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 static inline unsigned int pe_order(enum page_entry_size pe_size
)
43 if (pe_size
== PE_SIZE_PTE
)
44 return PAGE_SHIFT
- PAGE_SHIFT
;
45 if (pe_size
== PE_SIZE_PMD
)
46 return PMD_SHIFT
- PAGE_SHIFT
;
47 if (pe_size
== PE_SIZE_PUD
)
48 return PUD_SHIFT
- PAGE_SHIFT
;
52 /* We choose 4096 entries - same as per-zone page wait tables */
53 #define DAX_WAIT_TABLE_BITS 12
54 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
56 /* The 'colour' (ie low bits) within a PMD of a page offset. */
57 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
58 #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
60 /* The order of a PMD entry */
61 #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT)
63 static wait_queue_head_t wait_table
[DAX_WAIT_TABLE_ENTRIES
];
65 static int __init
init_dax_wait_table(void)
69 for (i
= 0; i
< DAX_WAIT_TABLE_ENTRIES
; i
++)
70 init_waitqueue_head(wait_table
+ i
);
73 fs_initcall(init_dax_wait_table
);
76 * DAX pagecache entries use XArray value entries so they can't be mistaken
77 * for pages. We use one bit for locking, one bit for the entry size (PMD)
78 * and two more to tell us if the entry is a zero page or an empty entry that
79 * is just used for locking. In total four special bits.
81 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
82 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
86 #define DAX_LOCKED (1UL << 0)
87 #define DAX_PMD (1UL << 1)
88 #define DAX_ZERO_PAGE (1UL << 2)
89 #define DAX_EMPTY (1UL << 3)
91 static unsigned long dax_to_pfn(void *entry
)
93 return xa_to_value(entry
) >> DAX_SHIFT
;
96 static void *dax_make_entry(pfn_t pfn
, unsigned long flags
)
98 return xa_mk_value(flags
| (pfn_t_to_pfn(pfn
) << DAX_SHIFT
));
101 static bool dax_is_locked(void *entry
)
103 return xa_to_value(entry
) & DAX_LOCKED
;
106 static unsigned int dax_entry_order(void *entry
)
108 if (xa_to_value(entry
) & DAX_PMD
)
113 static unsigned long dax_is_pmd_entry(void *entry
)
115 return xa_to_value(entry
) & DAX_PMD
;
118 static bool dax_is_pte_entry(void *entry
)
120 return !(xa_to_value(entry
) & DAX_PMD
);
123 static int dax_is_zero_entry(void *entry
)
125 return xa_to_value(entry
) & DAX_ZERO_PAGE
;
128 static int dax_is_empty_entry(void *entry
)
130 return xa_to_value(entry
) & DAX_EMPTY
;
134 * DAX page cache entry locking
136 struct exceptional_entry_key
{
141 struct wait_exceptional_entry_queue
{
142 wait_queue_entry_t wait
;
143 struct exceptional_entry_key key
;
146 static wait_queue_head_t
*dax_entry_waitqueue(struct xa_state
*xas
,
147 void *entry
, struct exceptional_entry_key
*key
)
150 unsigned long index
= xas
->xa_index
;
153 * If 'entry' is a PMD, align the 'index' that we use for the wait
154 * queue to the start of that PMD. This ensures that all offsets in
155 * the range covered by the PMD map to the same bit lock.
157 if (dax_is_pmd_entry(entry
))
158 index
&= ~PG_PMD_COLOUR
;
160 key
->entry_start
= index
;
162 hash
= hash_long((unsigned long)xas
->xa
^ index
, DAX_WAIT_TABLE_BITS
);
163 return wait_table
+ hash
;
166 static int wake_exceptional_entry_func(wait_queue_entry_t
*wait
,
167 unsigned int mode
, int sync
, void *keyp
)
169 struct exceptional_entry_key
*key
= keyp
;
170 struct wait_exceptional_entry_queue
*ewait
=
171 container_of(wait
, struct wait_exceptional_entry_queue
, wait
);
173 if (key
->xa
!= ewait
->key
.xa
||
174 key
->entry_start
!= ewait
->key
.entry_start
)
176 return autoremove_wake_function(wait
, mode
, sync
, NULL
);
180 * @entry may no longer be the entry at the index in the mapping.
181 * The important information it's conveying is whether the entry at
182 * this index used to be a PMD entry.
184 static void dax_wake_entry(struct xa_state
*xas
, void *entry
, bool wake_all
)
186 struct exceptional_entry_key key
;
187 wait_queue_head_t
*wq
;
189 wq
= dax_entry_waitqueue(xas
, entry
, &key
);
192 * Checking for locked entry and prepare_to_wait_exclusive() happens
193 * under the i_pages lock, ditto for entry handling in our callers.
194 * So at this point all tasks that could have seen our entry locked
195 * must be in the waitqueue and the following check will see them.
197 if (waitqueue_active(wq
))
198 __wake_up(wq
, TASK_NORMAL
, wake_all
? 0 : 1, &key
);
202 * Look up entry in page cache, wait for it to become unlocked if it
203 * is a DAX entry and return it. The caller must subsequently call
204 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
207 * Must be called with the i_pages lock held.
209 static void *get_unlocked_entry(struct xa_state
*xas
)
212 struct wait_exceptional_entry_queue ewait
;
213 wait_queue_head_t
*wq
;
215 init_wait(&ewait
.wait
);
216 ewait
.wait
.func
= wake_exceptional_entry_func
;
219 entry
= xas_find_conflict(xas
);
220 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)) ||
221 !dax_is_locked(entry
))
224 wq
= dax_entry_waitqueue(xas
, entry
, &ewait
.key
);
225 prepare_to_wait_exclusive(wq
, &ewait
.wait
,
226 TASK_UNINTERRUPTIBLE
);
230 finish_wait(wq
, &ewait
.wait
);
236 * The only thing keeping the address space around is the i_pages lock
237 * (it's cycled in clear_inode() after removing the entries from i_pages)
238 * After we call xas_unlock_irq(), we cannot touch xas->xa.
240 static void wait_entry_unlocked(struct xa_state
*xas
, void *entry
)
242 struct wait_exceptional_entry_queue ewait
;
243 wait_queue_head_t
*wq
;
245 init_wait(&ewait
.wait
);
246 ewait
.wait
.func
= wake_exceptional_entry_func
;
248 wq
= dax_entry_waitqueue(xas
, entry
, &ewait
.key
);
250 * Unlike get_unlocked_entry() there is no guarantee that this
251 * path ever successfully retrieves an unlocked entry before an
252 * inode dies. Perform a non-exclusive wait in case this path
253 * never successfully performs its own wake up.
255 prepare_to_wait(wq
, &ewait
.wait
, TASK_UNINTERRUPTIBLE
);
258 finish_wait(wq
, &ewait
.wait
);
261 static void put_unlocked_entry(struct xa_state
*xas
, void *entry
)
263 /* If we were the only waiter woken, wake the next one */
265 dax_wake_entry(xas
, entry
, false);
269 * We used the xa_state to get the entry, but then we locked the entry and
270 * dropped the xa_lock, so we know the xa_state is stale and must be reset
273 static void dax_unlock_entry(struct xa_state
*xas
, void *entry
)
277 BUG_ON(dax_is_locked(entry
));
280 old
= xas_store(xas
, entry
);
282 BUG_ON(!dax_is_locked(old
));
283 dax_wake_entry(xas
, entry
, false);
287 * Return: The entry stored at this location before it was locked.
289 static void *dax_lock_entry(struct xa_state
*xas
, void *entry
)
291 unsigned long v
= xa_to_value(entry
);
292 return xas_store(xas
, xa_mk_value(v
| DAX_LOCKED
));
295 static unsigned long dax_entry_size(void *entry
)
297 if (dax_is_zero_entry(entry
))
299 else if (dax_is_empty_entry(entry
))
301 else if (dax_is_pmd_entry(entry
))
307 static unsigned long dax_end_pfn(void *entry
)
309 return dax_to_pfn(entry
) + dax_entry_size(entry
) / PAGE_SIZE
;
313 * Iterate through all mapped pfns represented by an entry, i.e. skip
314 * 'empty' and 'zero' entries.
316 #define for_each_mapped_pfn(entry, pfn) \
317 for (pfn = dax_to_pfn(entry); \
318 pfn < dax_end_pfn(entry); pfn++)
321 * TODO: for reflink+dax we need a way to associate a single page with
322 * multiple address_space instances at different linear_page_index()
325 static void dax_associate_entry(void *entry
, struct address_space
*mapping
,
326 struct vm_area_struct
*vma
, unsigned long address
)
328 unsigned long size
= dax_entry_size(entry
), pfn
, index
;
331 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
334 index
= linear_page_index(vma
, address
& ~(size
- 1));
335 for_each_mapped_pfn(entry
, pfn
) {
336 struct page
*page
= pfn_to_page(pfn
);
338 WARN_ON_ONCE(page
->mapping
);
339 page
->mapping
= mapping
;
340 page
->index
= index
+ i
++;
344 static void dax_disassociate_entry(void *entry
, struct address_space
*mapping
,
349 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
352 for_each_mapped_pfn(entry
, pfn
) {
353 struct page
*page
= pfn_to_page(pfn
);
355 WARN_ON_ONCE(trunc
&& page_ref_count(page
) > 1);
356 WARN_ON_ONCE(page
->mapping
&& page
->mapping
!= mapping
);
357 page
->mapping
= NULL
;
362 static struct page
*dax_busy_page(void *entry
)
366 for_each_mapped_pfn(entry
, pfn
) {
367 struct page
*page
= pfn_to_page(pfn
);
369 if (page_ref_count(page
) > 1)
376 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page
377 * @page: The page whose entry we want to lock
379 * Context: Process context.
380 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
383 dax_entry_t
dax_lock_page(struct page
*page
)
385 XA_STATE(xas
, NULL
, 0);
388 /* Ensure page->mapping isn't freed while we look at it */
391 struct address_space
*mapping
= READ_ONCE(page
->mapping
);
394 if (!mapping
|| !dax_mapping(mapping
))
398 * In the device-dax case there's no need to lock, a
399 * struct dev_pagemap pin is sufficient to keep the
400 * inode alive, and we assume we have dev_pagemap pin
401 * otherwise we would not have a valid pfn_to_page()
404 entry
= (void *)~0UL;
405 if (S_ISCHR(mapping
->host
->i_mode
))
408 xas
.xa
= &mapping
->i_pages
;
410 if (mapping
!= page
->mapping
) {
411 xas_unlock_irq(&xas
);
414 xas_set(&xas
, page
->index
);
415 entry
= xas_load(&xas
);
416 if (dax_is_locked(entry
)) {
418 wait_entry_unlocked(&xas
, entry
);
422 dax_lock_entry(&xas
, entry
);
423 xas_unlock_irq(&xas
);
427 return (dax_entry_t
)entry
;
430 void dax_unlock_page(struct page
*page
, dax_entry_t cookie
)
432 struct address_space
*mapping
= page
->mapping
;
433 XA_STATE(xas
, &mapping
->i_pages
, page
->index
);
435 if (S_ISCHR(mapping
->host
->i_mode
))
438 dax_unlock_entry(&xas
, (void *)cookie
);
442 * Find page cache entry at given index. If it is a DAX entry, return it
443 * with the entry locked. If the page cache doesn't contain an entry at
444 * that index, add a locked empty entry.
446 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
447 * either return that locked entry or will return VM_FAULT_FALLBACK.
448 * This will happen if there are any PTE entries within the PMD range
449 * that we are requesting.
451 * We always favor PTE entries over PMD entries. There isn't a flow where we
452 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
453 * insertion will fail if it finds any PTE entries already in the tree, and a
454 * PTE insertion will cause an existing PMD entry to be unmapped and
455 * downgraded to PTE entries. This happens for both PMD zero pages as
456 * well as PMD empty entries.
458 * The exception to this downgrade path is for PMD entries that have
459 * real storage backing them. We will leave these real PMD entries in
460 * the tree, and PTE writes will simply dirty the entire PMD entry.
462 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
463 * persistent memory the benefit is doubtful. We can add that later if we can
466 * On error, this function does not return an ERR_PTR. Instead it returns
467 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
468 * overlap with xarray value entries.
470 static void *grab_mapping_entry(struct xa_state
*xas
,
471 struct address_space
*mapping
, unsigned long size_flag
)
473 unsigned long index
= xas
->xa_index
;
474 bool pmd_downgrade
= false; /* splitting PMD entry into PTE entries? */
479 entry
= get_unlocked_entry(xas
);
482 if (!xa_is_value(entry
)) {
483 xas_set_err(xas
, EIO
);
487 if (size_flag
& DAX_PMD
) {
488 if (dax_is_pte_entry(entry
)) {
489 put_unlocked_entry(xas
, entry
);
492 } else { /* trying to grab a PTE entry */
493 if (dax_is_pmd_entry(entry
) &&
494 (dax_is_zero_entry(entry
) ||
495 dax_is_empty_entry(entry
))) {
496 pmd_downgrade
= true;
503 * Make sure 'entry' remains valid while we drop
506 dax_lock_entry(xas
, entry
);
509 * Besides huge zero pages the only other thing that gets
510 * downgraded are empty entries which don't need to be
513 if (dax_is_zero_entry(entry
)) {
515 unmap_mapping_pages(mapping
,
516 xas
->xa_index
& ~PG_PMD_COLOUR
,
522 dax_disassociate_entry(entry
, mapping
, false);
523 xas_store(xas
, NULL
); /* undo the PMD join */
524 dax_wake_entry(xas
, entry
, true);
525 mapping
->nrexceptional
--;
531 dax_lock_entry(xas
, entry
);
533 entry
= dax_make_entry(pfn_to_pfn_t(0), size_flag
| DAX_EMPTY
);
534 dax_lock_entry(xas
, entry
);
537 mapping
->nrexceptional
++;
542 if (xas_nomem(xas
, mapping_gfp_mask(mapping
) & ~__GFP_HIGHMEM
))
544 if (xas
->xa_node
== XA_ERROR(-ENOMEM
))
545 return xa_mk_internal(VM_FAULT_OOM
);
547 return xa_mk_internal(VM_FAULT_SIGBUS
);
551 return xa_mk_internal(VM_FAULT_FALLBACK
);
555 * dax_layout_busy_page - find first pinned page in @mapping
556 * @mapping: address space to scan for a page with ref count > 1
558 * DAX requires ZONE_DEVICE mapped pages. These pages are never
559 * 'onlined' to the page allocator so they are considered idle when
560 * page->count == 1. A filesystem uses this interface to determine if
561 * any page in the mapping is busy, i.e. for DMA, or other
562 * get_user_pages() usages.
564 * It is expected that the filesystem is holding locks to block the
565 * establishment of new mappings in this address_space. I.e. it expects
566 * to be able to run unmap_mapping_range() and subsequently not race
567 * mapping_mapped() becoming true.
569 struct page
*dax_layout_busy_page(struct address_space
*mapping
)
571 XA_STATE(xas
, &mapping
->i_pages
, 0);
573 unsigned int scanned
= 0;
574 struct page
*page
= NULL
;
577 * In the 'limited' case get_user_pages() for dax is disabled.
579 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
582 if (!dax_mapping(mapping
) || !mapping_mapped(mapping
))
586 * If we race get_user_pages_fast() here either we'll see the
587 * elevated page count in the iteration and wait, or
588 * get_user_pages_fast() will see that the page it took a reference
589 * against is no longer mapped in the page tables and bail to the
590 * get_user_pages() slow path. The slow path is protected by
591 * pte_lock() and pmd_lock(). New references are not taken without
592 * holding those locks, and unmap_mapping_range() will not zero the
593 * pte or pmd without holding the respective lock, so we are
594 * guaranteed to either see new references or prevent new
595 * references from being established.
597 unmap_mapping_range(mapping
, 0, 0, 1);
600 xas_for_each(&xas
, entry
, ULONG_MAX
) {
601 if (WARN_ON_ONCE(!xa_is_value(entry
)))
603 if (unlikely(dax_is_locked(entry
)))
604 entry
= get_unlocked_entry(&xas
);
606 page
= dax_busy_page(entry
);
607 put_unlocked_entry(&xas
, entry
);
610 if (++scanned
% XA_CHECK_SCHED
)
614 xas_unlock_irq(&xas
);
618 xas_unlock_irq(&xas
);
621 EXPORT_SYMBOL_GPL(dax_layout_busy_page
);
623 static int __dax_invalidate_entry(struct address_space
*mapping
,
624 pgoff_t index
, bool trunc
)
626 XA_STATE(xas
, &mapping
->i_pages
, index
);
631 entry
= get_unlocked_entry(&xas
);
632 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
635 (xas_get_mark(&xas
, PAGECACHE_TAG_DIRTY
) ||
636 xas_get_mark(&xas
, PAGECACHE_TAG_TOWRITE
)))
638 dax_disassociate_entry(entry
, mapping
, trunc
);
639 xas_store(&xas
, NULL
);
640 mapping
->nrexceptional
--;
643 put_unlocked_entry(&xas
, entry
);
644 xas_unlock_irq(&xas
);
649 * Delete DAX entry at @index from @mapping. Wait for it
650 * to be unlocked before deleting it.
652 int dax_delete_mapping_entry(struct address_space
*mapping
, pgoff_t index
)
654 int ret
= __dax_invalidate_entry(mapping
, index
, true);
657 * This gets called from truncate / punch_hole path. As such, the caller
658 * must hold locks protecting against concurrent modifications of the
659 * page cache (usually fs-private i_mmap_sem for writing). Since the
660 * caller has seen a DAX entry for this index, we better find it
661 * at that index as well...
668 * Invalidate DAX entry if it is clean.
670 int dax_invalidate_mapping_entry_sync(struct address_space
*mapping
,
673 return __dax_invalidate_entry(mapping
, index
, false);
676 static int copy_user_dax(struct block_device
*bdev
, struct dax_device
*dax_dev
,
677 sector_t sector
, size_t size
, struct page
*to
,
685 rc
= bdev_dax_pgoff(bdev
, sector
, size
, &pgoff
);
689 id
= dax_read_lock();
690 rc
= dax_direct_access(dax_dev
, pgoff
, PHYS_PFN(size
), &kaddr
, NULL
);
695 vto
= kmap_atomic(to
);
696 copy_user_page(vto
, (void __force
*)kaddr
, vaddr
, to
);
703 * By this point grab_mapping_entry() has ensured that we have a locked entry
704 * of the appropriate size so we don't have to worry about downgrading PMDs to
705 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
706 * already in the tree, we will skip the insertion and just dirty the PMD as
709 static void *dax_insert_entry(struct xa_state
*xas
,
710 struct address_space
*mapping
, struct vm_fault
*vmf
,
711 void *entry
, pfn_t pfn
, unsigned long flags
, bool dirty
)
713 void *new_entry
= dax_make_entry(pfn
, flags
);
716 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
718 if (dax_is_zero_entry(entry
) && !(flags
& DAX_ZERO_PAGE
)) {
719 unsigned long index
= xas
->xa_index
;
720 /* we are replacing a zero page with block mapping */
721 if (dax_is_pmd_entry(entry
))
722 unmap_mapping_pages(mapping
, index
& ~PG_PMD_COLOUR
,
725 unmap_mapping_pages(mapping
, index
, 1, false);
730 if (dax_entry_size(entry
) != dax_entry_size(new_entry
)) {
731 dax_disassociate_entry(entry
, mapping
, false);
732 dax_associate_entry(new_entry
, mapping
, vmf
->vma
, vmf
->address
);
735 if (dax_is_zero_entry(entry
) || dax_is_empty_entry(entry
)) {
737 * Only swap our new entry into the page cache if the current
738 * entry is a zero page or an empty entry. If a normal PTE or
739 * PMD entry is already in the cache, we leave it alone. This
740 * means that if we are trying to insert a PTE and the
741 * existing entry is a PMD, we will just leave the PMD in the
742 * tree and dirty it if necessary.
744 void *old
= dax_lock_entry(xas
, new_entry
);
745 WARN_ON_ONCE(old
!= xa_mk_value(xa_to_value(entry
) |
749 xas_load(xas
); /* Walk the xa_state */
753 xas_set_mark(xas
, PAGECACHE_TAG_DIRTY
);
760 unsigned long pgoff_address(pgoff_t pgoff
, struct vm_area_struct
*vma
)
762 unsigned long address
;
764 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
765 VM_BUG_ON_VMA(address
< vma
->vm_start
|| address
>= vma
->vm_end
, vma
);
769 /* Walk all mappings of a given index of a file and writeprotect them */
770 static void dax_entry_mkclean(struct address_space
*mapping
, pgoff_t index
,
773 struct vm_area_struct
*vma
;
774 pte_t pte
, *ptep
= NULL
;
778 i_mmap_lock_read(mapping
);
779 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, index
, index
) {
780 struct mmu_notifier_range range
;
781 unsigned long address
;
785 if (!(vma
->vm_flags
& VM_SHARED
))
788 address
= pgoff_address(index
, vma
);
791 * Note because we provide range to follow_pte_pmd it will
792 * call mmu_notifier_invalidate_range_start() on our behalf
793 * before taking any lock.
795 if (follow_pte_pmd(vma
->vm_mm
, address
, &range
,
800 * No need to call mmu_notifier_invalidate_range() as we are
801 * downgrading page table protection not changing it to point
804 * See Documentation/vm/mmu_notifier.rst
807 #ifdef CONFIG_FS_DAX_PMD
810 if (pfn
!= pmd_pfn(*pmdp
))
812 if (!pmd_dirty(*pmdp
) && !pmd_write(*pmdp
))
815 flush_cache_page(vma
, address
, pfn
);
816 pmd
= pmdp_huge_clear_flush(vma
, address
, pmdp
);
817 pmd
= pmd_wrprotect(pmd
);
818 pmd
= pmd_mkclean(pmd
);
819 set_pmd_at(vma
->vm_mm
, address
, pmdp
, pmd
);
824 if (pfn
!= pte_pfn(*ptep
))
826 if (!pte_dirty(*ptep
) && !pte_write(*ptep
))
829 flush_cache_page(vma
, address
, pfn
);
830 pte
= ptep_clear_flush(vma
, address
, ptep
);
831 pte
= pte_wrprotect(pte
);
832 pte
= pte_mkclean(pte
);
833 set_pte_at(vma
->vm_mm
, address
, ptep
, pte
);
835 pte_unmap_unlock(ptep
, ptl
);
838 mmu_notifier_invalidate_range_end(&range
);
840 i_mmap_unlock_read(mapping
);
843 static int dax_writeback_one(struct xa_state
*xas
, struct dax_device
*dax_dev
,
844 struct address_space
*mapping
, void *entry
)
846 unsigned long pfn
, index
, count
;
850 * A page got tagged dirty in DAX mapping? Something is seriously
853 if (WARN_ON(!xa_is_value(entry
)))
856 if (unlikely(dax_is_locked(entry
))) {
857 void *old_entry
= entry
;
859 entry
= get_unlocked_entry(xas
);
861 /* Entry got punched out / reallocated? */
862 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
865 * Entry got reallocated elsewhere? No need to writeback.
866 * We have to compare pfns as we must not bail out due to
867 * difference in lockbit or entry type.
869 if (dax_to_pfn(old_entry
) != dax_to_pfn(entry
))
871 if (WARN_ON_ONCE(dax_is_empty_entry(entry
) ||
872 dax_is_zero_entry(entry
))) {
877 /* Another fsync thread may have already done this entry */
878 if (!xas_get_mark(xas
, PAGECACHE_TAG_TOWRITE
))
882 /* Lock the entry to serialize with page faults */
883 dax_lock_entry(xas
, entry
);
886 * We can clear the tag now but we have to be careful so that concurrent
887 * dax_writeback_one() calls for the same index cannot finish before we
888 * actually flush the caches. This is achieved as the calls will look
889 * at the entry only under the i_pages lock and once they do that
890 * they will see the entry locked and wait for it to unlock.
892 xas_clear_mark(xas
, PAGECACHE_TAG_TOWRITE
);
896 * If dax_writeback_mapping_range() was given a wbc->range_start
897 * in the middle of a PMD, the 'index' we use needs to be
898 * aligned to the start of the PMD.
899 * This allows us to flush for PMD_SIZE and not have to worry about
900 * partial PMD writebacks.
902 pfn
= dax_to_pfn(entry
);
903 count
= 1UL << dax_entry_order(entry
);
904 index
= xas
->xa_index
& ~(count
- 1);
906 dax_entry_mkclean(mapping
, index
, pfn
);
907 dax_flush(dax_dev
, page_address(pfn_to_page(pfn
)), count
* PAGE_SIZE
);
909 * After we have flushed the cache, we can clear the dirty tag. There
910 * cannot be new dirty data in the pfn after the flush has completed as
911 * the pfn mappings are writeprotected and fault waits for mapping
916 xas_store(xas
, entry
);
917 xas_clear_mark(xas
, PAGECACHE_TAG_DIRTY
);
918 dax_wake_entry(xas
, entry
, false);
920 trace_dax_writeback_one(mapping
->host
, index
, count
);
924 put_unlocked_entry(xas
, entry
);
929 * Flush the mapping to the persistent domain within the byte range of [start,
930 * end]. This is required by data integrity operations to ensure file data is
931 * on persistent storage prior to completion of the operation.
933 int dax_writeback_mapping_range(struct address_space
*mapping
,
934 struct block_device
*bdev
, struct writeback_control
*wbc
)
936 XA_STATE(xas
, &mapping
->i_pages
, wbc
->range_start
>> PAGE_SHIFT
);
937 struct inode
*inode
= mapping
->host
;
938 pgoff_t end_index
= wbc
->range_end
>> PAGE_SHIFT
;
939 struct dax_device
*dax_dev
;
942 unsigned int scanned
= 0;
944 if (WARN_ON_ONCE(inode
->i_blkbits
!= PAGE_SHIFT
))
947 if (!mapping
->nrexceptional
|| wbc
->sync_mode
!= WB_SYNC_ALL
)
950 dax_dev
= dax_get_by_host(bdev
->bd_disk
->disk_name
);
954 trace_dax_writeback_range(inode
, xas
.xa_index
, end_index
);
956 tag_pages_for_writeback(mapping
, xas
.xa_index
, end_index
);
959 xas_for_each_marked(&xas
, entry
, end_index
, PAGECACHE_TAG_TOWRITE
) {
960 ret
= dax_writeback_one(&xas
, dax_dev
, mapping
, entry
);
962 mapping_set_error(mapping
, ret
);
965 if (++scanned
% XA_CHECK_SCHED
)
969 xas_unlock_irq(&xas
);
973 xas_unlock_irq(&xas
);
975 trace_dax_writeback_range_done(inode
, xas
.xa_index
, end_index
);
978 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range
);
980 static sector_t
dax_iomap_sector(struct iomap
*iomap
, loff_t pos
)
982 return (iomap
->addr
+ (pos
& PAGE_MASK
) - iomap
->offset
) >> 9;
985 static int dax_iomap_pfn(struct iomap
*iomap
, loff_t pos
, size_t size
,
988 const sector_t sector
= dax_iomap_sector(iomap
, pos
);
993 rc
= bdev_dax_pgoff(iomap
->bdev
, sector
, size
, &pgoff
);
996 id
= dax_read_lock();
997 length
= dax_direct_access(iomap
->dax_dev
, pgoff
, PHYS_PFN(size
),
1004 if (PFN_PHYS(length
) < size
)
1006 if (pfn_t_to_pfn(*pfnp
) & (PHYS_PFN(size
)-1))
1008 /* For larger pages we need devmap */
1009 if (length
> 1 && !pfn_t_devmap(*pfnp
))
1013 dax_read_unlock(id
);
1018 * The user has performed a load from a hole in the file. Allocating a new
1019 * page in the file would cause excessive storage usage for workloads with
1020 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1021 * If this page is ever written to we will re-fault and change the mapping to
1022 * point to real DAX storage instead.
1024 static vm_fault_t
dax_load_hole(struct xa_state
*xas
,
1025 struct address_space
*mapping
, void **entry
,
1026 struct vm_fault
*vmf
)
1028 struct inode
*inode
= mapping
->host
;
1029 unsigned long vaddr
= vmf
->address
;
1030 pfn_t pfn
= pfn_to_pfn_t(my_zero_pfn(vaddr
));
1033 *entry
= dax_insert_entry(xas
, mapping
, vmf
, *entry
, pfn
,
1034 DAX_ZERO_PAGE
, false);
1036 ret
= vmf_insert_mixed(vmf
->vma
, vaddr
, pfn
);
1037 trace_dax_load_hole(inode
, vmf
, ret
);
1041 static bool dax_range_is_aligned(struct block_device
*bdev
,
1042 unsigned int offset
, unsigned int length
)
1044 unsigned short sector_size
= bdev_logical_block_size(bdev
);
1046 if (!IS_ALIGNED(offset
, sector_size
))
1048 if (!IS_ALIGNED(length
, sector_size
))
1054 int __dax_zero_page_range(struct block_device
*bdev
,
1055 struct dax_device
*dax_dev
, sector_t sector
,
1056 unsigned int offset
, unsigned int size
)
1058 if (dax_range_is_aligned(bdev
, offset
, size
)) {
1059 sector_t start_sector
= sector
+ (offset
>> 9);
1061 return blkdev_issue_zeroout(bdev
, start_sector
,
1062 size
>> 9, GFP_NOFS
, 0);
1068 rc
= bdev_dax_pgoff(bdev
, sector
, PAGE_SIZE
, &pgoff
);
1072 id
= dax_read_lock();
1073 rc
= dax_direct_access(dax_dev
, pgoff
, 1, &kaddr
, NULL
);
1075 dax_read_unlock(id
);
1078 memset(kaddr
+ offset
, 0, size
);
1079 dax_flush(dax_dev
, kaddr
+ offset
, size
);
1080 dax_read_unlock(id
);
1084 EXPORT_SYMBOL_GPL(__dax_zero_page_range
);
1087 dax_iomap_actor(struct inode
*inode
, loff_t pos
, loff_t length
, void *data
,
1088 struct iomap
*iomap
)
1090 struct block_device
*bdev
= iomap
->bdev
;
1091 struct dax_device
*dax_dev
= iomap
->dax_dev
;
1092 struct iov_iter
*iter
= data
;
1093 loff_t end
= pos
+ length
, done
= 0;
1098 if (iov_iter_rw(iter
) == READ
) {
1099 end
= min(end
, i_size_read(inode
));
1103 if (iomap
->type
== IOMAP_HOLE
|| iomap
->type
== IOMAP_UNWRITTEN
)
1104 return iov_iter_zero(min(length
, end
- pos
), iter
);
1107 if (WARN_ON_ONCE(iomap
->type
!= IOMAP_MAPPED
))
1111 * Write can allocate block for an area which has a hole page mapped
1112 * into page tables. We have to tear down these mappings so that data
1113 * written by write(2) is visible in mmap.
1115 if (iomap
->flags
& IOMAP_F_NEW
) {
1116 invalidate_inode_pages2_range(inode
->i_mapping
,
1118 (end
- 1) >> PAGE_SHIFT
);
1121 id
= dax_read_lock();
1123 unsigned offset
= pos
& (PAGE_SIZE
- 1);
1124 const size_t size
= ALIGN(length
+ offset
, PAGE_SIZE
);
1125 const sector_t sector
= dax_iomap_sector(iomap
, pos
);
1130 if (fatal_signal_pending(current
)) {
1135 ret
= bdev_dax_pgoff(bdev
, sector
, size
, &pgoff
);
1139 map_len
= dax_direct_access(dax_dev
, pgoff
, PHYS_PFN(size
),
1146 map_len
= PFN_PHYS(map_len
);
1149 if (map_len
> end
- pos
)
1150 map_len
= end
- pos
;
1153 * The userspace address for the memory copy has already been
1154 * validated via access_ok() in either vfs_read() or
1155 * vfs_write(), depending on which operation we are doing.
1157 if (iov_iter_rw(iter
) == WRITE
)
1158 xfer
= dax_copy_from_iter(dax_dev
, pgoff
, kaddr
,
1161 xfer
= dax_copy_to_iter(dax_dev
, pgoff
, kaddr
,
1173 dax_read_unlock(id
);
1175 return done
? done
: ret
;
1179 * dax_iomap_rw - Perform I/O to a DAX file
1180 * @iocb: The control block for this I/O
1181 * @iter: The addresses to do I/O from or to
1182 * @ops: iomap ops passed from the file system
1184 * This function performs read and write operations to directly mapped
1185 * persistent memory. The callers needs to take care of read/write exclusion
1186 * and evicting any page cache pages in the region under I/O.
1189 dax_iomap_rw(struct kiocb
*iocb
, struct iov_iter
*iter
,
1190 const struct iomap_ops
*ops
)
1192 struct address_space
*mapping
= iocb
->ki_filp
->f_mapping
;
1193 struct inode
*inode
= mapping
->host
;
1194 loff_t pos
= iocb
->ki_pos
, ret
= 0, done
= 0;
1197 if (iov_iter_rw(iter
) == WRITE
) {
1198 lockdep_assert_held_exclusive(&inode
->i_rwsem
);
1199 flags
|= IOMAP_WRITE
;
1201 lockdep_assert_held(&inode
->i_rwsem
);
1204 while (iov_iter_count(iter
)) {
1205 ret
= iomap_apply(inode
, pos
, iov_iter_count(iter
), flags
, ops
,
1206 iter
, dax_iomap_actor
);
1213 iocb
->ki_pos
+= done
;
1214 return done
? done
: ret
;
1216 EXPORT_SYMBOL_GPL(dax_iomap_rw
);
1218 static vm_fault_t
dax_fault_return(int error
)
1221 return VM_FAULT_NOPAGE
;
1222 return vmf_error(error
);
1226 * MAP_SYNC on a dax mapping guarantees dirty metadata is
1227 * flushed on write-faults (non-cow), but not read-faults.
1229 static bool dax_fault_is_synchronous(unsigned long flags
,
1230 struct vm_area_struct
*vma
, struct iomap
*iomap
)
1232 return (flags
& IOMAP_WRITE
) && (vma
->vm_flags
& VM_SYNC
)
1233 && (iomap
->flags
& IOMAP_F_DIRTY
);
1236 static vm_fault_t
dax_iomap_pte_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1237 int *iomap_errp
, const struct iomap_ops
*ops
)
1239 struct vm_area_struct
*vma
= vmf
->vma
;
1240 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1241 XA_STATE(xas
, &mapping
->i_pages
, vmf
->pgoff
);
1242 struct inode
*inode
= mapping
->host
;
1243 unsigned long vaddr
= vmf
->address
;
1244 loff_t pos
= (loff_t
)vmf
->pgoff
<< PAGE_SHIFT
;
1245 struct iomap iomap
= { 0 };
1246 unsigned flags
= IOMAP_FAULT
;
1247 int error
, major
= 0;
1248 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1254 trace_dax_pte_fault(inode
, vmf
, ret
);
1256 * Check whether offset isn't beyond end of file now. Caller is supposed
1257 * to hold locks serializing us with truncate / punch hole so this is
1260 if (pos
>= i_size_read(inode
)) {
1261 ret
= VM_FAULT_SIGBUS
;
1265 if (write
&& !vmf
->cow_page
)
1266 flags
|= IOMAP_WRITE
;
1268 entry
= grab_mapping_entry(&xas
, mapping
, 0);
1269 if (xa_is_internal(entry
)) {
1270 ret
= xa_to_internal(entry
);
1275 * It is possible, particularly with mixed reads & writes to private
1276 * mappings, that we have raced with a PMD fault that overlaps with
1277 * the PTE we need to set up. If so just return and the fault will be
1280 if (pmd_trans_huge(*vmf
->pmd
) || pmd_devmap(*vmf
->pmd
)) {
1281 ret
= VM_FAULT_NOPAGE
;
1286 * Note that we don't bother to use iomap_apply here: DAX required
1287 * the file system block size to be equal the page size, which means
1288 * that we never have to deal with more than a single extent here.
1290 error
= ops
->iomap_begin(inode
, pos
, PAGE_SIZE
, flags
, &iomap
);
1292 *iomap_errp
= error
;
1294 ret
= dax_fault_return(error
);
1297 if (WARN_ON_ONCE(iomap
.offset
+ iomap
.length
< pos
+ PAGE_SIZE
)) {
1298 error
= -EIO
; /* fs corruption? */
1299 goto error_finish_iomap
;
1302 if (vmf
->cow_page
) {
1303 sector_t sector
= dax_iomap_sector(&iomap
, pos
);
1305 switch (iomap
.type
) {
1307 case IOMAP_UNWRITTEN
:
1308 clear_user_highpage(vmf
->cow_page
, vaddr
);
1311 error
= copy_user_dax(iomap
.bdev
, iomap
.dax_dev
,
1312 sector
, PAGE_SIZE
, vmf
->cow_page
, vaddr
);
1321 goto error_finish_iomap
;
1323 __SetPageUptodate(vmf
->cow_page
);
1324 ret
= finish_fault(vmf
);
1326 ret
= VM_FAULT_DONE_COW
;
1330 sync
= dax_fault_is_synchronous(flags
, vma
, &iomap
);
1332 switch (iomap
.type
) {
1334 if (iomap
.flags
& IOMAP_F_NEW
) {
1335 count_vm_event(PGMAJFAULT
);
1336 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
1337 major
= VM_FAULT_MAJOR
;
1339 error
= dax_iomap_pfn(&iomap
, pos
, PAGE_SIZE
, &pfn
);
1341 goto error_finish_iomap
;
1343 entry
= dax_insert_entry(&xas
, mapping
, vmf
, entry
, pfn
,
1347 * If we are doing synchronous page fault and inode needs fsync,
1348 * we can insert PTE into page tables only after that happens.
1349 * Skip insertion for now and return the pfn so that caller can
1350 * insert it after fsync is done.
1353 if (WARN_ON_ONCE(!pfnp
)) {
1355 goto error_finish_iomap
;
1358 ret
= VM_FAULT_NEEDDSYNC
| major
;
1361 trace_dax_insert_mapping(inode
, vmf
, entry
);
1363 ret
= vmf_insert_mixed_mkwrite(vma
, vaddr
, pfn
);
1365 ret
= vmf_insert_mixed(vma
, vaddr
, pfn
);
1368 case IOMAP_UNWRITTEN
:
1371 ret
= dax_load_hole(&xas
, mapping
, &entry
, vmf
);
1382 ret
= dax_fault_return(error
);
1384 if (ops
->iomap_end
) {
1385 int copied
= PAGE_SIZE
;
1387 if (ret
& VM_FAULT_ERROR
)
1390 * The fault is done by now and there's no way back (other
1391 * thread may be already happily using PTE we have installed).
1392 * Just ignore error from ->iomap_end since we cannot do much
1395 ops
->iomap_end(inode
, pos
, PAGE_SIZE
, copied
, flags
, &iomap
);
1398 dax_unlock_entry(&xas
, entry
);
1400 trace_dax_pte_fault_done(inode
, vmf
, ret
);
1404 #ifdef CONFIG_FS_DAX_PMD
1405 static vm_fault_t
dax_pmd_load_hole(struct xa_state
*xas
, struct vm_fault
*vmf
,
1406 struct iomap
*iomap
, void **entry
)
1408 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1409 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
1410 struct inode
*inode
= mapping
->host
;
1411 struct page
*zero_page
;
1416 zero_page
= mm_get_huge_zero_page(vmf
->vma
->vm_mm
);
1418 if (unlikely(!zero_page
))
1421 pfn
= page_to_pfn_t(zero_page
);
1422 *entry
= dax_insert_entry(xas
, mapping
, vmf
, *entry
, pfn
,
1423 DAX_PMD
| DAX_ZERO_PAGE
, false);
1425 ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1426 if (!pmd_none(*(vmf
->pmd
))) {
1431 pmd_entry
= mk_pmd(zero_page
, vmf
->vma
->vm_page_prot
);
1432 pmd_entry
= pmd_mkhuge(pmd_entry
);
1433 set_pmd_at(vmf
->vma
->vm_mm
, pmd_addr
, vmf
->pmd
, pmd_entry
);
1435 trace_dax_pmd_load_hole(inode
, vmf
, zero_page
, *entry
);
1436 return VM_FAULT_NOPAGE
;
1439 trace_dax_pmd_load_hole_fallback(inode
, vmf
, zero_page
, *entry
);
1440 return VM_FAULT_FALLBACK
;
1443 static vm_fault_t
dax_iomap_pmd_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1444 const struct iomap_ops
*ops
)
1446 struct vm_area_struct
*vma
= vmf
->vma
;
1447 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1448 XA_STATE_ORDER(xas
, &mapping
->i_pages
, vmf
->pgoff
, PMD_ORDER
);
1449 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
1450 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1452 unsigned int iomap_flags
= (write
? IOMAP_WRITE
: 0) | IOMAP_FAULT
;
1453 struct inode
*inode
= mapping
->host
;
1454 vm_fault_t result
= VM_FAULT_FALLBACK
;
1455 struct iomap iomap
= { 0 };
1463 * Check whether offset isn't beyond end of file now. Caller is
1464 * supposed to hold locks serializing us with truncate / punch hole so
1465 * this is a reliable test.
1467 max_pgoff
= DIV_ROUND_UP(i_size_read(inode
), PAGE_SIZE
);
1469 trace_dax_pmd_fault(inode
, vmf
, max_pgoff
, 0);
1472 * Make sure that the faulting address's PMD offset (color) matches
1473 * the PMD offset from the start of the file. This is necessary so
1474 * that a PMD range in the page table overlaps exactly with a PMD
1475 * range in the page cache.
1477 if ((vmf
->pgoff
& PG_PMD_COLOUR
) !=
1478 ((vmf
->address
>> PAGE_SHIFT
) & PG_PMD_COLOUR
))
1481 /* Fall back to PTEs if we're going to COW */
1482 if (write
&& !(vma
->vm_flags
& VM_SHARED
))
1485 /* If the PMD would extend outside the VMA */
1486 if (pmd_addr
< vma
->vm_start
)
1488 if ((pmd_addr
+ PMD_SIZE
) > vma
->vm_end
)
1491 if (xas
.xa_index
>= max_pgoff
) {
1492 result
= VM_FAULT_SIGBUS
;
1496 /* If the PMD would extend beyond the file size */
1497 if ((xas
.xa_index
| PG_PMD_COLOUR
) >= max_pgoff
)
1501 * grab_mapping_entry() will make sure we get an empty PMD entry,
1502 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1503 * entry is already in the array, for instance), it will return
1504 * VM_FAULT_FALLBACK.
1506 entry
= grab_mapping_entry(&xas
, mapping
, DAX_PMD
);
1507 if (xa_is_internal(entry
)) {
1508 result
= xa_to_internal(entry
);
1513 * It is possible, particularly with mixed reads & writes to private
1514 * mappings, that we have raced with a PTE fault that overlaps with
1515 * the PMD we need to set up. If so just return and the fault will be
1518 if (!pmd_none(*vmf
->pmd
) && !pmd_trans_huge(*vmf
->pmd
) &&
1519 !pmd_devmap(*vmf
->pmd
)) {
1525 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1526 * setting up a mapping, so really we're using iomap_begin() as a way
1527 * to look up our filesystem block.
1529 pos
= (loff_t
)xas
.xa_index
<< PAGE_SHIFT
;
1530 error
= ops
->iomap_begin(inode
, pos
, PMD_SIZE
, iomap_flags
, &iomap
);
1534 if (iomap
.offset
+ iomap
.length
< pos
+ PMD_SIZE
)
1537 sync
= dax_fault_is_synchronous(iomap_flags
, vma
, &iomap
);
1539 switch (iomap
.type
) {
1541 error
= dax_iomap_pfn(&iomap
, pos
, PMD_SIZE
, &pfn
);
1545 entry
= dax_insert_entry(&xas
, mapping
, vmf
, entry
, pfn
,
1546 DAX_PMD
, write
&& !sync
);
1549 * If we are doing synchronous page fault and inode needs fsync,
1550 * we can insert PMD into page tables only after that happens.
1551 * Skip insertion for now and return the pfn so that caller can
1552 * insert it after fsync is done.
1555 if (WARN_ON_ONCE(!pfnp
))
1558 result
= VM_FAULT_NEEDDSYNC
;
1562 trace_dax_pmd_insert_mapping(inode
, vmf
, PMD_SIZE
, pfn
, entry
);
1563 result
= vmf_insert_pfn_pmd(vma
, vmf
->address
, vmf
->pmd
, pfn
,
1566 case IOMAP_UNWRITTEN
:
1568 if (WARN_ON_ONCE(write
))
1570 result
= dax_pmd_load_hole(&xas
, vmf
, &iomap
, &entry
);
1578 if (ops
->iomap_end
) {
1579 int copied
= PMD_SIZE
;
1581 if (result
== VM_FAULT_FALLBACK
)
1584 * The fault is done by now and there's no way back (other
1585 * thread may be already happily using PMD we have installed).
1586 * Just ignore error from ->iomap_end since we cannot do much
1589 ops
->iomap_end(inode
, pos
, PMD_SIZE
, copied
, iomap_flags
,
1593 dax_unlock_entry(&xas
, entry
);
1595 if (result
== VM_FAULT_FALLBACK
) {
1596 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1597 count_vm_event(THP_FAULT_FALLBACK
);
1600 trace_dax_pmd_fault_done(inode
, vmf
, max_pgoff
, result
);
1604 static vm_fault_t
dax_iomap_pmd_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1605 const struct iomap_ops
*ops
)
1607 return VM_FAULT_FALLBACK
;
1609 #endif /* CONFIG_FS_DAX_PMD */
1612 * dax_iomap_fault - handle a page fault on a DAX file
1613 * @vmf: The description of the fault
1614 * @pe_size: Size of the page to fault in
1615 * @pfnp: PFN to insert for synchronous faults if fsync is required
1616 * @iomap_errp: Storage for detailed error code in case of error
1617 * @ops: Iomap ops passed from the file system
1619 * When a page fault occurs, filesystems may call this helper in
1620 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1621 * has done all the necessary locking for page fault to proceed
1624 vm_fault_t
dax_iomap_fault(struct vm_fault
*vmf
, enum page_entry_size pe_size
,
1625 pfn_t
*pfnp
, int *iomap_errp
, const struct iomap_ops
*ops
)
1629 return dax_iomap_pte_fault(vmf
, pfnp
, iomap_errp
, ops
);
1631 return dax_iomap_pmd_fault(vmf
, pfnp
, ops
);
1633 return VM_FAULT_FALLBACK
;
1636 EXPORT_SYMBOL_GPL(dax_iomap_fault
);
1639 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1640 * @vmf: The description of the fault
1641 * @pfn: PFN to insert
1642 * @order: Order of entry to insert.
1644 * This function inserts a writeable PTE or PMD entry into the page tables
1645 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1648 dax_insert_pfn_mkwrite(struct vm_fault
*vmf
, pfn_t pfn
, unsigned int order
)
1650 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1651 XA_STATE_ORDER(xas
, &mapping
->i_pages
, vmf
->pgoff
, order
);
1656 entry
= get_unlocked_entry(&xas
);
1657 /* Did we race with someone splitting entry or so? */
1659 (order
== 0 && !dax_is_pte_entry(entry
)) ||
1660 (order
== PMD_ORDER
&& !dax_is_pmd_entry(entry
))) {
1661 put_unlocked_entry(&xas
, entry
);
1662 xas_unlock_irq(&xas
);
1663 trace_dax_insert_pfn_mkwrite_no_entry(mapping
->host
, vmf
,
1665 return VM_FAULT_NOPAGE
;
1667 xas_set_mark(&xas
, PAGECACHE_TAG_DIRTY
);
1668 dax_lock_entry(&xas
, entry
);
1669 xas_unlock_irq(&xas
);
1671 ret
= vmf_insert_mixed_mkwrite(vmf
->vma
, vmf
->address
, pfn
);
1672 #ifdef CONFIG_FS_DAX_PMD
1673 else if (order
== PMD_ORDER
)
1674 ret
= vmf_insert_pfn_pmd(vmf
->vma
, vmf
->address
, vmf
->pmd
,
1678 ret
= VM_FAULT_FALLBACK
;
1679 dax_unlock_entry(&xas
, entry
);
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
;
1699 unsigned int order
= pe_order(pe_size
);
1700 size_t len
= PAGE_SIZE
<< order
;
1702 err
= vfs_fsync_range(vmf
->vma
->vm_file
, start
, start
+ len
- 1, 1);
1704 return VM_FAULT_SIGBUS
;
1705 return dax_insert_pfn_mkwrite(vmf
, pfn
, order
);
1707 EXPORT_SYMBOL_GPL(dax_finish_sync_fault
);