2 * fs/dax.c - Direct Access filesystem code
3 * Copyright (c) 2013-2014 Intel Corporation
4 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
5 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms and conditions of the GNU General Public License,
9 * version 2, as published by the Free Software Foundation.
11 * This program is distributed in the hope it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
17 #include <linux/atomic.h>
18 #include <linux/blkdev.h>
19 #include <linux/buffer_head.h>
20 #include <linux/dax.h>
22 #include <linux/genhd.h>
23 #include <linux/highmem.h>
24 #include <linux/memcontrol.h>
26 #include <linux/mutex.h>
27 #include <linux/pagevec.h>
28 #include <linux/sched.h>
29 #include <linux/sched/signal.h>
30 #include <linux/uio.h>
31 #include <linux/vmstat.h>
32 #include <linux/pfn_t.h>
33 #include <linux/sizes.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/iomap.h>
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/fs_dax.h>
41 static inline unsigned int pe_order(enum page_entry_size pe_size
)
43 if (pe_size
== PE_SIZE_PTE
)
44 return PAGE_SHIFT
- PAGE_SHIFT
;
45 if (pe_size
== PE_SIZE_PMD
)
46 return PMD_SHIFT
- PAGE_SHIFT
;
47 if (pe_size
== PE_SIZE_PUD
)
48 return PUD_SHIFT
- PAGE_SHIFT
;
52 /* We choose 4096 entries - same as per-zone page wait tables */
53 #define DAX_WAIT_TABLE_BITS 12
54 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
56 /* The 'colour' (ie low bits) within a PMD of a page offset. */
57 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
58 #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
60 /* The order of a PMD entry */
61 #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT)
63 static wait_queue_head_t wait_table
[DAX_WAIT_TABLE_ENTRIES
];
65 static int __init
init_dax_wait_table(void)
69 for (i
= 0; i
< DAX_WAIT_TABLE_ENTRIES
; i
++)
70 init_waitqueue_head(wait_table
+ i
);
73 fs_initcall(init_dax_wait_table
);
76 * DAX pagecache entries use XArray value entries so they can't be mistaken
77 * for pages. We use one bit for locking, one bit for the entry size (PMD)
78 * and two more to tell us if the entry is a zero page or an empty entry that
79 * is just used for locking. In total four special bits.
81 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
82 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
86 #define DAX_LOCKED (1UL << 0)
87 #define DAX_PMD (1UL << 1)
88 #define DAX_ZERO_PAGE (1UL << 2)
89 #define DAX_EMPTY (1UL << 3)
91 static unsigned long dax_to_pfn(void *entry
)
93 return xa_to_value(entry
) >> DAX_SHIFT
;
96 static void *dax_make_entry(pfn_t pfn
, unsigned long flags
)
98 return xa_mk_value(flags
| (pfn_t_to_pfn(pfn
) << DAX_SHIFT
));
101 static bool dax_is_locked(void *entry
)
103 return xa_to_value(entry
) & DAX_LOCKED
;
106 static unsigned int dax_entry_order(void *entry
)
108 if (xa_to_value(entry
) & DAX_PMD
)
113 static unsigned long dax_is_pmd_entry(void *entry
)
115 return xa_to_value(entry
) & DAX_PMD
;
118 static bool dax_is_pte_entry(void *entry
)
120 return !(xa_to_value(entry
) & DAX_PMD
);
123 static int dax_is_zero_entry(void *entry
)
125 return xa_to_value(entry
) & DAX_ZERO_PAGE
;
128 static int dax_is_empty_entry(void *entry
)
130 return xa_to_value(entry
) & DAX_EMPTY
;
134 * DAX page cache entry locking
136 struct exceptional_entry_key
{
141 struct wait_exceptional_entry_queue
{
142 wait_queue_entry_t wait
;
143 struct exceptional_entry_key key
;
146 static wait_queue_head_t
*dax_entry_waitqueue(struct xa_state
*xas
,
147 void *entry
, struct exceptional_entry_key
*key
)
150 unsigned long index
= xas
->xa_index
;
153 * If 'entry' is a PMD, align the 'index' that we use for the wait
154 * queue to the start of that PMD. This ensures that all offsets in
155 * the range covered by the PMD map to the same bit lock.
157 if (dax_is_pmd_entry(entry
))
158 index
&= ~PG_PMD_COLOUR
;
160 key
->entry_start
= index
;
162 hash
= hash_long((unsigned long)xas
->xa
^ index
, DAX_WAIT_TABLE_BITS
);
163 return wait_table
+ hash
;
166 static int wake_exceptional_entry_func(wait_queue_entry_t
*wait
,
167 unsigned int mode
, int sync
, void *keyp
)
169 struct exceptional_entry_key
*key
= keyp
;
170 struct wait_exceptional_entry_queue
*ewait
=
171 container_of(wait
, struct wait_exceptional_entry_queue
, wait
);
173 if (key
->xa
!= ewait
->key
.xa
||
174 key
->entry_start
!= ewait
->key
.entry_start
)
176 return autoremove_wake_function(wait
, mode
, sync
, NULL
);
180 * @entry may no longer be the entry at the index in the mapping.
181 * The important information it's conveying is whether the entry at
182 * this index used to be a PMD entry.
184 static void dax_wake_entry(struct xa_state
*xas
, void *entry
, bool wake_all
)
186 struct exceptional_entry_key key
;
187 wait_queue_head_t
*wq
;
189 wq
= dax_entry_waitqueue(xas
, entry
, &key
);
192 * Checking for locked entry and prepare_to_wait_exclusive() happens
193 * under the i_pages lock, ditto for entry handling in our callers.
194 * So at this point all tasks that could have seen our entry locked
195 * must be in the waitqueue and the following check will see them.
197 if (waitqueue_active(wq
))
198 __wake_up(wq
, TASK_NORMAL
, wake_all
? 0 : 1, &key
);
202 * Look up entry in page cache, wait for it to become unlocked if it
203 * is a DAX entry and return it. The caller must subsequently call
204 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
207 * Must be called with the i_pages lock held.
209 static void *get_unlocked_entry(struct xa_state
*xas
)
212 struct wait_exceptional_entry_queue ewait
;
213 wait_queue_head_t
*wq
;
215 init_wait(&ewait
.wait
);
216 ewait
.wait
.func
= wake_exceptional_entry_func
;
219 entry
= xas_find_conflict(xas
);
220 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)) ||
221 !dax_is_locked(entry
))
224 wq
= dax_entry_waitqueue(xas
, entry
, &ewait
.key
);
225 prepare_to_wait_exclusive(wq
, &ewait
.wait
,
226 TASK_UNINTERRUPTIBLE
);
230 finish_wait(wq
, &ewait
.wait
);
235 static void put_unlocked_entry(struct xa_state
*xas
, void *entry
)
237 /* If we were the only waiter woken, wake the next one */
239 dax_wake_entry(xas
, entry
, false);
243 * We used the xa_state to get the entry, but then we locked the entry and
244 * dropped the xa_lock, so we know the xa_state is stale and must be reset
247 static void dax_unlock_entry(struct xa_state
*xas
, void *entry
)
251 BUG_ON(dax_is_locked(entry
));
254 old
= xas_store(xas
, entry
);
256 BUG_ON(!dax_is_locked(old
));
257 dax_wake_entry(xas
, entry
, false);
261 * Return: The entry stored at this location before it was locked.
263 static void *dax_lock_entry(struct xa_state
*xas
, void *entry
)
265 unsigned long v
= xa_to_value(entry
);
266 return xas_store(xas
, xa_mk_value(v
| DAX_LOCKED
));
269 static unsigned long dax_entry_size(void *entry
)
271 if (dax_is_zero_entry(entry
))
273 else if (dax_is_empty_entry(entry
))
275 else if (dax_is_pmd_entry(entry
))
281 static unsigned long dax_end_pfn(void *entry
)
283 return dax_to_pfn(entry
) + dax_entry_size(entry
) / PAGE_SIZE
;
287 * Iterate through all mapped pfns represented by an entry, i.e. skip
288 * 'empty' and 'zero' entries.
290 #define for_each_mapped_pfn(entry, pfn) \
291 for (pfn = dax_to_pfn(entry); \
292 pfn < dax_end_pfn(entry); pfn++)
295 * TODO: for reflink+dax we need a way to associate a single page with
296 * multiple address_space instances at different linear_page_index()
299 static void dax_associate_entry(void *entry
, struct address_space
*mapping
,
300 struct vm_area_struct
*vma
, unsigned long address
)
302 unsigned long size
= dax_entry_size(entry
), pfn
, index
;
305 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
308 index
= linear_page_index(vma
, address
& ~(size
- 1));
309 for_each_mapped_pfn(entry
, pfn
) {
310 struct page
*page
= pfn_to_page(pfn
);
312 WARN_ON_ONCE(page
->mapping
);
313 page
->mapping
= mapping
;
314 page
->index
= index
+ i
++;
318 static void dax_disassociate_entry(void *entry
, struct address_space
*mapping
,
323 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
326 for_each_mapped_pfn(entry
, pfn
) {
327 struct page
*page
= pfn_to_page(pfn
);
329 WARN_ON_ONCE(trunc
&& page_ref_count(page
) > 1);
330 WARN_ON_ONCE(page
->mapping
&& page
->mapping
!= mapping
);
331 page
->mapping
= NULL
;
336 static struct page
*dax_busy_page(void *entry
)
340 for_each_mapped_pfn(entry
, pfn
) {
341 struct page
*page
= pfn_to_page(pfn
);
343 if (page_ref_count(page
) > 1)
350 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page
351 * @page: The page whose entry we want to lock
353 * Context: Process context.
354 * Return: %true if the entry was locked or does not need to be locked.
356 bool dax_lock_mapping_entry(struct page
*page
)
358 XA_STATE(xas
, NULL
, 0);
362 /* Ensure page->mapping isn't freed while we look at it */
365 struct address_space
*mapping
= READ_ONCE(page
->mapping
);
368 if (!dax_mapping(mapping
))
372 * In the device-dax case there's no need to lock, a
373 * struct dev_pagemap pin is sufficient to keep the
374 * inode alive, and we assume we have dev_pagemap pin
375 * otherwise we would not have a valid pfn_to_page()
379 if (S_ISCHR(mapping
->host
->i_mode
))
382 xas
.xa
= &mapping
->i_pages
;
384 if (mapping
!= page
->mapping
) {
385 xas_unlock_irq(&xas
);
388 xas_set(&xas
, page
->index
);
389 entry
= xas_load(&xas
);
390 if (dax_is_locked(entry
)) {
392 entry
= get_unlocked_entry(&xas
);
393 xas_unlock_irq(&xas
);
394 put_unlocked_entry(&xas
, entry
);
398 dax_lock_entry(&xas
, entry
);
399 xas_unlock_irq(&xas
);
406 void dax_unlock_mapping_entry(struct page
*page
)
408 struct address_space
*mapping
= page
->mapping
;
409 XA_STATE(xas
, &mapping
->i_pages
, page
->index
);
412 if (S_ISCHR(mapping
->host
->i_mode
))
416 entry
= xas_load(&xas
);
418 entry
= dax_make_entry(page_to_pfn_t(page
), dax_is_pmd_entry(entry
));
419 dax_unlock_entry(&xas
, entry
);
423 * Find page cache entry at given index. If it is a DAX entry, return it
424 * with the entry locked. If the page cache doesn't contain an entry at
425 * that index, add a locked empty entry.
427 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
428 * either return that locked entry or will return VM_FAULT_FALLBACK.
429 * This will happen if there are any PTE entries within the PMD range
430 * that we are requesting.
432 * We always favor PTE entries over PMD entries. There isn't a flow where we
433 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
434 * insertion will fail if it finds any PTE entries already in the tree, and a
435 * PTE insertion will cause an existing PMD entry to be unmapped and
436 * downgraded to PTE entries. This happens for both PMD zero pages as
437 * well as PMD empty entries.
439 * The exception to this downgrade path is for PMD entries that have
440 * real storage backing them. We will leave these real PMD entries in
441 * the tree, and PTE writes will simply dirty the entire PMD entry.
443 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
444 * persistent memory the benefit is doubtful. We can add that later if we can
447 * On error, this function does not return an ERR_PTR. Instead it returns
448 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
449 * overlap with xarray value entries.
451 static void *grab_mapping_entry(struct xa_state
*xas
,
452 struct address_space
*mapping
, unsigned long size_flag
)
454 unsigned long index
= xas
->xa_index
;
455 bool pmd_downgrade
= false; /* splitting PMD entry into PTE entries? */
460 entry
= get_unlocked_entry(xas
);
463 if (!xa_is_value(entry
)) {
464 xas_set_err(xas
, EIO
);
468 if (size_flag
& DAX_PMD
) {
469 if (dax_is_pte_entry(entry
)) {
470 put_unlocked_entry(xas
, entry
);
473 } else { /* trying to grab a PTE entry */
474 if (dax_is_pmd_entry(entry
) &&
475 (dax_is_zero_entry(entry
) ||
476 dax_is_empty_entry(entry
))) {
477 pmd_downgrade
= true;
484 * Make sure 'entry' remains valid while we drop
487 dax_lock_entry(xas
, entry
);
490 * Besides huge zero pages the only other thing that gets
491 * downgraded are empty entries which don't need to be
494 if (dax_is_zero_entry(entry
)) {
496 unmap_mapping_pages(mapping
,
497 xas
->xa_index
& ~PG_PMD_COLOUR
,
503 dax_disassociate_entry(entry
, mapping
, false);
504 xas_store(xas
, NULL
); /* undo the PMD join */
505 dax_wake_entry(xas
, entry
, true);
506 mapping
->nrexceptional
--;
512 dax_lock_entry(xas
, entry
);
514 entry
= dax_make_entry(pfn_to_pfn_t(0), size_flag
| DAX_EMPTY
);
515 dax_lock_entry(xas
, entry
);
518 mapping
->nrexceptional
++;
523 if (xas_nomem(xas
, mapping_gfp_mask(mapping
) & ~__GFP_HIGHMEM
))
525 if (xas
->xa_node
== XA_ERROR(-ENOMEM
))
526 return xa_mk_internal(VM_FAULT_OOM
);
528 return xa_mk_internal(VM_FAULT_SIGBUS
);
532 return xa_mk_internal(VM_FAULT_FALLBACK
);
536 * dax_layout_busy_page - find first pinned page in @mapping
537 * @mapping: address space to scan for a page with ref count > 1
539 * DAX requires ZONE_DEVICE mapped pages. These pages are never
540 * 'onlined' to the page allocator so they are considered idle when
541 * page->count == 1. A filesystem uses this interface to determine if
542 * any page in the mapping is busy, i.e. for DMA, or other
543 * get_user_pages() usages.
545 * It is expected that the filesystem is holding locks to block the
546 * establishment of new mappings in this address_space. I.e. it expects
547 * to be able to run unmap_mapping_range() and subsequently not race
548 * mapping_mapped() becoming true.
550 struct page
*dax_layout_busy_page(struct address_space
*mapping
)
552 XA_STATE(xas
, &mapping
->i_pages
, 0);
554 unsigned int scanned
= 0;
555 struct page
*page
= NULL
;
558 * In the 'limited' case get_user_pages() for dax is disabled.
560 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED
))
563 if (!dax_mapping(mapping
) || !mapping_mapped(mapping
))
567 * If we race get_user_pages_fast() here either we'll see the
568 * elevated page count in the iteration and wait, or
569 * get_user_pages_fast() will see that the page it took a reference
570 * against is no longer mapped in the page tables and bail to the
571 * get_user_pages() slow path. The slow path is protected by
572 * pte_lock() and pmd_lock(). New references are not taken without
573 * holding those locks, and unmap_mapping_range() will not zero the
574 * pte or pmd without holding the respective lock, so we are
575 * guaranteed to either see new references or prevent new
576 * references from being established.
578 unmap_mapping_range(mapping
, 0, 0, 1);
581 xas_for_each(&xas
, entry
, ULONG_MAX
) {
582 if (WARN_ON_ONCE(!xa_is_value(entry
)))
584 if (unlikely(dax_is_locked(entry
)))
585 entry
= get_unlocked_entry(&xas
);
587 page
= dax_busy_page(entry
);
588 put_unlocked_entry(&xas
, entry
);
591 if (++scanned
% XA_CHECK_SCHED
)
595 xas_unlock_irq(&xas
);
599 xas_unlock_irq(&xas
);
602 EXPORT_SYMBOL_GPL(dax_layout_busy_page
);
604 static int __dax_invalidate_entry(struct address_space
*mapping
,
605 pgoff_t index
, bool trunc
)
607 XA_STATE(xas
, &mapping
->i_pages
, index
);
612 entry
= get_unlocked_entry(&xas
);
613 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
616 (xas_get_mark(&xas
, PAGECACHE_TAG_DIRTY
) ||
617 xas_get_mark(&xas
, PAGECACHE_TAG_TOWRITE
)))
619 dax_disassociate_entry(entry
, mapping
, trunc
);
620 xas_store(&xas
, NULL
);
621 mapping
->nrexceptional
--;
624 put_unlocked_entry(&xas
, entry
);
625 xas_unlock_irq(&xas
);
630 * Delete DAX entry at @index from @mapping. Wait for it
631 * to be unlocked before deleting it.
633 int dax_delete_mapping_entry(struct address_space
*mapping
, pgoff_t index
)
635 int ret
= __dax_invalidate_entry(mapping
, index
, true);
638 * This gets called from truncate / punch_hole path. As such, the caller
639 * must hold locks protecting against concurrent modifications of the
640 * page cache (usually fs-private i_mmap_sem for writing). Since the
641 * caller has seen a DAX entry for this index, we better find it
642 * at that index as well...
649 * Invalidate DAX entry if it is clean.
651 int dax_invalidate_mapping_entry_sync(struct address_space
*mapping
,
654 return __dax_invalidate_entry(mapping
, index
, false);
657 static int copy_user_dax(struct block_device
*bdev
, struct dax_device
*dax_dev
,
658 sector_t sector
, size_t size
, struct page
*to
,
666 rc
= bdev_dax_pgoff(bdev
, sector
, size
, &pgoff
);
670 id
= dax_read_lock();
671 rc
= dax_direct_access(dax_dev
, pgoff
, PHYS_PFN(size
), &kaddr
, NULL
);
676 vto
= kmap_atomic(to
);
677 copy_user_page(vto
, (void __force
*)kaddr
, vaddr
, to
);
684 * By this point grab_mapping_entry() has ensured that we have a locked entry
685 * of the appropriate size so we don't have to worry about downgrading PMDs to
686 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
687 * already in the tree, we will skip the insertion and just dirty the PMD as
690 static void *dax_insert_entry(struct xa_state
*xas
,
691 struct address_space
*mapping
, struct vm_fault
*vmf
,
692 void *entry
, pfn_t pfn
, unsigned long flags
, bool dirty
)
694 void *new_entry
= dax_make_entry(pfn
, flags
);
697 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
699 if (dax_is_zero_entry(entry
) && !(flags
& DAX_ZERO_PAGE
)) {
700 unsigned long index
= xas
->xa_index
;
701 /* we are replacing a zero page with block mapping */
702 if (dax_is_pmd_entry(entry
))
703 unmap_mapping_pages(mapping
, index
& ~PG_PMD_COLOUR
,
706 unmap_mapping_pages(mapping
, index
, 1, false);
711 if (dax_entry_size(entry
) != dax_entry_size(new_entry
)) {
712 dax_disassociate_entry(entry
, mapping
, false);
713 dax_associate_entry(new_entry
, mapping
, vmf
->vma
, vmf
->address
);
716 if (dax_is_zero_entry(entry
) || dax_is_empty_entry(entry
)) {
718 * Only swap our new entry into the page cache if the current
719 * entry is a zero page or an empty entry. If a normal PTE or
720 * PMD entry is already in the cache, we leave it alone. This
721 * means that if we are trying to insert a PTE and the
722 * existing entry is a PMD, we will just leave the PMD in the
723 * tree and dirty it if necessary.
725 void *old
= dax_lock_entry(xas
, new_entry
);
726 WARN_ON_ONCE(old
!= xa_mk_value(xa_to_value(entry
) |
730 xas_load(xas
); /* Walk the xa_state */
734 xas_set_mark(xas
, PAGECACHE_TAG_DIRTY
);
741 unsigned long pgoff_address(pgoff_t pgoff
, struct vm_area_struct
*vma
)
743 unsigned long address
;
745 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
746 VM_BUG_ON_VMA(address
< vma
->vm_start
|| address
>= vma
->vm_end
, vma
);
750 /* Walk all mappings of a given index of a file and writeprotect them */
751 static void dax_entry_mkclean(struct address_space
*mapping
, pgoff_t index
,
754 struct vm_area_struct
*vma
;
755 pte_t pte
, *ptep
= NULL
;
759 i_mmap_lock_read(mapping
);
760 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, index
, index
) {
761 unsigned long address
, start
, end
;
765 if (!(vma
->vm_flags
& VM_SHARED
))
768 address
= pgoff_address(index
, vma
);
771 * Note because we provide start/end to follow_pte_pmd it will
772 * call mmu_notifier_invalidate_range_start() on our behalf
773 * before taking any lock.
775 if (follow_pte_pmd(vma
->vm_mm
, address
, &start
, &end
, &ptep
, &pmdp
, &ptl
))
779 * No need to call mmu_notifier_invalidate_range() as we are
780 * downgrading page table protection not changing it to point
783 * See Documentation/vm/mmu_notifier.rst
786 #ifdef CONFIG_FS_DAX_PMD
789 if (pfn
!= pmd_pfn(*pmdp
))
791 if (!pmd_dirty(*pmdp
) && !pmd_write(*pmdp
))
794 flush_cache_page(vma
, address
, pfn
);
795 pmd
= pmdp_huge_clear_flush(vma
, address
, pmdp
);
796 pmd
= pmd_wrprotect(pmd
);
797 pmd
= pmd_mkclean(pmd
);
798 set_pmd_at(vma
->vm_mm
, address
, pmdp
, pmd
);
803 if (pfn
!= pte_pfn(*ptep
))
805 if (!pte_dirty(*ptep
) && !pte_write(*ptep
))
808 flush_cache_page(vma
, address
, pfn
);
809 pte
= ptep_clear_flush(vma
, address
, ptep
);
810 pte
= pte_wrprotect(pte
);
811 pte
= pte_mkclean(pte
);
812 set_pte_at(vma
->vm_mm
, address
, ptep
, pte
);
814 pte_unmap_unlock(ptep
, ptl
);
817 mmu_notifier_invalidate_range_end(vma
->vm_mm
, start
, end
);
819 i_mmap_unlock_read(mapping
);
822 static int dax_writeback_one(struct xa_state
*xas
, struct dax_device
*dax_dev
,
823 struct address_space
*mapping
, void *entry
)
830 * A page got tagged dirty in DAX mapping? Something is seriously
833 if (WARN_ON(!xa_is_value(entry
)))
836 if (unlikely(dax_is_locked(entry
))) {
837 void *old_entry
= entry
;
839 entry
= get_unlocked_entry(xas
);
841 /* Entry got punched out / reallocated? */
842 if (!entry
|| WARN_ON_ONCE(!xa_is_value(entry
)))
845 * Entry got reallocated elsewhere? No need to writeback.
846 * We have to compare pfns as we must not bail out due to
847 * difference in lockbit or entry type.
849 if (dax_to_pfn(old_entry
) != dax_to_pfn(entry
))
851 if (WARN_ON_ONCE(dax_is_empty_entry(entry
) ||
852 dax_is_zero_entry(entry
))) {
857 /* Another fsync thread may have already done this entry */
858 if (!xas_get_mark(xas
, PAGECACHE_TAG_TOWRITE
))
862 /* Lock the entry to serialize with page faults */
863 dax_lock_entry(xas
, entry
);
866 * We can clear the tag now but we have to be careful so that concurrent
867 * dax_writeback_one() calls for the same index cannot finish before we
868 * actually flush the caches. This is achieved as the calls will look
869 * at the entry only under the i_pages lock and once they do that
870 * they will see the entry locked and wait for it to unlock.
872 xas_clear_mark(xas
, PAGECACHE_TAG_TOWRITE
);
876 * Even if dax_writeback_mapping_range() was given a wbc->range_start
877 * in the middle of a PMD, the 'index' we are given will be aligned to
878 * the start index of the PMD, as will the pfn we pull from 'entry'.
879 * This allows us to flush for PMD_SIZE and not have to worry about
880 * partial PMD writebacks.
882 pfn
= dax_to_pfn(entry
);
883 size
= PAGE_SIZE
<< dax_entry_order(entry
);
885 dax_entry_mkclean(mapping
, xas
->xa_index
, pfn
);
886 dax_flush(dax_dev
, page_address(pfn_to_page(pfn
)), size
);
888 * After we have flushed the cache, we can clear the dirty tag. There
889 * cannot be new dirty data in the pfn after the flush has completed as
890 * the pfn mappings are writeprotected and fault waits for mapping
895 xas_store(xas
, entry
);
896 xas_clear_mark(xas
, PAGECACHE_TAG_DIRTY
);
897 dax_wake_entry(xas
, entry
, false);
899 trace_dax_writeback_one(mapping
->host
, xas
->xa_index
,
904 put_unlocked_entry(xas
, entry
);
909 * Flush the mapping to the persistent domain within the byte range of [start,
910 * end]. This is required by data integrity operations to ensure file data is
911 * on persistent storage prior to completion of the operation.
913 int dax_writeback_mapping_range(struct address_space
*mapping
,
914 struct block_device
*bdev
, struct writeback_control
*wbc
)
916 XA_STATE(xas
, &mapping
->i_pages
, wbc
->range_start
>> PAGE_SHIFT
);
917 struct inode
*inode
= mapping
->host
;
918 pgoff_t end_index
= wbc
->range_end
>> PAGE_SHIFT
;
919 struct dax_device
*dax_dev
;
922 unsigned int scanned
= 0;
924 if (WARN_ON_ONCE(inode
->i_blkbits
!= PAGE_SHIFT
))
927 if (!mapping
->nrexceptional
|| wbc
->sync_mode
!= WB_SYNC_ALL
)
930 dax_dev
= dax_get_by_host(bdev
->bd_disk
->disk_name
);
934 trace_dax_writeback_range(inode
, xas
.xa_index
, end_index
);
936 tag_pages_for_writeback(mapping
, xas
.xa_index
, end_index
);
939 xas_for_each_marked(&xas
, entry
, end_index
, PAGECACHE_TAG_TOWRITE
) {
940 ret
= dax_writeback_one(&xas
, dax_dev
, mapping
, entry
);
942 mapping_set_error(mapping
, ret
);
945 if (++scanned
% XA_CHECK_SCHED
)
949 xas_unlock_irq(&xas
);
953 xas_unlock_irq(&xas
);
955 trace_dax_writeback_range_done(inode
, xas
.xa_index
, end_index
);
958 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range
);
960 static sector_t
dax_iomap_sector(struct iomap
*iomap
, loff_t pos
)
962 return (iomap
->addr
+ (pos
& PAGE_MASK
) - iomap
->offset
) >> 9;
965 static int dax_iomap_pfn(struct iomap
*iomap
, loff_t pos
, size_t size
,
968 const sector_t sector
= dax_iomap_sector(iomap
, pos
);
973 rc
= bdev_dax_pgoff(iomap
->bdev
, sector
, size
, &pgoff
);
976 id
= dax_read_lock();
977 length
= dax_direct_access(iomap
->dax_dev
, pgoff
, PHYS_PFN(size
),
984 if (PFN_PHYS(length
) < size
)
986 if (pfn_t_to_pfn(*pfnp
) & (PHYS_PFN(size
)-1))
988 /* For larger pages we need devmap */
989 if (length
> 1 && !pfn_t_devmap(*pfnp
))
998 * The user has performed a load from a hole in the file. Allocating a new
999 * page in the file would cause excessive storage usage for workloads with
1000 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1001 * If this page is ever written to we will re-fault and change the mapping to
1002 * point to real DAX storage instead.
1004 static vm_fault_t
dax_load_hole(struct xa_state
*xas
,
1005 struct address_space
*mapping
, void **entry
,
1006 struct vm_fault
*vmf
)
1008 struct inode
*inode
= mapping
->host
;
1009 unsigned long vaddr
= vmf
->address
;
1010 pfn_t pfn
= pfn_to_pfn_t(my_zero_pfn(vaddr
));
1013 *entry
= dax_insert_entry(xas
, mapping
, vmf
, *entry
, pfn
,
1014 DAX_ZERO_PAGE
, false);
1016 ret
= vmf_insert_mixed(vmf
->vma
, vaddr
, pfn
);
1017 trace_dax_load_hole(inode
, vmf
, ret
);
1021 static bool dax_range_is_aligned(struct block_device
*bdev
,
1022 unsigned int offset
, unsigned int length
)
1024 unsigned short sector_size
= bdev_logical_block_size(bdev
);
1026 if (!IS_ALIGNED(offset
, sector_size
))
1028 if (!IS_ALIGNED(length
, sector_size
))
1034 int __dax_zero_page_range(struct block_device
*bdev
,
1035 struct dax_device
*dax_dev
, sector_t sector
,
1036 unsigned int offset
, unsigned int size
)
1038 if (dax_range_is_aligned(bdev
, offset
, size
)) {
1039 sector_t start_sector
= sector
+ (offset
>> 9);
1041 return blkdev_issue_zeroout(bdev
, start_sector
,
1042 size
>> 9, GFP_NOFS
, 0);
1048 rc
= bdev_dax_pgoff(bdev
, sector
, PAGE_SIZE
, &pgoff
);
1052 id
= dax_read_lock();
1053 rc
= dax_direct_access(dax_dev
, pgoff
, 1, &kaddr
, NULL
);
1055 dax_read_unlock(id
);
1058 memset(kaddr
+ offset
, 0, size
);
1059 dax_flush(dax_dev
, kaddr
+ offset
, size
);
1060 dax_read_unlock(id
);
1064 EXPORT_SYMBOL_GPL(__dax_zero_page_range
);
1067 dax_iomap_actor(struct inode
*inode
, loff_t pos
, loff_t length
, void *data
,
1068 struct iomap
*iomap
)
1070 struct block_device
*bdev
= iomap
->bdev
;
1071 struct dax_device
*dax_dev
= iomap
->dax_dev
;
1072 struct iov_iter
*iter
= data
;
1073 loff_t end
= pos
+ length
, done
= 0;
1078 if (iov_iter_rw(iter
) == READ
) {
1079 end
= min(end
, i_size_read(inode
));
1083 if (iomap
->type
== IOMAP_HOLE
|| iomap
->type
== IOMAP_UNWRITTEN
)
1084 return iov_iter_zero(min(length
, end
- pos
), iter
);
1087 if (WARN_ON_ONCE(iomap
->type
!= IOMAP_MAPPED
))
1091 * Write can allocate block for an area which has a hole page mapped
1092 * into page tables. We have to tear down these mappings so that data
1093 * written by write(2) is visible in mmap.
1095 if (iomap
->flags
& IOMAP_F_NEW
) {
1096 invalidate_inode_pages2_range(inode
->i_mapping
,
1098 (end
- 1) >> PAGE_SHIFT
);
1101 id
= dax_read_lock();
1103 unsigned offset
= pos
& (PAGE_SIZE
- 1);
1104 const size_t size
= ALIGN(length
+ offset
, PAGE_SIZE
);
1105 const sector_t sector
= dax_iomap_sector(iomap
, pos
);
1110 if (fatal_signal_pending(current
)) {
1115 ret
= bdev_dax_pgoff(bdev
, sector
, size
, &pgoff
);
1119 map_len
= dax_direct_access(dax_dev
, pgoff
, PHYS_PFN(size
),
1126 map_len
= PFN_PHYS(map_len
);
1129 if (map_len
> end
- pos
)
1130 map_len
= end
- pos
;
1133 * The userspace address for the memory copy has already been
1134 * validated via access_ok() in either vfs_read() or
1135 * vfs_write(), depending on which operation we are doing.
1137 if (iov_iter_rw(iter
) == WRITE
)
1138 xfer
= dax_copy_from_iter(dax_dev
, pgoff
, kaddr
,
1141 xfer
= dax_copy_to_iter(dax_dev
, pgoff
, kaddr
,
1153 dax_read_unlock(id
);
1155 return done
? done
: ret
;
1159 * dax_iomap_rw - Perform I/O to a DAX file
1160 * @iocb: The control block for this I/O
1161 * @iter: The addresses to do I/O from or to
1162 * @ops: iomap ops passed from the file system
1164 * This function performs read and write operations to directly mapped
1165 * persistent memory. The callers needs to take care of read/write exclusion
1166 * and evicting any page cache pages in the region under I/O.
1169 dax_iomap_rw(struct kiocb
*iocb
, struct iov_iter
*iter
,
1170 const struct iomap_ops
*ops
)
1172 struct address_space
*mapping
= iocb
->ki_filp
->f_mapping
;
1173 struct inode
*inode
= mapping
->host
;
1174 loff_t pos
= iocb
->ki_pos
, ret
= 0, done
= 0;
1177 if (iov_iter_rw(iter
) == WRITE
) {
1178 lockdep_assert_held_exclusive(&inode
->i_rwsem
);
1179 flags
|= IOMAP_WRITE
;
1181 lockdep_assert_held(&inode
->i_rwsem
);
1184 while (iov_iter_count(iter
)) {
1185 ret
= iomap_apply(inode
, pos
, iov_iter_count(iter
), flags
, ops
,
1186 iter
, dax_iomap_actor
);
1193 iocb
->ki_pos
+= done
;
1194 return done
? done
: ret
;
1196 EXPORT_SYMBOL_GPL(dax_iomap_rw
);
1198 static vm_fault_t
dax_fault_return(int error
)
1201 return VM_FAULT_NOPAGE
;
1202 if (error
== -ENOMEM
)
1203 return VM_FAULT_OOM
;
1204 return VM_FAULT_SIGBUS
;
1208 * MAP_SYNC on a dax mapping guarantees dirty metadata is
1209 * flushed on write-faults (non-cow), but not read-faults.
1211 static bool dax_fault_is_synchronous(unsigned long flags
,
1212 struct vm_area_struct
*vma
, struct iomap
*iomap
)
1214 return (flags
& IOMAP_WRITE
) && (vma
->vm_flags
& VM_SYNC
)
1215 && (iomap
->flags
& IOMAP_F_DIRTY
);
1218 static vm_fault_t
dax_iomap_pte_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1219 int *iomap_errp
, const struct iomap_ops
*ops
)
1221 struct vm_area_struct
*vma
= vmf
->vma
;
1222 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1223 XA_STATE(xas
, &mapping
->i_pages
, vmf
->pgoff
);
1224 struct inode
*inode
= mapping
->host
;
1225 unsigned long vaddr
= vmf
->address
;
1226 loff_t pos
= (loff_t
)vmf
->pgoff
<< PAGE_SHIFT
;
1227 struct iomap iomap
= { 0 };
1228 unsigned flags
= IOMAP_FAULT
;
1229 int error
, major
= 0;
1230 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1236 trace_dax_pte_fault(inode
, vmf
, ret
);
1238 * Check whether offset isn't beyond end of file now. Caller is supposed
1239 * to hold locks serializing us with truncate / punch hole so this is
1242 if (pos
>= i_size_read(inode
)) {
1243 ret
= VM_FAULT_SIGBUS
;
1247 if (write
&& !vmf
->cow_page
)
1248 flags
|= IOMAP_WRITE
;
1250 entry
= grab_mapping_entry(&xas
, mapping
, 0);
1251 if (xa_is_internal(entry
)) {
1252 ret
= xa_to_internal(entry
);
1257 * It is possible, particularly with mixed reads & writes to private
1258 * mappings, that we have raced with a PMD fault that overlaps with
1259 * the PTE we need to set up. If so just return and the fault will be
1262 if (pmd_trans_huge(*vmf
->pmd
) || pmd_devmap(*vmf
->pmd
)) {
1263 ret
= VM_FAULT_NOPAGE
;
1268 * Note that we don't bother to use iomap_apply here: DAX required
1269 * the file system block size to be equal the page size, which means
1270 * that we never have to deal with more than a single extent here.
1272 error
= ops
->iomap_begin(inode
, pos
, PAGE_SIZE
, flags
, &iomap
);
1274 *iomap_errp
= error
;
1276 ret
= dax_fault_return(error
);
1279 if (WARN_ON_ONCE(iomap
.offset
+ iomap
.length
< pos
+ PAGE_SIZE
)) {
1280 error
= -EIO
; /* fs corruption? */
1281 goto error_finish_iomap
;
1284 if (vmf
->cow_page
) {
1285 sector_t sector
= dax_iomap_sector(&iomap
, pos
);
1287 switch (iomap
.type
) {
1289 case IOMAP_UNWRITTEN
:
1290 clear_user_highpage(vmf
->cow_page
, vaddr
);
1293 error
= copy_user_dax(iomap
.bdev
, iomap
.dax_dev
,
1294 sector
, PAGE_SIZE
, vmf
->cow_page
, vaddr
);
1303 goto error_finish_iomap
;
1305 __SetPageUptodate(vmf
->cow_page
);
1306 ret
= finish_fault(vmf
);
1308 ret
= VM_FAULT_DONE_COW
;
1312 sync
= dax_fault_is_synchronous(flags
, vma
, &iomap
);
1314 switch (iomap
.type
) {
1316 if (iomap
.flags
& IOMAP_F_NEW
) {
1317 count_vm_event(PGMAJFAULT
);
1318 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
1319 major
= VM_FAULT_MAJOR
;
1321 error
= dax_iomap_pfn(&iomap
, pos
, PAGE_SIZE
, &pfn
);
1323 goto error_finish_iomap
;
1325 entry
= dax_insert_entry(&xas
, mapping
, vmf
, entry
, pfn
,
1329 * If we are doing synchronous page fault and inode needs fsync,
1330 * we can insert PTE into page tables only after that happens.
1331 * Skip insertion for now and return the pfn so that caller can
1332 * insert it after fsync is done.
1335 if (WARN_ON_ONCE(!pfnp
)) {
1337 goto error_finish_iomap
;
1340 ret
= VM_FAULT_NEEDDSYNC
| major
;
1343 trace_dax_insert_mapping(inode
, vmf
, entry
);
1345 ret
= vmf_insert_mixed_mkwrite(vma
, vaddr
, pfn
);
1347 ret
= vmf_insert_mixed(vma
, vaddr
, pfn
);
1350 case IOMAP_UNWRITTEN
:
1353 ret
= dax_load_hole(&xas
, mapping
, &entry
, vmf
);
1364 ret
= dax_fault_return(error
);
1366 if (ops
->iomap_end
) {
1367 int copied
= PAGE_SIZE
;
1369 if (ret
& VM_FAULT_ERROR
)
1372 * The fault is done by now and there's no way back (other
1373 * thread may be already happily using PTE we have installed).
1374 * Just ignore error from ->iomap_end since we cannot do much
1377 ops
->iomap_end(inode
, pos
, PAGE_SIZE
, copied
, flags
, &iomap
);
1380 dax_unlock_entry(&xas
, entry
);
1382 trace_dax_pte_fault_done(inode
, vmf
, ret
);
1386 #ifdef CONFIG_FS_DAX_PMD
1387 static vm_fault_t
dax_pmd_load_hole(struct xa_state
*xas
, struct vm_fault
*vmf
,
1388 struct iomap
*iomap
, void **entry
)
1390 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1391 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
1392 struct inode
*inode
= mapping
->host
;
1393 struct page
*zero_page
;
1398 zero_page
= mm_get_huge_zero_page(vmf
->vma
->vm_mm
);
1400 if (unlikely(!zero_page
))
1403 pfn
= page_to_pfn_t(zero_page
);
1404 *entry
= dax_insert_entry(xas
, mapping
, vmf
, *entry
, pfn
,
1405 DAX_PMD
| DAX_ZERO_PAGE
, false);
1407 ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1408 if (!pmd_none(*(vmf
->pmd
))) {
1413 pmd_entry
= mk_pmd(zero_page
, vmf
->vma
->vm_page_prot
);
1414 pmd_entry
= pmd_mkhuge(pmd_entry
);
1415 set_pmd_at(vmf
->vma
->vm_mm
, pmd_addr
, vmf
->pmd
, pmd_entry
);
1417 trace_dax_pmd_load_hole(inode
, vmf
, zero_page
, *entry
);
1418 return VM_FAULT_NOPAGE
;
1421 trace_dax_pmd_load_hole_fallback(inode
, vmf
, zero_page
, *entry
);
1422 return VM_FAULT_FALLBACK
;
1425 static vm_fault_t
dax_iomap_pmd_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1426 const struct iomap_ops
*ops
)
1428 struct vm_area_struct
*vma
= vmf
->vma
;
1429 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1430 XA_STATE_ORDER(xas
, &mapping
->i_pages
, vmf
->pgoff
, PMD_ORDER
);
1431 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
1432 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1434 unsigned int iomap_flags
= (write
? IOMAP_WRITE
: 0) | IOMAP_FAULT
;
1435 struct inode
*inode
= mapping
->host
;
1436 vm_fault_t result
= VM_FAULT_FALLBACK
;
1437 struct iomap iomap
= { 0 };
1445 * Check whether offset isn't beyond end of file now. Caller is
1446 * supposed to hold locks serializing us with truncate / punch hole so
1447 * this is a reliable test.
1449 max_pgoff
= DIV_ROUND_UP(i_size_read(inode
), PAGE_SIZE
);
1451 trace_dax_pmd_fault(inode
, vmf
, max_pgoff
, 0);
1454 * Make sure that the faulting address's PMD offset (color) matches
1455 * the PMD offset from the start of the file. This is necessary so
1456 * that a PMD range in the page table overlaps exactly with a PMD
1457 * range in the page cache.
1459 if ((vmf
->pgoff
& PG_PMD_COLOUR
) !=
1460 ((vmf
->address
>> PAGE_SHIFT
) & PG_PMD_COLOUR
))
1463 /* Fall back to PTEs if we're going to COW */
1464 if (write
&& !(vma
->vm_flags
& VM_SHARED
))
1467 /* If the PMD would extend outside the VMA */
1468 if (pmd_addr
< vma
->vm_start
)
1470 if ((pmd_addr
+ PMD_SIZE
) > vma
->vm_end
)
1473 if (xas
.xa_index
>= max_pgoff
) {
1474 result
= VM_FAULT_SIGBUS
;
1478 /* If the PMD would extend beyond the file size */
1479 if ((xas
.xa_index
| PG_PMD_COLOUR
) >= max_pgoff
)
1483 * grab_mapping_entry() will make sure we get an empty PMD entry,
1484 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1485 * entry is already in the array, for instance), it will return
1486 * VM_FAULT_FALLBACK.
1488 entry
= grab_mapping_entry(&xas
, mapping
, DAX_PMD
);
1489 if (xa_is_internal(entry
)) {
1490 result
= xa_to_internal(entry
);
1495 * It is possible, particularly with mixed reads & writes to private
1496 * mappings, that we have raced with a PTE fault that overlaps with
1497 * the PMD we need to set up. If so just return and the fault will be
1500 if (!pmd_none(*vmf
->pmd
) && !pmd_trans_huge(*vmf
->pmd
) &&
1501 !pmd_devmap(*vmf
->pmd
)) {
1507 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1508 * setting up a mapping, so really we're using iomap_begin() as a way
1509 * to look up our filesystem block.
1511 pos
= (loff_t
)xas
.xa_index
<< PAGE_SHIFT
;
1512 error
= ops
->iomap_begin(inode
, pos
, PMD_SIZE
, iomap_flags
, &iomap
);
1516 if (iomap
.offset
+ iomap
.length
< pos
+ PMD_SIZE
)
1519 sync
= dax_fault_is_synchronous(iomap_flags
, vma
, &iomap
);
1521 switch (iomap
.type
) {
1523 error
= dax_iomap_pfn(&iomap
, pos
, PMD_SIZE
, &pfn
);
1527 entry
= dax_insert_entry(&xas
, mapping
, vmf
, entry
, pfn
,
1528 DAX_PMD
, write
&& !sync
);
1531 * If we are doing synchronous page fault and inode needs fsync,
1532 * we can insert PMD into page tables only after that happens.
1533 * Skip insertion for now and return the pfn so that caller can
1534 * insert it after fsync is done.
1537 if (WARN_ON_ONCE(!pfnp
))
1540 result
= VM_FAULT_NEEDDSYNC
;
1544 trace_dax_pmd_insert_mapping(inode
, vmf
, PMD_SIZE
, pfn
, entry
);
1545 result
= vmf_insert_pfn_pmd(vma
, vmf
->address
, vmf
->pmd
, pfn
,
1548 case IOMAP_UNWRITTEN
:
1550 if (WARN_ON_ONCE(write
))
1552 result
= dax_pmd_load_hole(&xas
, vmf
, &iomap
, &entry
);
1560 if (ops
->iomap_end
) {
1561 int copied
= PMD_SIZE
;
1563 if (result
== VM_FAULT_FALLBACK
)
1566 * The fault is done by now and there's no way back (other
1567 * thread may be already happily using PMD we have installed).
1568 * Just ignore error from ->iomap_end since we cannot do much
1571 ops
->iomap_end(inode
, pos
, PMD_SIZE
, copied
, iomap_flags
,
1575 dax_unlock_entry(&xas
, entry
);
1577 if (result
== VM_FAULT_FALLBACK
) {
1578 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1579 count_vm_event(THP_FAULT_FALLBACK
);
1582 trace_dax_pmd_fault_done(inode
, vmf
, max_pgoff
, result
);
1586 static vm_fault_t
dax_iomap_pmd_fault(struct vm_fault
*vmf
, pfn_t
*pfnp
,
1587 const struct iomap_ops
*ops
)
1589 return VM_FAULT_FALLBACK
;
1591 #endif /* CONFIG_FS_DAX_PMD */
1594 * dax_iomap_fault - handle a page fault on a DAX file
1595 * @vmf: The description of the fault
1596 * @pe_size: Size of the page to fault in
1597 * @pfnp: PFN to insert for synchronous faults if fsync is required
1598 * @iomap_errp: Storage for detailed error code in case of error
1599 * @ops: Iomap ops passed from the file system
1601 * When a page fault occurs, filesystems may call this helper in
1602 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1603 * has done all the necessary locking for page fault to proceed
1606 vm_fault_t
dax_iomap_fault(struct vm_fault
*vmf
, enum page_entry_size pe_size
,
1607 pfn_t
*pfnp
, int *iomap_errp
, const struct iomap_ops
*ops
)
1611 return dax_iomap_pte_fault(vmf
, pfnp
, iomap_errp
, ops
);
1613 return dax_iomap_pmd_fault(vmf
, pfnp
, ops
);
1615 return VM_FAULT_FALLBACK
;
1618 EXPORT_SYMBOL_GPL(dax_iomap_fault
);
1621 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1622 * @vmf: The description of the fault
1623 * @pfn: PFN to insert
1624 * @order: Order of entry to insert.
1626 * This function inserts a writeable PTE or PMD entry into the page tables
1627 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1630 dax_insert_pfn_mkwrite(struct vm_fault
*vmf
, pfn_t pfn
, unsigned int order
)
1632 struct address_space
*mapping
= vmf
->vma
->vm_file
->f_mapping
;
1633 XA_STATE_ORDER(xas
, &mapping
->i_pages
, vmf
->pgoff
, order
);
1638 entry
= get_unlocked_entry(&xas
);
1639 /* Did we race with someone splitting entry or so? */
1641 (order
== 0 && !dax_is_pte_entry(entry
)) ||
1642 (order
== PMD_ORDER
&& !dax_is_pmd_entry(entry
))) {
1643 put_unlocked_entry(&xas
, entry
);
1644 xas_unlock_irq(&xas
);
1645 trace_dax_insert_pfn_mkwrite_no_entry(mapping
->host
, vmf
,
1647 return VM_FAULT_NOPAGE
;
1649 xas_set_mark(&xas
, PAGECACHE_TAG_DIRTY
);
1650 dax_lock_entry(&xas
, entry
);
1651 xas_unlock_irq(&xas
);
1653 ret
= vmf_insert_mixed_mkwrite(vmf
->vma
, vmf
->address
, pfn
);
1654 #ifdef CONFIG_FS_DAX_PMD
1655 else if (order
== PMD_ORDER
)
1656 ret
= vmf_insert_pfn_pmd(vmf
->vma
, vmf
->address
, vmf
->pmd
,
1660 ret
= VM_FAULT_FALLBACK
;
1661 dax_unlock_entry(&xas
, entry
);
1662 trace_dax_insert_pfn_mkwrite(mapping
->host
, vmf
, ret
);
1667 * dax_finish_sync_fault - finish synchronous page fault
1668 * @vmf: The description of the fault
1669 * @pe_size: Size of entry to be inserted
1670 * @pfn: PFN to insert
1672 * This function ensures that the file range touched by the page fault is
1673 * stored persistently on the media and handles inserting of appropriate page
1676 vm_fault_t
dax_finish_sync_fault(struct vm_fault
*vmf
,
1677 enum page_entry_size pe_size
, pfn_t pfn
)
1680 loff_t start
= ((loff_t
)vmf
->pgoff
) << PAGE_SHIFT
;
1681 unsigned int order
= pe_order(pe_size
);
1682 size_t len
= PAGE_SIZE
<< order
;
1684 err
= vfs_fsync_range(vmf
->vma
->vm_file
, start
, start
+ len
- 1, 1);
1686 return VM_FAULT_SIGBUS
;
1687 return dax_insert_pfn_mkwrite(vmf
, pfn
, order
);
1689 EXPORT_SYMBOL_GPL(dax_finish_sync_fault
);