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_mapping_is_cow(struct address_space
*mapping
)
339 return (unsigned long)mapping
== PAGE_MAPPING_DAX_COW
;
343 * Set the page->mapping with FS_DAX_MAPPING_COW flag, increase the refcount.
345 static inline void dax_mapping_set_cow(struct page
*page
)
347 if ((uintptr_t)page
->mapping
!= PAGE_MAPPING_DAX_COW
) {
349 * Reset the index if the page was already mapped
354 page
->mapping
= (void *)PAGE_MAPPING_DAX_COW
;
360 * When it is called in dax_insert_entry(), the cow flag will indicate that
361 * whether this entry is shared by multiple files. If so, set the page->mapping
362 * FS_DAX_MAPPING_COW, and use page->index as refcount.
364 static void dax_associate_entry(void *entry
, struct address_space
*mapping
,
365 struct vm_area_struct
*vma
, unsigned long address
, bool cow
)
367 unsigned long size
= dax_entry_size(entry
), pfn
, index
;
370 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
373 index
= linear_page_index(vma
, address
& ~(size
- 1));
374 for_each_mapped_pfn(entry
, pfn
) {
375 struct page
*page
= pfn_to_page(pfn
);
378 dax_mapping_set_cow(page
);
380 WARN_ON_ONCE(page
->mapping
);
381 page
->mapping
= mapping
;
382 page
->index
= index
+ i
++;
387 static void dax_disassociate_entry(void *entry
, struct address_space
*mapping
,
392 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
395 for_each_mapped_pfn(entry
, pfn
) {
396 struct page
*page
= pfn_to_page(pfn
);
398 WARN_ON_ONCE(trunc
&& page_ref_count(page
) > 1);
399 if (dax_mapping_is_cow(page
->mapping
)) {
400 /* keep the CoW flag if this page is still shared */
401 if (page
->index
-- > 0)
404 WARN_ON_ONCE(page
->mapping
&& page
->mapping
!= mapping
);
405 page
->mapping
= NULL
;
410 static struct page
*dax_busy_page(void *entry
)
414 for_each_mapped_pfn(entry
, pfn
) {
415 struct page
*page
= pfn_to_page(pfn
);
417 if (page_ref_count(page
) > 1)
424 * dax_lock_page - Lock the DAX entry corresponding to a page
425 * @page: The page whose entry we want to lock
427 * Context: Process context.
428 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
431 dax_entry_t
dax_lock_page(struct page
*page
)
433 XA_STATE(xas
, NULL
, 0);
436 /* Ensure page->mapping isn't freed while we look at it */
439 struct address_space
*mapping
= READ_ONCE(page
->mapping
);
442 if (!mapping
|| !dax_mapping(mapping
))
446 * In the device-dax case there's no need to lock, a
447 * struct dev_pagemap pin is sufficient to keep the
448 * inode alive, and we assume we have dev_pagemap pin
449 * otherwise we would not have a valid pfn_to_page()
452 entry
= (void *)~0UL;
453 if (S_ISCHR(mapping
->host
->i_mode
))
456 xas
.xa
= &mapping
->i_pages
;
458 if (mapping
!= page
->mapping
) {
459 xas_unlock_irq(&xas
);
462 xas_set(&xas
, page
->index
);
463 entry
= xas_load(&xas
);
464 if (dax_is_locked(entry
)) {
466 wait_entry_unlocked(&xas
, entry
);
470 dax_lock_entry(&xas
, entry
);
471 xas_unlock_irq(&xas
);
475 return (dax_entry_t
)entry
;
478 void dax_unlock_page(struct page
*page
, dax_entry_t cookie
)
480 struct address_space
*mapping
= page
->mapping
;
481 XA_STATE(xas
, &mapping
->i_pages
, page
->index
);
483 if (S_ISCHR(mapping
->host
->i_mode
))
486 dax_unlock_entry(&xas
, (void *)cookie
);
490 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping
491 * @mapping: the file's mapping whose entry we want to lock
492 * @index: the offset within this file
493 * @page: output the dax page corresponding to this dax entry
495 * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry
496 * could not be locked.
498 dax_entry_t
dax_lock_mapping_entry(struct address_space
*mapping
, pgoff_t index
,
501 XA_STATE(xas
, NULL
, 0);
507 if (!dax_mapping(mapping
))
510 xas
.xa
= &mapping
->i_pages
;
512 xas_set(&xas
, index
);
513 entry
= xas_load(&xas
);
514 if (dax_is_locked(entry
)) {
516 wait_entry_unlocked(&xas
, entry
);
521 dax_is_zero_entry(entry
) || dax_is_empty_entry(entry
)) {
523 * Because we are looking for entry from file's mapping
524 * and index, so the entry may not be inserted for now,
525 * or even a zero/empty entry. We don't think this is
526 * an error case. So, return a special value and do
529 entry
= (void *)~0UL;
531 *page
= pfn_to_page(dax_to_pfn(entry
));
532 dax_lock_entry(&xas
, entry
);
534 xas_unlock_irq(&xas
);
538 return (dax_entry_t
)entry
;
541 void dax_unlock_mapping_entry(struct address_space
*mapping
, pgoff_t index
,
544 XA_STATE(xas
, &mapping
->i_pages
, index
);
549 dax_unlock_entry(&xas
, (void *)cookie
);
553 * Find page cache entry at given index. If it is a DAX entry, return it
554 * with the entry locked. If the page cache doesn't contain an entry at
555 * that index, add a locked empty entry.
557 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
558 * either return that locked entry or will return VM_FAULT_FALLBACK.
559 * This will happen if there are any PTE entries within the PMD range
560 * that we are requesting.
562 * We always favor PTE entries over PMD entries. There isn't a flow where we
563 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
564 * insertion will fail if it finds any PTE entries already in the tree, and a
565 * PTE insertion will cause an existing PMD entry to be unmapped and
566 * downgraded to PTE entries. This happens for both PMD zero pages as
567 * well as PMD empty entries.
569 * The exception to this downgrade path is for PMD entries that have
570 * real storage backing them. We will leave these real PMD entries in
571 * the tree, and PTE writes will simply dirty the entire PMD entry.
573 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
574 * persistent memory the benefit is doubtful. We can add that later if we can
577 * On error, this function does not return an ERR_PTR. Instead it returns
578 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
579 * overlap with xarray value entries.
581 static void *grab_mapping_entry(struct xa_state
*xas
,
582 struct address_space
*mapping
, unsigned int order
)
584 unsigned long index
= xas
->xa_index
;
585 bool pmd_downgrade
; /* splitting PMD entry into PTE entries? */
589 pmd_downgrade
= false;
591 entry
= get_unlocked_entry(xas
, order
);
594 if (dax_is_conflict(entry
))
596 if (!xa_is_value(entry
)) {
597 xas_set_err(xas
, -EIO
);
602 if (dax_is_pmd_entry(entry
) &&
603 (dax_is_zero_entry(entry
) ||
604 dax_is_empty_entry(entry
))) {
605 pmd_downgrade
= true;
612 * Make sure 'entry' remains valid while we drop
615 dax_lock_entry(xas
, entry
);
618 * Besides huge zero pages the only other thing that gets
619 * downgraded are empty entries which don't need to be
622 if (dax_is_zero_entry(entry
)) {
624 unmap_mapping_pages(mapping
,
625 xas
->xa_index
& ~PG_PMD_COLOUR
,
631 dax_disassociate_entry(entry
, mapping
, false);
632 xas_store(xas
, NULL
); /* undo the PMD join */
633 dax_wake_entry(xas
, entry
, WAKE_ALL
);
634 mapping
->nrpages
-= PG_PMD_NR
;
640 dax_lock_entry(xas
, entry
);
642 unsigned long flags
= DAX_EMPTY
;
646 entry
= dax_make_entry(pfn_to_pfn_t(0), flags
);
647 dax_lock_entry(xas
, entry
);
650 mapping
->nrpages
+= 1UL << order
;
655 if (xas_nomem(xas
, mapping_gfp_mask(mapping
) & ~__GFP_HIGHMEM
))
657 if (xas
->xa_node
== XA_ERROR(-ENOMEM
))
658 return xa_mk_internal(VM_FAULT_OOM
);
660 return xa_mk_internal(VM_FAULT_SIGBUS
);
664 return xa_mk_internal(VM_FAULT_FALLBACK
);
668 * dax_layout_busy_page_range - find first pinned page in @mapping
669 * @mapping: address space to scan for a page with ref count > 1
670 * @start: Starting offset. Page containing 'start' is included.
671 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
672 * pages from 'start' till the end of file are included.
674 * DAX requires ZONE_DEVICE mapped pages. These pages are never
675 * 'onlined' to the page allocator so they are considered idle when
676 * page->count == 1. A filesystem uses this interface to determine if
677 * any page in the mapping is busy, i.e. for DMA, or other
678 * get_user_pages() usages.
680 * It is expected that the filesystem is holding locks to block the
681 * establishment of new mappings in this address_space. I.e. it expects
682 * to be able to run unmap_mapping_range() and subsequently not race
683 * mapping_mapped() becoming true.
685 struct page
*dax_layout_busy_page_range(struct address_space
*mapping
,
686 loff_t start
, loff_t end
)
689 unsigned int scanned
= 0;
690 struct page
*page
= NULL
;
691 pgoff_t start_idx
= start
>> PAGE_SHIFT
;
693 XA_STATE(xas
, &mapping
->i_pages
, start_idx
);
696 * In the 'limited' case get_user_pages() for dax is disabled.
698 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
701 if (!dax_mapping(mapping
) || !mapping_mapped(mapping
))
704 /* If end == LLONG_MAX, all pages from start to till end of file */
705 if (end
== LLONG_MAX
)
708 end_idx
= end
>> PAGE_SHIFT
;
710 * If we race get_user_pages_fast() here either we'll see the
711 * elevated page count in the iteration and wait, or
712 * get_user_pages_fast() will see that the page it took a reference
713 * against is no longer mapped in the page tables and bail to the
714 * get_user_pages() slow path. The slow path is protected by
715 * pte_lock() and pmd_lock(). New references are not taken without
716 * holding those locks, and unmap_mapping_pages() will not zero the
717 * pte or pmd without holding the respective lock, so we are
718 * guaranteed to either see new references or prevent new
719 * references from being established.
721 unmap_mapping_pages(mapping
, start_idx
, end_idx
- start_idx
+ 1, 0);
724 xas_for_each(&xas
, entry
, end_idx
) {
725 if (WARN_ON_ONCE(!xa_is_value(entry
)))
727 if (unlikely(dax_is_locked(entry
)))
728 entry
= get_unlocked_entry(&xas
, 0);
730 page
= dax_busy_page(entry
);
731 put_unlocked_entry(&xas
, entry
, WAKE_NEXT
);
734 if (++scanned
% XA_CHECK_SCHED
)
738 xas_unlock_irq(&xas
);
742 xas_unlock_irq(&xas
);
745 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range
);
747 struct page
*dax_layout_busy_page(struct address_space
*mapping
)
749 return dax_layout_busy_page_range(mapping
, 0, LLONG_MAX
);
751 EXPORT_SYMBOL_GPL(dax_layout_busy_page
);
753 static int __dax_invalidate_entry(struct address_space
*mapping
,
754 pgoff_t index
, bool trunc
)
756 XA_STATE(xas
, &mapping
->i_pages
, index
);
761 entry
= get_unlocked_entry(&xas
, 0);
762 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
765 (xas_get_mark(&xas
, PAGECACHE_TAG_DIRTY
) ||
766 xas_get_mark(&xas
, PAGECACHE_TAG_TOWRITE
)))
768 dax_disassociate_entry(entry
, mapping
, trunc
);
769 xas_store(&xas
, NULL
);
770 mapping
->nrpages
-= 1UL << dax_entry_order(entry
);
773 put_unlocked_entry(&xas
, entry
, WAKE_ALL
);
774 xas_unlock_irq(&xas
);
779 * Delete DAX entry at @index from @mapping. Wait for it
780 * to be unlocked before deleting it.
782 int dax_delete_mapping_entry(struct address_space
*mapping
, pgoff_t index
)
784 int ret
= __dax_invalidate_entry(mapping
, index
, true);
787 * This gets called from truncate / punch_hole path. As such, the caller
788 * must hold locks protecting against concurrent modifications of the
789 * page cache (usually fs-private i_mmap_sem for writing). Since the
790 * caller has seen a DAX entry for this index, we better find it
791 * at that index as well...
798 * Invalidate DAX entry if it is clean.
800 int dax_invalidate_mapping_entry_sync(struct address_space
*mapping
,
803 return __dax_invalidate_entry(mapping
, index
, false);
806 static pgoff_t
dax_iomap_pgoff(const struct iomap
*iomap
, loff_t pos
)
808 return PHYS_PFN(iomap
->addr
+ (pos
& PAGE_MASK
) - iomap
->offset
);
811 static int copy_cow_page_dax(struct vm_fault
*vmf
, const struct iomap_iter
*iter
)
813 pgoff_t pgoff
= dax_iomap_pgoff(&iter
->iomap
, iter
->pos
);
818 id
= dax_read_lock();
819 rc
= dax_direct_access(iter
->iomap
.dax_dev
, pgoff
, 1, DAX_ACCESS
,
825 vto
= kmap_atomic(vmf
->cow_page
);
826 copy_user_page(vto
, kaddr
, vmf
->address
, vmf
->cow_page
);
833 * MAP_SYNC on a dax mapping guarantees dirty metadata is
834 * flushed on write-faults (non-cow), but not read-faults.
836 static bool dax_fault_is_synchronous(const struct iomap_iter
*iter
,
837 struct vm_area_struct
*vma
)
839 return (iter
->flags
& IOMAP_WRITE
) && (vma
->vm_flags
& VM_SYNC
) &&
840 (iter
->iomap
.flags
& IOMAP_F_DIRTY
);
843 static bool dax_fault_is_cow(const struct iomap_iter
*iter
)
845 return (iter
->flags
& IOMAP_WRITE
) &&
846 (iter
->iomap
.flags
& IOMAP_F_SHARED
);
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 dirty
= !dax_fault_is_synchronous(iter
, vmf
->vma
);
863 bool cow
= dax_fault_is_cow(iter
);
866 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
868 if (cow
|| (dax_is_zero_entry(entry
) && !(flags
& DAX_ZERO_PAGE
))) {
869 unsigned long index
= xas
->xa_index
;
870 /* we are replacing a zero page with block mapping */
871 if (dax_is_pmd_entry(entry
))
872 unmap_mapping_pages(mapping
, index
& ~PG_PMD_COLOUR
,
875 unmap_mapping_pages(mapping
, index
, 1, false);
880 if (cow
|| dax_is_zero_entry(entry
) || dax_is_empty_entry(entry
)) {
883 dax_disassociate_entry(entry
, mapping
, false);
884 dax_associate_entry(new_entry
, mapping
, vmf
->vma
, vmf
->address
,
887 * Only swap our new entry into the page cache if the current
888 * entry is a zero page or an empty entry. If a normal PTE or
889 * PMD entry is already in the cache, we leave it alone. This
890 * means that if we are trying to insert a PTE and the
891 * existing entry is a PMD, we will just leave the PMD in the
892 * tree and dirty it if necessary.
894 old
= dax_lock_entry(xas
, new_entry
);
895 WARN_ON_ONCE(old
!= xa_mk_value(xa_to_value(entry
) |
899 xas_load(xas
); /* Walk the xa_state */
903 xas_set_mark(xas
, PAGECACHE_TAG_DIRTY
);
906 xas_set_mark(xas
, PAGECACHE_TAG_TOWRITE
);
912 static int dax_writeback_one(struct xa_state
*xas
, struct dax_device
*dax_dev
,
913 struct address_space
*mapping
, void *entry
)
915 unsigned long pfn
, index
, count
, end
;
917 struct vm_area_struct
*vma
;
920 * A page got tagged dirty in DAX mapping? Something is seriously
923 if (WARN_ON(!xa_is_value(entry
)))
926 if (unlikely(dax_is_locked(entry
))) {
927 void *old_entry
= entry
;
929 entry
= get_unlocked_entry(xas
, 0);
931 /* Entry got punched out / reallocated? */
932 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
935 * Entry got reallocated elsewhere? No need to writeback.
936 * We have to compare pfns as we must not bail out due to
937 * difference in lockbit or entry type.
939 if (dax_to_pfn(old_entry
) != dax_to_pfn(entry
))
941 if (WARN_ON_ONCE(dax_is_empty_entry(entry
) ||
942 dax_is_zero_entry(entry
))) {
947 /* Another fsync thread may have already done this entry */
948 if (!xas_get_mark(xas
, PAGECACHE_TAG_TOWRITE
))
952 /* Lock the entry to serialize with page faults */
953 dax_lock_entry(xas
, entry
);
956 * We can clear the tag now but we have to be careful so that concurrent
957 * dax_writeback_one() calls for the same index cannot finish before we
958 * actually flush the caches. This is achieved as the calls will look
959 * at the entry only under the i_pages lock and once they do that
960 * they will see the entry locked and wait for it to unlock.
962 xas_clear_mark(xas
, PAGECACHE_TAG_TOWRITE
);
966 * If dax_writeback_mapping_range() was given a wbc->range_start
967 * in the middle of a PMD, the 'index' we use needs to be
968 * aligned to the start of the PMD.
969 * This allows us to flush for PMD_SIZE and not have to worry about
970 * partial PMD writebacks.
972 pfn
= dax_to_pfn(entry
);
973 count
= 1UL << dax_entry_order(entry
);
974 index
= xas
->xa_index
& ~(count
- 1);
975 end
= index
+ count
- 1;
977 /* Walk all mappings of a given index of a file and writeprotect them */
978 i_mmap_lock_read(mapping
);
979 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, index
, end
) {
980 pfn_mkclean_range(pfn
, count
, index
, vma
);
983 i_mmap_unlock_read(mapping
);
985 dax_flush(dax_dev
, page_address(pfn_to_page(pfn
)), count
* PAGE_SIZE
);
987 * After we have flushed the cache, we can clear the dirty tag. There
988 * cannot be new dirty data in the pfn after the flush has completed as
989 * the pfn mappings are writeprotected and fault waits for mapping
994 xas_store(xas
, entry
);
995 xas_clear_mark(xas
, PAGECACHE_TAG_DIRTY
);
996 dax_wake_entry(xas
, entry
, WAKE_NEXT
);
998 trace_dax_writeback_one(mapping
->host
, index
, count
);
1002 put_unlocked_entry(xas
, entry
, WAKE_NEXT
);
1007 * Flush the mapping to the persistent domain within the byte range of [start,
1008 * end]. This is required by data integrity operations to ensure file data is
1009 * on persistent storage prior to completion of the operation.
1011 int dax_writeback_mapping_range(struct address_space
*mapping
,
1012 struct dax_device
*dax_dev
, struct writeback_control
*wbc
)
1014 XA_STATE(xas
, &mapping
->i_pages
, wbc
->range_start
>> PAGE_SHIFT
);
1015 struct inode
*inode
= mapping
->host
;
1016 pgoff_t end_index
= wbc
->range_end
>> PAGE_SHIFT
;
1019 unsigned int scanned
= 0;
1021 if (WARN_ON_ONCE(inode
->i_blkbits
!= PAGE_SHIFT
))
1024 if (mapping_empty(mapping
) || wbc
->sync_mode
!= WB_SYNC_ALL
)
1027 trace_dax_writeback_range(inode
, xas
.xa_index
, end_index
);
1029 tag_pages_for_writeback(mapping
, xas
.xa_index
, end_index
);
1032 xas_for_each_marked(&xas
, entry
, end_index
, PAGECACHE_TAG_TOWRITE
) {
1033 ret
= dax_writeback_one(&xas
, dax_dev
, mapping
, entry
);
1035 mapping_set_error(mapping
, ret
);
1038 if (++scanned
% XA_CHECK_SCHED
)
1042 xas_unlock_irq(&xas
);
1046 xas_unlock_irq(&xas
);
1047 trace_dax_writeback_range_done(inode
, xas
.xa_index
, end_index
);
1050 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range
);
1052 static int dax_iomap_direct_access(const struct iomap
*iomap
, loff_t pos
,
1053 size_t size
, void **kaddr
, pfn_t
*pfnp
)
1055 pgoff_t pgoff
= dax_iomap_pgoff(iomap
, pos
);
1059 id
= dax_read_lock();
1060 length
= dax_direct_access(iomap
->dax_dev
, pgoff
, PHYS_PFN(size
),
1061 DAX_ACCESS
, kaddr
, pfnp
);
1067 goto out_check_addr
;
1069 if (PFN_PHYS(length
) < size
)
1071 if (pfn_t_to_pfn(*pfnp
) & (PHYS_PFN(size
)-1))
1073 /* For larger pages we need devmap */
1074 if (length
> 1 && !pfn_t_devmap(*pfnp
))
1084 dax_read_unlock(id
);
1089 * dax_iomap_cow_copy - Copy the data from source to destination before write
1090 * @pos: address to do copy from.
1091 * @length: size of copy operation.
1092 * @align_size: aligned w.r.t align_size (either PMD_SIZE or PAGE_SIZE)
1093 * @srcmap: iomap srcmap
1094 * @daddr: destination address to copy to.
1096 * This can be called from two places. Either during DAX write fault (page
1097 * aligned), to copy the length size data to daddr. Or, while doing normal DAX
1098 * write operation, dax_iomap_actor() might call this to do the copy of either
1099 * start or end unaligned address. In the latter case the rest of the copy of
1100 * aligned ranges is taken care by dax_iomap_actor() itself.
1102 static int dax_iomap_cow_copy(loff_t pos
, uint64_t length
, size_t align_size
,
1103 const struct iomap
*srcmap
, void *daddr
)
1105 loff_t head_off
= pos
& (align_size
- 1);
1106 size_t size
= ALIGN(head_off
+ length
, align_size
);
1107 loff_t end
= pos
+ length
;
1108 loff_t pg_end
= round_up(end
, align_size
);
1109 bool copy_all
= head_off
== 0 && end
== pg_end
;
1113 ret
= dax_iomap_direct_access(srcmap
, pos
, size
, &saddr
, NULL
);
1118 ret
= copy_mc_to_kernel(daddr
, saddr
, length
);
1119 return ret
? -EIO
: 0;
1122 /* Copy the head part of the range */
1124 ret
= copy_mc_to_kernel(daddr
, saddr
, head_off
);
1129 /* Copy the tail part of the range */
1131 loff_t tail_off
= head_off
+ length
;
1132 loff_t tail_len
= pg_end
- end
;
1134 ret
= copy_mc_to_kernel(daddr
+ tail_off
, saddr
+ tail_off
,
1143 * The user has performed a load from a hole in the file. Allocating a new
1144 * page in the file would cause excessive storage usage for workloads with
1145 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1146 * If this page is ever written to we will re-fault and change the mapping to
1147 * point to real DAX storage instead.
1149 static vm_fault_t
dax_load_hole(struct xa_state
*xas
, struct vm_fault
*vmf
,
1150 const struct iomap_iter
*iter
, void **entry
)
1152 struct inode
*inode
= iter
->inode
;
1153 unsigned long vaddr
= vmf
->address
;
1154 pfn_t pfn
= pfn_to_pfn_t(my_zero_pfn(vaddr
));
1157 *entry
= dax_insert_entry(xas
, vmf
, iter
, *entry
, pfn
, DAX_ZERO_PAGE
);
1159 ret
= vmf_insert_mixed(vmf
->vma
, vaddr
, pfn
);
1160 trace_dax_load_hole(inode
, vmf
, ret
);
1164 #ifdef CONFIG_FS_DAX_PMD
1165 static vm_fault_t
dax_pmd_load_hole(struct xa_state
*xas
, struct vm_fault
*vmf
,
1166 const struct iomap_iter
*iter
, void **entry
)
1168 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1169 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
1170 struct vm_area_struct
*vma
= vmf
->vma
;
1171 struct inode
*inode
= mapping
->host
;
1172 pgtable_t pgtable
= NULL
;
1173 struct page
*zero_page
;
1178 zero_page
= mm_get_huge_zero_page(vmf
->vma
->vm_mm
);
1180 if (unlikely(!zero_page
))
1183 pfn
= page_to_pfn_t(zero_page
);
1184 *entry
= dax_insert_entry(xas
, vmf
, iter
, *entry
, pfn
,
1185 DAX_PMD
| DAX_ZERO_PAGE
);
1187 if (arch_needs_pgtable_deposit()) {
1188 pgtable
= pte_alloc_one(vma
->vm_mm
);
1190 return VM_FAULT_OOM
;
1193 ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1194 if (!pmd_none(*(vmf
->pmd
))) {
1200 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, pgtable
);
1201 mm_inc_nr_ptes(vma
->vm_mm
);
1203 pmd_entry
= mk_pmd(zero_page
, vmf
->vma
->vm_page_prot
);
1204 pmd_entry
= pmd_mkhuge(pmd_entry
);
1205 set_pmd_at(vmf
->vma
->vm_mm
, pmd_addr
, vmf
->pmd
, pmd_entry
);
1207 trace_dax_pmd_load_hole(inode
, vmf
, zero_page
, *entry
);
1208 return VM_FAULT_NOPAGE
;
1212 pte_free(vma
->vm_mm
, pgtable
);
1213 trace_dax_pmd_load_hole_fallback(inode
, vmf
, zero_page
, *entry
);
1214 return VM_FAULT_FALLBACK
;
1217 static vm_fault_t
dax_pmd_load_hole(struct xa_state
*xas
, struct vm_fault
*vmf
,
1218 const struct iomap_iter
*iter
, void **entry
)
1220 return VM_FAULT_FALLBACK
;
1222 #endif /* CONFIG_FS_DAX_PMD */
1224 static int dax_memzero(struct iomap_iter
*iter
, loff_t pos
, size_t size
)
1226 const struct iomap
*iomap
= &iter
->iomap
;
1227 const struct iomap
*srcmap
= iomap_iter_srcmap(iter
);
1228 unsigned offset
= offset_in_page(pos
);
1229 pgoff_t pgoff
= dax_iomap_pgoff(iomap
, pos
);
1233 ret
= dax_direct_access(iomap
->dax_dev
, pgoff
, 1, DAX_ACCESS
, &kaddr
,
1237 memset(kaddr
+ offset
, 0, size
);
1238 if (srcmap
->addr
!= iomap
->addr
) {
1239 ret
= dax_iomap_cow_copy(pos
, size
, PAGE_SIZE
, srcmap
,
1243 dax_flush(iomap
->dax_dev
, kaddr
, PAGE_SIZE
);
1245 dax_flush(iomap
->dax_dev
, kaddr
+ offset
, size
);
1249 static s64
dax_zero_iter(struct iomap_iter
*iter
, bool *did_zero
)
1251 const struct iomap
*iomap
= &iter
->iomap
;
1252 const struct iomap
*srcmap
= iomap_iter_srcmap(iter
);
1253 loff_t pos
= iter
->pos
;
1254 u64 length
= iomap_length(iter
);
1257 /* already zeroed? we're done. */
1258 if (srcmap
->type
== IOMAP_HOLE
|| srcmap
->type
== IOMAP_UNWRITTEN
)
1262 unsigned offset
= offset_in_page(pos
);
1263 unsigned size
= min_t(u64
, PAGE_SIZE
- offset
, length
);
1264 pgoff_t pgoff
= dax_iomap_pgoff(iomap
, pos
);
1268 id
= dax_read_lock();
1269 if (IS_ALIGNED(pos
, PAGE_SIZE
) && size
== PAGE_SIZE
)
1270 rc
= dax_zero_page_range(iomap
->dax_dev
, pgoff
, 1);
1272 rc
= dax_memzero(iter
, pos
, size
);
1273 dax_read_unlock(id
);
1280 } while (length
> 0);
1287 int dax_zero_range(struct inode
*inode
, loff_t pos
, loff_t len
, bool *did_zero
,
1288 const struct iomap_ops
*ops
)
1290 struct iomap_iter iter
= {
1294 .flags
= IOMAP_DAX
| IOMAP_ZERO
,
1298 while ((ret
= iomap_iter(&iter
, ops
)) > 0)
1299 iter
.processed
= dax_zero_iter(&iter
, did_zero
);
1302 EXPORT_SYMBOL_GPL(dax_zero_range
);
1304 int dax_truncate_page(struct inode
*inode
, loff_t pos
, bool *did_zero
,
1305 const struct iomap_ops
*ops
)
1307 unsigned int blocksize
= i_blocksize(inode
);
1308 unsigned int off
= pos
& (blocksize
- 1);
1310 /* Block boundary? Nothing to do */
1313 return dax_zero_range(inode
, pos
, blocksize
- off
, did_zero
, ops
);
1315 EXPORT_SYMBOL_GPL(dax_truncate_page
);
1317 static loff_t
dax_iomap_iter(const struct iomap_iter
*iomi
,
1318 struct iov_iter
*iter
)
1320 const struct iomap
*iomap
= &iomi
->iomap
;
1321 const struct iomap
*srcmap
= &iomi
->srcmap
;
1322 loff_t length
= iomap_length(iomi
);
1323 loff_t pos
= iomi
->pos
;
1324 struct dax_device
*dax_dev
= iomap
->dax_dev
;
1325 loff_t end
= pos
+ length
, done
= 0;
1326 bool write
= iov_iter_rw(iter
) == WRITE
;
1332 end
= min(end
, i_size_read(iomi
->inode
));
1336 if (iomap
->type
== IOMAP_HOLE
|| iomap
->type
== IOMAP_UNWRITTEN
)
1337 return iov_iter_zero(min(length
, end
- pos
), iter
);
1341 * In DAX mode, enforce either pure overwrites of written extents, or
1342 * writes to unwritten extents as part of a copy-on-write operation.
1344 if (WARN_ON_ONCE(iomap
->type
!= IOMAP_MAPPED
&&
1345 !(iomap
->flags
& IOMAP_F_SHARED
)))
1349 * Write can allocate block for an area which has a hole page mapped
1350 * into page tables. We have to tear down these mappings so that data
1351 * written by write(2) is visible in mmap.
1353 if (iomap
->flags
& IOMAP_F_NEW
) {
1354 invalidate_inode_pages2_range(iomi
->inode
->i_mapping
,
1356 (end
- 1) >> PAGE_SHIFT
);
1359 id
= dax_read_lock();
1361 unsigned offset
= pos
& (PAGE_SIZE
- 1);
1362 const size_t size
= ALIGN(length
+ offset
, PAGE_SIZE
);
1363 pgoff_t pgoff
= dax_iomap_pgoff(iomap
, pos
);
1365 bool recovery
= false;
1368 if (fatal_signal_pending(current
)) {
1373 map_len
= dax_direct_access(dax_dev
, pgoff
, PHYS_PFN(size
),
1374 DAX_ACCESS
, &kaddr
, NULL
);
1375 if (map_len
== -EIO
&& iov_iter_rw(iter
) == WRITE
) {
1376 map_len
= dax_direct_access(dax_dev
, pgoff
,
1377 PHYS_PFN(size
), DAX_RECOVERY_WRITE
,
1388 srcmap
->type
!= IOMAP_HOLE
&& srcmap
->addr
!= iomap
->addr
) {
1389 ret
= dax_iomap_cow_copy(pos
, length
, PAGE_SIZE
, srcmap
,
1395 map_len
= PFN_PHYS(map_len
);
1398 if (map_len
> end
- pos
)
1399 map_len
= end
- pos
;
1402 xfer
= dax_recovery_write(dax_dev
, pgoff
, kaddr
,
1405 xfer
= dax_copy_from_iter(dax_dev
, pgoff
, kaddr
,
1408 xfer
= dax_copy_to_iter(dax_dev
, pgoff
, kaddr
,
1420 dax_read_unlock(id
);
1422 return done
? done
: ret
;
1426 * dax_iomap_rw - Perform I/O to a DAX file
1427 * @iocb: The control block for this I/O
1428 * @iter: The addresses to do I/O from or to
1429 * @ops: iomap ops passed from the file system
1431 * This function performs read and write operations to directly mapped
1432 * persistent memory. The callers needs to take care of read/write exclusion
1433 * and evicting any page cache pages in the region under I/O.
1436 dax_iomap_rw(struct kiocb
*iocb
, struct iov_iter
*iter
,
1437 const struct iomap_ops
*ops
)
1439 struct iomap_iter iomi
= {
1440 .inode
= iocb
->ki_filp
->f_mapping
->host
,
1441 .pos
= iocb
->ki_pos
,
1442 .len
= iov_iter_count(iter
),
1451 if (iov_iter_rw(iter
) == WRITE
) {
1452 lockdep_assert_held_write(&iomi
.inode
->i_rwsem
);
1453 iomi
.flags
|= IOMAP_WRITE
;
1455 lockdep_assert_held(&iomi
.inode
->i_rwsem
);
1458 if (iocb
->ki_flags
& IOCB_NOWAIT
)
1459 iomi
.flags
|= IOMAP_NOWAIT
;
1461 while ((ret
= iomap_iter(&iomi
, ops
)) > 0)
1462 iomi
.processed
= dax_iomap_iter(&iomi
, iter
);
1464 done
= iomi
.pos
- iocb
->ki_pos
;
1465 iocb
->ki_pos
= iomi
.pos
;
1466 return done
? done
: ret
;
1468 EXPORT_SYMBOL_GPL(dax_iomap_rw
);
1470 static vm_fault_t
dax_fault_return(int error
)
1473 return VM_FAULT_NOPAGE
;
1474 return vmf_error(error
);
1478 * When handling a synchronous page fault and the inode need a fsync, we can
1479 * insert the PTE/PMD into page tables only after that fsync happened. Skip
1480 * insertion for now and return the pfn so that caller can insert it after the
1483 static vm_fault_t
dax_fault_synchronous_pfnp(pfn_t
*pfnp
, pfn_t pfn
)
1485 if (WARN_ON_ONCE(!pfnp
))
1486 return VM_FAULT_SIGBUS
;
1488 return VM_FAULT_NEEDDSYNC
;
1491 static vm_fault_t
dax_fault_cow_page(struct vm_fault
*vmf
,
1492 const struct iomap_iter
*iter
)
1497 switch (iter
->iomap
.type
) {
1499 case IOMAP_UNWRITTEN
:
1500 clear_user_highpage(vmf
->cow_page
, vmf
->address
);
1503 error
= copy_cow_page_dax(vmf
, iter
);
1512 return dax_fault_return(error
);
1514 __SetPageUptodate(vmf
->cow_page
);
1515 ret
= finish_fault(vmf
);
1517 return VM_FAULT_DONE_COW
;
1522 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault.
1523 * @vmf: vm fault instance
1525 * @pfnp: pfn to be returned
1526 * @xas: the dax mapping tree of a file
1527 * @entry: an unlocked dax entry to be inserted
1528 * @pmd: distinguish whether it is a pmd fault
1530 static vm_fault_t
dax_fault_iter(struct vm_fault
*vmf
,
1531 const struct iomap_iter
*iter
, pfn_t
*pfnp
,
1532 struct xa_state
*xas
, void **entry
, bool pmd
)
1534 const struct iomap
*iomap
= &iter
->iomap
;
1535 const struct iomap
*srcmap
= &iter
->srcmap
;
1536 size_t size
= pmd
? PMD_SIZE
: PAGE_SIZE
;
1537 loff_t pos
= (loff_t
)xas
->xa_index
<< PAGE_SHIFT
;
1538 bool write
= iter
->flags
& IOMAP_WRITE
;
1539 unsigned long entry_flags
= pmd
? DAX_PMD
: 0;
1544 if (!pmd
&& vmf
->cow_page
)
1545 return dax_fault_cow_page(vmf
, iter
);
1547 /* if we are reading UNWRITTEN and HOLE, return a hole. */
1549 (iomap
->type
== IOMAP_UNWRITTEN
|| iomap
->type
== IOMAP_HOLE
)) {
1551 return dax_load_hole(xas
, vmf
, iter
, entry
);
1552 return dax_pmd_load_hole(xas
, vmf
, iter
, entry
);
1555 if (iomap
->type
!= IOMAP_MAPPED
&& !(iomap
->flags
& IOMAP_F_SHARED
)) {
1557 return pmd
? VM_FAULT_FALLBACK
: VM_FAULT_SIGBUS
;
1560 err
= dax_iomap_direct_access(iomap
, pos
, size
, &kaddr
, &pfn
);
1562 return pmd
? VM_FAULT_FALLBACK
: dax_fault_return(err
);
1564 *entry
= dax_insert_entry(xas
, vmf
, iter
, *entry
, pfn
, entry_flags
);
1567 srcmap
->type
!= IOMAP_HOLE
&& srcmap
->addr
!= iomap
->addr
) {
1568 err
= dax_iomap_cow_copy(pos
, size
, size
, srcmap
, kaddr
);
1570 return dax_fault_return(err
);
1573 if (dax_fault_is_synchronous(iter
, vmf
->vma
))
1574 return dax_fault_synchronous_pfnp(pfnp
, pfn
);
1576 /* insert PMD pfn */
1578 return vmf_insert_pfn_pmd(vmf
, pfn
, write
);
1580 /* insert PTE pfn */
1582 return vmf_insert_mixed_mkwrite(vmf
->vma
, vmf
->address
, pfn
);
1583 return vmf_insert_mixed(vmf
->vma
, vmf
->address
, pfn
);
1586 static vm_fault_t
dax_iomap_pte_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1587 int *iomap_errp
, const struct iomap_ops
*ops
)
1589 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1590 XA_STATE(xas
, &mapping
->i_pages
, vmf
->pgoff
);
1591 struct iomap_iter iter
= {
1592 .inode
= mapping
->host
,
1593 .pos
= (loff_t
)vmf
->pgoff
<< PAGE_SHIFT
,
1595 .flags
= IOMAP_DAX
| IOMAP_FAULT
,
1601 trace_dax_pte_fault(iter
.inode
, vmf
, ret
);
1603 * Check whether offset isn't beyond end of file now. Caller is supposed
1604 * to hold locks serializing us with truncate / punch hole so this is
1607 if (iter
.pos
>= i_size_read(iter
.inode
)) {
1608 ret
= VM_FAULT_SIGBUS
;
1612 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && !vmf
->cow_page
)
1613 iter
.flags
|= IOMAP_WRITE
;
1615 entry
= grab_mapping_entry(&xas
, mapping
, 0);
1616 if (xa_is_internal(entry
)) {
1617 ret
= xa_to_internal(entry
);
1622 * It is possible, particularly with mixed reads & writes to private
1623 * mappings, that we have raced with a PMD fault that overlaps with
1624 * the PTE we need to set up. If so just return and the fault will be
1627 if (pmd_trans_huge(*vmf
->pmd
) || pmd_devmap(*vmf
->pmd
)) {
1628 ret
= VM_FAULT_NOPAGE
;
1632 while ((error
= iomap_iter(&iter
, ops
)) > 0) {
1633 if (WARN_ON_ONCE(iomap_length(&iter
) < PAGE_SIZE
)) {
1634 iter
.processed
= -EIO
; /* fs corruption? */
1638 ret
= dax_fault_iter(vmf
, &iter
, pfnp
, &xas
, &entry
, false);
1639 if (ret
!= VM_FAULT_SIGBUS
&&
1640 (iter
.iomap
.flags
& IOMAP_F_NEW
)) {
1641 count_vm_event(PGMAJFAULT
);
1642 count_memcg_event_mm(vmf
->vma
->vm_mm
, PGMAJFAULT
);
1643 ret
|= VM_FAULT_MAJOR
;
1646 if (!(ret
& VM_FAULT_ERROR
))
1647 iter
.processed
= PAGE_SIZE
;
1651 *iomap_errp
= error
;
1653 ret
= dax_fault_return(error
);
1656 dax_unlock_entry(&xas
, entry
);
1658 trace_dax_pte_fault_done(iter
.inode
, vmf
, ret
);
1662 #ifdef CONFIG_FS_DAX_PMD
1663 static bool dax_fault_check_fallback(struct vm_fault
*vmf
, struct xa_state
*xas
,
1666 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
1667 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1670 * Make sure that the faulting address's PMD offset (color) matches
1671 * the PMD offset from the start of the file. This is necessary so
1672 * that a PMD range in the page table overlaps exactly with a PMD
1673 * range in the page cache.
1675 if ((vmf
->pgoff
& PG_PMD_COLOUR
) !=
1676 ((vmf
->address
>> PAGE_SHIFT
) & PG_PMD_COLOUR
))
1679 /* Fall back to PTEs if we're going to COW */
1680 if (write
&& !(vmf
->vma
->vm_flags
& VM_SHARED
))
1683 /* If the PMD would extend outside the VMA */
1684 if (pmd_addr
< vmf
->vma
->vm_start
)
1686 if ((pmd_addr
+ PMD_SIZE
) > vmf
->vma
->vm_end
)
1689 /* If the PMD would extend beyond the file size */
1690 if ((xas
->xa_index
| PG_PMD_COLOUR
) >= max_pgoff
)
1696 static vm_fault_t
dax_iomap_pmd_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1697 const struct iomap_ops
*ops
)
1699 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1700 XA_STATE_ORDER(xas
, &mapping
->i_pages
, vmf
->pgoff
, PMD_ORDER
);
1701 struct iomap_iter iter
= {
1702 .inode
= mapping
->host
,
1704 .flags
= IOMAP_DAX
| IOMAP_FAULT
,
1706 vm_fault_t ret
= VM_FAULT_FALLBACK
;
1711 if (vmf
->flags
& FAULT_FLAG_WRITE
)
1712 iter
.flags
|= IOMAP_WRITE
;
1715 * Check whether offset isn't beyond end of file now. Caller is
1716 * supposed to hold locks serializing us with truncate / punch hole so
1717 * this is a reliable test.
1719 max_pgoff
= DIV_ROUND_UP(i_size_read(iter
.inode
), PAGE_SIZE
);
1721 trace_dax_pmd_fault(iter
.inode
, vmf
, max_pgoff
, 0);
1723 if (xas
.xa_index
>= max_pgoff
) {
1724 ret
= VM_FAULT_SIGBUS
;
1728 if (dax_fault_check_fallback(vmf
, &xas
, max_pgoff
))
1732 * grab_mapping_entry() will make sure we get an empty PMD entry,
1733 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1734 * entry is already in the array, for instance), it will return
1735 * VM_FAULT_FALLBACK.
1737 entry
= grab_mapping_entry(&xas
, mapping
, PMD_ORDER
);
1738 if (xa_is_internal(entry
)) {
1739 ret
= xa_to_internal(entry
);
1744 * It is possible, particularly with mixed reads & writes to private
1745 * mappings, that we have raced with a PTE fault that overlaps with
1746 * the PMD we need to set up. If so just return and the fault will be
1749 if (!pmd_none(*vmf
->pmd
) && !pmd_trans_huge(*vmf
->pmd
) &&
1750 !pmd_devmap(*vmf
->pmd
)) {
1755 iter
.pos
= (loff_t
)xas
.xa_index
<< PAGE_SHIFT
;
1756 while ((error
= iomap_iter(&iter
, ops
)) > 0) {
1757 if (iomap_length(&iter
) < PMD_SIZE
)
1758 continue; /* actually breaks out of the loop */
1760 ret
= dax_fault_iter(vmf
, &iter
, pfnp
, &xas
, &entry
, true);
1761 if (ret
!= VM_FAULT_FALLBACK
)
1762 iter
.processed
= PMD_SIZE
;
1766 dax_unlock_entry(&xas
, entry
);
1768 if (ret
== VM_FAULT_FALLBACK
) {
1769 split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
);
1770 count_vm_event(THP_FAULT_FALLBACK
);
1773 trace_dax_pmd_fault_done(iter
.inode
, vmf
, max_pgoff
, ret
);
1777 static vm_fault_t
dax_iomap_pmd_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1778 const struct iomap_ops
*ops
)
1780 return VM_FAULT_FALLBACK
;
1782 #endif /* CONFIG_FS_DAX_PMD */
1785 * dax_iomap_fault - handle a page fault on a DAX file
1786 * @vmf: The description of the fault
1787 * @pe_size: Size of the page to fault in
1788 * @pfnp: PFN to insert for synchronous faults if fsync is required
1789 * @iomap_errp: Storage for detailed error code in case of error
1790 * @ops: Iomap ops passed from the file system
1792 * When a page fault occurs, filesystems may call this helper in
1793 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1794 * has done all the necessary locking for page fault to proceed
1797 vm_fault_t
dax_iomap_fault(struct vm_fault
*vmf
, enum page_entry_size pe_size
,
1798 pfn_t
*pfnp
, int *iomap_errp
, const struct iomap_ops
*ops
)
1802 return dax_iomap_pte_fault(vmf
, pfnp
, iomap_errp
, ops
);
1804 return dax_iomap_pmd_fault(vmf
, pfnp
, ops
);
1806 return VM_FAULT_FALLBACK
;
1809 EXPORT_SYMBOL_GPL(dax_iomap_fault
);
1812 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1813 * @vmf: The description of the fault
1814 * @pfn: PFN to insert
1815 * @order: Order of entry to insert.
1817 * This function inserts a writeable PTE or PMD entry into the page tables
1818 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1821 dax_insert_pfn_mkwrite(struct vm_fault
*vmf
, pfn_t pfn
, unsigned int order
)
1823 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1824 XA_STATE_ORDER(xas
, &mapping
->i_pages
, vmf
->pgoff
, order
);
1829 entry
= get_unlocked_entry(&xas
, order
);
1830 /* Did we race with someone splitting entry or so? */
1831 if (!entry
|| dax_is_conflict(entry
) ||
1832 (order
== 0 && !dax_is_pte_entry(entry
))) {
1833 put_unlocked_entry(&xas
, entry
, WAKE_NEXT
);
1834 xas_unlock_irq(&xas
);
1835 trace_dax_insert_pfn_mkwrite_no_entry(mapping
->host
, vmf
,
1837 return VM_FAULT_NOPAGE
;
1839 xas_set_mark(&xas
, PAGECACHE_TAG_DIRTY
);
1840 dax_lock_entry(&xas
, entry
);
1841 xas_unlock_irq(&xas
);
1843 ret
= vmf_insert_mixed_mkwrite(vmf
->vma
, vmf
->address
, pfn
);
1844 #ifdef CONFIG_FS_DAX_PMD
1845 else if (order
== PMD_ORDER
)
1846 ret
= vmf_insert_pfn_pmd(vmf
, pfn
, FAULT_FLAG_WRITE
);
1849 ret
= VM_FAULT_FALLBACK
;
1850 dax_unlock_entry(&xas
, entry
);
1851 trace_dax_insert_pfn_mkwrite(mapping
->host
, vmf
, ret
);
1856 * dax_finish_sync_fault - finish synchronous page fault
1857 * @vmf: The description of the fault
1858 * @pe_size: Size of entry to be inserted
1859 * @pfn: PFN to insert
1861 * This function ensures that the file range touched by the page fault is
1862 * stored persistently on the media and handles inserting of appropriate page
1865 vm_fault_t
dax_finish_sync_fault(struct vm_fault
*vmf
,
1866 enum page_entry_size pe_size
, pfn_t pfn
)
1869 loff_t start
= ((loff_t
)vmf
->pgoff
) << PAGE_SHIFT
;
1870 unsigned int order
= pe_order(pe_size
);
1871 size_t len
= PAGE_SIZE
<< order
;
1873 err
= vfs_fsync_range(vmf
->vma
->vm_file
, start
, start
+ len
- 1, 1);
1875 return VM_FAULT_SIGBUS
;
1876 return dax_insert_pfn_mkwrite(vmf
, pfn
, order
);
1878 EXPORT_SYMBOL_GPL(dax_finish_sync_fault
);
1880 static loff_t
dax_range_compare_iter(struct iomap_iter
*it_src
,
1881 struct iomap_iter
*it_dest
, u64 len
, bool *same
)
1883 const struct iomap
*smap
= &it_src
->iomap
;
1884 const struct iomap
*dmap
= &it_dest
->iomap
;
1885 loff_t pos1
= it_src
->pos
, pos2
= it_dest
->pos
;
1886 void *saddr
, *daddr
;
1889 len
= min(len
, min(smap
->length
, dmap
->length
));
1891 if (smap
->type
== IOMAP_HOLE
&& dmap
->type
== IOMAP_HOLE
) {
1896 if (smap
->type
== IOMAP_HOLE
|| dmap
->type
== IOMAP_HOLE
) {
1901 id
= dax_read_lock();
1902 ret
= dax_iomap_direct_access(smap
, pos1
, ALIGN(pos1
+ len
, PAGE_SIZE
),
1907 ret
= dax_iomap_direct_access(dmap
, pos2
, ALIGN(pos2
+ len
, PAGE_SIZE
),
1912 *same
= !memcmp(saddr
, daddr
, len
);
1915 dax_read_unlock(id
);
1919 dax_read_unlock(id
);
1923 int dax_dedupe_file_range_compare(struct inode
*src
, loff_t srcoff
,
1924 struct inode
*dst
, loff_t dstoff
, loff_t len
, bool *same
,
1925 const struct iomap_ops
*ops
)
1927 struct iomap_iter src_iter
= {
1933 struct iomap_iter dst_iter
= {
1941 while ((ret
= iomap_iter(&src_iter
, ops
)) > 0) {
1942 while ((ret
= iomap_iter(&dst_iter
, ops
)) > 0) {
1943 dst_iter
.processed
= dax_range_compare_iter(&src_iter
,
1944 &dst_iter
, len
, same
);
1947 src_iter
.processed
= ret
;
1952 int dax_remap_file_range_prep(struct file
*file_in
, loff_t pos_in
,
1953 struct file
*file_out
, loff_t pos_out
,
1954 loff_t
*len
, unsigned int remap_flags
,
1955 const struct iomap_ops
*ops
)
1957 return __generic_remap_file_range_prep(file_in
, pos_in
, file_out
,
1958 pos_out
, len
, remap_flags
, ops
);
1960 EXPORT_SYMBOL_GPL(dax_remap_file_range_prep
);