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;
220 static void put_unlocked_mapping_entry(struct address_space *mapping,
221 pgoff_t index, void *entry);
224 * Lookup entry in radix tree, wait for it to become unlocked if it is
225 * exceptional entry and return it. The caller must call
226 * put_unlocked_mapping_entry() when he decided not to lock the entry or
227 * put_locked_mapping_entry() when he locked the entry and now wants to
230 * Must be called with the i_pages lock held.
232 static void *get_unlocked_mapping_entry(struct address_space *mapping,
233 pgoff_t index, void ***slotp)
236 struct wait_exceptional_entry_queue ewait;
237 wait_queue_head_t *wq;
239 init_wait(&ewait.wait);
240 ewait.wait.func = wake_exceptional_entry_func;
243 entry = __radix_tree_lookup(&mapping->i_pages, index, NULL,
246 WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)) ||
247 !slot_locked(mapping, slot)) {
253 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
254 prepare_to_wait_exclusive(wq, &ewait.wait,
255 TASK_UNINTERRUPTIBLE);
256 xa_unlock_irq(&mapping->i_pages);
258 finish_wait(wq, &ewait.wait);
259 xa_lock_irq(&mapping->i_pages);
264 * The only thing keeping the address space around is the i_pages lock
265 * (it's cycled in clear_inode() after removing the entries from i_pages)
266 * After we call xas_unlock_irq(), we cannot touch xas->xa.
268 static void wait_entry_unlocked(struct address_space *mapping, pgoff_t index,
269 void ***slotp, void *entry)
271 struct wait_exceptional_entry_queue ewait;
272 wait_queue_head_t *wq;
274 init_wait(&ewait.wait);
275 ewait.wait.func = wake_exceptional_entry_func;
277 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
279 * Unlike get_unlocked_entry() there is no guarantee that this
280 * path ever successfully retrieves an unlocked entry before an
281 * inode dies. Perform a non-exclusive wait in case this path
282 * never successfully performs its own wake up.
284 prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
285 xa_unlock_irq(&mapping->i_pages);
287 finish_wait(wq, &ewait.wait);
290 static void unlock_mapping_entry(struct address_space *mapping, pgoff_t index)
294 xa_lock_irq(&mapping->i_pages);
295 entry = __radix_tree_lookup(&mapping->i_pages, index, NULL, &slot);
296 if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
297 !slot_locked(mapping, slot))) {
298 xa_unlock_irq(&mapping->i_pages);
301 unlock_slot(mapping, slot);
302 xa_unlock_irq(&mapping->i_pages);
303 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
306 static void put_locked_mapping_entry(struct address_space *mapping,
309 unlock_mapping_entry(mapping, index);
313 * Called when we are done with radix tree entry we looked up via
314 * get_unlocked_mapping_entry() and which we didn't lock in the end.
316 static void put_unlocked_mapping_entry(struct address_space *mapping,
317 pgoff_t index, void *entry)
322 /* We have to wake up next waiter for the radix tree entry lock */
323 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
326 static unsigned long dax_entry_size(void *entry)
328 if (dax_is_zero_entry(entry))
330 else if (dax_is_empty_entry(entry))
332 else if (dax_is_pmd_entry(entry))
338 static unsigned long dax_radix_end_pfn(void *entry)
340 return dax_radix_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
344 * Iterate through all mapped pfns represented by an entry, i.e. skip
345 * 'empty' and 'zero' entries.
347 #define for_each_mapped_pfn(entry, pfn) \
348 for (pfn = dax_radix_pfn(entry); \
349 pfn < dax_radix_end_pfn(entry); pfn++)
352 * TODO: for reflink+dax we need a way to associate a single page with
353 * multiple address_space instances at different linear_page_index()
356 static void dax_associate_entry(void *entry, struct address_space *mapping,
357 struct vm_area_struct *vma, unsigned long address)
359 unsigned long size = dax_entry_size(entry), pfn, index;
362 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
365 index = linear_page_index(vma, address & ~(size - 1));
366 for_each_mapped_pfn(entry, pfn) {
367 struct page *page = pfn_to_page(pfn);
369 WARN_ON_ONCE(page->mapping);
370 page->mapping = mapping;
371 page->index = index + i++;
375 static void dax_disassociate_entry(void *entry, struct address_space *mapping,
380 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
383 for_each_mapped_pfn(entry, pfn) {
384 struct page *page = pfn_to_page(pfn);
386 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
387 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
388 page->mapping = NULL;
393 static struct page *dax_busy_page(void *entry)
397 for_each_mapped_pfn(entry, pfn) {
398 struct page *page = pfn_to_page(pfn);
400 if (page_ref_count(page) > 1)
406 bool dax_lock_mapping_entry(struct page *page)
410 bool did_lock = false;
411 void *entry = NULL, **slot;
412 struct address_space *mapping;
416 mapping = READ_ONCE(page->mapping);
418 if (!mapping || !dax_mapping(mapping))
422 * In the device-dax case there's no need to lock, a
423 * struct dev_pagemap pin is sufficient to keep the
424 * inode alive, and we assume we have dev_pagemap pin
425 * otherwise we would not have a valid pfn_to_page()
428 inode = mapping->host;
429 if (S_ISCHR(inode->i_mode)) {
434 xa_lock_irq(&mapping->i_pages);
435 if (mapping != page->mapping) {
436 xa_unlock_irq(&mapping->i_pages);
441 entry = __radix_tree_lookup(&mapping->i_pages, index,
444 xa_unlock_irq(&mapping->i_pages);
446 } else if (slot_locked(mapping, slot)) {
448 wait_entry_unlocked(mapping, index, &slot, entry);
452 lock_slot(mapping, slot);
454 xa_unlock_irq(&mapping->i_pages);
462 void dax_unlock_mapping_entry(struct page *page)
464 struct address_space *mapping = page->mapping;
465 struct inode *inode = mapping->host;
467 if (S_ISCHR(inode->i_mode))
470 unlock_mapping_entry(mapping, page->index);
474 * Find radix tree entry at given index. If it points to an exceptional entry,
475 * return it with the radix tree entry locked. If the radix tree doesn't
476 * contain given index, create an empty exceptional entry for the index and
477 * return with it locked.
479 * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
480 * either return that locked entry or will return an error. This error will
481 * happen if there are any 4k entries within the 2MiB range that we are
484 * We always favor 4k entries over 2MiB entries. There isn't a flow where we
485 * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB
486 * insertion will fail if it finds any 4k entries already in the tree, and a
487 * 4k insertion will cause an existing 2MiB entry to be unmapped and
488 * downgraded to 4k entries. This happens for both 2MiB huge zero pages as
489 * well as 2MiB empty entries.
491 * The exception to this downgrade path is for 2MiB DAX PMD entries that have
492 * real storage backing them. We will leave these real 2MiB DAX entries in
493 * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
495 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
496 * persistent memory the benefit is doubtful. We can add that later if we can
499 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
500 unsigned long size_flag)
502 bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
506 xa_lock_irq(&mapping->i_pages);
507 entry = get_unlocked_mapping_entry(mapping, index, &slot);
509 if (WARN_ON_ONCE(entry && !radix_tree_exceptional_entry(entry))) {
510 entry = ERR_PTR(-EIO);
515 if (size_flag & RADIX_DAX_PMD) {
516 if (dax_is_pte_entry(entry)) {
517 put_unlocked_mapping_entry(mapping, index,
519 entry = ERR_PTR(-EEXIST);
522 } else { /* trying to grab a PTE entry */
523 if (dax_is_pmd_entry(entry) &&
524 (dax_is_zero_entry(entry) ||
525 dax_is_empty_entry(entry))) {
526 pmd_downgrade = true;
531 /* No entry for given index? Make sure radix tree is big enough. */
532 if (!entry || pmd_downgrade) {
537 * Make sure 'entry' remains valid while we drop
540 entry = lock_slot(mapping, slot);
543 xa_unlock_irq(&mapping->i_pages);
545 * Besides huge zero pages the only other thing that gets
546 * downgraded are empty entries which don't need to be
549 if (pmd_downgrade && dax_is_zero_entry(entry))
550 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
553 err = radix_tree_preload(
554 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
557 put_locked_mapping_entry(mapping, index);
560 xa_lock_irq(&mapping->i_pages);
564 * We needed to drop the i_pages lock while calling
565 * radix_tree_preload() and we didn't have an entry to
566 * lock. See if another thread inserted an entry at
567 * our index during this time.
569 entry = __radix_tree_lookup(&mapping->i_pages, index,
572 radix_tree_preload_end();
573 xa_unlock_irq(&mapping->i_pages);
579 dax_disassociate_entry(entry, mapping, false);
580 radix_tree_delete(&mapping->i_pages, index);
581 mapping->nrexceptional--;
582 dax_wake_mapping_entry_waiter(mapping, index, entry,
586 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
588 err = __radix_tree_insert(&mapping->i_pages, index,
589 dax_radix_order(entry), entry);
590 radix_tree_preload_end();
592 xa_unlock_irq(&mapping->i_pages);
594 * Our insertion of a DAX entry failed, most likely
595 * because we were inserting a PMD entry and it
596 * collided with a PTE sized entry at a different
597 * index in the PMD range. We haven't inserted
598 * anything into the radix tree and have no waiters to
603 /* Good, we have inserted empty locked entry into the tree. */
604 mapping->nrexceptional++;
605 xa_unlock_irq(&mapping->i_pages);
608 entry = lock_slot(mapping, slot);
610 xa_unlock_irq(&mapping->i_pages);
615 * dax_layout_busy_page - find first pinned page in @mapping
616 * @mapping: address space to scan for a page with ref count > 1
618 * DAX requires ZONE_DEVICE mapped pages. These pages are never
619 * 'onlined' to the page allocator so they are considered idle when
620 * page->count == 1. A filesystem uses this interface to determine if
621 * any page in the mapping is busy, i.e. for DMA, or other
622 * get_user_pages() usages.
624 * It is expected that the filesystem is holding locks to block the
625 * establishment of new mappings in this address_space. I.e. it expects
626 * to be able to run unmap_mapping_range() and subsequently not race
627 * mapping_mapped() becoming true.
629 struct page *dax_layout_busy_page(struct address_space *mapping)
631 pgoff_t indices[PAGEVEC_SIZE];
632 struct page *page = NULL;
638 * In the 'limited' case get_user_pages() for dax is disabled.
640 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
643 if (!dax_mapping(mapping) || !mapping_mapped(mapping))
651 * If we race get_user_pages_fast() here either we'll see the
652 * elevated page count in the pagevec_lookup and wait, or
653 * get_user_pages_fast() will see that the page it took a reference
654 * against is no longer mapped in the page tables and bail to the
655 * get_user_pages() slow path. The slow path is protected by
656 * pte_lock() and pmd_lock(). New references are not taken without
657 * holding those locks, and unmap_mapping_range() will not zero the
658 * pte or pmd without holding the respective lock, so we are
659 * guaranteed to either see new references or prevent new
660 * references from being established.
662 unmap_mapping_range(mapping, 0, 0, 0);
664 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
665 min(end - index, (pgoff_t)PAGEVEC_SIZE),
667 pgoff_t nr_pages = 1;
669 for (i = 0; i < pagevec_count(&pvec); i++) {
670 struct page *pvec_ent = pvec.pages[i];
678 !radix_tree_exceptional_entry(pvec_ent)))
681 xa_lock_irq(&mapping->i_pages);
682 entry = get_unlocked_mapping_entry(mapping, index, NULL);
684 page = dax_busy_page(entry);
686 * Account for multi-order entries at
687 * the end of the pagevec.
689 if (i + 1 >= pagevec_count(&pvec))
690 nr_pages = 1UL << dax_radix_order(entry);
692 put_unlocked_mapping_entry(mapping, index, entry);
693 xa_unlock_irq(&mapping->i_pages);
699 * We don't expect normal struct page entries to exist in our
700 * tree, but we keep these pagevec calls so that this code is
701 * consistent with the common pattern for handling pagevecs
702 * throughout the kernel.
704 pagevec_remove_exceptionals(&pvec);
705 pagevec_release(&pvec);
713 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
715 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
716 pgoff_t index, bool trunc)
720 struct radix_tree_root *pages = &mapping->i_pages;
723 entry = get_unlocked_mapping_entry(mapping, index, NULL);
724 if (!entry || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)))
727 (radix_tree_tag_get(pages, index, PAGECACHE_TAG_DIRTY) ||
728 radix_tree_tag_get(pages, index, PAGECACHE_TAG_TOWRITE)))
730 dax_disassociate_entry(entry, mapping, trunc);
731 radix_tree_delete(pages, index);
732 mapping->nrexceptional--;
735 put_unlocked_mapping_entry(mapping, index, entry);
736 xa_unlock_irq(pages);
740 * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
741 * entry to get unlocked before deleting it.
743 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
745 int ret = __dax_invalidate_mapping_entry(mapping, index, true);
748 * This gets called from truncate / punch_hole path. As such, the caller
749 * must hold locks protecting against concurrent modifications of the
750 * radix tree (usually fs-private i_mmap_sem for writing). Since the
751 * caller has seen exceptional entry for this index, we better find it
752 * at that index as well...
759 * Invalidate exceptional DAX entry if it is clean.
761 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
764 return __dax_invalidate_mapping_entry(mapping, index, false);
767 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
768 sector_t sector, size_t size, struct page *to,
776 rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
780 id = dax_read_lock();
781 rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, NULL);
786 vto = kmap_atomic(to);
787 copy_user_page(vto, (void __force *)kaddr, vaddr, to);
794 * By this point grab_mapping_entry() has ensured that we have a locked entry
795 * of the appropriate size so we don't have to worry about downgrading PMDs to
796 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
797 * already in the tree, we will skip the insertion and just dirty the PMD as
800 static void *dax_insert_mapping_entry(struct address_space *mapping,
801 struct vm_fault *vmf,
802 void *entry, pfn_t pfn_t,
803 unsigned long flags, bool dirty)
805 struct radix_tree_root *pages = &mapping->i_pages;
806 unsigned long pfn = pfn_t_to_pfn(pfn_t);
807 pgoff_t index = vmf->pgoff;
811 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
813 if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_ZERO_PAGE)) {
814 /* we are replacing a zero page with block mapping */
815 if (dax_is_pmd_entry(entry))
816 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
819 unmap_mapping_pages(mapping, vmf->pgoff, 1, false);
823 new_entry = dax_radix_locked_entry(pfn, flags);
824 if (dax_entry_size(entry) != dax_entry_size(new_entry)) {
825 dax_disassociate_entry(entry, mapping, false);
826 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address);
829 if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
831 * Only swap our new entry into the radix tree if the current
832 * entry is a zero page or an empty entry. If a normal PTE or
833 * PMD entry is already in the tree, we leave it alone. This
834 * means that if we are trying to insert a PTE and the
835 * existing entry is a PMD, we will just leave the PMD in the
836 * tree and dirty it if necessary.
838 struct radix_tree_node *node;
842 ret = __radix_tree_lookup(pages, index, &node, &slot);
843 WARN_ON_ONCE(ret != entry);
844 __radix_tree_replace(pages, node, slot,
850 radix_tree_tag_set(pages, index, PAGECACHE_TAG_DIRTY);
852 xa_unlock_irq(pages);
856 static inline unsigned long
857 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
859 unsigned long address;
861 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
862 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
866 /* Walk all mappings of a given index of a file and writeprotect them */
867 static void dax_mapping_entry_mkclean(struct address_space *mapping,
868 pgoff_t index, unsigned long pfn)
870 struct vm_area_struct *vma;
871 pte_t pte, *ptep = NULL;
875 i_mmap_lock_read(mapping);
876 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
877 unsigned long address, start, end;
881 if (!(vma->vm_flags & VM_SHARED))
884 address = pgoff_address(index, vma);
887 * Note because we provide start/end to follow_pte_pmd it will
888 * call mmu_notifier_invalidate_range_start() on our behalf
889 * before taking any lock.
891 if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl))
895 * No need to call mmu_notifier_invalidate_range() as we are
896 * downgrading page table protection not changing it to point
899 * See Documentation/vm/mmu_notifier.rst
902 #ifdef CONFIG_FS_DAX_PMD
905 if (pfn != pmd_pfn(*pmdp))
907 if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
910 flush_cache_range(vma, address,
911 address + HPAGE_PMD_SIZE);
912 pmd = pmdp_invalidate(vma, address, pmdp);
913 pmd = pmd_wrprotect(pmd);
914 pmd = pmd_mkclean(pmd);
915 set_pmd_at(vma->vm_mm, address, pmdp, pmd);
920 if (pfn != pte_pfn(*ptep))
922 if (!pte_dirty(*ptep) && !pte_write(*ptep))
925 flush_cache_page(vma, address, pfn);
926 pte = ptep_clear_flush(vma, address, ptep);
927 pte = pte_wrprotect(pte);
928 pte = pte_mkclean(pte);
929 set_pte_at(vma->vm_mm, address, ptep, pte);
931 pte_unmap_unlock(ptep, ptl);
934 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
936 i_mmap_unlock_read(mapping);
939 static int dax_writeback_one(struct dax_device *dax_dev,
940 struct address_space *mapping, pgoff_t index, void *entry)
942 struct radix_tree_root *pages = &mapping->i_pages;
943 void *entry2, **slot;
949 * A page got tagged dirty in DAX mapping? Something is seriously
952 if (WARN_ON(!radix_tree_exceptional_entry(entry)))
956 entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
957 /* Entry got punched out / reallocated? */
958 if (!entry2 || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry2)))
961 * Entry got reallocated elsewhere? No need to writeback. We have to
962 * compare pfns as we must not bail out due to difference in lockbit
965 if (dax_radix_pfn(entry2) != dax_radix_pfn(entry))
967 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
968 dax_is_zero_entry(entry))) {
973 /* Another fsync thread may have already written back this entry */
974 if (!radix_tree_tag_get(pages, index, PAGECACHE_TAG_TOWRITE))
976 /* Lock the entry to serialize with page faults */
977 entry = lock_slot(mapping, slot);
979 * We can clear the tag now but we have to be careful so that concurrent
980 * dax_writeback_one() calls for the same index cannot finish before we
981 * actually flush the caches. This is achieved as the calls will look
982 * at the entry only under the i_pages lock and once they do that
983 * they will see the entry locked and wait for it to unlock.
985 radix_tree_tag_clear(pages, index, PAGECACHE_TAG_TOWRITE);
986 xa_unlock_irq(pages);
989 * Even if dax_writeback_mapping_range() was given a wbc->range_start
990 * in the middle of a PMD, the 'index' we are given will be aligned to
991 * the start index of the PMD, as will the pfn we pull from 'entry'.
992 * This allows us to flush for PMD_SIZE and not have to worry about
993 * partial PMD writebacks.
995 pfn = dax_radix_pfn(entry);
996 size = PAGE_SIZE << dax_radix_order(entry);
998 dax_mapping_entry_mkclean(mapping, index, pfn);
999 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), size);
1001 * After we have flushed the cache, we can clear the dirty tag. There
1002 * cannot be new dirty data in the pfn after the flush has completed as
1003 * the pfn mappings are writeprotected and fault waits for mapping
1007 radix_tree_tag_clear(pages, index, PAGECACHE_TAG_DIRTY);
1008 xa_unlock_irq(pages);
1009 trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT);
1010 put_locked_mapping_entry(mapping, index);
1014 put_unlocked_mapping_entry(mapping, index, entry2);
1015 xa_unlock_irq(pages);
1020 * Flush the mapping to the persistent domain within the byte range of [start,
1021 * end]. This is required by data integrity operations to ensure file data is
1022 * on persistent storage prior to completion of the operation.
1024 int dax_writeback_mapping_range(struct address_space *mapping,
1025 struct block_device *bdev, struct writeback_control *wbc)
1027 struct inode *inode = mapping->host;
1028 pgoff_t start_index, end_index;
1029 pgoff_t indices[PAGEVEC_SIZE];
1030 struct dax_device *dax_dev;
1031 struct pagevec pvec;
1035 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
1038 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
1041 dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
1045 start_index = wbc->range_start >> PAGE_SHIFT;
1046 end_index = wbc->range_end >> PAGE_SHIFT;
1048 trace_dax_writeback_range(inode, start_index, end_index);
1050 tag_pages_for_writeback(mapping, start_index, end_index);
1052 pagevec_init(&pvec);
1054 pvec.nr = find_get_entries_tag(mapping, start_index,
1055 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
1056 pvec.pages, indices);
1061 for (i = 0; i < pvec.nr; i++) {
1062 if (indices[i] > end_index) {
1067 ret = dax_writeback_one(dax_dev, mapping, indices[i],
1070 mapping_set_error(mapping, ret);
1074 start_index = indices[pvec.nr - 1] + 1;
1078 trace_dax_writeback_range_done(inode, start_index, end_index);
1079 return (ret < 0 ? ret : 0);
1081 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1083 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
1085 return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
1088 static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
1091 const sector_t sector = dax_iomap_sector(iomap, pos);
1096 rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
1099 id = dax_read_lock();
1100 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1107 if (PFN_PHYS(length) < size)
1109 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1111 /* For larger pages we need devmap */
1112 if (length > 1 && !pfn_t_devmap(*pfnp))
1116 dax_read_unlock(id);
1121 * The user has performed a load from a hole in the file. Allocating a new
1122 * page in the file would cause excessive storage usage for workloads with
1123 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1124 * If this page is ever written to we will re-fault and change the mapping to
1125 * point to real DAX storage instead.
1127 static vm_fault_t dax_load_hole(struct address_space *mapping, void *entry,
1128 struct vm_fault *vmf)
1130 struct inode *inode = mapping->host;
1131 unsigned long vaddr = vmf->address;
1132 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1135 dax_insert_mapping_entry(mapping, vmf, entry, pfn, RADIX_DAX_ZERO_PAGE,
1137 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1138 trace_dax_load_hole(inode, vmf, ret);
1142 static bool dax_range_is_aligned(struct block_device *bdev,
1143 unsigned int offset, unsigned int length)
1145 unsigned short sector_size = bdev_logical_block_size(bdev);
1147 if (!IS_ALIGNED(offset, sector_size))
1149 if (!IS_ALIGNED(length, sector_size))
1155 int __dax_zero_page_range(struct block_device *bdev,
1156 struct dax_device *dax_dev, sector_t sector,
1157 unsigned int offset, unsigned int size)
1159 if (dax_range_is_aligned(bdev, offset, size)) {
1160 sector_t start_sector = sector + (offset >> 9);
1162 return blkdev_issue_zeroout(bdev, start_sector,
1163 size >> 9, GFP_NOFS, 0);
1169 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
1173 id = dax_read_lock();
1174 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, NULL);
1176 dax_read_unlock(id);
1179 memset(kaddr + offset, 0, size);
1180 dax_flush(dax_dev, kaddr + offset, size);
1181 dax_read_unlock(id);
1185 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1188 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1189 struct iomap *iomap)
1191 struct block_device *bdev = iomap->bdev;
1192 struct dax_device *dax_dev = iomap->dax_dev;
1193 struct iov_iter *iter = data;
1194 loff_t end = pos + length, done = 0;
1199 if (iov_iter_rw(iter) == READ) {
1200 end = min(end, i_size_read(inode));
1204 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1205 return iov_iter_zero(min(length, end - pos), iter);
1208 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1212 * Write can allocate block for an area which has a hole page mapped
1213 * into page tables. We have to tear down these mappings so that data
1214 * written by write(2) is visible in mmap.
1216 if (iomap->flags & IOMAP_F_NEW) {
1217 invalidate_inode_pages2_range(inode->i_mapping,
1219 (end - 1) >> PAGE_SHIFT);
1222 id = dax_read_lock();
1224 unsigned offset = pos & (PAGE_SIZE - 1);
1225 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1226 const sector_t sector = dax_iomap_sector(iomap, pos);
1231 if (fatal_signal_pending(current)) {
1236 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1240 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1247 map_len = PFN_PHYS(map_len);
1250 if (map_len > end - pos)
1251 map_len = end - pos;
1254 * The userspace address for the memory copy has already been
1255 * validated via access_ok() in either vfs_read() or
1256 * vfs_write(), depending on which operation we are doing.
1258 if (iov_iter_rw(iter) == WRITE)
1259 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1262 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1274 dax_read_unlock(id);
1276 return done ? done : ret;
1280 * dax_iomap_rw - Perform I/O to a DAX file
1281 * @iocb: The control block for this I/O
1282 * @iter: The addresses to do I/O from or to
1283 * @ops: iomap ops passed from the file system
1285 * This function performs read and write operations to directly mapped
1286 * persistent memory. The callers needs to take care of read/write exclusion
1287 * and evicting any page cache pages in the region under I/O.
1290 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1291 const struct iomap_ops *ops)
1293 struct address_space *mapping = iocb->ki_filp->f_mapping;
1294 struct inode *inode = mapping->host;
1295 loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1298 if (iov_iter_rw(iter) == WRITE) {
1299 lockdep_assert_held_exclusive(&inode->i_rwsem);
1300 flags |= IOMAP_WRITE;
1302 lockdep_assert_held(&inode->i_rwsem);
1305 if (iocb->ki_flags & IOCB_NOWAIT)
1306 flags |= IOMAP_NOWAIT;
1308 while (iov_iter_count(iter)) {
1309 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1310 iter, dax_iomap_actor);
1317 iocb->ki_pos += done;
1318 return done ? done : ret;
1320 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1322 static vm_fault_t dax_fault_return(int error)
1325 return VM_FAULT_NOPAGE;
1326 if (error == -ENOMEM)
1327 return VM_FAULT_OOM;
1328 return VM_FAULT_SIGBUS;
1332 * MAP_SYNC on a dax mapping guarantees dirty metadata is
1333 * flushed on write-faults (non-cow), but not read-faults.
1335 static bool dax_fault_is_synchronous(unsigned long flags,
1336 struct vm_area_struct *vma, struct iomap *iomap)
1338 return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1339 && (iomap->flags & IOMAP_F_DIRTY);
1342 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1343 int *iomap_errp, const struct iomap_ops *ops)
1345 struct vm_area_struct *vma = vmf->vma;
1346 struct address_space *mapping = vma->vm_file->f_mapping;
1347 struct inode *inode = mapping->host;
1348 unsigned long vaddr = vmf->address;
1349 loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1350 struct iomap iomap = { 0 };
1351 unsigned flags = IOMAP_FAULT;
1352 int error, major = 0;
1353 bool write = vmf->flags & FAULT_FLAG_WRITE;
1359 trace_dax_pte_fault(inode, vmf, ret);
1361 * Check whether offset isn't beyond end of file now. Caller is supposed
1362 * to hold locks serializing us with truncate / punch hole so this is
1365 if (pos >= i_size_read(inode)) {
1366 ret = VM_FAULT_SIGBUS;
1370 if (write && !vmf->cow_page)
1371 flags |= IOMAP_WRITE;
1373 entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1374 if (IS_ERR(entry)) {
1375 ret = dax_fault_return(PTR_ERR(entry));
1380 * It is possible, particularly with mixed reads & writes to private
1381 * mappings, that we have raced with a PMD fault that overlaps with
1382 * the PTE we need to set up. If so just return and the fault will be
1385 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1386 ret = VM_FAULT_NOPAGE;
1391 * Note that we don't bother to use iomap_apply here: DAX required
1392 * the file system block size to be equal the page size, which means
1393 * that we never have to deal with more than a single extent here.
1395 error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1397 *iomap_errp = error;
1399 ret = dax_fault_return(error);
1402 if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1403 error = -EIO; /* fs corruption? */
1404 goto error_finish_iomap;
1407 if (vmf->cow_page) {
1408 sector_t sector = dax_iomap_sector(&iomap, pos);
1410 switch (iomap.type) {
1412 case IOMAP_UNWRITTEN:
1413 clear_user_highpage(vmf->cow_page, vaddr);
1416 error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1417 sector, PAGE_SIZE, vmf->cow_page, vaddr);
1426 goto error_finish_iomap;
1428 __SetPageUptodate(vmf->cow_page);
1429 ret = finish_fault(vmf);
1431 ret = VM_FAULT_DONE_COW;
1435 sync = dax_fault_is_synchronous(flags, vma, &iomap);
1437 switch (iomap.type) {
1439 if (iomap.flags & IOMAP_F_NEW) {
1440 count_vm_event(PGMAJFAULT);
1441 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1442 major = VM_FAULT_MAJOR;
1444 error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1446 goto error_finish_iomap;
1448 entry = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1452 * If we are doing synchronous page fault and inode needs fsync,
1453 * we can insert PTE into page tables only after that happens.
1454 * Skip insertion for now and return the pfn so that caller can
1455 * insert it after fsync is done.
1458 if (WARN_ON_ONCE(!pfnp)) {
1460 goto error_finish_iomap;
1463 ret = VM_FAULT_NEEDDSYNC | major;
1466 trace_dax_insert_mapping(inode, vmf, entry);
1468 ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn);
1470 ret = vmf_insert_mixed(vma, vaddr, pfn);
1473 case IOMAP_UNWRITTEN:
1476 ret = dax_load_hole(mapping, entry, vmf);
1487 ret = dax_fault_return(error);
1489 if (ops->iomap_end) {
1490 int copied = PAGE_SIZE;
1492 if (ret & VM_FAULT_ERROR)
1495 * The fault is done by now and there's no way back (other
1496 * thread may be already happily using PTE we have installed).
1497 * Just ignore error from ->iomap_end since we cannot do much
1500 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1503 put_locked_mapping_entry(mapping, vmf->pgoff);
1505 trace_dax_pte_fault_done(inode, vmf, ret);
1509 #ifdef CONFIG_FS_DAX_PMD
1510 static vm_fault_t dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
1513 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1514 unsigned long pmd_addr = vmf->address & PMD_MASK;
1515 struct inode *inode = mapping->host;
1516 struct page *zero_page;
1522 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1524 if (unlikely(!zero_page))
1527 pfn = page_to_pfn_t(zero_page);
1528 ret = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1529 RADIX_DAX_PMD | RADIX_DAX_ZERO_PAGE, false);
1531 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1532 if (!pmd_none(*(vmf->pmd))) {
1537 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1538 pmd_entry = pmd_mkhuge(pmd_entry);
1539 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1541 trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
1542 return VM_FAULT_NOPAGE;
1545 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
1546 return VM_FAULT_FALLBACK;
1549 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1550 const struct iomap_ops *ops)
1552 struct vm_area_struct *vma = vmf->vma;
1553 struct address_space *mapping = vma->vm_file->f_mapping;
1554 unsigned long pmd_addr = vmf->address & PMD_MASK;
1555 bool write = vmf->flags & FAULT_FLAG_WRITE;
1557 unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1558 struct inode *inode = mapping->host;
1559 vm_fault_t result = VM_FAULT_FALLBACK;
1560 struct iomap iomap = { 0 };
1561 pgoff_t max_pgoff, pgoff;
1568 * Check whether offset isn't beyond end of file now. Caller is
1569 * supposed to hold locks serializing us with truncate / punch hole so
1570 * this is a reliable test.
1572 pgoff = linear_page_index(vma, pmd_addr);
1573 max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1575 trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1578 * Make sure that the faulting address's PMD offset (color) matches
1579 * the PMD offset from the start of the file. This is necessary so
1580 * that a PMD range in the page table overlaps exactly with a PMD
1581 * range in the radix tree.
1583 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1584 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1587 /* Fall back to PTEs if we're going to COW */
1588 if (write && !(vma->vm_flags & VM_SHARED))
1591 /* If the PMD would extend outside the VMA */
1592 if (pmd_addr < vma->vm_start)
1594 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1597 if (pgoff >= max_pgoff) {
1598 result = VM_FAULT_SIGBUS;
1602 /* If the PMD would extend beyond the file size */
1603 if ((pgoff | PG_PMD_COLOUR) >= max_pgoff)
1607 * grab_mapping_entry() will make sure we get a 2MiB empty entry, a
1608 * 2MiB zero page entry or a DAX PMD. If it can't (because a 4k page
1609 * is already in the tree, for instance), it will return -EEXIST and
1610 * we just fall back to 4k entries.
1612 entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1617 * It is possible, particularly with mixed reads & writes to private
1618 * mappings, that we have raced with a PTE fault that overlaps with
1619 * the PMD we need to set up. If so just return and the fault will be
1622 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1623 !pmd_devmap(*vmf->pmd)) {
1629 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1630 * setting up a mapping, so really we're using iomap_begin() as a way
1631 * to look up our filesystem block.
1633 pos = (loff_t)pgoff << PAGE_SHIFT;
1634 error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1638 if (iomap.offset + iomap.length < pos + PMD_SIZE)
1641 sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1643 switch (iomap.type) {
1645 error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1649 entry = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1650 RADIX_DAX_PMD, write && !sync);
1653 * If we are doing synchronous page fault and inode needs fsync,
1654 * we can insert PMD into page tables only after that happens.
1655 * Skip insertion for now and return the pfn so that caller can
1656 * insert it after fsync is done.
1659 if (WARN_ON_ONCE(!pfnp))
1662 result = VM_FAULT_NEEDDSYNC;
1666 trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
1667 result = vmf_insert_pfn_pmd(vmf, pfn, write);
1669 case IOMAP_UNWRITTEN:
1671 if (WARN_ON_ONCE(write))
1673 result = dax_pmd_load_hole(vmf, &iomap, entry);
1681 if (ops->iomap_end) {
1682 int copied = PMD_SIZE;
1684 if (result == VM_FAULT_FALLBACK)
1687 * The fault is done by now and there's no way back (other
1688 * thread may be already happily using PMD we have installed).
1689 * Just ignore error from ->iomap_end since we cannot do much
1692 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1696 put_locked_mapping_entry(mapping, pgoff);
1698 if (result == VM_FAULT_FALLBACK) {
1699 split_huge_pmd(vma, vmf->pmd, vmf->address);
1700 count_vm_event(THP_FAULT_FALLBACK);
1703 trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1707 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1708 const struct iomap_ops *ops)
1710 return VM_FAULT_FALLBACK;
1712 #endif /* CONFIG_FS_DAX_PMD */
1715 * dax_iomap_fault - handle a page fault on a DAX file
1716 * @vmf: The description of the fault
1717 * @pe_size: Size of the page to fault in
1718 * @pfnp: PFN to insert for synchronous faults if fsync is required
1719 * @iomap_errp: Storage for detailed error code in case of error
1720 * @ops: Iomap ops passed from the file system
1722 * When a page fault occurs, filesystems may call this helper in
1723 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1724 * has done all the necessary locking for page fault to proceed
1727 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1728 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1732 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1734 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1736 return VM_FAULT_FALLBACK;
1739 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1742 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1743 * @vmf: The description of the fault
1744 * @pe_size: Size of entry to be inserted
1745 * @pfn: PFN to insert
1747 * This function inserts writeable PTE or PMD entry into page tables for mmaped
1748 * DAX file. It takes care of marking corresponding radix tree entry as dirty
1751 static vm_fault_t dax_insert_pfn_mkwrite(struct vm_fault *vmf,
1752 enum page_entry_size pe_size,
1755 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1756 void *entry, **slot;
1757 pgoff_t index = vmf->pgoff;
1760 xa_lock_irq(&mapping->i_pages);
1761 entry = get_unlocked_mapping_entry(mapping, index, &slot);
1762 /* Did we race with someone splitting entry or so? */
1764 (pe_size == PE_SIZE_PTE && !dax_is_pte_entry(entry)) ||
1765 (pe_size == PE_SIZE_PMD && !dax_is_pmd_entry(entry))) {
1766 put_unlocked_mapping_entry(mapping, index, entry);
1767 xa_unlock_irq(&mapping->i_pages);
1768 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1770 return VM_FAULT_NOPAGE;
1772 radix_tree_tag_set(&mapping->i_pages, index, PAGECACHE_TAG_DIRTY);
1773 entry = lock_slot(mapping, slot);
1774 xa_unlock_irq(&mapping->i_pages);
1777 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1779 #ifdef CONFIG_FS_DAX_PMD
1781 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1785 ret = VM_FAULT_FALLBACK;
1787 put_locked_mapping_entry(mapping, index);
1788 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1793 * dax_finish_sync_fault - finish synchronous page fault
1794 * @vmf: The description of the fault
1795 * @pe_size: Size of entry to be inserted
1796 * @pfn: PFN to insert
1798 * This function ensures that the file range touched by the page fault is
1799 * stored persistently on the media and handles inserting of appropriate page
1802 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1803 enum page_entry_size pe_size, pfn_t pfn)
1806 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1809 if (pe_size == PE_SIZE_PTE)
1811 else if (pe_size == PE_SIZE_PMD)
1815 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1817 return VM_FAULT_SIGBUS;
1818 return dax_insert_pfn_mkwrite(vmf, pe_size, pfn);
1820 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);