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/genhd.h>
15 #include <linux/highmem.h>
16 #include <linux/memcontrol.h>
18 #include <linux/mutex.h>
19 #include <linux/pagevec.h>
20 #include <linux/sched.h>
21 #include <linux/sched/signal.h>
22 #include <linux/uio.h>
23 #include <linux/vmstat.h>
24 #include <linux/pfn_t.h>
25 #include <linux/sizes.h>
26 #include <linux/mmu_notifier.h>
27 #include <linux/iomap.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++)
338 * TODO: for reflink+dax we need a way to associate a single page with
339 * multiple address_space instances at different linear_page_index()
342 static void dax_associate_entry(void *entry
, struct address_space
*mapping
,
343 struct vm_area_struct
*vma
, unsigned long address
)
345 unsigned long size
= dax_entry_size(entry
), pfn
, index
;
348 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
351 index
= linear_page_index(vma
, address
& ~(size
- 1));
352 for_each_mapped_pfn(entry
, pfn
) {
353 struct page
*page
= pfn_to_page(pfn
);
355 WARN_ON_ONCE(page
->mapping
);
356 page
->mapping
= mapping
;
357 page
->index
= index
+ i
++;
361 static void dax_disassociate_entry(void *entry
, struct address_space
*mapping
,
366 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
369 for_each_mapped_pfn(entry
, pfn
) {
370 struct page
*page
= pfn_to_page(pfn
);
372 WARN_ON_ONCE(trunc
&& page_ref_count(page
) > 1);
373 WARN_ON_ONCE(page
->mapping
&& page
->mapping
!= mapping
);
374 page
->mapping
= NULL
;
379 static struct page
*dax_busy_page(void *entry
)
383 for_each_mapped_pfn(entry
, pfn
) {
384 struct page
*page
= pfn_to_page(pfn
);
386 if (page_ref_count(page
) > 1)
393 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page
394 * @page: The page whose entry we want to lock
396 * Context: Process context.
397 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
400 dax_entry_t
dax_lock_page(struct page
*page
)
402 XA_STATE(xas
, NULL
, 0);
405 /* Ensure page->mapping isn't freed while we look at it */
408 struct address_space
*mapping
= READ_ONCE(page
->mapping
);
411 if (!mapping
|| !dax_mapping(mapping
))
415 * In the device-dax case there's no need to lock, a
416 * struct dev_pagemap pin is sufficient to keep the
417 * inode alive, and we assume we have dev_pagemap pin
418 * otherwise we would not have a valid pfn_to_page()
421 entry
= (void *)~0UL;
422 if (S_ISCHR(mapping
->host
->i_mode
))
425 xas
.xa
= &mapping
->i_pages
;
427 if (mapping
!= page
->mapping
) {
428 xas_unlock_irq(&xas
);
431 xas_set(&xas
, page
->index
);
432 entry
= xas_load(&xas
);
433 if (dax_is_locked(entry
)) {
435 wait_entry_unlocked(&xas
, entry
);
439 dax_lock_entry(&xas
, entry
);
440 xas_unlock_irq(&xas
);
444 return (dax_entry_t
)entry
;
447 void dax_unlock_page(struct page
*page
, dax_entry_t cookie
)
449 struct address_space
*mapping
= page
->mapping
;
450 XA_STATE(xas
, &mapping
->i_pages
, page
->index
);
452 if (S_ISCHR(mapping
->host
->i_mode
))
455 dax_unlock_entry(&xas
, (void *)cookie
);
459 * Find page cache entry at given index. If it is a DAX entry, return it
460 * with the entry locked. If the page cache doesn't contain an entry at
461 * that index, add a locked empty entry.
463 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
464 * either return that locked entry or will return VM_FAULT_FALLBACK.
465 * This will happen if there are any PTE entries within the PMD range
466 * that we are requesting.
468 * We always favor PTE entries over PMD entries. There isn't a flow where we
469 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
470 * insertion will fail if it finds any PTE entries already in the tree, and a
471 * PTE insertion will cause an existing PMD entry to be unmapped and
472 * downgraded to PTE entries. This happens for both PMD zero pages as
473 * well as PMD empty entries.
475 * The exception to this downgrade path is for PMD entries that have
476 * real storage backing them. We will leave these real PMD entries in
477 * the tree, and PTE writes will simply dirty the entire PMD entry.
479 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
480 * persistent memory the benefit is doubtful. We can add that later if we can
483 * On error, this function does not return an ERR_PTR. Instead it returns
484 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
485 * overlap with xarray value entries.
487 static void *grab_mapping_entry(struct xa_state
*xas
,
488 struct address_space
*mapping
, unsigned int order
)
490 unsigned long index
= xas
->xa_index
;
491 bool pmd_downgrade
; /* splitting PMD entry into PTE entries? */
495 pmd_downgrade
= false;
497 entry
= get_unlocked_entry(xas
, order
);
500 if (dax_is_conflict(entry
))
502 if (!xa_is_value(entry
)) {
503 xas_set_err(xas
, -EIO
);
508 if (dax_is_pmd_entry(entry
) &&
509 (dax_is_zero_entry(entry
) ||
510 dax_is_empty_entry(entry
))) {
511 pmd_downgrade
= true;
518 * Make sure 'entry' remains valid while we drop
521 dax_lock_entry(xas
, entry
);
524 * Besides huge zero pages the only other thing that gets
525 * downgraded are empty entries which don't need to be
528 if (dax_is_zero_entry(entry
)) {
530 unmap_mapping_pages(mapping
,
531 xas
->xa_index
& ~PG_PMD_COLOUR
,
537 dax_disassociate_entry(entry
, mapping
, false);
538 xas_store(xas
, NULL
); /* undo the PMD join */
539 dax_wake_entry(xas
, entry
, WAKE_ALL
);
540 mapping
->nrpages
-= PG_PMD_NR
;
546 dax_lock_entry(xas
, entry
);
548 unsigned long flags
= DAX_EMPTY
;
552 entry
= dax_make_entry(pfn_to_pfn_t(0), flags
);
553 dax_lock_entry(xas
, entry
);
556 mapping
->nrpages
+= 1UL << order
;
561 if (xas_nomem(xas
, mapping_gfp_mask(mapping
) & ~__GFP_HIGHMEM
))
563 if (xas
->xa_node
== XA_ERROR(-ENOMEM
))
564 return xa_mk_internal(VM_FAULT_OOM
);
566 return xa_mk_internal(VM_FAULT_SIGBUS
);
570 return xa_mk_internal(VM_FAULT_FALLBACK
);
574 * dax_layout_busy_page_range - find first pinned page in @mapping
575 * @mapping: address space to scan for a page with ref count > 1
576 * @start: Starting offset. Page containing 'start' is included.
577 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
578 * pages from 'start' till the end of file are included.
580 * DAX requires ZONE_DEVICE mapped pages. These pages are never
581 * 'onlined' to the page allocator so they are considered idle when
582 * page->count == 1. A filesystem uses this interface to determine if
583 * any page in the mapping is busy, i.e. for DMA, or other
584 * get_user_pages() usages.
586 * It is expected that the filesystem is holding locks to block the
587 * establishment of new mappings in this address_space. I.e. it expects
588 * to be able to run unmap_mapping_range() and subsequently not race
589 * mapping_mapped() becoming true.
591 struct page
*dax_layout_busy_page_range(struct address_space
*mapping
,
592 loff_t start
, loff_t end
)
595 unsigned int scanned
= 0;
596 struct page
*page
= NULL
;
597 pgoff_t start_idx
= start
>> PAGE_SHIFT
;
599 XA_STATE(xas
, &mapping
->i_pages
, start_idx
);
602 * In the 'limited' case get_user_pages() for dax is disabled.
604 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
607 if (!dax_mapping(mapping
) || !mapping_mapped(mapping
))
610 /* If end == LLONG_MAX, all pages from start to till end of file */
611 if (end
== LLONG_MAX
)
614 end_idx
= end
>> PAGE_SHIFT
;
616 * If we race get_user_pages_fast() here either we'll see the
617 * elevated page count in the iteration and wait, or
618 * get_user_pages_fast() will see that the page it took a reference
619 * against is no longer mapped in the page tables and bail to the
620 * get_user_pages() slow path. The slow path is protected by
621 * pte_lock() and pmd_lock(). New references are not taken without
622 * holding those locks, and unmap_mapping_pages() will not zero the
623 * pte or pmd without holding the respective lock, so we are
624 * guaranteed to either see new references or prevent new
625 * references from being established.
627 unmap_mapping_pages(mapping
, start_idx
, end_idx
- start_idx
+ 1, 0);
630 xas_for_each(&xas
, entry
, end_idx
) {
631 if (WARN_ON_ONCE(!xa_is_value(entry
)))
633 if (unlikely(dax_is_locked(entry
)))
634 entry
= get_unlocked_entry(&xas
, 0);
636 page
= dax_busy_page(entry
);
637 put_unlocked_entry(&xas
, entry
, WAKE_NEXT
);
640 if (++scanned
% XA_CHECK_SCHED
)
644 xas_unlock_irq(&xas
);
648 xas_unlock_irq(&xas
);
651 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range
);
653 struct page
*dax_layout_busy_page(struct address_space
*mapping
)
655 return dax_layout_busy_page_range(mapping
, 0, LLONG_MAX
);
657 EXPORT_SYMBOL_GPL(dax_layout_busy_page
);
659 static int __dax_invalidate_entry(struct address_space
*mapping
,
660 pgoff_t index
, bool trunc
)
662 XA_STATE(xas
, &mapping
->i_pages
, index
);
667 entry
= get_unlocked_entry(&xas
, 0);
668 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
671 (xas_get_mark(&xas
, PAGECACHE_TAG_DIRTY
) ||
672 xas_get_mark(&xas
, PAGECACHE_TAG_TOWRITE
)))
674 dax_disassociate_entry(entry
, mapping
, trunc
);
675 xas_store(&xas
, NULL
);
676 mapping
->nrpages
-= 1UL << dax_entry_order(entry
);
679 put_unlocked_entry(&xas
, entry
, WAKE_ALL
);
680 xas_unlock_irq(&xas
);
685 * Delete DAX entry at @index from @mapping. Wait for it
686 * to be unlocked before deleting it.
688 int dax_delete_mapping_entry(struct address_space
*mapping
, pgoff_t index
)
690 int ret
= __dax_invalidate_entry(mapping
, index
, true);
693 * This gets called from truncate / punch_hole path. As such, the caller
694 * must hold locks protecting against concurrent modifications of the
695 * page cache (usually fs-private i_mmap_sem for writing). Since the
696 * caller has seen a DAX entry for this index, we better find it
697 * at that index as well...
704 * Invalidate DAX entry if it is clean.
706 int dax_invalidate_mapping_entry_sync(struct address_space
*mapping
,
709 return __dax_invalidate_entry(mapping
, index
, false);
712 static int copy_cow_page_dax(struct block_device
*bdev
, struct dax_device
*dax_dev
,
713 sector_t sector
, struct page
*to
, unsigned long vaddr
)
720 rc
= bdev_dax_pgoff(bdev
, sector
, PAGE_SIZE
, &pgoff
);
724 id
= dax_read_lock();
725 rc
= dax_direct_access(dax_dev
, pgoff
, 1, &kaddr
, NULL
);
730 vto
= kmap_atomic(to
);
731 copy_user_page(vto
, (void __force
*)kaddr
, vaddr
, to
);
738 * By this point grab_mapping_entry() has ensured that we have a locked entry
739 * of the appropriate size so we don't have to worry about downgrading PMDs to
740 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
741 * already in the tree, we will skip the insertion and just dirty the PMD as
744 static void *dax_insert_entry(struct xa_state
*xas
,
745 struct address_space
*mapping
, struct vm_fault
*vmf
,
746 void *entry
, pfn_t pfn
, unsigned long flags
, bool dirty
)
748 void *new_entry
= dax_make_entry(pfn
, flags
);
751 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
753 if (dax_is_zero_entry(entry
) && !(flags
& DAX_ZERO_PAGE
)) {
754 unsigned long index
= xas
->xa_index
;
755 /* we are replacing a zero page with block mapping */
756 if (dax_is_pmd_entry(entry
))
757 unmap_mapping_pages(mapping
, index
& ~PG_PMD_COLOUR
,
760 unmap_mapping_pages(mapping
, index
, 1, false);
765 if (dax_is_zero_entry(entry
) || dax_is_empty_entry(entry
)) {
768 dax_disassociate_entry(entry
, mapping
, false);
769 dax_associate_entry(new_entry
, mapping
, vmf
->vma
, vmf
->address
);
771 * Only swap our new entry into the page cache if the current
772 * entry is a zero page or an empty entry. If a normal PTE or
773 * PMD entry is already in the cache, we leave it alone. This
774 * means that if we are trying to insert a PTE and the
775 * existing entry is a PMD, we will just leave the PMD in the
776 * tree and dirty it if necessary.
778 old
= dax_lock_entry(xas
, new_entry
);
779 WARN_ON_ONCE(old
!= xa_mk_value(xa_to_value(entry
) |
783 xas_load(xas
); /* Walk the xa_state */
787 xas_set_mark(xas
, PAGECACHE_TAG_DIRTY
);
794 unsigned long pgoff_address(pgoff_t pgoff
, struct vm_area_struct
*vma
)
796 unsigned long address
;
798 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
799 VM_BUG_ON_VMA(address
< vma
->vm_start
|| address
>= vma
->vm_end
, vma
);
803 /* Walk all mappings of a given index of a file and writeprotect them */
804 static void dax_entry_mkclean(struct address_space
*mapping
, pgoff_t index
,
807 struct vm_area_struct
*vma
;
808 pte_t pte
, *ptep
= NULL
;
812 i_mmap_lock_read(mapping
);
813 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, index
, index
) {
814 struct mmu_notifier_range range
;
815 unsigned long address
;
819 if (!(vma
->vm_flags
& VM_SHARED
))
822 address
= pgoff_address(index
, vma
);
825 * follow_invalidate_pte() will use the range to call
826 * mmu_notifier_invalidate_range_start() on our behalf before
829 if (follow_invalidate_pte(vma
->vm_mm
, address
, &range
, &ptep
,
834 * No need to call mmu_notifier_invalidate_range() as we are
835 * downgrading page table protection not changing it to point
838 * See Documentation/vm/mmu_notifier.rst
841 #ifdef CONFIG_FS_DAX_PMD
844 if (pfn
!= pmd_pfn(*pmdp
))
846 if (!pmd_dirty(*pmdp
) && !pmd_write(*pmdp
))
849 flush_cache_page(vma
, address
, pfn
);
850 pmd
= pmdp_invalidate(vma
, address
, pmdp
);
851 pmd
= pmd_wrprotect(pmd
);
852 pmd
= pmd_mkclean(pmd
);
853 set_pmd_at(vma
->vm_mm
, address
, pmdp
, pmd
);
858 if (pfn
!= pte_pfn(*ptep
))
860 if (!pte_dirty(*ptep
) && !pte_write(*ptep
))
863 flush_cache_page(vma
, address
, pfn
);
864 pte
= ptep_clear_flush(vma
, address
, ptep
);
865 pte
= pte_wrprotect(pte
);
866 pte
= pte_mkclean(pte
);
867 set_pte_at(vma
->vm_mm
, address
, ptep
, pte
);
869 pte_unmap_unlock(ptep
, ptl
);
872 mmu_notifier_invalidate_range_end(&range
);
874 i_mmap_unlock_read(mapping
);
877 static int dax_writeback_one(struct xa_state
*xas
, struct dax_device
*dax_dev
,
878 struct address_space
*mapping
, void *entry
)
880 unsigned long pfn
, index
, count
;
884 * A page got tagged dirty in DAX mapping? Something is seriously
887 if (WARN_ON(!xa_is_value(entry
)))
890 if (unlikely(dax_is_locked(entry
))) {
891 void *old_entry
= entry
;
893 entry
= get_unlocked_entry(xas
, 0);
895 /* Entry got punched out / reallocated? */
896 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
899 * Entry got reallocated elsewhere? No need to writeback.
900 * We have to compare pfns as we must not bail out due to
901 * difference in lockbit or entry type.
903 if (dax_to_pfn(old_entry
) != dax_to_pfn(entry
))
905 if (WARN_ON_ONCE(dax_is_empty_entry(entry
) ||
906 dax_is_zero_entry(entry
))) {
911 /* Another fsync thread may have already done this entry */
912 if (!xas_get_mark(xas
, PAGECACHE_TAG_TOWRITE
))
916 /* Lock the entry to serialize with page faults */
917 dax_lock_entry(xas
, entry
);
920 * We can clear the tag now but we have to be careful so that concurrent
921 * dax_writeback_one() calls for the same index cannot finish before we
922 * actually flush the caches. This is achieved as the calls will look
923 * at the entry only under the i_pages lock and once they do that
924 * they will see the entry locked and wait for it to unlock.
926 xas_clear_mark(xas
, PAGECACHE_TAG_TOWRITE
);
930 * If dax_writeback_mapping_range() was given a wbc->range_start
931 * in the middle of a PMD, the 'index' we use needs to be
932 * aligned to the start of the PMD.
933 * This allows us to flush for PMD_SIZE and not have to worry about
934 * partial PMD writebacks.
936 pfn
= dax_to_pfn(entry
);
937 count
= 1UL << dax_entry_order(entry
);
938 index
= xas
->xa_index
& ~(count
- 1);
940 dax_entry_mkclean(mapping
, index
, pfn
);
941 dax_flush(dax_dev
, page_address(pfn_to_page(pfn
)), count
* PAGE_SIZE
);
943 * After we have flushed the cache, we can clear the dirty tag. There
944 * cannot be new dirty data in the pfn after the flush has completed as
945 * the pfn mappings are writeprotected and fault waits for mapping
950 xas_store(xas
, entry
);
951 xas_clear_mark(xas
, PAGECACHE_TAG_DIRTY
);
952 dax_wake_entry(xas
, entry
, WAKE_NEXT
);
954 trace_dax_writeback_one(mapping
->host
, index
, count
);
958 put_unlocked_entry(xas
, entry
, WAKE_NEXT
);
963 * Flush the mapping to the persistent domain within the byte range of [start,
964 * end]. This is required by data integrity operations to ensure file data is
965 * on persistent storage prior to completion of the operation.
967 int dax_writeback_mapping_range(struct address_space
*mapping
,
968 struct dax_device
*dax_dev
, struct writeback_control
*wbc
)
970 XA_STATE(xas
, &mapping
->i_pages
, wbc
->range_start
>> PAGE_SHIFT
);
971 struct inode
*inode
= mapping
->host
;
972 pgoff_t end_index
= wbc
->range_end
>> PAGE_SHIFT
;
975 unsigned int scanned
= 0;
977 if (WARN_ON_ONCE(inode
->i_blkbits
!= PAGE_SHIFT
))
980 if (mapping_empty(mapping
) || wbc
->sync_mode
!= WB_SYNC_ALL
)
983 trace_dax_writeback_range(inode
, xas
.xa_index
, end_index
);
985 tag_pages_for_writeback(mapping
, xas
.xa_index
, end_index
);
988 xas_for_each_marked(&xas
, entry
, end_index
, PAGECACHE_TAG_TOWRITE
) {
989 ret
= dax_writeback_one(&xas
, dax_dev
, mapping
, entry
);
991 mapping_set_error(mapping
, ret
);
994 if (++scanned
% XA_CHECK_SCHED
)
998 xas_unlock_irq(&xas
);
1002 xas_unlock_irq(&xas
);
1003 trace_dax_writeback_range_done(inode
, xas
.xa_index
, end_index
);
1006 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range
);
1008 static sector_t
dax_iomap_sector(const struct iomap
*iomap
, loff_t pos
)
1010 return (iomap
->addr
+ (pos
& PAGE_MASK
) - iomap
->offset
) >> 9;
1013 static int dax_iomap_pfn(const struct iomap
*iomap
, loff_t pos
, size_t size
,
1016 const sector_t sector
= dax_iomap_sector(iomap
, pos
);
1021 rc
= bdev_dax_pgoff(iomap
->bdev
, sector
, size
, &pgoff
);
1024 id
= dax_read_lock();
1025 length
= dax_direct_access(iomap
->dax_dev
, pgoff
, PHYS_PFN(size
),
1032 if (PFN_PHYS(length
) < size
)
1034 if (pfn_t_to_pfn(*pfnp
) & (PHYS_PFN(size
)-1))
1036 /* For larger pages we need devmap */
1037 if (length
> 1 && !pfn_t_devmap(*pfnp
))
1041 dax_read_unlock(id
);
1046 * The user has performed a load from a hole in the file. Allocating a new
1047 * page in the file would cause excessive storage usage for workloads with
1048 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1049 * If this page is ever written to we will re-fault and change the mapping to
1050 * point to real DAX storage instead.
1052 static vm_fault_t
dax_load_hole(struct xa_state
*xas
,
1053 struct address_space
*mapping
, void **entry
,
1054 struct vm_fault
*vmf
)
1056 struct inode
*inode
= mapping
->host
;
1057 unsigned long vaddr
= vmf
->address
;
1058 pfn_t pfn
= pfn_to_pfn_t(my_zero_pfn(vaddr
));
1061 *entry
= dax_insert_entry(xas
, mapping
, vmf
, *entry
, pfn
,
1062 DAX_ZERO_PAGE
, false);
1064 ret
= vmf_insert_mixed(vmf
->vma
, vaddr
, pfn
);
1065 trace_dax_load_hole(inode
, vmf
, ret
);
1069 #ifdef CONFIG_FS_DAX_PMD
1070 static vm_fault_t
dax_pmd_load_hole(struct xa_state
*xas
, struct vm_fault
*vmf
,
1071 const struct iomap
*iomap
, void **entry
)
1073 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1074 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
1075 struct vm_area_struct
*vma
= vmf
->vma
;
1076 struct inode
*inode
= mapping
->host
;
1077 pgtable_t pgtable
= NULL
;
1078 struct page
*zero_page
;
1083 zero_page
= mm_get_huge_zero_page(vmf
->vma
->vm_mm
);
1085 if (unlikely(!zero_page
))
1088 pfn
= page_to_pfn_t(zero_page
);
1089 *entry
= dax_insert_entry(xas
, mapping
, vmf
, *entry
, pfn
,
1090 DAX_PMD
| DAX_ZERO_PAGE
, false);
1092 if (arch_needs_pgtable_deposit()) {
1093 pgtable
= pte_alloc_one(vma
->vm_mm
);
1095 return VM_FAULT_OOM
;
1098 ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1099 if (!pmd_none(*(vmf
->pmd
))) {
1105 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, pgtable
);
1106 mm_inc_nr_ptes(vma
->vm_mm
);
1108 pmd_entry
= mk_pmd(zero_page
, vmf
->vma
->vm_page_prot
);
1109 pmd_entry
= pmd_mkhuge(pmd_entry
);
1110 set_pmd_at(vmf
->vma
->vm_mm
, pmd_addr
, vmf
->pmd
, pmd_entry
);
1112 trace_dax_pmd_load_hole(inode
, vmf
, zero_page
, *entry
);
1113 return VM_FAULT_NOPAGE
;
1117 pte_free(vma
->vm_mm
, pgtable
);
1118 trace_dax_pmd_load_hole_fallback(inode
, vmf
, zero_page
, *entry
);
1119 return VM_FAULT_FALLBACK
;
1122 static vm_fault_t
dax_pmd_load_hole(struct xa_state
*xas
, struct vm_fault
*vmf
,
1123 const struct iomap
*iomap
, void **entry
)
1125 return VM_FAULT_FALLBACK
;
1127 #endif /* CONFIG_FS_DAX_PMD */
1129 s64
dax_iomap_zero(loff_t pos
, u64 length
, struct iomap
*iomap
)
1131 sector_t sector
= iomap_sector(iomap
, pos
& PAGE_MASK
);
1135 bool page_aligned
= false;
1136 unsigned offset
= offset_in_page(pos
);
1137 unsigned size
= min_t(u64
, PAGE_SIZE
- offset
, length
);
1139 if (IS_ALIGNED(sector
<< SECTOR_SHIFT
, PAGE_SIZE
) &&
1140 (size
== PAGE_SIZE
))
1141 page_aligned
= true;
1143 rc
= bdev_dax_pgoff(iomap
->bdev
, sector
, PAGE_SIZE
, &pgoff
);
1147 id
= dax_read_lock();
1150 rc
= dax_zero_page_range(iomap
->dax_dev
, pgoff
, 1);
1152 rc
= dax_direct_access(iomap
->dax_dev
, pgoff
, 1, &kaddr
, NULL
);
1154 dax_read_unlock(id
);
1158 if (!page_aligned
) {
1159 memset(kaddr
+ offset
, 0, size
);
1160 dax_flush(iomap
->dax_dev
, kaddr
+ offset
, size
);
1162 dax_read_unlock(id
);
1166 static loff_t
dax_iomap_iter(const struct iomap_iter
*iomi
,
1167 struct iov_iter
*iter
)
1169 const struct iomap
*iomap
= &iomi
->iomap
;
1170 loff_t length
= iomap_length(iomi
);
1171 loff_t pos
= iomi
->pos
;
1172 struct block_device
*bdev
= iomap
->bdev
;
1173 struct dax_device
*dax_dev
= iomap
->dax_dev
;
1174 loff_t end
= pos
+ length
, done
= 0;
1179 if (iov_iter_rw(iter
) == READ
) {
1180 end
= min(end
, i_size_read(iomi
->inode
));
1184 if (iomap
->type
== IOMAP_HOLE
|| iomap
->type
== IOMAP_UNWRITTEN
)
1185 return iov_iter_zero(min(length
, end
- pos
), iter
);
1188 if (WARN_ON_ONCE(iomap
->type
!= IOMAP_MAPPED
))
1192 * Write can allocate block for an area which has a hole page mapped
1193 * into page tables. We have to tear down these mappings so that data
1194 * written by write(2) is visible in mmap.
1196 if (iomap
->flags
& IOMAP_F_NEW
) {
1197 invalidate_inode_pages2_range(iomi
->inode
->i_mapping
,
1199 (end
- 1) >> PAGE_SHIFT
);
1202 id
= dax_read_lock();
1204 unsigned offset
= pos
& (PAGE_SIZE
- 1);
1205 const size_t size
= ALIGN(length
+ offset
, PAGE_SIZE
);
1206 const sector_t sector
= dax_iomap_sector(iomap
, pos
);
1211 if (fatal_signal_pending(current
)) {
1216 ret
= bdev_dax_pgoff(bdev
, sector
, size
, &pgoff
);
1220 map_len
= dax_direct_access(dax_dev
, pgoff
, PHYS_PFN(size
),
1227 map_len
= PFN_PHYS(map_len
);
1230 if (map_len
> end
- pos
)
1231 map_len
= end
- pos
;
1234 * The userspace address for the memory copy has already been
1235 * validated via access_ok() in either vfs_read() or
1236 * vfs_write(), depending on which operation we are doing.
1238 if (iov_iter_rw(iter
) == WRITE
)
1239 xfer
= dax_copy_from_iter(dax_dev
, pgoff
, kaddr
,
1242 xfer
= dax_copy_to_iter(dax_dev
, pgoff
, kaddr
,
1254 dax_read_unlock(id
);
1256 return done
? done
: ret
;
1260 * dax_iomap_rw - Perform I/O to a DAX file
1261 * @iocb: The control block for this I/O
1262 * @iter: The addresses to do I/O from or to
1263 * @ops: iomap ops passed from the file system
1265 * This function performs read and write operations to directly mapped
1266 * persistent memory. The callers needs to take care of read/write exclusion
1267 * and evicting any page cache pages in the region under I/O.
1270 dax_iomap_rw(struct kiocb
*iocb
, struct iov_iter
*iter
,
1271 const struct iomap_ops
*ops
)
1273 struct iomap_iter iomi
= {
1274 .inode
= iocb
->ki_filp
->f_mapping
->host
,
1275 .pos
= iocb
->ki_pos
,
1276 .len
= iov_iter_count(iter
),
1281 if (iov_iter_rw(iter
) == WRITE
) {
1282 lockdep_assert_held_write(&iomi
.inode
->i_rwsem
);
1283 iomi
.flags
|= IOMAP_WRITE
;
1285 lockdep_assert_held(&iomi
.inode
->i_rwsem
);
1288 if (iocb
->ki_flags
& IOCB_NOWAIT
)
1289 iomi
.flags
|= IOMAP_NOWAIT
;
1291 while ((ret
= iomap_iter(&iomi
, ops
)) > 0)
1292 iomi
.processed
= dax_iomap_iter(&iomi
, iter
);
1294 done
= iomi
.pos
- iocb
->ki_pos
;
1295 iocb
->ki_pos
= iomi
.pos
;
1296 return done
? done
: ret
;
1298 EXPORT_SYMBOL_GPL(dax_iomap_rw
);
1300 static vm_fault_t
dax_fault_return(int error
)
1303 return VM_FAULT_NOPAGE
;
1304 return vmf_error(error
);
1308 * MAP_SYNC on a dax mapping guarantees dirty metadata is
1309 * flushed on write-faults (non-cow), but not read-faults.
1311 static bool dax_fault_is_synchronous(unsigned long flags
,
1312 struct vm_area_struct
*vma
, const struct iomap
*iomap
)
1314 return (flags
& IOMAP_WRITE
) && (vma
->vm_flags
& VM_SYNC
)
1315 && (iomap
->flags
& IOMAP_F_DIRTY
);
1319 * When handling a synchronous page fault and the inode need a fsync, we can
1320 * insert the PTE/PMD into page tables only after that fsync happened. Skip
1321 * insertion for now and return the pfn so that caller can insert it after the
1324 static vm_fault_t
dax_fault_synchronous_pfnp(pfn_t
*pfnp
, pfn_t pfn
)
1326 if (WARN_ON_ONCE(!pfnp
))
1327 return VM_FAULT_SIGBUS
;
1329 return VM_FAULT_NEEDDSYNC
;
1332 static vm_fault_t
dax_fault_cow_page(struct vm_fault
*vmf
,
1333 const struct iomap_iter
*iter
)
1335 sector_t sector
= dax_iomap_sector(&iter
->iomap
, iter
->pos
);
1336 unsigned long vaddr
= vmf
->address
;
1340 switch (iter
->iomap
.type
) {
1342 case IOMAP_UNWRITTEN
:
1343 clear_user_highpage(vmf
->cow_page
, vaddr
);
1346 error
= copy_cow_page_dax(iter
->iomap
.bdev
, iter
->iomap
.dax_dev
,
1347 sector
, vmf
->cow_page
, vaddr
);
1356 return dax_fault_return(error
);
1358 __SetPageUptodate(vmf
->cow_page
);
1359 ret
= finish_fault(vmf
);
1361 return VM_FAULT_DONE_COW
;
1366 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault.
1367 * @vmf: vm fault instance
1369 * @pfnp: pfn to be returned
1370 * @xas: the dax mapping tree of a file
1371 * @entry: an unlocked dax entry to be inserted
1372 * @pmd: distinguish whether it is a pmd fault
1374 static vm_fault_t
dax_fault_iter(struct vm_fault
*vmf
,
1375 const struct iomap_iter
*iter
, pfn_t
*pfnp
,
1376 struct xa_state
*xas
, void **entry
, bool pmd
)
1378 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1379 const struct iomap
*iomap
= &iter
->iomap
;
1380 size_t size
= pmd
? PMD_SIZE
: PAGE_SIZE
;
1381 loff_t pos
= (loff_t
)xas
->xa_index
<< PAGE_SHIFT
;
1382 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1383 bool sync
= dax_fault_is_synchronous(iter
->flags
, vmf
->vma
, iomap
);
1384 unsigned long entry_flags
= pmd
? DAX_PMD
: 0;
1388 if (!pmd
&& vmf
->cow_page
)
1389 return dax_fault_cow_page(vmf
, iter
);
1391 /* if we are reading UNWRITTEN and HOLE, return a hole. */
1393 (iomap
->type
== IOMAP_UNWRITTEN
|| iomap
->type
== IOMAP_HOLE
)) {
1395 return dax_load_hole(xas
, mapping
, entry
, vmf
);
1396 return dax_pmd_load_hole(xas
, vmf
, iomap
, entry
);
1399 if (iomap
->type
!= IOMAP_MAPPED
) {
1401 return pmd
? VM_FAULT_FALLBACK
: VM_FAULT_SIGBUS
;
1404 err
= dax_iomap_pfn(&iter
->iomap
, pos
, size
, &pfn
);
1406 return pmd
? VM_FAULT_FALLBACK
: dax_fault_return(err
);
1408 *entry
= dax_insert_entry(xas
, mapping
, vmf
, *entry
, pfn
, entry_flags
,
1412 return dax_fault_synchronous_pfnp(pfnp
, pfn
);
1414 /* insert PMD pfn */
1416 return vmf_insert_pfn_pmd(vmf
, pfn
, write
);
1418 /* insert PTE pfn */
1420 return vmf_insert_mixed_mkwrite(vmf
->vma
, vmf
->address
, pfn
);
1421 return vmf_insert_mixed(vmf
->vma
, vmf
->address
, pfn
);
1424 static vm_fault_t
dax_iomap_pte_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1425 int *iomap_errp
, const struct iomap_ops
*ops
)
1427 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1428 XA_STATE(xas
, &mapping
->i_pages
, vmf
->pgoff
);
1429 struct iomap_iter iter
= {
1430 .inode
= mapping
->host
,
1431 .pos
= (loff_t
)vmf
->pgoff
<< PAGE_SHIFT
,
1433 .flags
= IOMAP_FAULT
,
1439 trace_dax_pte_fault(iter
.inode
, vmf
, ret
);
1441 * Check whether offset isn't beyond end of file now. Caller is supposed
1442 * to hold locks serializing us with truncate / punch hole so this is
1445 if (iter
.pos
>= i_size_read(iter
.inode
)) {
1446 ret
= VM_FAULT_SIGBUS
;
1450 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && !vmf
->cow_page
)
1451 iter
.flags
|= IOMAP_WRITE
;
1453 entry
= grab_mapping_entry(&xas
, mapping
, 0);
1454 if (xa_is_internal(entry
)) {
1455 ret
= xa_to_internal(entry
);
1460 * It is possible, particularly with mixed reads & writes to private
1461 * mappings, that we have raced with a PMD fault that overlaps with
1462 * the PTE we need to set up. If so just return and the fault will be
1465 if (pmd_trans_huge(*vmf
->pmd
) || pmd_devmap(*vmf
->pmd
)) {
1466 ret
= VM_FAULT_NOPAGE
;
1470 while ((error
= iomap_iter(&iter
, ops
)) > 0) {
1471 if (WARN_ON_ONCE(iomap_length(&iter
) < PAGE_SIZE
)) {
1472 iter
.processed
= -EIO
; /* fs corruption? */
1476 ret
= dax_fault_iter(vmf
, &iter
, pfnp
, &xas
, &entry
, false);
1477 if (ret
!= VM_FAULT_SIGBUS
&&
1478 (iter
.iomap
.flags
& IOMAP_F_NEW
)) {
1479 count_vm_event(PGMAJFAULT
);
1480 count_memcg_event_mm(vmf
->vma
->vm_mm
, PGMAJFAULT
);
1481 ret
|= VM_FAULT_MAJOR
;
1484 if (!(ret
& VM_FAULT_ERROR
))
1485 iter
.processed
= PAGE_SIZE
;
1489 *iomap_errp
= error
;
1491 ret
= dax_fault_return(error
);
1494 dax_unlock_entry(&xas
, entry
);
1496 trace_dax_pte_fault_done(iter
.inode
, vmf
, ret
);
1500 #ifdef CONFIG_FS_DAX_PMD
1501 static bool dax_fault_check_fallback(struct vm_fault
*vmf
, struct xa_state
*xas
,
1504 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
1505 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1508 * Make sure that the faulting address's PMD offset (color) matches
1509 * the PMD offset from the start of the file. This is necessary so
1510 * that a PMD range in the page table overlaps exactly with a PMD
1511 * range in the page cache.
1513 if ((vmf
->pgoff
& PG_PMD_COLOUR
) !=
1514 ((vmf
->address
>> PAGE_SHIFT
) & PG_PMD_COLOUR
))
1517 /* Fall back to PTEs if we're going to COW */
1518 if (write
&& !(vmf
->vma
->vm_flags
& VM_SHARED
))
1521 /* If the PMD would extend outside the VMA */
1522 if (pmd_addr
< vmf
->vma
->vm_start
)
1524 if ((pmd_addr
+ PMD_SIZE
) > vmf
->vma
->vm_end
)
1527 /* If the PMD would extend beyond the file size */
1528 if ((xas
->xa_index
| PG_PMD_COLOUR
) >= max_pgoff
)
1534 static vm_fault_t
dax_iomap_pmd_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1535 const struct iomap_ops
*ops
)
1537 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1538 XA_STATE_ORDER(xas
, &mapping
->i_pages
, vmf
->pgoff
, PMD_ORDER
);
1539 struct iomap_iter iter
= {
1540 .inode
= mapping
->host
,
1542 .flags
= IOMAP_FAULT
,
1544 vm_fault_t ret
= VM_FAULT_FALLBACK
;
1549 if (vmf
->flags
& FAULT_FLAG_WRITE
)
1550 iter
.flags
|= IOMAP_WRITE
;
1553 * Check whether offset isn't beyond end of file now. Caller is
1554 * supposed to hold locks serializing us with truncate / punch hole so
1555 * this is a reliable test.
1557 max_pgoff
= DIV_ROUND_UP(i_size_read(iter
.inode
), PAGE_SIZE
);
1559 trace_dax_pmd_fault(iter
.inode
, vmf
, max_pgoff
, 0);
1561 if (xas
.xa_index
>= max_pgoff
) {
1562 ret
= VM_FAULT_SIGBUS
;
1566 if (dax_fault_check_fallback(vmf
, &xas
, max_pgoff
))
1570 * grab_mapping_entry() will make sure we get an empty PMD entry,
1571 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1572 * entry is already in the array, for instance), it will return
1573 * VM_FAULT_FALLBACK.
1575 entry
= grab_mapping_entry(&xas
, mapping
, PMD_ORDER
);
1576 if (xa_is_internal(entry
)) {
1577 ret
= xa_to_internal(entry
);
1582 * It is possible, particularly with mixed reads & writes to private
1583 * mappings, that we have raced with a PTE fault that overlaps with
1584 * the PMD we need to set up. If so just return and the fault will be
1587 if (!pmd_none(*vmf
->pmd
) && !pmd_trans_huge(*vmf
->pmd
) &&
1588 !pmd_devmap(*vmf
->pmd
)) {
1593 iter
.pos
= (loff_t
)xas
.xa_index
<< PAGE_SHIFT
;
1594 while ((error
= iomap_iter(&iter
, ops
)) > 0) {
1595 if (iomap_length(&iter
) < PMD_SIZE
)
1596 continue; /* actually breaks out of the loop */
1598 ret
= dax_fault_iter(vmf
, &iter
, pfnp
, &xas
, &entry
, true);
1599 if (ret
!= VM_FAULT_FALLBACK
)
1600 iter
.processed
= PMD_SIZE
;
1604 dax_unlock_entry(&xas
, entry
);
1606 if (ret
== VM_FAULT_FALLBACK
) {
1607 split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
);
1608 count_vm_event(THP_FAULT_FALLBACK
);
1611 trace_dax_pmd_fault_done(iter
.inode
, vmf
, max_pgoff
, ret
);
1615 static vm_fault_t
dax_iomap_pmd_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1616 const struct iomap_ops
*ops
)
1618 return VM_FAULT_FALLBACK
;
1620 #endif /* CONFIG_FS_DAX_PMD */
1623 * dax_iomap_fault - handle a page fault on a DAX file
1624 * @vmf: The description of the fault
1625 * @pe_size: Size of the page to fault in
1626 * @pfnp: PFN to insert for synchronous faults if fsync is required
1627 * @iomap_errp: Storage for detailed error code in case of error
1628 * @ops: Iomap ops passed from the file system
1630 * When a page fault occurs, filesystems may call this helper in
1631 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1632 * has done all the necessary locking for page fault to proceed
1635 vm_fault_t
dax_iomap_fault(struct vm_fault
*vmf
, enum page_entry_size pe_size
,
1636 pfn_t
*pfnp
, int *iomap_errp
, const struct iomap_ops
*ops
)
1640 return dax_iomap_pte_fault(vmf
, pfnp
, iomap_errp
, ops
);
1642 return dax_iomap_pmd_fault(vmf
, pfnp
, ops
);
1644 return VM_FAULT_FALLBACK
;
1647 EXPORT_SYMBOL_GPL(dax_iomap_fault
);
1650 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1651 * @vmf: The description of the fault
1652 * @pfn: PFN to insert
1653 * @order: Order of entry to insert.
1655 * This function inserts a writeable PTE or PMD entry into the page tables
1656 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1659 dax_insert_pfn_mkwrite(struct vm_fault
*vmf
, pfn_t pfn
, unsigned int order
)
1661 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1662 XA_STATE_ORDER(xas
, &mapping
->i_pages
, vmf
->pgoff
, order
);
1667 entry
= get_unlocked_entry(&xas
, order
);
1668 /* Did we race with someone splitting entry or so? */
1669 if (!entry
|| dax_is_conflict(entry
) ||
1670 (order
== 0 && !dax_is_pte_entry(entry
))) {
1671 put_unlocked_entry(&xas
, entry
, WAKE_NEXT
);
1672 xas_unlock_irq(&xas
);
1673 trace_dax_insert_pfn_mkwrite_no_entry(mapping
->host
, vmf
,
1675 return VM_FAULT_NOPAGE
;
1677 xas_set_mark(&xas
, PAGECACHE_TAG_DIRTY
);
1678 dax_lock_entry(&xas
, entry
);
1679 xas_unlock_irq(&xas
);
1681 ret
= vmf_insert_mixed_mkwrite(vmf
->vma
, vmf
->address
, pfn
);
1682 #ifdef CONFIG_FS_DAX_PMD
1683 else if (order
== PMD_ORDER
)
1684 ret
= vmf_insert_pfn_pmd(vmf
, pfn
, FAULT_FLAG_WRITE
);
1687 ret
= VM_FAULT_FALLBACK
;
1688 dax_unlock_entry(&xas
, entry
);
1689 trace_dax_insert_pfn_mkwrite(mapping
->host
, vmf
, ret
);
1694 * dax_finish_sync_fault - finish synchronous page fault
1695 * @vmf: The description of the fault
1696 * @pe_size: Size of entry to be inserted
1697 * @pfn: PFN to insert
1699 * This function ensures that the file range touched by the page fault is
1700 * stored persistently on the media and handles inserting of appropriate page
1703 vm_fault_t
dax_finish_sync_fault(struct vm_fault
*vmf
,
1704 enum page_entry_size pe_size
, pfn_t pfn
)
1707 loff_t start
= ((loff_t
)vmf
->pgoff
) << PAGE_SHIFT
;
1708 unsigned int order
= pe_order(pe_size
);
1709 size_t len
= PAGE_SIZE
<< order
;
1711 err
= vfs_fsync_range(vmf
->vma
->vm_file
, start
, start
+ len
- 1, 1);
1713 return VM_FAULT_SIGBUS
;
1714 return dax_insert_pfn_mkwrite(vmf
, pfn
, order
);
1716 EXPORT_SYMBOL_GPL(dax_finish_sync_fault
);