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>
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.
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
17 #include <linux/atomic.h>
18 #include <linux/blkdev.h>
19 #include <linux/buffer_head.h>
20 #include <linux/dax.h>
22 #include <linux/genhd.h>
23 #include <linux/highmem.h>
24 #include <linux/memcontrol.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>
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/fs_dax.h>
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)
45 /* The 'colour' (ie low bits) within a PMD of a page offset. */
46 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
47 #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
49 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
51 static int __init init_dax_wait_table(void)
55 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
56 init_waitqueue_head(wait_table + i);
59 fs_initcall(init_dax_wait_table);
62 * We use lowest available bit in exceptional entry for locking, one bit for
63 * the entry size (PMD) and two more to tell us if the entry is a zero page or
64 * an empty entry that is just used for locking. In total four special bits.
66 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
67 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
70 #define RADIX_DAX_SHIFT (RADIX_TREE_EXCEPTIONAL_SHIFT + 4)
71 #define RADIX_DAX_ENTRY_LOCK (1 << RADIX_TREE_EXCEPTIONAL_SHIFT)
72 #define RADIX_DAX_PMD (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1))
73 #define RADIX_DAX_ZERO_PAGE (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2))
74 #define RADIX_DAX_EMPTY (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 3))
76 static unsigned long dax_radix_pfn(void *entry)
78 return (unsigned long)entry >> RADIX_DAX_SHIFT;
81 static void *dax_radix_locked_entry(unsigned long pfn, unsigned long flags)
83 return (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY | flags |
84 (pfn << RADIX_DAX_SHIFT) | RADIX_DAX_ENTRY_LOCK);
87 static unsigned int dax_radix_order(void *entry)
89 if ((unsigned long)entry & RADIX_DAX_PMD)
90 return PMD_SHIFT - PAGE_SHIFT;
94 static int dax_is_pmd_entry(void *entry)
96 return (unsigned long)entry & RADIX_DAX_PMD;
99 static int dax_is_pte_entry(void *entry)
101 return !((unsigned long)entry & RADIX_DAX_PMD);
104 static int dax_is_zero_entry(void *entry)
106 return (unsigned long)entry & RADIX_DAX_ZERO_PAGE;
109 static int dax_is_empty_entry(void *entry)
111 return (unsigned long)entry & RADIX_DAX_EMPTY;
115 * DAX radix tree locking
117 struct exceptional_entry_key {
118 struct address_space *mapping;
122 struct wait_exceptional_entry_queue {
123 wait_queue_entry_t wait;
124 struct exceptional_entry_key key;
127 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
128 pgoff_t index, void *entry, struct exceptional_entry_key *key)
133 * If 'entry' is a PMD, align the 'index' that we use for the wait
134 * queue to the start of that PMD. This ensures that all offsets in
135 * the range covered by the PMD map to the same bit lock.
137 if (dax_is_pmd_entry(entry))
138 index &= ~PG_PMD_COLOUR;
140 key->mapping = mapping;
141 key->entry_start = index;
143 hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
144 return wait_table + hash;
147 static int wake_exceptional_entry_func(wait_queue_entry_t *wait, unsigned int mode,
148 int sync, void *keyp)
150 struct exceptional_entry_key *key = keyp;
151 struct wait_exceptional_entry_queue *ewait =
152 container_of(wait, struct wait_exceptional_entry_queue, wait);
154 if (key->mapping != ewait->key.mapping ||
155 key->entry_start != ewait->key.entry_start)
157 return autoremove_wake_function(wait, mode, sync, NULL);
161 * @entry may no longer be the entry at the index in the mapping.
162 * The important information it's conveying is whether the entry at
163 * this index used to be a PMD entry.
165 static void dax_wake_mapping_entry_waiter(struct address_space *mapping,
166 pgoff_t index, void *entry, bool wake_all)
168 struct exceptional_entry_key key;
169 wait_queue_head_t *wq;
171 wq = dax_entry_waitqueue(mapping, index, entry, &key);
174 * Checking for locked entry and prepare_to_wait_exclusive() happens
175 * under the i_pages lock, ditto for entry handling in our callers.
176 * So at this point all tasks that could have seen our entry locked
177 * must be in the waitqueue and the following check will see them.
179 if (waitqueue_active(wq))
180 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
184 * Check whether the given slot is locked. Must be called with the i_pages
187 static inline int slot_locked(struct address_space *mapping, void **slot)
189 unsigned long entry = (unsigned long)
190 radix_tree_deref_slot_protected(slot, &mapping->i_pages.xa_lock);
191 return entry & RADIX_DAX_ENTRY_LOCK;
195 * Mark the given slot as locked. Must be called with the i_pages lock held.
197 static inline void *lock_slot(struct address_space *mapping, void **slot)
199 unsigned long entry = (unsigned long)
200 radix_tree_deref_slot_protected(slot, &mapping->i_pages.xa_lock);
202 entry |= RADIX_DAX_ENTRY_LOCK;
203 radix_tree_replace_slot(&mapping->i_pages, slot, (void *)entry);
204 return (void *)entry;
208 * Mark the given slot as unlocked. Must be called with the i_pages lock held.
210 static inline void *unlock_slot(struct address_space *mapping, void **slot)
212 unsigned long entry = (unsigned long)
213 radix_tree_deref_slot_protected(slot, &mapping->i_pages.xa_lock);
215 entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
216 radix_tree_replace_slot(&mapping->i_pages, slot, (void *)entry);
217 return (void *)entry;
221 * Lookup entry in radix tree, wait for it to become unlocked if it is
222 * exceptional entry and return it. The caller must call
223 * put_unlocked_mapping_entry() when he decided not to lock the entry or
224 * put_locked_mapping_entry() when he locked the entry and now wants to
227 * Must be called with the i_pages lock held.
229 static void *get_unlocked_mapping_entry(struct address_space *mapping,
230 pgoff_t index, void ***slotp)
233 struct wait_exceptional_entry_queue ewait;
234 wait_queue_head_t *wq;
236 init_wait(&ewait.wait);
237 ewait.wait.func = wake_exceptional_entry_func;
240 entry = __radix_tree_lookup(&mapping->i_pages, index, NULL,
243 WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)) ||
244 !slot_locked(mapping, slot)) {
250 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
251 prepare_to_wait_exclusive(wq, &ewait.wait,
252 TASK_UNINTERRUPTIBLE);
253 xa_unlock_irq(&mapping->i_pages);
255 finish_wait(wq, &ewait.wait);
256 xa_lock_irq(&mapping->i_pages);
260 static void dax_unlock_mapping_entry(struct address_space *mapping,
265 xa_lock_irq(&mapping->i_pages);
266 entry = __radix_tree_lookup(&mapping->i_pages, index, NULL, &slot);
267 if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
268 !slot_locked(mapping, slot))) {
269 xa_unlock_irq(&mapping->i_pages);
272 unlock_slot(mapping, slot);
273 xa_unlock_irq(&mapping->i_pages);
274 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
277 static void put_locked_mapping_entry(struct address_space *mapping,
280 dax_unlock_mapping_entry(mapping, index);
284 * Called when we are done with radix tree entry we looked up via
285 * get_unlocked_mapping_entry() and which we didn't lock in the end.
287 static void put_unlocked_mapping_entry(struct address_space *mapping,
288 pgoff_t index, void *entry)
293 /* We have to wake up next waiter for the radix tree entry lock */
294 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
297 static unsigned long dax_entry_size(void *entry)
299 if (dax_is_zero_entry(entry))
301 else if (dax_is_empty_entry(entry))
303 else if (dax_is_pmd_entry(entry))
309 static unsigned long dax_radix_end_pfn(void *entry)
311 return dax_radix_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
315 * Iterate through all mapped pfns represented by an entry, i.e. skip
316 * 'empty' and 'zero' entries.
318 #define for_each_mapped_pfn(entry, pfn) \
319 for (pfn = dax_radix_pfn(entry); \
320 pfn < dax_radix_end_pfn(entry); pfn++)
322 static void dax_associate_entry(void *entry, struct address_space *mapping)
326 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
329 for_each_mapped_pfn(entry, pfn) {
330 struct page *page = pfn_to_page(pfn);
332 WARN_ON_ONCE(page->mapping);
333 page->mapping = mapping;
337 static void dax_disassociate_entry(void *entry, struct address_space *mapping,
342 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
345 for_each_mapped_pfn(entry, pfn) {
346 struct page *page = pfn_to_page(pfn);
348 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
349 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
350 page->mapping = NULL;
354 static struct page *dax_busy_page(void *entry)
358 for_each_mapped_pfn(entry, pfn) {
359 struct page *page = pfn_to_page(pfn);
361 if (page_ref_count(page) > 1)
368 * Find radix tree entry at given index. If it points to an exceptional entry,
369 * return it with the radix tree entry locked. If the radix tree doesn't
370 * contain given index, create an empty exceptional entry for the index and
371 * return with it locked.
373 * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
374 * either return that locked entry or will return an error. This error will
375 * happen if there are any 4k entries within the 2MiB range that we are
378 * We always favor 4k entries over 2MiB entries. There isn't a flow where we
379 * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB
380 * insertion will fail if it finds any 4k entries already in the tree, and a
381 * 4k insertion will cause an existing 2MiB entry to be unmapped and
382 * downgraded to 4k entries. This happens for both 2MiB huge zero pages as
383 * well as 2MiB empty entries.
385 * The exception to this downgrade path is for 2MiB DAX PMD entries that have
386 * real storage backing them. We will leave these real 2MiB DAX entries in
387 * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
389 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
390 * persistent memory the benefit is doubtful. We can add that later if we can
393 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
394 unsigned long size_flag)
396 bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
400 xa_lock_irq(&mapping->i_pages);
401 entry = get_unlocked_mapping_entry(mapping, index, &slot);
403 if (WARN_ON_ONCE(entry && !radix_tree_exceptional_entry(entry))) {
404 entry = ERR_PTR(-EIO);
409 if (size_flag & RADIX_DAX_PMD) {
410 if (dax_is_pte_entry(entry)) {
411 put_unlocked_mapping_entry(mapping, index,
413 entry = ERR_PTR(-EEXIST);
416 } else { /* trying to grab a PTE entry */
417 if (dax_is_pmd_entry(entry) &&
418 (dax_is_zero_entry(entry) ||
419 dax_is_empty_entry(entry))) {
420 pmd_downgrade = true;
425 /* No entry for given index? Make sure radix tree is big enough. */
426 if (!entry || pmd_downgrade) {
431 * Make sure 'entry' remains valid while we drop
434 entry = lock_slot(mapping, slot);
437 xa_unlock_irq(&mapping->i_pages);
439 * Besides huge zero pages the only other thing that gets
440 * downgraded are empty entries which don't need to be
443 if (pmd_downgrade && dax_is_zero_entry(entry))
444 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
447 err = radix_tree_preload(
448 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
451 put_locked_mapping_entry(mapping, index);
454 xa_lock_irq(&mapping->i_pages);
458 * We needed to drop the i_pages lock while calling
459 * radix_tree_preload() and we didn't have an entry to
460 * lock. See if another thread inserted an entry at
461 * our index during this time.
463 entry = __radix_tree_lookup(&mapping->i_pages, index,
466 radix_tree_preload_end();
467 xa_unlock_irq(&mapping->i_pages);
473 dax_disassociate_entry(entry, mapping, false);
474 radix_tree_delete(&mapping->i_pages, index);
475 mapping->nrexceptional--;
476 dax_wake_mapping_entry_waiter(mapping, index, entry,
480 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
482 err = __radix_tree_insert(&mapping->i_pages, index,
483 dax_radix_order(entry), entry);
484 radix_tree_preload_end();
486 xa_unlock_irq(&mapping->i_pages);
488 * Our insertion of a DAX entry failed, most likely
489 * because we were inserting a PMD entry and it
490 * collided with a PTE sized entry at a different
491 * index in the PMD range. We haven't inserted
492 * anything into the radix tree and have no waiters to
497 /* Good, we have inserted empty locked entry into the tree. */
498 mapping->nrexceptional++;
499 xa_unlock_irq(&mapping->i_pages);
502 entry = lock_slot(mapping, slot);
504 xa_unlock_irq(&mapping->i_pages);
509 * dax_layout_busy_page - find first pinned page in @mapping
510 * @mapping: address space to scan for a page with ref count > 1
512 * DAX requires ZONE_DEVICE mapped pages. These pages are never
513 * 'onlined' to the page allocator so they are considered idle when
514 * page->count == 1. A filesystem uses this interface to determine if
515 * any page in the mapping is busy, i.e. for DMA, or other
516 * get_user_pages() usages.
518 * It is expected that the filesystem is holding locks to block the
519 * establishment of new mappings in this address_space. I.e. it expects
520 * to be able to run unmap_mapping_range() and subsequently not race
521 * mapping_mapped() becoming true.
523 struct page *dax_layout_busy_page(struct address_space *mapping)
525 pgoff_t indices[PAGEVEC_SIZE];
526 struct page *page = NULL;
532 * In the 'limited' case get_user_pages() for dax is disabled.
534 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
537 if (!dax_mapping(mapping) || !mapping_mapped(mapping))
545 * If we race get_user_pages_fast() here either we'll see the
546 * elevated page count in the pagevec_lookup and wait, or
547 * get_user_pages_fast() will see that the page it took a reference
548 * against is no longer mapped in the page tables and bail to the
549 * get_user_pages() slow path. The slow path is protected by
550 * pte_lock() and pmd_lock(). New references are not taken without
551 * holding those locks, and unmap_mapping_range() will not zero the
552 * pte or pmd without holding the respective lock, so we are
553 * guaranteed to either see new references or prevent new
554 * references from being established.
556 unmap_mapping_range(mapping, 0, 0, 1);
558 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
559 min(end - index, (pgoff_t)PAGEVEC_SIZE),
561 for (i = 0; i < pagevec_count(&pvec); i++) {
562 struct page *pvec_ent = pvec.pages[i];
570 !radix_tree_exceptional_entry(pvec_ent)))
573 xa_lock_irq(&mapping->i_pages);
574 entry = get_unlocked_mapping_entry(mapping, index, NULL);
576 page = dax_busy_page(entry);
577 put_unlocked_mapping_entry(mapping, index, entry);
578 xa_unlock_irq(&mapping->i_pages);
584 * We don't expect normal struct page entries to exist in our
585 * tree, but we keep these pagevec calls so that this code is
586 * consistent with the common pattern for handling pagevecs
587 * throughout the kernel.
589 pagevec_remove_exceptionals(&pvec);
590 pagevec_release(&pvec);
598 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
600 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
601 pgoff_t index, bool trunc)
605 struct radix_tree_root *pages = &mapping->i_pages;
608 entry = get_unlocked_mapping_entry(mapping, index, NULL);
609 if (!entry || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)))
612 (radix_tree_tag_get(pages, index, PAGECACHE_TAG_DIRTY) ||
613 radix_tree_tag_get(pages, index, PAGECACHE_TAG_TOWRITE)))
615 dax_disassociate_entry(entry, mapping, trunc);
616 radix_tree_delete(pages, index);
617 mapping->nrexceptional--;
620 put_unlocked_mapping_entry(mapping, index, entry);
621 xa_unlock_irq(pages);
625 * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
626 * entry to get unlocked before deleting it.
628 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
630 int ret = __dax_invalidate_mapping_entry(mapping, index, true);
633 * This gets called from truncate / punch_hole path. As such, the caller
634 * must hold locks protecting against concurrent modifications of the
635 * radix tree (usually fs-private i_mmap_sem for writing). Since the
636 * caller has seen exceptional entry for this index, we better find it
637 * at that index as well...
644 * Invalidate exceptional DAX entry if it is clean.
646 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
649 return __dax_invalidate_mapping_entry(mapping, index, false);
652 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
653 sector_t sector, size_t size, struct page *to,
661 rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
665 id = dax_read_lock();
666 rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, NULL);
671 vto = kmap_atomic(to);
672 copy_user_page(vto, (void __force *)kaddr, vaddr, to);
679 * By this point grab_mapping_entry() has ensured that we have a locked entry
680 * of the appropriate size so we don't have to worry about downgrading PMDs to
681 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
682 * already in the tree, we will skip the insertion and just dirty the PMD as
685 static void *dax_insert_mapping_entry(struct address_space *mapping,
686 struct vm_fault *vmf,
687 void *entry, pfn_t pfn_t,
688 unsigned long flags, bool dirty)
690 struct radix_tree_root *pages = &mapping->i_pages;
691 unsigned long pfn = pfn_t_to_pfn(pfn_t);
692 pgoff_t index = vmf->pgoff;
696 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
698 if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_ZERO_PAGE)) {
699 /* we are replacing a zero page with block mapping */
700 if (dax_is_pmd_entry(entry))
701 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
704 unmap_mapping_pages(mapping, vmf->pgoff, 1, false);
708 new_entry = dax_radix_locked_entry(pfn, flags);
709 if (dax_entry_size(entry) != dax_entry_size(new_entry)) {
710 dax_disassociate_entry(entry, mapping, false);
711 dax_associate_entry(new_entry, mapping);
714 if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
716 * Only swap our new entry into the radix tree if the current
717 * entry is a zero page or an empty entry. If a normal PTE or
718 * PMD entry is already in the tree, we leave it alone. This
719 * means that if we are trying to insert a PTE and the
720 * existing entry is a PMD, we will just leave the PMD in the
721 * tree and dirty it if necessary.
723 struct radix_tree_node *node;
727 ret = __radix_tree_lookup(pages, index, &node, &slot);
728 WARN_ON_ONCE(ret != entry);
729 __radix_tree_replace(pages, node, slot,
735 radix_tree_tag_set(pages, index, PAGECACHE_TAG_DIRTY);
737 xa_unlock_irq(pages);
741 static inline unsigned long
742 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
744 unsigned long address;
746 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
747 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
751 /* Walk all mappings of a given index of a file and writeprotect them */
752 static void dax_mapping_entry_mkclean(struct address_space *mapping,
753 pgoff_t index, unsigned long pfn)
755 struct vm_area_struct *vma;
756 pte_t pte, *ptep = NULL;
760 i_mmap_lock_read(mapping);
761 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
762 unsigned long address, start, end;
766 if (!(vma->vm_flags & VM_SHARED))
769 address = pgoff_address(index, vma);
772 * Note because we provide start/end to follow_pte_pmd it will
773 * call mmu_notifier_invalidate_range_start() on our behalf
774 * before taking any lock.
776 if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl))
780 * No need to call mmu_notifier_invalidate_range() as we are
781 * downgrading page table protection not changing it to point
784 * See Documentation/vm/mmu_notifier.rst
787 #ifdef CONFIG_FS_DAX_PMD
790 if (pfn != pmd_pfn(*pmdp))
792 if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
795 flush_cache_page(vma, address, pfn);
796 pmd = pmdp_huge_clear_flush(vma, address, pmdp);
797 pmd = pmd_wrprotect(pmd);
798 pmd = pmd_mkclean(pmd);
799 set_pmd_at(vma->vm_mm, address, pmdp, pmd);
804 if (pfn != pte_pfn(*ptep))
806 if (!pte_dirty(*ptep) && !pte_write(*ptep))
809 flush_cache_page(vma, address, pfn);
810 pte = ptep_clear_flush(vma, address, ptep);
811 pte = pte_wrprotect(pte);
812 pte = pte_mkclean(pte);
813 set_pte_at(vma->vm_mm, address, ptep, pte);
815 pte_unmap_unlock(ptep, ptl);
818 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
820 i_mmap_unlock_read(mapping);
823 static int dax_writeback_one(struct dax_device *dax_dev,
824 struct address_space *mapping, pgoff_t index, void *entry)
826 struct radix_tree_root *pages = &mapping->i_pages;
827 void *entry2, **slot;
833 * A page got tagged dirty in DAX mapping? Something is seriously
836 if (WARN_ON(!radix_tree_exceptional_entry(entry)))
840 entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
841 /* Entry got punched out / reallocated? */
842 if (!entry2 || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry2)))
845 * Entry got reallocated elsewhere? No need to writeback. We have to
846 * compare pfns as we must not bail out due to difference in lockbit
849 if (dax_radix_pfn(entry2) != dax_radix_pfn(entry))
851 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
852 dax_is_zero_entry(entry))) {
857 /* Another fsync thread may have already written back this entry */
858 if (!radix_tree_tag_get(pages, index, PAGECACHE_TAG_TOWRITE))
860 /* Lock the entry to serialize with page faults */
861 entry = lock_slot(mapping, slot);
863 * We can clear the tag now but we have to be careful so that concurrent
864 * dax_writeback_one() calls for the same index cannot finish before we
865 * actually flush the caches. This is achieved as the calls will look
866 * at the entry only under the i_pages lock and once they do that
867 * they will see the entry locked and wait for it to unlock.
869 radix_tree_tag_clear(pages, index, PAGECACHE_TAG_TOWRITE);
870 xa_unlock_irq(pages);
873 * Even if dax_writeback_mapping_range() was given a wbc->range_start
874 * in the middle of a PMD, the 'index' we are given will be aligned to
875 * the start index of the PMD, as will the pfn we pull from 'entry'.
876 * This allows us to flush for PMD_SIZE and not have to worry about
877 * partial PMD writebacks.
879 pfn = dax_radix_pfn(entry);
880 size = PAGE_SIZE << dax_radix_order(entry);
882 dax_mapping_entry_mkclean(mapping, index, pfn);
883 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), size);
885 * After we have flushed the cache, we can clear the dirty tag. There
886 * cannot be new dirty data in the pfn after the flush has completed as
887 * the pfn mappings are writeprotected and fault waits for mapping
891 radix_tree_tag_clear(pages, index, PAGECACHE_TAG_DIRTY);
892 xa_unlock_irq(pages);
893 trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT);
894 put_locked_mapping_entry(mapping, index);
898 put_unlocked_mapping_entry(mapping, index, entry2);
899 xa_unlock_irq(pages);
904 * Flush the mapping to the persistent domain within the byte range of [start,
905 * end]. This is required by data integrity operations to ensure file data is
906 * on persistent storage prior to completion of the operation.
908 int dax_writeback_mapping_range(struct address_space *mapping,
909 struct block_device *bdev, struct writeback_control *wbc)
911 struct inode *inode = mapping->host;
912 pgoff_t start_index, end_index;
913 pgoff_t indices[PAGEVEC_SIZE];
914 struct dax_device *dax_dev;
919 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
922 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
925 dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
929 start_index = wbc->range_start >> PAGE_SHIFT;
930 end_index = wbc->range_end >> PAGE_SHIFT;
932 trace_dax_writeback_range(inode, start_index, end_index);
934 tag_pages_for_writeback(mapping, start_index, end_index);
938 pvec.nr = find_get_entries_tag(mapping, start_index,
939 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
940 pvec.pages, indices);
945 for (i = 0; i < pvec.nr; i++) {
946 if (indices[i] > end_index) {
951 ret = dax_writeback_one(dax_dev, mapping, indices[i],
954 mapping_set_error(mapping, ret);
958 start_index = indices[pvec.nr - 1] + 1;
962 trace_dax_writeback_range_done(inode, start_index, end_index);
963 return (ret < 0 ? ret : 0);
965 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
967 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
969 return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
972 static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
975 const sector_t sector = dax_iomap_sector(iomap, pos);
980 rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
983 id = dax_read_lock();
984 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
991 if (PFN_PHYS(length) < size)
993 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
995 /* For larger pages we need devmap */
996 if (length > 1 && !pfn_t_devmap(*pfnp))
1000 dax_read_unlock(id);
1005 * The user has performed a load from a hole in the file. Allocating a new
1006 * page in the file would cause excessive storage usage for workloads with
1007 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1008 * If this page is ever written to we will re-fault and change the mapping to
1009 * point to real DAX storage instead.
1011 static vm_fault_t dax_load_hole(struct address_space *mapping, void *entry,
1012 struct vm_fault *vmf)
1014 struct inode *inode = mapping->host;
1015 unsigned long vaddr = vmf->address;
1016 vm_fault_t ret = VM_FAULT_NOPAGE;
1017 struct page *zero_page;
1020 zero_page = ZERO_PAGE(0);
1021 if (unlikely(!zero_page)) {
1026 pfn = page_to_pfn_t(zero_page);
1027 dax_insert_mapping_entry(mapping, vmf, entry, pfn, RADIX_DAX_ZERO_PAGE,
1029 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1031 trace_dax_load_hole(inode, vmf, ret);
1035 static bool dax_range_is_aligned(struct block_device *bdev,
1036 unsigned int offset, unsigned int length)
1038 unsigned short sector_size = bdev_logical_block_size(bdev);
1040 if (!IS_ALIGNED(offset, sector_size))
1042 if (!IS_ALIGNED(length, sector_size))
1048 int __dax_zero_page_range(struct block_device *bdev,
1049 struct dax_device *dax_dev, sector_t sector,
1050 unsigned int offset, unsigned int size)
1052 if (dax_range_is_aligned(bdev, offset, size)) {
1053 sector_t start_sector = sector + (offset >> 9);
1055 return blkdev_issue_zeroout(bdev, start_sector,
1056 size >> 9, GFP_NOFS, 0);
1062 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
1066 id = dax_read_lock();
1067 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, NULL);
1069 dax_read_unlock(id);
1072 memset(kaddr + offset, 0, size);
1073 dax_flush(dax_dev, kaddr + offset, size);
1074 dax_read_unlock(id);
1078 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1081 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1082 struct iomap *iomap)
1084 struct block_device *bdev = iomap->bdev;
1085 struct dax_device *dax_dev = iomap->dax_dev;
1086 struct iov_iter *iter = data;
1087 loff_t end = pos + length, done = 0;
1092 if (iov_iter_rw(iter) == READ) {
1093 end = min(end, i_size_read(inode));
1097 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1098 return iov_iter_zero(min(length, end - pos), iter);
1101 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1105 * Write can allocate block for an area which has a hole page mapped
1106 * into page tables. We have to tear down these mappings so that data
1107 * written by write(2) is visible in mmap.
1109 if (iomap->flags & IOMAP_F_NEW) {
1110 invalidate_inode_pages2_range(inode->i_mapping,
1112 (end - 1) >> PAGE_SHIFT);
1115 id = dax_read_lock();
1117 unsigned offset = pos & (PAGE_SIZE - 1);
1118 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1119 const sector_t sector = dax_iomap_sector(iomap, pos);
1124 if (fatal_signal_pending(current)) {
1129 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1133 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1140 map_len = PFN_PHYS(map_len);
1143 if (map_len > end - pos)
1144 map_len = end - pos;
1147 * The userspace address for the memory copy has already been
1148 * validated via access_ok() in either vfs_read() or
1149 * vfs_write(), depending on which operation we are doing.
1151 if (iov_iter_rw(iter) == WRITE)
1152 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1155 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1167 dax_read_unlock(id);
1169 return done ? done : ret;
1173 * dax_iomap_rw - Perform I/O to a DAX file
1174 * @iocb: The control block for this I/O
1175 * @iter: The addresses to do I/O from or to
1176 * @ops: iomap ops passed from the file system
1178 * This function performs read and write operations to directly mapped
1179 * persistent memory. The callers needs to take care of read/write exclusion
1180 * and evicting any page cache pages in the region under I/O.
1183 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1184 const struct iomap_ops *ops)
1186 struct address_space *mapping = iocb->ki_filp->f_mapping;
1187 struct inode *inode = mapping->host;
1188 loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1191 if (iov_iter_rw(iter) == WRITE) {
1192 lockdep_assert_held_exclusive(&inode->i_rwsem);
1193 flags |= IOMAP_WRITE;
1195 lockdep_assert_held(&inode->i_rwsem);
1198 while (iov_iter_count(iter)) {
1199 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1200 iter, dax_iomap_actor);
1207 iocb->ki_pos += done;
1208 return done ? done : ret;
1210 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1212 static vm_fault_t dax_fault_return(int error)
1215 return VM_FAULT_NOPAGE;
1216 if (error == -ENOMEM)
1217 return VM_FAULT_OOM;
1218 return VM_FAULT_SIGBUS;
1222 * MAP_SYNC on a dax mapping guarantees dirty metadata is
1223 * flushed on write-faults (non-cow), but not read-faults.
1225 static bool dax_fault_is_synchronous(unsigned long flags,
1226 struct vm_area_struct *vma, struct iomap *iomap)
1228 return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1229 && (iomap->flags & IOMAP_F_DIRTY);
1232 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1233 int *iomap_errp, const struct iomap_ops *ops)
1235 struct vm_area_struct *vma = vmf->vma;
1236 struct address_space *mapping = vma->vm_file->f_mapping;
1237 struct inode *inode = mapping->host;
1238 unsigned long vaddr = vmf->address;
1239 loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1240 struct iomap iomap = { 0 };
1241 unsigned flags = IOMAP_FAULT;
1242 int error, major = 0;
1243 bool write = vmf->flags & FAULT_FLAG_WRITE;
1249 trace_dax_pte_fault(inode, vmf, ret);
1251 * Check whether offset isn't beyond end of file now. Caller is supposed
1252 * to hold locks serializing us with truncate / punch hole so this is
1255 if (pos >= i_size_read(inode)) {
1256 ret = VM_FAULT_SIGBUS;
1260 if (write && !vmf->cow_page)
1261 flags |= IOMAP_WRITE;
1263 entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1264 if (IS_ERR(entry)) {
1265 ret = dax_fault_return(PTR_ERR(entry));
1270 * It is possible, particularly with mixed reads & writes to private
1271 * mappings, that we have raced with a PMD fault that overlaps with
1272 * the PTE we need to set up. If so just return and the fault will be
1275 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1276 ret = VM_FAULT_NOPAGE;
1281 * Note that we don't bother to use iomap_apply here: DAX required
1282 * the file system block size to be equal the page size, which means
1283 * that we never have to deal with more than a single extent here.
1285 error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1287 *iomap_errp = error;
1289 ret = dax_fault_return(error);
1292 if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1293 error = -EIO; /* fs corruption? */
1294 goto error_finish_iomap;
1297 if (vmf->cow_page) {
1298 sector_t sector = dax_iomap_sector(&iomap, pos);
1300 switch (iomap.type) {
1302 case IOMAP_UNWRITTEN:
1303 clear_user_highpage(vmf->cow_page, vaddr);
1306 error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1307 sector, PAGE_SIZE, vmf->cow_page, vaddr);
1316 goto error_finish_iomap;
1318 __SetPageUptodate(vmf->cow_page);
1319 ret = finish_fault(vmf);
1321 ret = VM_FAULT_DONE_COW;
1325 sync = dax_fault_is_synchronous(flags, vma, &iomap);
1327 switch (iomap.type) {
1329 if (iomap.flags & IOMAP_F_NEW) {
1330 count_vm_event(PGMAJFAULT);
1331 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1332 major = VM_FAULT_MAJOR;
1334 error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1336 goto error_finish_iomap;
1338 entry = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1342 * If we are doing synchronous page fault and inode needs fsync,
1343 * we can insert PTE into page tables only after that happens.
1344 * Skip insertion for now and return the pfn so that caller can
1345 * insert it after fsync is done.
1348 if (WARN_ON_ONCE(!pfnp)) {
1350 goto error_finish_iomap;
1353 ret = VM_FAULT_NEEDDSYNC | major;
1356 trace_dax_insert_mapping(inode, vmf, entry);
1358 ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn);
1360 ret = vmf_insert_mixed(vma, vaddr, pfn);
1363 case IOMAP_UNWRITTEN:
1366 ret = dax_load_hole(mapping, entry, vmf);
1377 ret = dax_fault_return(error);
1379 if (ops->iomap_end) {
1380 int copied = PAGE_SIZE;
1382 if (ret & VM_FAULT_ERROR)
1385 * The fault is done by now and there's no way back (other
1386 * thread may be already happily using PTE we have installed).
1387 * Just ignore error from ->iomap_end since we cannot do much
1390 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1393 put_locked_mapping_entry(mapping, vmf->pgoff);
1395 trace_dax_pte_fault_done(inode, vmf, ret);
1399 #ifdef CONFIG_FS_DAX_PMD
1400 static vm_fault_t dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
1403 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1404 unsigned long pmd_addr = vmf->address & PMD_MASK;
1405 struct inode *inode = mapping->host;
1406 struct page *zero_page;
1412 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1414 if (unlikely(!zero_page))
1417 pfn = page_to_pfn_t(zero_page);
1418 ret = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1419 RADIX_DAX_PMD | RADIX_DAX_ZERO_PAGE, false);
1421 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1422 if (!pmd_none(*(vmf->pmd))) {
1427 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1428 pmd_entry = pmd_mkhuge(pmd_entry);
1429 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1431 trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
1432 return VM_FAULT_NOPAGE;
1435 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
1436 return VM_FAULT_FALLBACK;
1439 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1440 const struct iomap_ops *ops)
1442 struct vm_area_struct *vma = vmf->vma;
1443 struct address_space *mapping = vma->vm_file->f_mapping;
1444 unsigned long pmd_addr = vmf->address & PMD_MASK;
1445 bool write = vmf->flags & FAULT_FLAG_WRITE;
1447 unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1448 struct inode *inode = mapping->host;
1449 vm_fault_t result = VM_FAULT_FALLBACK;
1450 struct iomap iomap = { 0 };
1451 pgoff_t max_pgoff, pgoff;
1458 * Check whether offset isn't beyond end of file now. Caller is
1459 * supposed to hold locks serializing us with truncate / punch hole so
1460 * this is a reliable test.
1462 pgoff = linear_page_index(vma, pmd_addr);
1463 max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1465 trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1468 * Make sure that the faulting address's PMD offset (color) matches
1469 * the PMD offset from the start of the file. This is necessary so
1470 * that a PMD range in the page table overlaps exactly with a PMD
1471 * range in the radix tree.
1473 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1474 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1477 /* Fall back to PTEs if we're going to COW */
1478 if (write && !(vma->vm_flags & VM_SHARED))
1481 /* If the PMD would extend outside the VMA */
1482 if (pmd_addr < vma->vm_start)
1484 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1487 if (pgoff >= max_pgoff) {
1488 result = VM_FAULT_SIGBUS;
1492 /* If the PMD would extend beyond the file size */
1493 if ((pgoff | PG_PMD_COLOUR) >= max_pgoff)
1497 * grab_mapping_entry() will make sure we get a 2MiB empty entry, a
1498 * 2MiB zero page entry or a DAX PMD. If it can't (because a 4k page
1499 * is already in the tree, for instance), it will return -EEXIST and
1500 * we just fall back to 4k entries.
1502 entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1507 * It is possible, particularly with mixed reads & writes to private
1508 * mappings, that we have raced with a PTE fault that overlaps with
1509 * the PMD we need to set up. If so just return and the fault will be
1512 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1513 !pmd_devmap(*vmf->pmd)) {
1519 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1520 * setting up a mapping, so really we're using iomap_begin() as a way
1521 * to look up our filesystem block.
1523 pos = (loff_t)pgoff << PAGE_SHIFT;
1524 error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1528 if (iomap.offset + iomap.length < pos + PMD_SIZE)
1531 sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1533 switch (iomap.type) {
1535 error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1539 entry = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1540 RADIX_DAX_PMD, write && !sync);
1543 * If we are doing synchronous page fault and inode needs fsync,
1544 * we can insert PMD into page tables only after that happens.
1545 * Skip insertion for now and return the pfn so that caller can
1546 * insert it after fsync is done.
1549 if (WARN_ON_ONCE(!pfnp))
1552 result = VM_FAULT_NEEDDSYNC;
1556 trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
1557 result = vmf_insert_pfn_pmd(vma, vmf->address, vmf->pmd, pfn,
1560 case IOMAP_UNWRITTEN:
1562 if (WARN_ON_ONCE(write))
1564 result = dax_pmd_load_hole(vmf, &iomap, entry);
1572 if (ops->iomap_end) {
1573 int copied = PMD_SIZE;
1575 if (result == VM_FAULT_FALLBACK)
1578 * The fault is done by now and there's no way back (other
1579 * thread may be already happily using PMD we have installed).
1580 * Just ignore error from ->iomap_end since we cannot do much
1583 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1587 put_locked_mapping_entry(mapping, pgoff);
1589 if (result == VM_FAULT_FALLBACK) {
1590 split_huge_pmd(vma, vmf->pmd, vmf->address);
1591 count_vm_event(THP_FAULT_FALLBACK);
1594 trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1598 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1599 const struct iomap_ops *ops)
1601 return VM_FAULT_FALLBACK;
1603 #endif /* CONFIG_FS_DAX_PMD */
1606 * dax_iomap_fault - handle a page fault on a DAX file
1607 * @vmf: The description of the fault
1608 * @pe_size: Size of the page to fault in
1609 * @pfnp: PFN to insert for synchronous faults if fsync is required
1610 * @iomap_errp: Storage for detailed error code in case of error
1611 * @ops: Iomap ops passed from the file system
1613 * When a page fault occurs, filesystems may call this helper in
1614 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1615 * has done all the necessary locking for page fault to proceed
1618 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1619 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1623 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1625 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1627 return VM_FAULT_FALLBACK;
1630 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1633 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1634 * @vmf: The description of the fault
1635 * @pe_size: Size of entry to be inserted
1636 * @pfn: PFN to insert
1638 * This function inserts writeable PTE or PMD entry into page tables for mmaped
1639 * DAX file. It takes care of marking corresponding radix tree entry as dirty
1642 static vm_fault_t dax_insert_pfn_mkwrite(struct vm_fault *vmf,
1643 enum page_entry_size pe_size,
1646 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1647 void *entry, **slot;
1648 pgoff_t index = vmf->pgoff;
1651 xa_lock_irq(&mapping->i_pages);
1652 entry = get_unlocked_mapping_entry(mapping, index, &slot);
1653 /* Did we race with someone splitting entry or so? */
1655 (pe_size == PE_SIZE_PTE && !dax_is_pte_entry(entry)) ||
1656 (pe_size == PE_SIZE_PMD && !dax_is_pmd_entry(entry))) {
1657 put_unlocked_mapping_entry(mapping, index, entry);
1658 xa_unlock_irq(&mapping->i_pages);
1659 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1661 return VM_FAULT_NOPAGE;
1663 radix_tree_tag_set(&mapping->i_pages, index, PAGECACHE_TAG_DIRTY);
1664 entry = lock_slot(mapping, slot);
1665 xa_unlock_irq(&mapping->i_pages);
1668 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1670 #ifdef CONFIG_FS_DAX_PMD
1672 ret = vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
1677 ret = VM_FAULT_FALLBACK;
1679 put_locked_mapping_entry(mapping, index);
1680 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1685 * dax_finish_sync_fault - finish synchronous page fault
1686 * @vmf: The description of the fault
1687 * @pe_size: Size of entry to be inserted
1688 * @pfn: PFN to insert
1690 * This function ensures that the file range touched by the page fault is
1691 * stored persistently on the media and handles inserting of appropriate page
1694 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1695 enum page_entry_size pe_size, pfn_t pfn)
1698 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1701 if (pe_size == PE_SIZE_PTE)
1703 else if (pe_size == PE_SIZE_PMD)
1707 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1709 return VM_FAULT_SIGBUS;
1710 return dax_insert_pfn_mkwrite(vmf, pe_size, pfn);
1712 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);