1 // SPDX-License-Identifier: GPL-2.0-only
3 * fs/dax.c - Direct Access filesystem code
4 * Copyright (c) 2013-2014 Intel Corporation
5 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
6 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
9 #include <linux/atomic.h>
10 #include <linux/blkdev.h>
11 #include <linux/buffer_head.h>
12 #include <linux/dax.h>
14 #include <linux/highmem.h>
15 #include <linux/memcontrol.h>
17 #include <linux/mutex.h>
18 #include <linux/pagevec.h>
19 #include <linux/sched.h>
20 #include <linux/sched/signal.h>
21 #include <linux/uio.h>
22 #include <linux/vmstat.h>
23 #include <linux/pfn_t.h>
24 #include <linux/sizes.h>
25 #include <linux/mmu_notifier.h>
26 #include <linux/iomap.h>
27 #include <linux/rmap.h>
28 #include <asm/pgalloc.h>
30 #define CREATE_TRACE_POINTS
31 #include <trace/events/fs_dax.h>
33 static inline unsigned int pe_order(enum page_entry_size pe_size
)
35 if (pe_size
== PE_SIZE_PTE
)
36 return PAGE_SHIFT
- PAGE_SHIFT
;
37 if (pe_size
== PE_SIZE_PMD
)
38 return PMD_SHIFT
- PAGE_SHIFT
;
39 if (pe_size
== PE_SIZE_PUD
)
40 return PUD_SHIFT
- PAGE_SHIFT
;
44 /* We choose 4096 entries - same as per-zone page wait tables */
45 #define DAX_WAIT_TABLE_BITS 12
46 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
48 /* The 'colour' (ie low bits) within a PMD of a page offset. */
49 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
50 #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
52 /* The order of a PMD entry */
53 #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT)
55 static wait_queue_head_t wait_table
[DAX_WAIT_TABLE_ENTRIES
];
57 static int __init
init_dax_wait_table(void)
61 for (i
= 0; i
< DAX_WAIT_TABLE_ENTRIES
; i
++)
62 init_waitqueue_head(wait_table
+ i
);
65 fs_initcall(init_dax_wait_table
);
68 * DAX pagecache entries use XArray value entries so they can't be mistaken
69 * for pages. We use one bit for locking, one bit for the entry size (PMD)
70 * and two more to tell us if the entry is a zero page or an empty entry that
71 * is just used for locking. In total four special bits.
73 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
74 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
78 #define DAX_LOCKED (1UL << 0)
79 #define DAX_PMD (1UL << 1)
80 #define DAX_ZERO_PAGE (1UL << 2)
81 #define DAX_EMPTY (1UL << 3)
83 static unsigned long dax_to_pfn(void *entry
)
85 return xa_to_value(entry
) >> DAX_SHIFT
;
88 static void *dax_make_entry(pfn_t pfn
, unsigned long flags
)
90 return xa_mk_value(flags
| (pfn_t_to_pfn(pfn
) << DAX_SHIFT
));
93 static bool dax_is_locked(void *entry
)
95 return xa_to_value(entry
) & DAX_LOCKED
;
98 static unsigned int dax_entry_order(void *entry
)
100 if (xa_to_value(entry
) & DAX_PMD
)
105 static unsigned long dax_is_pmd_entry(void *entry
)
107 return xa_to_value(entry
) & DAX_PMD
;
110 static bool dax_is_pte_entry(void *entry
)
112 return !(xa_to_value(entry
) & DAX_PMD
);
115 static int dax_is_zero_entry(void *entry
)
117 return xa_to_value(entry
) & DAX_ZERO_PAGE
;
120 static int dax_is_empty_entry(void *entry
)
122 return xa_to_value(entry
) & DAX_EMPTY
;
126 * true if the entry that was found is of a smaller order than the entry
127 * we were looking for
129 static bool dax_is_conflict(void *entry
)
131 return entry
== XA_RETRY_ENTRY
;
135 * DAX page cache entry locking
137 struct exceptional_entry_key
{
142 struct wait_exceptional_entry_queue
{
143 wait_queue_entry_t wait
;
144 struct exceptional_entry_key key
;
148 * enum dax_wake_mode: waitqueue wakeup behaviour
149 * @WAKE_ALL: wake all waiters in the waitqueue
150 * @WAKE_NEXT: wake only the first waiter in the waitqueue
157 static wait_queue_head_t
*dax_entry_waitqueue(struct xa_state
*xas
,
158 void *entry
, struct exceptional_entry_key
*key
)
161 unsigned long index
= xas
->xa_index
;
164 * If 'entry' is a PMD, align the 'index' that we use for the wait
165 * queue to the start of that PMD. This ensures that all offsets in
166 * the range covered by the PMD map to the same bit lock.
168 if (dax_is_pmd_entry(entry
))
169 index
&= ~PG_PMD_COLOUR
;
171 key
->entry_start
= index
;
173 hash
= hash_long((unsigned long)xas
->xa
^ index
, DAX_WAIT_TABLE_BITS
);
174 return wait_table
+ hash
;
177 static int wake_exceptional_entry_func(wait_queue_entry_t
*wait
,
178 unsigned int mode
, int sync
, void *keyp
)
180 struct exceptional_entry_key
*key
= keyp
;
181 struct wait_exceptional_entry_queue
*ewait
=
182 container_of(wait
, struct wait_exceptional_entry_queue
, wait
);
184 if (key
->xa
!= ewait
->key
.xa
||
185 key
->entry_start
!= ewait
->key
.entry_start
)
187 return autoremove_wake_function(wait
, mode
, sync
, NULL
);
191 * @entry may no longer be the entry at the index in the mapping.
192 * The important information it's conveying is whether the entry at
193 * this index used to be a PMD entry.
195 static void dax_wake_entry(struct xa_state
*xas
, void *entry
,
196 enum dax_wake_mode mode
)
198 struct exceptional_entry_key key
;
199 wait_queue_head_t
*wq
;
201 wq
= dax_entry_waitqueue(xas
, entry
, &key
);
204 * Checking for locked entry and prepare_to_wait_exclusive() happens
205 * under the i_pages lock, ditto for entry handling in our callers.
206 * So at this point all tasks that could have seen our entry locked
207 * must be in the waitqueue and the following check will see them.
209 if (waitqueue_active(wq
))
210 __wake_up(wq
, TASK_NORMAL
, mode
== WAKE_ALL
? 0 : 1, &key
);
214 * Look up entry in page cache, wait for it to become unlocked if it
215 * is a DAX entry and return it. The caller must subsequently call
216 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
217 * if it did. The entry returned may have a larger order than @order.
218 * If @order is larger than the order of the entry found in i_pages, this
219 * function returns a dax_is_conflict entry.
221 * Must be called with the i_pages lock held.
223 static void *get_unlocked_entry(struct xa_state
*xas
, unsigned int order
)
226 struct wait_exceptional_entry_queue ewait
;
227 wait_queue_head_t
*wq
;
229 init_wait(&ewait
.wait
);
230 ewait
.wait
.func
= wake_exceptional_entry_func
;
233 entry
= xas_find_conflict(xas
);
234 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
236 if (dax_entry_order(entry
) < order
)
237 return XA_RETRY_ENTRY
;
238 if (!dax_is_locked(entry
))
241 wq
= dax_entry_waitqueue(xas
, entry
, &ewait
.key
);
242 prepare_to_wait_exclusive(wq
, &ewait
.wait
,
243 TASK_UNINTERRUPTIBLE
);
247 finish_wait(wq
, &ewait
.wait
);
253 * The only thing keeping the address space around is the i_pages lock
254 * (it's cycled in clear_inode() after removing the entries from i_pages)
255 * After we call xas_unlock_irq(), we cannot touch xas->xa.
257 static void wait_entry_unlocked(struct xa_state
*xas
, void *entry
)
259 struct wait_exceptional_entry_queue ewait
;
260 wait_queue_head_t
*wq
;
262 init_wait(&ewait
.wait
);
263 ewait
.wait
.func
= wake_exceptional_entry_func
;
265 wq
= dax_entry_waitqueue(xas
, entry
, &ewait
.key
);
267 * Unlike get_unlocked_entry() there is no guarantee that this
268 * path ever successfully retrieves an unlocked entry before an
269 * inode dies. Perform a non-exclusive wait in case this path
270 * never successfully performs its own wake up.
272 prepare_to_wait(wq
, &ewait
.wait
, TASK_UNINTERRUPTIBLE
);
275 finish_wait(wq
, &ewait
.wait
);
278 static void put_unlocked_entry(struct xa_state
*xas
, void *entry
,
279 enum dax_wake_mode mode
)
281 if (entry
&& !dax_is_conflict(entry
))
282 dax_wake_entry(xas
, entry
, mode
);
286 * We used the xa_state to get the entry, but then we locked the entry and
287 * dropped the xa_lock, so we know the xa_state is stale and must be reset
290 static void dax_unlock_entry(struct xa_state
*xas
, void *entry
)
294 BUG_ON(dax_is_locked(entry
));
297 old
= xas_store(xas
, entry
);
299 BUG_ON(!dax_is_locked(old
));
300 dax_wake_entry(xas
, entry
, WAKE_NEXT
);
304 * Return: The entry stored at this location before it was locked.
306 static void *dax_lock_entry(struct xa_state
*xas
, void *entry
)
308 unsigned long v
= xa_to_value(entry
);
309 return xas_store(xas
, xa_mk_value(v
| DAX_LOCKED
));
312 static unsigned long dax_entry_size(void *entry
)
314 if (dax_is_zero_entry(entry
))
316 else if (dax_is_empty_entry(entry
))
318 else if (dax_is_pmd_entry(entry
))
324 static unsigned long dax_end_pfn(void *entry
)
326 return dax_to_pfn(entry
) + dax_entry_size(entry
) / PAGE_SIZE
;
330 * Iterate through all mapped pfns represented by an entry, i.e. skip
331 * 'empty' and 'zero' entries.
333 #define for_each_mapped_pfn(entry, pfn) \
334 for (pfn = dax_to_pfn(entry); \
335 pfn < dax_end_pfn(entry); pfn++)
337 static inline bool dax_page_is_shared(struct page
*page
)
339 return page
->mapping
== PAGE_MAPPING_DAX_SHARED
;
343 * Set the page->mapping with PAGE_MAPPING_DAX_SHARED flag, increase the
346 static inline void dax_page_share_get(struct page
*page
)
348 if (page
->mapping
!= PAGE_MAPPING_DAX_SHARED
) {
350 * Reset the index if the page was already mapped
355 page
->mapping
= PAGE_MAPPING_DAX_SHARED
;
360 static inline unsigned long dax_page_share_put(struct page
*page
)
362 return --page
->share
;
366 * When it is called in dax_insert_entry(), the shared flag will indicate that
367 * whether this entry is shared by multiple files. If so, set the page->mapping
368 * PAGE_MAPPING_DAX_SHARED, and use page->share as refcount.
370 static void dax_associate_entry(void *entry
, struct address_space
*mapping
,
371 struct vm_area_struct
*vma
, unsigned long address
, bool shared
)
373 unsigned long size
= dax_entry_size(entry
), pfn
, index
;
376 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
379 index
= linear_page_index(vma
, address
& ~(size
- 1));
380 for_each_mapped_pfn(entry
, pfn
) {
381 struct page
*page
= pfn_to_page(pfn
);
384 dax_page_share_get(page
);
386 WARN_ON_ONCE(page
->mapping
);
387 page
->mapping
= mapping
;
388 page
->index
= index
+ i
++;
393 static void dax_disassociate_entry(void *entry
, struct address_space
*mapping
,
398 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
401 for_each_mapped_pfn(entry
, pfn
) {
402 struct page
*page
= pfn_to_page(pfn
);
404 WARN_ON_ONCE(trunc
&& page_ref_count(page
) > 1);
405 if (dax_page_is_shared(page
)) {
406 /* keep the shared flag if this page is still shared */
407 if (dax_page_share_put(page
) > 0)
410 WARN_ON_ONCE(page
->mapping
&& page
->mapping
!= mapping
);
411 page
->mapping
= NULL
;
416 static struct page
*dax_busy_page(void *entry
)
420 for_each_mapped_pfn(entry
, pfn
) {
421 struct page
*page
= pfn_to_page(pfn
);
423 if (page_ref_count(page
) > 1)
430 * dax_lock_page - Lock the DAX entry corresponding to a page
431 * @page: The page whose entry we want to lock
433 * Context: Process context.
434 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
437 dax_entry_t
dax_lock_page(struct page
*page
)
439 XA_STATE(xas
, NULL
, 0);
442 /* Ensure page->mapping isn't freed while we look at it */
445 struct address_space
*mapping
= READ_ONCE(page
->mapping
);
448 if (!mapping
|| !dax_mapping(mapping
))
452 * In the device-dax case there's no need to lock, a
453 * struct dev_pagemap pin is sufficient to keep the
454 * inode alive, and we assume we have dev_pagemap pin
455 * otherwise we would not have a valid pfn_to_page()
458 entry
= (void *)~0UL;
459 if (S_ISCHR(mapping
->host
->i_mode
))
462 xas
.xa
= &mapping
->i_pages
;
464 if (mapping
!= page
->mapping
) {
465 xas_unlock_irq(&xas
);
468 xas_set(&xas
, page
->index
);
469 entry
= xas_load(&xas
);
470 if (dax_is_locked(entry
)) {
472 wait_entry_unlocked(&xas
, entry
);
476 dax_lock_entry(&xas
, entry
);
477 xas_unlock_irq(&xas
);
481 return (dax_entry_t
)entry
;
484 void dax_unlock_page(struct page
*page
, dax_entry_t cookie
)
486 struct address_space
*mapping
= page
->mapping
;
487 XA_STATE(xas
, &mapping
->i_pages
, page
->index
);
489 if (S_ISCHR(mapping
->host
->i_mode
))
492 dax_unlock_entry(&xas
, (void *)cookie
);
496 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping
497 * @mapping: the file's mapping whose entry we want to lock
498 * @index: the offset within this file
499 * @page: output the dax page corresponding to this dax entry
501 * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry
502 * could not be locked.
504 dax_entry_t
dax_lock_mapping_entry(struct address_space
*mapping
, pgoff_t index
,
507 XA_STATE(xas
, NULL
, 0);
513 if (!dax_mapping(mapping
))
516 xas
.xa
= &mapping
->i_pages
;
518 xas_set(&xas
, index
);
519 entry
= xas_load(&xas
);
520 if (dax_is_locked(entry
)) {
522 wait_entry_unlocked(&xas
, entry
);
527 dax_is_zero_entry(entry
) || dax_is_empty_entry(entry
)) {
529 * Because we are looking for entry from file's mapping
530 * and index, so the entry may not be inserted for now,
531 * or even a zero/empty entry. We don't think this is
532 * an error case. So, return a special value and do
535 entry
= (void *)~0UL;
537 *page
= pfn_to_page(dax_to_pfn(entry
));
538 dax_lock_entry(&xas
, entry
);
540 xas_unlock_irq(&xas
);
544 return (dax_entry_t
)entry
;
547 void dax_unlock_mapping_entry(struct address_space
*mapping
, pgoff_t index
,
550 XA_STATE(xas
, &mapping
->i_pages
, index
);
555 dax_unlock_entry(&xas
, (void *)cookie
);
559 * Find page cache entry at given index. If it is a DAX entry, return it
560 * with the entry locked. If the page cache doesn't contain an entry at
561 * that index, add a locked empty entry.
563 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
564 * either return that locked entry or will return VM_FAULT_FALLBACK.
565 * This will happen if there are any PTE entries within the PMD range
566 * that we are requesting.
568 * We always favor PTE entries over PMD entries. There isn't a flow where we
569 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
570 * insertion will fail if it finds any PTE entries already in the tree, and a
571 * PTE insertion will cause an existing PMD entry to be unmapped and
572 * downgraded to PTE entries. This happens for both PMD zero pages as
573 * well as PMD empty entries.
575 * The exception to this downgrade path is for PMD entries that have
576 * real storage backing them. We will leave these real PMD entries in
577 * the tree, and PTE writes will simply dirty the entire PMD entry.
579 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
580 * persistent memory the benefit is doubtful. We can add that later if we can
583 * On error, this function does not return an ERR_PTR. Instead it returns
584 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
585 * overlap with xarray value entries.
587 static void *grab_mapping_entry(struct xa_state
*xas
,
588 struct address_space
*mapping
, unsigned int order
)
590 unsigned long index
= xas
->xa_index
;
591 bool pmd_downgrade
; /* splitting PMD entry into PTE entries? */
595 pmd_downgrade
= false;
597 entry
= get_unlocked_entry(xas
, order
);
600 if (dax_is_conflict(entry
))
602 if (!xa_is_value(entry
)) {
603 xas_set_err(xas
, -EIO
);
608 if (dax_is_pmd_entry(entry
) &&
609 (dax_is_zero_entry(entry
) ||
610 dax_is_empty_entry(entry
))) {
611 pmd_downgrade
= true;
618 * Make sure 'entry' remains valid while we drop
621 dax_lock_entry(xas
, entry
);
624 * Besides huge zero pages the only other thing that gets
625 * downgraded are empty entries which don't need to be
628 if (dax_is_zero_entry(entry
)) {
630 unmap_mapping_pages(mapping
,
631 xas
->xa_index
& ~PG_PMD_COLOUR
,
637 dax_disassociate_entry(entry
, mapping
, false);
638 xas_store(xas
, NULL
); /* undo the PMD join */
639 dax_wake_entry(xas
, entry
, WAKE_ALL
);
640 mapping
->nrpages
-= PG_PMD_NR
;
646 dax_lock_entry(xas
, entry
);
648 unsigned long flags
= DAX_EMPTY
;
652 entry
= dax_make_entry(pfn_to_pfn_t(0), flags
);
653 dax_lock_entry(xas
, entry
);
656 mapping
->nrpages
+= 1UL << order
;
661 if (xas_nomem(xas
, mapping_gfp_mask(mapping
) & ~__GFP_HIGHMEM
))
663 if (xas
->xa_node
== XA_ERROR(-ENOMEM
))
664 return xa_mk_internal(VM_FAULT_OOM
);
666 return xa_mk_internal(VM_FAULT_SIGBUS
);
670 return xa_mk_internal(VM_FAULT_FALLBACK
);
674 * dax_layout_busy_page_range - find first pinned page in @mapping
675 * @mapping: address space to scan for a page with ref count > 1
676 * @start: Starting offset. Page containing 'start' is included.
677 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
678 * pages from 'start' till the end of file are included.
680 * DAX requires ZONE_DEVICE mapped pages. These pages are never
681 * 'onlined' to the page allocator so they are considered idle when
682 * page->count == 1. A filesystem uses this interface to determine if
683 * any page in the mapping is busy, i.e. for DMA, or other
684 * get_user_pages() usages.
686 * It is expected that the filesystem is holding locks to block the
687 * establishment of new mappings in this address_space. I.e. it expects
688 * to be able to run unmap_mapping_range() and subsequently not race
689 * mapping_mapped() becoming true.
691 struct page
*dax_layout_busy_page_range(struct address_space
*mapping
,
692 loff_t start
, loff_t end
)
695 unsigned int scanned
= 0;
696 struct page
*page
= NULL
;
697 pgoff_t start_idx
= start
>> PAGE_SHIFT
;
699 XA_STATE(xas
, &mapping
->i_pages
, start_idx
);
702 * In the 'limited' case get_user_pages() for dax is disabled.
704 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
707 if (!dax_mapping(mapping
) || !mapping_mapped(mapping
))
710 /* If end == LLONG_MAX, all pages from start to till end of file */
711 if (end
== LLONG_MAX
)
714 end_idx
= end
>> PAGE_SHIFT
;
716 * If we race get_user_pages_fast() here either we'll see the
717 * elevated page count in the iteration and wait, or
718 * get_user_pages_fast() will see that the page it took a reference
719 * against is no longer mapped in the page tables and bail to the
720 * get_user_pages() slow path. The slow path is protected by
721 * pte_lock() and pmd_lock(). New references are not taken without
722 * holding those locks, and unmap_mapping_pages() will not zero the
723 * pte or pmd without holding the respective lock, so we are
724 * guaranteed to either see new references or prevent new
725 * references from being established.
727 unmap_mapping_pages(mapping
, start_idx
, end_idx
- start_idx
+ 1, 0);
730 xas_for_each(&xas
, entry
, end_idx
) {
731 if (WARN_ON_ONCE(!xa_is_value(entry
)))
733 if (unlikely(dax_is_locked(entry
)))
734 entry
= get_unlocked_entry(&xas
, 0);
736 page
= dax_busy_page(entry
);
737 put_unlocked_entry(&xas
, entry
, WAKE_NEXT
);
740 if (++scanned
% XA_CHECK_SCHED
)
744 xas_unlock_irq(&xas
);
748 xas_unlock_irq(&xas
);
751 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range
);
753 struct page
*dax_layout_busy_page(struct address_space
*mapping
)
755 return dax_layout_busy_page_range(mapping
, 0, LLONG_MAX
);
757 EXPORT_SYMBOL_GPL(dax_layout_busy_page
);
759 static int __dax_invalidate_entry(struct address_space
*mapping
,
760 pgoff_t index
, bool trunc
)
762 XA_STATE(xas
, &mapping
->i_pages
, index
);
767 entry
= get_unlocked_entry(&xas
, 0);
768 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
771 (xas_get_mark(&xas
, PAGECACHE_TAG_DIRTY
) ||
772 xas_get_mark(&xas
, PAGECACHE_TAG_TOWRITE
)))
774 dax_disassociate_entry(entry
, mapping
, trunc
);
775 xas_store(&xas
, NULL
);
776 mapping
->nrpages
-= 1UL << dax_entry_order(entry
);
779 put_unlocked_entry(&xas
, entry
, WAKE_ALL
);
780 xas_unlock_irq(&xas
);
785 * Delete DAX entry at @index from @mapping. Wait for it
786 * to be unlocked before deleting it.
788 int dax_delete_mapping_entry(struct address_space
*mapping
, pgoff_t index
)
790 int ret
= __dax_invalidate_entry(mapping
, index
, true);
793 * This gets called from truncate / punch_hole path. As such, the caller
794 * must hold locks protecting against concurrent modifications of the
795 * page cache (usually fs-private i_mmap_sem for writing). Since the
796 * caller has seen a DAX entry for this index, we better find it
797 * at that index as well...
804 * Invalidate DAX entry if it is clean.
806 int dax_invalidate_mapping_entry_sync(struct address_space
*mapping
,
809 return __dax_invalidate_entry(mapping
, index
, false);
812 static pgoff_t
dax_iomap_pgoff(const struct iomap
*iomap
, loff_t pos
)
814 return PHYS_PFN(iomap
->addr
+ (pos
& PAGE_MASK
) - iomap
->offset
);
817 static int copy_cow_page_dax(struct vm_fault
*vmf
, const struct iomap_iter
*iter
)
819 pgoff_t pgoff
= dax_iomap_pgoff(&iter
->iomap
, iter
->pos
);
824 id
= dax_read_lock();
825 rc
= dax_direct_access(iter
->iomap
.dax_dev
, pgoff
, 1, DAX_ACCESS
,
831 vto
= kmap_atomic(vmf
->cow_page
);
832 copy_user_page(vto
, kaddr
, vmf
->address
, vmf
->cow_page
);
839 * MAP_SYNC on a dax mapping guarantees dirty metadata is
840 * flushed on write-faults (non-cow), but not read-faults.
842 static bool dax_fault_is_synchronous(const struct iomap_iter
*iter
,
843 struct vm_area_struct
*vma
)
845 return (iter
->flags
& IOMAP_WRITE
) && (vma
->vm_flags
& VM_SYNC
) &&
846 (iter
->iomap
.flags
& IOMAP_F_DIRTY
);
850 * By this point grab_mapping_entry() has ensured that we have a locked entry
851 * of the appropriate size so we don't have to worry about downgrading PMDs to
852 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
853 * already in the tree, we will skip the insertion and just dirty the PMD as
856 static void *dax_insert_entry(struct xa_state
*xas
, struct vm_fault
*vmf
,
857 const struct iomap_iter
*iter
, void *entry
, pfn_t pfn
,
860 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
861 void *new_entry
= dax_make_entry(pfn
, flags
);
862 bool write
= iter
->flags
& IOMAP_WRITE
;
863 bool dirty
= write
&& !dax_fault_is_synchronous(iter
, vmf
->vma
);
864 bool shared
= iter
->iomap
.flags
& IOMAP_F_SHARED
;
867 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
869 if (shared
|| (dax_is_zero_entry(entry
) && !(flags
& DAX_ZERO_PAGE
))) {
870 unsigned long index
= xas
->xa_index
;
871 /* we are replacing a zero page with block mapping */
872 if (dax_is_pmd_entry(entry
))
873 unmap_mapping_pages(mapping
, index
& ~PG_PMD_COLOUR
,
876 unmap_mapping_pages(mapping
, index
, 1, false);
881 if (shared
|| dax_is_zero_entry(entry
) || dax_is_empty_entry(entry
)) {
884 dax_disassociate_entry(entry
, mapping
, false);
885 dax_associate_entry(new_entry
, mapping
, vmf
->vma
, vmf
->address
,
888 * Only swap our new entry into the page cache if the current
889 * entry is a zero page or an empty entry. If a normal PTE or
890 * PMD entry is already in the cache, we leave it alone. This
891 * means that if we are trying to insert a PTE and the
892 * existing entry is a PMD, we will just leave the PMD in the
893 * tree and dirty it if necessary.
895 old
= dax_lock_entry(xas
, new_entry
);
896 WARN_ON_ONCE(old
!= xa_mk_value(xa_to_value(entry
) |
900 xas_load(xas
); /* Walk the xa_state */
904 xas_set_mark(xas
, PAGECACHE_TAG_DIRTY
);
907 xas_set_mark(xas
, PAGECACHE_TAG_TOWRITE
);
913 static int dax_writeback_one(struct xa_state
*xas
, struct dax_device
*dax_dev
,
914 struct address_space
*mapping
, void *entry
)
916 unsigned long pfn
, index
, count
, end
;
918 struct vm_area_struct
*vma
;
921 * A page got tagged dirty in DAX mapping? Something is seriously
924 if (WARN_ON(!xa_is_value(entry
)))
927 if (unlikely(dax_is_locked(entry
))) {
928 void *old_entry
= entry
;
930 entry
= get_unlocked_entry(xas
, 0);
932 /* Entry got punched out / reallocated? */
933 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
936 * Entry got reallocated elsewhere? No need to writeback.
937 * We have to compare pfns as we must not bail out due to
938 * difference in lockbit or entry type.
940 if (dax_to_pfn(old_entry
) != dax_to_pfn(entry
))
942 if (WARN_ON_ONCE(dax_is_empty_entry(entry
) ||
943 dax_is_zero_entry(entry
))) {
948 /* Another fsync thread may have already done this entry */
949 if (!xas_get_mark(xas
, PAGECACHE_TAG_TOWRITE
))
953 /* Lock the entry to serialize with page faults */
954 dax_lock_entry(xas
, entry
);
957 * We can clear the tag now but we have to be careful so that concurrent
958 * dax_writeback_one() calls for the same index cannot finish before we
959 * actually flush the caches. This is achieved as the calls will look
960 * at the entry only under the i_pages lock and once they do that
961 * they will see the entry locked and wait for it to unlock.
963 xas_clear_mark(xas
, PAGECACHE_TAG_TOWRITE
);
967 * If dax_writeback_mapping_range() was given a wbc->range_start
968 * in the middle of a PMD, the 'index' we use needs to be
969 * aligned to the start of the PMD.
970 * This allows us to flush for PMD_SIZE and not have to worry about
971 * partial PMD writebacks.
973 pfn
= dax_to_pfn(entry
);
974 count
= 1UL << dax_entry_order(entry
);
975 index
= xas
->xa_index
& ~(count
- 1);
976 end
= index
+ count
- 1;
978 /* Walk all mappings of a given index of a file and writeprotect them */
979 i_mmap_lock_read(mapping
);
980 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, index
, end
) {
981 pfn_mkclean_range(pfn
, count
, index
, vma
);
984 i_mmap_unlock_read(mapping
);
986 dax_flush(dax_dev
, page_address(pfn_to_page(pfn
)), count
* PAGE_SIZE
);
988 * After we have flushed the cache, we can clear the dirty tag. There
989 * cannot be new dirty data in the pfn after the flush has completed as
990 * the pfn mappings are writeprotected and fault waits for mapping
995 xas_store(xas
, entry
);
996 xas_clear_mark(xas
, PAGECACHE_TAG_DIRTY
);
997 dax_wake_entry(xas
, entry
, WAKE_NEXT
);
999 trace_dax_writeback_one(mapping
->host
, index
, count
);
1003 put_unlocked_entry(xas
, entry
, WAKE_NEXT
);
1008 * Flush the mapping to the persistent domain within the byte range of [start,
1009 * end]. This is required by data integrity operations to ensure file data is
1010 * on persistent storage prior to completion of the operation.
1012 int dax_writeback_mapping_range(struct address_space
*mapping
,
1013 struct dax_device
*dax_dev
, struct writeback_control
*wbc
)
1015 XA_STATE(xas
, &mapping
->i_pages
, wbc
->range_start
>> PAGE_SHIFT
);
1016 struct inode
*inode
= mapping
->host
;
1017 pgoff_t end_index
= wbc
->range_end
>> PAGE_SHIFT
;
1020 unsigned int scanned
= 0;
1022 if (WARN_ON_ONCE(inode
->i_blkbits
!= PAGE_SHIFT
))
1025 if (mapping_empty(mapping
) || wbc
->sync_mode
!= WB_SYNC_ALL
)
1028 trace_dax_writeback_range(inode
, xas
.xa_index
, end_index
);
1030 tag_pages_for_writeback(mapping
, xas
.xa_index
, end_index
);
1033 xas_for_each_marked(&xas
, entry
, end_index
, PAGECACHE_TAG_TOWRITE
) {
1034 ret
= dax_writeback_one(&xas
, dax_dev
, mapping
, entry
);
1036 mapping_set_error(mapping
, ret
);
1039 if (++scanned
% XA_CHECK_SCHED
)
1043 xas_unlock_irq(&xas
);
1047 xas_unlock_irq(&xas
);
1048 trace_dax_writeback_range_done(inode
, xas
.xa_index
, end_index
);
1051 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range
);
1053 static int dax_iomap_direct_access(const struct iomap
*iomap
, loff_t pos
,
1054 size_t size
, void **kaddr
, pfn_t
*pfnp
)
1056 pgoff_t pgoff
= dax_iomap_pgoff(iomap
, pos
);
1060 id
= dax_read_lock();
1061 length
= dax_direct_access(iomap
->dax_dev
, pgoff
, PHYS_PFN(size
),
1062 DAX_ACCESS
, kaddr
, pfnp
);
1068 goto out_check_addr
;
1070 if (PFN_PHYS(length
) < size
)
1072 if (pfn_t_to_pfn(*pfnp
) & (PHYS_PFN(size
)-1))
1074 /* For larger pages we need devmap */
1075 if (length
> 1 && !pfn_t_devmap(*pfnp
))
1085 dax_read_unlock(id
);
1090 * dax_iomap_copy_around - Prepare for an unaligned write to a shared/cow page
1091 * by copying the data before and after the range to be written.
1092 * @pos: address to do copy from.
1093 * @length: size of copy operation.
1094 * @align_size: aligned w.r.t align_size (either PMD_SIZE or PAGE_SIZE)
1095 * @srcmap: iomap srcmap
1096 * @daddr: destination address to copy to.
1098 * This can be called from two places. Either during DAX write fault (page
1099 * aligned), to copy the length size data to daddr. Or, while doing normal DAX
1100 * write operation, dax_iomap_iter() might call this to do the copy of either
1101 * start or end unaligned address. In the latter case the rest of the copy of
1102 * aligned ranges is taken care by dax_iomap_iter() itself.
1103 * If the srcmap contains invalid data, such as HOLE and UNWRITTEN, zero the
1104 * area to make sure no old data remains.
1106 static int dax_iomap_copy_around(loff_t pos
, uint64_t length
, size_t align_size
,
1107 const struct iomap
*srcmap
, void *daddr
)
1109 loff_t head_off
= pos
& (align_size
- 1);
1110 size_t size
= ALIGN(head_off
+ length
, align_size
);
1111 loff_t end
= pos
+ length
;
1112 loff_t pg_end
= round_up(end
, align_size
);
1113 /* copy_all is usually in page fault case */
1114 bool copy_all
= head_off
== 0 && end
== pg_end
;
1115 /* zero the edges if srcmap is a HOLE or IOMAP_UNWRITTEN */
1116 bool zero_edge
= srcmap
->flags
& IOMAP_F_SHARED
||
1117 srcmap
->type
== IOMAP_UNWRITTEN
;
1122 ret
= dax_iomap_direct_access(srcmap
, pos
, size
, &saddr
, NULL
);
1129 memset(daddr
, 0, size
);
1131 ret
= copy_mc_to_kernel(daddr
, saddr
, length
);
1135 /* Copy the head part of the range */
1138 memset(daddr
, 0, head_off
);
1140 ret
= copy_mc_to_kernel(daddr
, saddr
, head_off
);
1146 /* Copy the tail part of the range */
1148 loff_t tail_off
= head_off
+ length
;
1149 loff_t tail_len
= pg_end
- end
;
1152 memset(daddr
+ tail_off
, 0, tail_len
);
1154 ret
= copy_mc_to_kernel(daddr
+ tail_off
,
1155 saddr
+ tail_off
, tail_len
);
1162 dax_flush(srcmap
->dax_dev
, daddr
, size
);
1163 return ret
? -EIO
: 0;
1167 * The user has performed a load from a hole in the file. Allocating a new
1168 * page in the file would cause excessive storage usage for workloads with
1169 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1170 * If this page is ever written to we will re-fault and change the mapping to
1171 * point to real DAX storage instead.
1173 static vm_fault_t
dax_load_hole(struct xa_state
*xas
, struct vm_fault
*vmf
,
1174 const struct iomap_iter
*iter
, void **entry
)
1176 struct inode
*inode
= iter
->inode
;
1177 unsigned long vaddr
= vmf
->address
;
1178 pfn_t pfn
= pfn_to_pfn_t(my_zero_pfn(vaddr
));
1181 *entry
= dax_insert_entry(xas
, vmf
, iter
, *entry
, pfn
, DAX_ZERO_PAGE
);
1183 ret
= vmf_insert_mixed(vmf
->vma
, vaddr
, pfn
);
1184 trace_dax_load_hole(inode
, vmf
, ret
);
1188 #ifdef CONFIG_FS_DAX_PMD
1189 static vm_fault_t
dax_pmd_load_hole(struct xa_state
*xas
, struct vm_fault
*vmf
,
1190 const struct iomap_iter
*iter
, void **entry
)
1192 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1193 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
1194 struct vm_area_struct
*vma
= vmf
->vma
;
1195 struct inode
*inode
= mapping
->host
;
1196 pgtable_t pgtable
= NULL
;
1197 struct page
*zero_page
;
1202 zero_page
= mm_get_huge_zero_page(vmf
->vma
->vm_mm
);
1204 if (unlikely(!zero_page
))
1207 pfn
= page_to_pfn_t(zero_page
);
1208 *entry
= dax_insert_entry(xas
, vmf
, iter
, *entry
, pfn
,
1209 DAX_PMD
| DAX_ZERO_PAGE
);
1211 if (arch_needs_pgtable_deposit()) {
1212 pgtable
= pte_alloc_one(vma
->vm_mm
);
1214 return VM_FAULT_OOM
;
1217 ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1218 if (!pmd_none(*(vmf
->pmd
))) {
1224 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, pgtable
);
1225 mm_inc_nr_ptes(vma
->vm_mm
);
1227 pmd_entry
= mk_pmd(zero_page
, vmf
->vma
->vm_page_prot
);
1228 pmd_entry
= pmd_mkhuge(pmd_entry
);
1229 set_pmd_at(vmf
->vma
->vm_mm
, pmd_addr
, vmf
->pmd
, pmd_entry
);
1231 trace_dax_pmd_load_hole(inode
, vmf
, zero_page
, *entry
);
1232 return VM_FAULT_NOPAGE
;
1236 pte_free(vma
->vm_mm
, pgtable
);
1237 trace_dax_pmd_load_hole_fallback(inode
, vmf
, zero_page
, *entry
);
1238 return VM_FAULT_FALLBACK
;
1241 static vm_fault_t
dax_pmd_load_hole(struct xa_state
*xas
, struct vm_fault
*vmf
,
1242 const struct iomap_iter
*iter
, void **entry
)
1244 return VM_FAULT_FALLBACK
;
1246 #endif /* CONFIG_FS_DAX_PMD */
1248 static s64
dax_unshare_iter(struct iomap_iter
*iter
)
1250 struct iomap
*iomap
= &iter
->iomap
;
1251 const struct iomap
*srcmap
= iomap_iter_srcmap(iter
);
1252 loff_t pos
= iter
->pos
;
1253 loff_t length
= iomap_length(iter
);
1256 void *daddr
= NULL
, *saddr
= NULL
;
1258 /* don't bother with blocks that are not shared to start with */
1259 if (!(iomap
->flags
& IOMAP_F_SHARED
))
1261 /* don't bother with holes or unwritten extents */
1262 if (srcmap
->type
== IOMAP_HOLE
|| srcmap
->type
== IOMAP_UNWRITTEN
)
1265 id
= dax_read_lock();
1266 ret
= dax_iomap_direct_access(iomap
, pos
, length
, &daddr
, NULL
);
1270 ret
= dax_iomap_direct_access(srcmap
, pos
, length
, &saddr
, NULL
);
1274 if (copy_mc_to_kernel(daddr
, saddr
, length
) == 0)
1280 dax_read_unlock(id
);
1284 int dax_file_unshare(struct inode
*inode
, loff_t pos
, loff_t len
,
1285 const struct iomap_ops
*ops
)
1287 struct iomap_iter iter
= {
1291 .flags
= IOMAP_WRITE
| IOMAP_UNSHARE
| IOMAP_DAX
,
1295 while ((ret
= iomap_iter(&iter
, ops
)) > 0)
1296 iter
.processed
= dax_unshare_iter(&iter
);
1299 EXPORT_SYMBOL_GPL(dax_file_unshare
);
1301 static int dax_memzero(struct iomap_iter
*iter
, loff_t pos
, size_t size
)
1303 const struct iomap
*iomap
= &iter
->iomap
;
1304 const struct iomap
*srcmap
= iomap_iter_srcmap(iter
);
1305 unsigned offset
= offset_in_page(pos
);
1306 pgoff_t pgoff
= dax_iomap_pgoff(iomap
, pos
);
1310 ret
= dax_direct_access(iomap
->dax_dev
, pgoff
, 1, DAX_ACCESS
, &kaddr
,
1314 memset(kaddr
+ offset
, 0, size
);
1315 if (iomap
->flags
& IOMAP_F_SHARED
)
1316 ret
= dax_iomap_copy_around(pos
, size
, PAGE_SIZE
, srcmap
,
1319 dax_flush(iomap
->dax_dev
, kaddr
+ offset
, size
);
1323 static s64
dax_zero_iter(struct iomap_iter
*iter
, bool *did_zero
)
1325 const struct iomap
*iomap
= &iter
->iomap
;
1326 const struct iomap
*srcmap
= iomap_iter_srcmap(iter
);
1327 loff_t pos
= iter
->pos
;
1328 u64 length
= iomap_length(iter
);
1331 /* already zeroed? we're done. */
1332 if (srcmap
->type
== IOMAP_HOLE
|| srcmap
->type
== IOMAP_UNWRITTEN
)
1336 * invalidate the pages whose sharing state is to be changed
1339 if (iomap
->flags
& IOMAP_F_SHARED
)
1340 invalidate_inode_pages2_range(iter
->inode
->i_mapping
,
1342 (pos
+ length
- 1) >> PAGE_SHIFT
);
1345 unsigned offset
= offset_in_page(pos
);
1346 unsigned size
= min_t(u64
, PAGE_SIZE
- offset
, length
);
1347 pgoff_t pgoff
= dax_iomap_pgoff(iomap
, pos
);
1351 id
= dax_read_lock();
1352 if (IS_ALIGNED(pos
, PAGE_SIZE
) && size
== PAGE_SIZE
)
1353 rc
= dax_zero_page_range(iomap
->dax_dev
, pgoff
, 1);
1355 rc
= dax_memzero(iter
, pos
, size
);
1356 dax_read_unlock(id
);
1363 } while (length
> 0);
1370 int dax_zero_range(struct inode
*inode
, loff_t pos
, loff_t len
, bool *did_zero
,
1371 const struct iomap_ops
*ops
)
1373 struct iomap_iter iter
= {
1377 .flags
= IOMAP_DAX
| IOMAP_ZERO
,
1381 while ((ret
= iomap_iter(&iter
, ops
)) > 0)
1382 iter
.processed
= dax_zero_iter(&iter
, did_zero
);
1385 EXPORT_SYMBOL_GPL(dax_zero_range
);
1387 int dax_truncate_page(struct inode
*inode
, loff_t pos
, bool *did_zero
,
1388 const struct iomap_ops
*ops
)
1390 unsigned int blocksize
= i_blocksize(inode
);
1391 unsigned int off
= pos
& (blocksize
- 1);
1393 /* Block boundary? Nothing to do */
1396 return dax_zero_range(inode
, pos
, blocksize
- off
, did_zero
, ops
);
1398 EXPORT_SYMBOL_GPL(dax_truncate_page
);
1400 static loff_t
dax_iomap_iter(const struct iomap_iter
*iomi
,
1401 struct iov_iter
*iter
)
1403 const struct iomap
*iomap
= &iomi
->iomap
;
1404 const struct iomap
*srcmap
= iomap_iter_srcmap(iomi
);
1405 loff_t length
= iomap_length(iomi
);
1406 loff_t pos
= iomi
->pos
;
1407 struct dax_device
*dax_dev
= iomap
->dax_dev
;
1408 loff_t end
= pos
+ length
, done
= 0;
1409 bool write
= iov_iter_rw(iter
) == WRITE
;
1410 bool cow
= write
&& iomap
->flags
& IOMAP_F_SHARED
;
1416 end
= min(end
, i_size_read(iomi
->inode
));
1420 if (iomap
->type
== IOMAP_HOLE
|| iomap
->type
== IOMAP_UNWRITTEN
)
1421 return iov_iter_zero(min(length
, end
- pos
), iter
);
1425 * In DAX mode, enforce either pure overwrites of written extents, or
1426 * writes to unwritten extents as part of a copy-on-write operation.
1428 if (WARN_ON_ONCE(iomap
->type
!= IOMAP_MAPPED
&&
1429 !(iomap
->flags
& IOMAP_F_SHARED
)))
1433 * Write can allocate block for an area which has a hole page mapped
1434 * into page tables. We have to tear down these mappings so that data
1435 * written by write(2) is visible in mmap.
1437 if (iomap
->flags
& IOMAP_F_NEW
|| cow
) {
1438 invalidate_inode_pages2_range(iomi
->inode
->i_mapping
,
1440 (end
- 1) >> PAGE_SHIFT
);
1443 id
= dax_read_lock();
1445 unsigned offset
= pos
& (PAGE_SIZE
- 1);
1446 const size_t size
= ALIGN(length
+ offset
, PAGE_SIZE
);
1447 pgoff_t pgoff
= dax_iomap_pgoff(iomap
, pos
);
1449 bool recovery
= false;
1452 if (fatal_signal_pending(current
)) {
1457 map_len
= dax_direct_access(dax_dev
, pgoff
, PHYS_PFN(size
),
1458 DAX_ACCESS
, &kaddr
, NULL
);
1459 if (map_len
== -EIO
&& iov_iter_rw(iter
) == WRITE
) {
1460 map_len
= dax_direct_access(dax_dev
, pgoff
,
1461 PHYS_PFN(size
), DAX_RECOVERY_WRITE
,
1472 ret
= dax_iomap_copy_around(pos
, length
, PAGE_SIZE
,
1478 map_len
= PFN_PHYS(map_len
);
1481 if (map_len
> end
- pos
)
1482 map_len
= end
- pos
;
1485 xfer
= dax_recovery_write(dax_dev
, pgoff
, kaddr
,
1488 xfer
= dax_copy_from_iter(dax_dev
, pgoff
, kaddr
,
1491 xfer
= dax_copy_to_iter(dax_dev
, pgoff
, kaddr
,
1503 dax_read_unlock(id
);
1505 return done
? done
: ret
;
1509 * dax_iomap_rw - Perform I/O to a DAX file
1510 * @iocb: The control block for this I/O
1511 * @iter: The addresses to do I/O from or to
1512 * @ops: iomap ops passed from the file system
1514 * This function performs read and write operations to directly mapped
1515 * persistent memory. The callers needs to take care of read/write exclusion
1516 * and evicting any page cache pages in the region under I/O.
1519 dax_iomap_rw(struct kiocb
*iocb
, struct iov_iter
*iter
,
1520 const struct iomap_ops
*ops
)
1522 struct iomap_iter iomi
= {
1523 .inode
= iocb
->ki_filp
->f_mapping
->host
,
1524 .pos
= iocb
->ki_pos
,
1525 .len
= iov_iter_count(iter
),
1534 if (iov_iter_rw(iter
) == WRITE
) {
1535 lockdep_assert_held_write(&iomi
.inode
->i_rwsem
);
1536 iomi
.flags
|= IOMAP_WRITE
;
1538 lockdep_assert_held(&iomi
.inode
->i_rwsem
);
1541 if (iocb
->ki_flags
& IOCB_NOWAIT
)
1542 iomi
.flags
|= IOMAP_NOWAIT
;
1544 while ((ret
= iomap_iter(&iomi
, ops
)) > 0)
1545 iomi
.processed
= dax_iomap_iter(&iomi
, iter
);
1547 done
= iomi
.pos
- iocb
->ki_pos
;
1548 iocb
->ki_pos
= iomi
.pos
;
1549 return done
? done
: ret
;
1551 EXPORT_SYMBOL_GPL(dax_iomap_rw
);
1553 static vm_fault_t
dax_fault_return(int error
)
1556 return VM_FAULT_NOPAGE
;
1557 return vmf_error(error
);
1561 * When handling a synchronous page fault and the inode need a fsync, we can
1562 * insert the PTE/PMD into page tables only after that fsync happened. Skip
1563 * insertion for now and return the pfn so that caller can insert it after the
1566 static vm_fault_t
dax_fault_synchronous_pfnp(pfn_t
*pfnp
, pfn_t pfn
)
1568 if (WARN_ON_ONCE(!pfnp
))
1569 return VM_FAULT_SIGBUS
;
1571 return VM_FAULT_NEEDDSYNC
;
1574 static vm_fault_t
dax_fault_cow_page(struct vm_fault
*vmf
,
1575 const struct iomap_iter
*iter
)
1580 switch (iter
->iomap
.type
) {
1582 case IOMAP_UNWRITTEN
:
1583 clear_user_highpage(vmf
->cow_page
, vmf
->address
);
1586 error
= copy_cow_page_dax(vmf
, iter
);
1595 return dax_fault_return(error
);
1597 __SetPageUptodate(vmf
->cow_page
);
1598 ret
= finish_fault(vmf
);
1600 return VM_FAULT_DONE_COW
;
1605 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault.
1606 * @vmf: vm fault instance
1608 * @pfnp: pfn to be returned
1609 * @xas: the dax mapping tree of a file
1610 * @entry: an unlocked dax entry to be inserted
1611 * @pmd: distinguish whether it is a pmd fault
1613 static vm_fault_t
dax_fault_iter(struct vm_fault
*vmf
,
1614 const struct iomap_iter
*iter
, pfn_t
*pfnp
,
1615 struct xa_state
*xas
, void **entry
, bool pmd
)
1617 const struct iomap
*iomap
= &iter
->iomap
;
1618 const struct iomap
*srcmap
= iomap_iter_srcmap(iter
);
1619 size_t size
= pmd
? PMD_SIZE
: PAGE_SIZE
;
1620 loff_t pos
= (loff_t
)xas
->xa_index
<< PAGE_SHIFT
;
1621 bool write
= iter
->flags
& IOMAP_WRITE
;
1622 unsigned long entry_flags
= pmd
? DAX_PMD
: 0;
1627 if (!pmd
&& vmf
->cow_page
)
1628 return dax_fault_cow_page(vmf
, iter
);
1630 /* if we are reading UNWRITTEN and HOLE, return a hole. */
1632 (iomap
->type
== IOMAP_UNWRITTEN
|| iomap
->type
== IOMAP_HOLE
)) {
1634 return dax_load_hole(xas
, vmf
, iter
, entry
);
1635 return dax_pmd_load_hole(xas
, vmf
, iter
, entry
);
1638 if (iomap
->type
!= IOMAP_MAPPED
&& !(iomap
->flags
& IOMAP_F_SHARED
)) {
1640 return pmd
? VM_FAULT_FALLBACK
: VM_FAULT_SIGBUS
;
1643 err
= dax_iomap_direct_access(iomap
, pos
, size
, &kaddr
, &pfn
);
1645 return pmd
? VM_FAULT_FALLBACK
: dax_fault_return(err
);
1647 *entry
= dax_insert_entry(xas
, vmf
, iter
, *entry
, pfn
, entry_flags
);
1649 if (write
&& iomap
->flags
& IOMAP_F_SHARED
) {
1650 err
= dax_iomap_copy_around(pos
, size
, size
, srcmap
, kaddr
);
1652 return dax_fault_return(err
);
1655 if (dax_fault_is_synchronous(iter
, vmf
->vma
))
1656 return dax_fault_synchronous_pfnp(pfnp
, pfn
);
1658 /* insert PMD pfn */
1660 return vmf_insert_pfn_pmd(vmf
, pfn
, write
);
1662 /* insert PTE pfn */
1664 return vmf_insert_mixed_mkwrite(vmf
->vma
, vmf
->address
, pfn
);
1665 return vmf_insert_mixed(vmf
->vma
, vmf
->address
, pfn
);
1668 static vm_fault_t
dax_iomap_pte_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1669 int *iomap_errp
, const struct iomap_ops
*ops
)
1671 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1672 XA_STATE(xas
, &mapping
->i_pages
, vmf
->pgoff
);
1673 struct iomap_iter iter
= {
1674 .inode
= mapping
->host
,
1675 .pos
= (loff_t
)vmf
->pgoff
<< PAGE_SHIFT
,
1677 .flags
= IOMAP_DAX
| IOMAP_FAULT
,
1683 trace_dax_pte_fault(iter
.inode
, vmf
, ret
);
1685 * Check whether offset isn't beyond end of file now. Caller is supposed
1686 * to hold locks serializing us with truncate / punch hole so this is
1689 if (iter
.pos
>= i_size_read(iter
.inode
)) {
1690 ret
= VM_FAULT_SIGBUS
;
1694 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && !vmf
->cow_page
)
1695 iter
.flags
|= IOMAP_WRITE
;
1697 entry
= grab_mapping_entry(&xas
, mapping
, 0);
1698 if (xa_is_internal(entry
)) {
1699 ret
= xa_to_internal(entry
);
1704 * It is possible, particularly with mixed reads & writes to private
1705 * mappings, that we have raced with a PMD fault that overlaps with
1706 * the PTE we need to set up. If so just return and the fault will be
1709 if (pmd_trans_huge(*vmf
->pmd
) || pmd_devmap(*vmf
->pmd
)) {
1710 ret
= VM_FAULT_NOPAGE
;
1714 while ((error
= iomap_iter(&iter
, ops
)) > 0) {
1715 if (WARN_ON_ONCE(iomap_length(&iter
) < PAGE_SIZE
)) {
1716 iter
.processed
= -EIO
; /* fs corruption? */
1720 ret
= dax_fault_iter(vmf
, &iter
, pfnp
, &xas
, &entry
, false);
1721 if (ret
!= VM_FAULT_SIGBUS
&&
1722 (iter
.iomap
.flags
& IOMAP_F_NEW
)) {
1723 count_vm_event(PGMAJFAULT
);
1724 count_memcg_event_mm(vmf
->vma
->vm_mm
, PGMAJFAULT
);
1725 ret
|= VM_FAULT_MAJOR
;
1728 if (!(ret
& VM_FAULT_ERROR
))
1729 iter
.processed
= PAGE_SIZE
;
1733 *iomap_errp
= error
;
1735 ret
= dax_fault_return(error
);
1738 dax_unlock_entry(&xas
, entry
);
1740 trace_dax_pte_fault_done(iter
.inode
, vmf
, ret
);
1744 #ifdef CONFIG_FS_DAX_PMD
1745 static bool dax_fault_check_fallback(struct vm_fault
*vmf
, struct xa_state
*xas
,
1748 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
1749 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1752 * Make sure that the faulting address's PMD offset (color) matches
1753 * the PMD offset from the start of the file. This is necessary so
1754 * that a PMD range in the page table overlaps exactly with a PMD
1755 * range in the page cache.
1757 if ((vmf
->pgoff
& PG_PMD_COLOUR
) !=
1758 ((vmf
->address
>> PAGE_SHIFT
) & PG_PMD_COLOUR
))
1761 /* Fall back to PTEs if we're going to COW */
1762 if (write
&& !(vmf
->vma
->vm_flags
& VM_SHARED
))
1765 /* If the PMD would extend outside the VMA */
1766 if (pmd_addr
< vmf
->vma
->vm_start
)
1768 if ((pmd_addr
+ PMD_SIZE
) > vmf
->vma
->vm_end
)
1771 /* If the PMD would extend beyond the file size */
1772 if ((xas
->xa_index
| PG_PMD_COLOUR
) >= max_pgoff
)
1778 static vm_fault_t
dax_iomap_pmd_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1779 const struct iomap_ops
*ops
)
1781 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1782 XA_STATE_ORDER(xas
, &mapping
->i_pages
, vmf
->pgoff
, PMD_ORDER
);
1783 struct iomap_iter iter
= {
1784 .inode
= mapping
->host
,
1786 .flags
= IOMAP_DAX
| IOMAP_FAULT
,
1788 vm_fault_t ret
= VM_FAULT_FALLBACK
;
1793 if (vmf
->flags
& FAULT_FLAG_WRITE
)
1794 iter
.flags
|= IOMAP_WRITE
;
1797 * Check whether offset isn't beyond end of file now. Caller is
1798 * supposed to hold locks serializing us with truncate / punch hole so
1799 * this is a reliable test.
1801 max_pgoff
= DIV_ROUND_UP(i_size_read(iter
.inode
), PAGE_SIZE
);
1803 trace_dax_pmd_fault(iter
.inode
, vmf
, max_pgoff
, 0);
1805 if (xas
.xa_index
>= max_pgoff
) {
1806 ret
= VM_FAULT_SIGBUS
;
1810 if (dax_fault_check_fallback(vmf
, &xas
, max_pgoff
))
1814 * grab_mapping_entry() will make sure we get an empty PMD entry,
1815 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1816 * entry is already in the array, for instance), it will return
1817 * VM_FAULT_FALLBACK.
1819 entry
= grab_mapping_entry(&xas
, mapping
, PMD_ORDER
);
1820 if (xa_is_internal(entry
)) {
1821 ret
= xa_to_internal(entry
);
1826 * It is possible, particularly with mixed reads & writes to private
1827 * mappings, that we have raced with a PTE fault that overlaps with
1828 * the PMD we need to set up. If so just return and the fault will be
1831 if (!pmd_none(*vmf
->pmd
) && !pmd_trans_huge(*vmf
->pmd
) &&
1832 !pmd_devmap(*vmf
->pmd
)) {
1837 iter
.pos
= (loff_t
)xas
.xa_index
<< PAGE_SHIFT
;
1838 while ((error
= iomap_iter(&iter
, ops
)) > 0) {
1839 if (iomap_length(&iter
) < PMD_SIZE
)
1840 continue; /* actually breaks out of the loop */
1842 ret
= dax_fault_iter(vmf
, &iter
, pfnp
, &xas
, &entry
, true);
1843 if (ret
!= VM_FAULT_FALLBACK
)
1844 iter
.processed
= PMD_SIZE
;
1848 dax_unlock_entry(&xas
, entry
);
1850 if (ret
== VM_FAULT_FALLBACK
) {
1851 split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
);
1852 count_vm_event(THP_FAULT_FALLBACK
);
1855 trace_dax_pmd_fault_done(iter
.inode
, vmf
, max_pgoff
, ret
);
1859 static vm_fault_t
dax_iomap_pmd_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1860 const struct iomap_ops
*ops
)
1862 return VM_FAULT_FALLBACK
;
1864 #endif /* CONFIG_FS_DAX_PMD */
1867 * dax_iomap_fault - handle a page fault on a DAX file
1868 * @vmf: The description of the fault
1869 * @pe_size: Size of the page to fault in
1870 * @pfnp: PFN to insert for synchronous faults if fsync is required
1871 * @iomap_errp: Storage for detailed error code in case of error
1872 * @ops: Iomap ops passed from the file system
1874 * When a page fault occurs, filesystems may call this helper in
1875 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1876 * has done all the necessary locking for page fault to proceed
1879 vm_fault_t
dax_iomap_fault(struct vm_fault
*vmf
, enum page_entry_size pe_size
,
1880 pfn_t
*pfnp
, int *iomap_errp
, const struct iomap_ops
*ops
)
1884 return dax_iomap_pte_fault(vmf
, pfnp
, iomap_errp
, ops
);
1886 return dax_iomap_pmd_fault(vmf
, pfnp
, ops
);
1888 return VM_FAULT_FALLBACK
;
1891 EXPORT_SYMBOL_GPL(dax_iomap_fault
);
1894 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1895 * @vmf: The description of the fault
1896 * @pfn: PFN to insert
1897 * @order: Order of entry to insert.
1899 * This function inserts a writeable PTE or PMD entry into the page tables
1900 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1903 dax_insert_pfn_mkwrite(struct vm_fault
*vmf
, pfn_t pfn
, unsigned int order
)
1905 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1906 XA_STATE_ORDER(xas
, &mapping
->i_pages
, vmf
->pgoff
, order
);
1911 entry
= get_unlocked_entry(&xas
, order
);
1912 /* Did we race with someone splitting entry or so? */
1913 if (!entry
|| dax_is_conflict(entry
) ||
1914 (order
== 0 && !dax_is_pte_entry(entry
))) {
1915 put_unlocked_entry(&xas
, entry
, WAKE_NEXT
);
1916 xas_unlock_irq(&xas
);
1917 trace_dax_insert_pfn_mkwrite_no_entry(mapping
->host
, vmf
,
1919 return VM_FAULT_NOPAGE
;
1921 xas_set_mark(&xas
, PAGECACHE_TAG_DIRTY
);
1922 dax_lock_entry(&xas
, entry
);
1923 xas_unlock_irq(&xas
);
1925 ret
= vmf_insert_mixed_mkwrite(vmf
->vma
, vmf
->address
, pfn
);
1926 #ifdef CONFIG_FS_DAX_PMD
1927 else if (order
== PMD_ORDER
)
1928 ret
= vmf_insert_pfn_pmd(vmf
, pfn
, FAULT_FLAG_WRITE
);
1931 ret
= VM_FAULT_FALLBACK
;
1932 dax_unlock_entry(&xas
, entry
);
1933 trace_dax_insert_pfn_mkwrite(mapping
->host
, vmf
, ret
);
1938 * dax_finish_sync_fault - finish synchronous page fault
1939 * @vmf: The description of the fault
1940 * @pe_size: Size of entry to be inserted
1941 * @pfn: PFN to insert
1943 * This function ensures that the file range touched by the page fault is
1944 * stored persistently on the media and handles inserting of appropriate page
1947 vm_fault_t
dax_finish_sync_fault(struct vm_fault
*vmf
,
1948 enum page_entry_size pe_size
, pfn_t pfn
)
1951 loff_t start
= ((loff_t
)vmf
->pgoff
) << PAGE_SHIFT
;
1952 unsigned int order
= pe_order(pe_size
);
1953 size_t len
= PAGE_SIZE
<< order
;
1955 err
= vfs_fsync_range(vmf
->vma
->vm_file
, start
, start
+ len
- 1, 1);
1957 return VM_FAULT_SIGBUS
;
1958 return dax_insert_pfn_mkwrite(vmf
, pfn
, order
);
1960 EXPORT_SYMBOL_GPL(dax_finish_sync_fault
);
1962 static loff_t
dax_range_compare_iter(struct iomap_iter
*it_src
,
1963 struct iomap_iter
*it_dest
, u64 len
, bool *same
)
1965 const struct iomap
*smap
= &it_src
->iomap
;
1966 const struct iomap
*dmap
= &it_dest
->iomap
;
1967 loff_t pos1
= it_src
->pos
, pos2
= it_dest
->pos
;
1968 void *saddr
, *daddr
;
1971 len
= min(len
, min(smap
->length
, dmap
->length
));
1973 if (smap
->type
== IOMAP_HOLE
&& dmap
->type
== IOMAP_HOLE
) {
1978 if (smap
->type
== IOMAP_HOLE
|| dmap
->type
== IOMAP_HOLE
) {
1983 id
= dax_read_lock();
1984 ret
= dax_iomap_direct_access(smap
, pos1
, ALIGN(pos1
+ len
, PAGE_SIZE
),
1989 ret
= dax_iomap_direct_access(dmap
, pos2
, ALIGN(pos2
+ len
, PAGE_SIZE
),
1994 *same
= !memcmp(saddr
, daddr
, len
);
1997 dax_read_unlock(id
);
2001 dax_read_unlock(id
);
2005 int dax_dedupe_file_range_compare(struct inode
*src
, loff_t srcoff
,
2006 struct inode
*dst
, loff_t dstoff
, loff_t len
, bool *same
,
2007 const struct iomap_ops
*ops
)
2009 struct iomap_iter src_iter
= {
2015 struct iomap_iter dst_iter
= {
2021 int ret
, compared
= 0;
2023 while ((ret
= iomap_iter(&src_iter
, ops
)) > 0 &&
2024 (ret
= iomap_iter(&dst_iter
, ops
)) > 0) {
2025 compared
= dax_range_compare_iter(&src_iter
, &dst_iter
, len
,
2029 src_iter
.processed
= dst_iter
.processed
= compared
;
2034 int dax_remap_file_range_prep(struct file
*file_in
, loff_t pos_in
,
2035 struct file
*file_out
, loff_t pos_out
,
2036 loff_t
*len
, unsigned int remap_flags
,
2037 const struct iomap_ops
*ops
)
2039 return __generic_remap_file_range_prep(file_in
, pos_in
, file_out
,
2040 pos_out
, len
, remap_flags
, ops
);
2042 EXPORT_SYMBOL_GPL(dax_remap_file_range_prep
);