result = VM_FAULT_FALLBACK;
1480 struct iomap iomap = { 0 };
1488 * Check whether offset isn't beyond end of file now. Caller is
1489 * supposed to hold locks serializing us with truncate / punch hole so
1490 * this is a reliable test.
1492 max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1494 trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1497 * Make sure that the faulting address's PMD offset (color) matches
1498 * the PMD offset from the start of the file. This is necessary so
1499 * that a PMD range in the page table overlaps exactly with a PMD
1500 * range in the page cache.
1502 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1503 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1506 /* Fall back to PTEs if we're going to COW */
1507 if (write && !(vma->vm_flags & VM_SHARED))
1510 /* If the PMD would extend outside the VMA */
1511 if (pmd_addr < vma->vm_start)
1513 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1516 if (xas.xa_index >= max_pgoff) {
1517 result = VM_FAULT_SIGBUS;
1521 /* If the PMD would extend beyond the file size */
1522 if ((xas.xa_index | PG_PMD_COLOUR) >= max_pgoff)
1526 * grab_mapping_entry() will make sure we get an empty PMD entry,
1527 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1528 * entry is already in the array, for instance), it will return
1529 * VM_FAULT_FALLBACK.
1531 entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1532 if (xa_is_internal(entry)) {
1533 result = xa_to_internal(entry);
1538 * It is possible, particularly with mixed reads & writes to private
1539 * mappings, that we have raced with a PTE fault that overlaps with
1540 * the PMD we need to set up. If so just return and the fault will be
1543 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1544 !pmd_devmap(*vmf->pmd)) {
1550 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1551 * setting up a mapping, so really we're using iomap_begin() as a way
1552 * to look up our filesystem block.
1554 pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1555 error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1559 if (iomap.offset + iomap.length < pos + PMD_SIZE)
1562 sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1564 switch (iomap.type) {
1566 error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1570 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1571 DAX_PMD, write && !sync);
1574 * If we are doing synchronous page fault and inode needs fsync,
1575 * we can insert PMD into page tables only after that happens.
1576 * Skip insertion for now and return the pfn so that caller can
1577 * insert it after fsync is done.
1580 if (WARN_ON_ONCE(!pfnp))
1583 result = VM_FAULT_NEEDDSYNC;
1587 trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
1588 result = vmf_insert_pfn_pmd(vmf, pfn, write);
1590 case IOMAP_UNWRITTEN:
1592 if (WARN_ON_ONCE(write))
1594 result = dax_pmd_load_hole(&xas, vmf, &iomap, &entry);
1602 if (ops->iomap_end) {
1603 int copied = PMD_SIZE;
1605 if (result == VM_FAULT_FALLBACK)
1608 * The fault is done by now and there's no way back (other
1609 * thread may be already happily using PMD we have installed).
1610 * Just ignore error from ->iomap_end since we cannot do much
1613 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1617 dax_unlock_entry(&xas, entry);
1619 if (result == VM_FAULT_FALLBACK) {
1620 split_huge_pmd(vma, vmf->pmd, vmf->address);
1621 count_vm_event(THP_FAULT_FALLBACK);
1624 trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1628 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1629 const struct iomap_ops *ops)
1631 return VM_FAULT_FALLBACK;
1633 #endif /* CONFIG_FS_DAX_PMD */
1636 * dax_iomap_fault - handle a page fault on a DAX file
1637 * @vmf: The description of the fault
1638 * @pe_size: Size of the page to fault in
1639 * @pfnp: PFN to insert for synchronous faults if fsync is required
1640 * @iomap_errp: Storage for detailed error code in case of error
1641 * @ops: Iomap ops passed from the file system
1643 * When a page fault occurs, filesystems may call this helper in
1644 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1645 * has done all the necessary locking for page fault to proceed
1648 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1649 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1653 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1655 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1657 return VM_FAULT_FALLBACK;
1660 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1663 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1664 * @vmf: The description of the fault
1665 * @pfn: PFN to insert
1666 * @order: Order of entry to insert.
1668 * This function inserts a writeable PTE or PMD entry into the page tables
1669 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1672 dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1674 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1675 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1680 entry = get_unlocked_entry(&xas, order);
1681 /* Did we race with someone splitting entry or so? */
1682 if (!entry || dax_is_conflict(entry) ||
1683 (order == 0 && !dax_is_pte_entry(entry))) {
1684 put_unlocked_entry(&xas, entry);
1685 xas_unlock_irq(&xas);
1686 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1688 return VM_FAULT_NOPAGE;
1690 xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1691 dax_lock_entry(&xas, entry);
1692 xas_unlock_irq(&xas);
1694 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1695 #ifdef CONFIG_FS_DAX_PMD
1696 else if (order == PMD_ORDER)
1697 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1700 ret = VM_FAULT_FALLBACK;
1701 dax_unlock_entry(&xas, entry);
1702 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1707 * dax_finish_sync_fault - finish synchronous page fault
1708 * @vmf: The description of the fault
1709 * @pe_size: Size of entry to be inserted
1710 * @pfn: PFN to insert
1712 * This function ensures that the file range touched by the page fault is
1713 * stored persistently on the media and handles inserting of appropriate page
1716 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1717 enum page_entry_size pe_size, pfn_t pfn)
1720 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1721 unsigned int order = pe_order(pe_size);
1722 size_t len = PAGE_SIZE << order;
1724 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1726 return VM_FAULT_SIGBUS;
1727 return dax_insert_pfn_mkwrite(vmf, pfn, order);
1729 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);