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/pmem.h>
29 #include <linux/sched.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/iomap.h>
37 /* We choose 4096 entries - same as per-zone page wait tables */
38 #define DAX_WAIT_TABLE_BITS 12
39 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
41 static wait_queue_head_t wait_table
[DAX_WAIT_TABLE_ENTRIES
];
43 static int __init
init_dax_wait_table(void)
47 for (i
= 0; i
< DAX_WAIT_TABLE_ENTRIES
; i
++)
48 init_waitqueue_head(wait_table
+ i
);
51 fs_initcall(init_dax_wait_table
);
53 static long dax_map_atomic(struct block_device
*bdev
, struct blk_dax_ctl
*dax
)
55 struct request_queue
*q
= bdev
->bd_queue
;
58 dax
->addr
= ERR_PTR(-EIO
);
59 if (blk_queue_enter(q
, true) != 0)
62 rc
= bdev_direct_access(bdev
, dax
);
64 dax
->addr
= ERR_PTR(rc
);
71 static void dax_unmap_atomic(struct block_device
*bdev
,
72 const struct blk_dax_ctl
*dax
)
74 if (IS_ERR(dax
->addr
))
76 blk_queue_exit(bdev
->bd_queue
);
79 static int dax_is_pmd_entry(void *entry
)
81 return (unsigned long)entry
& RADIX_DAX_PMD
;
84 static int dax_is_pte_entry(void *entry
)
86 return !((unsigned long)entry
& RADIX_DAX_PMD
);
89 static int dax_is_zero_entry(void *entry
)
91 return (unsigned long)entry
& RADIX_DAX_HZP
;
94 static int dax_is_empty_entry(void *entry
)
96 return (unsigned long)entry
& RADIX_DAX_EMPTY
;
99 struct page
*read_dax_sector(struct block_device
*bdev
, sector_t n
)
101 struct page
*page
= alloc_pages(GFP_KERNEL
, 0);
102 struct blk_dax_ctl dax
= {
104 .sector
= n
& ~((((int) PAGE_SIZE
) / 512) - 1),
109 return ERR_PTR(-ENOMEM
);
111 rc
= dax_map_atomic(bdev
, &dax
);
114 memcpy_from_pmem(page_address(page
), dax
.addr
, PAGE_SIZE
);
115 dax_unmap_atomic(bdev
, &dax
);
120 * DAX radix tree locking
122 struct exceptional_entry_key
{
123 struct address_space
*mapping
;
127 struct wait_exceptional_entry_queue
{
129 struct exceptional_entry_key key
;
132 static wait_queue_head_t
*dax_entry_waitqueue(struct address_space
*mapping
,
133 pgoff_t index
, void *entry
, struct exceptional_entry_key
*key
)
138 * If 'entry' is a PMD, align the 'index' that we use for the wait
139 * queue to the start of that PMD. This ensures that all offsets in
140 * the range covered by the PMD map to the same bit lock.
142 if (dax_is_pmd_entry(entry
))
143 index
&= ~((1UL << (PMD_SHIFT
- PAGE_SHIFT
)) - 1);
145 key
->mapping
= mapping
;
146 key
->entry_start
= index
;
148 hash
= hash_long((unsigned long)mapping
^ index
, DAX_WAIT_TABLE_BITS
);
149 return wait_table
+ hash
;
152 static int wake_exceptional_entry_func(wait_queue_t
*wait
, unsigned int mode
,
153 int sync
, void *keyp
)
155 struct exceptional_entry_key
*key
= keyp
;
156 struct wait_exceptional_entry_queue
*ewait
=
157 container_of(wait
, struct wait_exceptional_entry_queue
, wait
);
159 if (key
->mapping
!= ewait
->key
.mapping
||
160 key
->entry_start
!= ewait
->key
.entry_start
)
162 return autoremove_wake_function(wait
, mode
, sync
, NULL
);
166 * Check whether the given slot is locked. The function must be called with
167 * mapping->tree_lock held
169 static inline int slot_locked(struct address_space
*mapping
, void **slot
)
171 unsigned long entry
= (unsigned long)
172 radix_tree_deref_slot_protected(slot
, &mapping
->tree_lock
);
173 return entry
& RADIX_DAX_ENTRY_LOCK
;
177 * Mark the given slot is locked. The function must be called with
178 * mapping->tree_lock held
180 static inline void *lock_slot(struct address_space
*mapping
, void **slot
)
182 unsigned long entry
= (unsigned long)
183 radix_tree_deref_slot_protected(slot
, &mapping
->tree_lock
);
185 entry
|= RADIX_DAX_ENTRY_LOCK
;
186 radix_tree_replace_slot(slot
, (void *)entry
);
187 return (void *)entry
;
191 * Mark the given slot is unlocked. The function must be called with
192 * mapping->tree_lock held
194 static inline void *unlock_slot(struct address_space
*mapping
, void **slot
)
196 unsigned long entry
= (unsigned long)
197 radix_tree_deref_slot_protected(slot
, &mapping
->tree_lock
);
199 entry
&= ~(unsigned long)RADIX_DAX_ENTRY_LOCK
;
200 radix_tree_replace_slot(slot
, (void *)entry
);
201 return (void *)entry
;
205 * Lookup entry in radix tree, wait for it to become unlocked if it is
206 * exceptional entry and return it. The caller must call
207 * put_unlocked_mapping_entry() when he decided not to lock the entry or
208 * put_locked_mapping_entry() when he locked the entry and now wants to
211 * The function must be called with mapping->tree_lock held.
213 static void *get_unlocked_mapping_entry(struct address_space
*mapping
,
214 pgoff_t index
, void ***slotp
)
217 struct wait_exceptional_entry_queue ewait
;
218 wait_queue_head_t
*wq
;
220 init_wait(&ewait
.wait
);
221 ewait
.wait
.func
= wake_exceptional_entry_func
;
224 entry
= __radix_tree_lookup(&mapping
->page_tree
, index
, NULL
,
226 if (!entry
|| !radix_tree_exceptional_entry(entry
) ||
227 !slot_locked(mapping
, slot
)) {
233 wq
= dax_entry_waitqueue(mapping
, index
, entry
, &ewait
.key
);
234 prepare_to_wait_exclusive(wq
, &ewait
.wait
,
235 TASK_UNINTERRUPTIBLE
);
236 spin_unlock_irq(&mapping
->tree_lock
);
238 finish_wait(wq
, &ewait
.wait
);
239 spin_lock_irq(&mapping
->tree_lock
);
243 static void put_locked_mapping_entry(struct address_space
*mapping
,
244 pgoff_t index
, void *entry
)
246 if (!radix_tree_exceptional_entry(entry
)) {
250 dax_unlock_mapping_entry(mapping
, index
);
255 * Called when we are done with radix tree entry we looked up via
256 * get_unlocked_mapping_entry() and which we didn't lock in the end.
258 static void put_unlocked_mapping_entry(struct address_space
*mapping
,
259 pgoff_t index
, void *entry
)
261 if (!radix_tree_exceptional_entry(entry
))
264 /* We have to wake up next waiter for the radix tree entry lock */
265 dax_wake_mapping_entry_waiter(mapping
, index
, entry
, false);
269 * Find radix tree entry at given index. If it points to a page, return with
270 * the page locked. If it points to the exceptional entry, return with the
271 * radix tree entry locked. If the radix tree doesn't contain given index,
272 * create empty exceptional entry for the index and return with it locked.
274 * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
275 * either return that locked entry or will return an error. This error will
276 * happen if there are any 4k entries (either zero pages or DAX entries)
277 * within the 2MiB range that we are requesting.
279 * We always favor 4k entries over 2MiB entries. There isn't a flow where we
280 * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB
281 * insertion will fail if it finds any 4k entries already in the tree, and a
282 * 4k insertion will cause an existing 2MiB entry to be unmapped and
283 * downgraded to 4k entries. This happens for both 2MiB huge zero pages as
284 * well as 2MiB empty entries.
286 * The exception to this downgrade path is for 2MiB DAX PMD entries that have
287 * real storage backing them. We will leave these real 2MiB DAX entries in
288 * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
290 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
291 * persistent memory the benefit is doubtful. We can add that later if we can
294 static void *grab_mapping_entry(struct address_space
*mapping
, pgoff_t index
,
295 unsigned long size_flag
)
297 bool pmd_downgrade
= false; /* splitting 2MiB entry into 4k entries? */
301 spin_lock_irq(&mapping
->tree_lock
);
302 entry
= get_unlocked_mapping_entry(mapping
, index
, &slot
);
305 if (size_flag
& RADIX_DAX_PMD
) {
306 if (!radix_tree_exceptional_entry(entry
) ||
307 dax_is_pte_entry(entry
)) {
308 put_unlocked_mapping_entry(mapping
, index
,
310 entry
= ERR_PTR(-EEXIST
);
313 } else { /* trying to grab a PTE entry */
314 if (radix_tree_exceptional_entry(entry
) &&
315 dax_is_pmd_entry(entry
) &&
316 (dax_is_zero_entry(entry
) ||
317 dax_is_empty_entry(entry
))) {
318 pmd_downgrade
= true;
323 /* No entry for given index? Make sure radix tree is big enough. */
324 if (!entry
|| pmd_downgrade
) {
329 * Make sure 'entry' remains valid while we drop
330 * mapping->tree_lock.
332 entry
= lock_slot(mapping
, slot
);
335 spin_unlock_irq(&mapping
->tree_lock
);
337 * Besides huge zero pages the only other thing that gets
338 * downgraded are empty entries which don't need to be
341 if (pmd_downgrade
&& dax_is_zero_entry(entry
))
342 unmap_mapping_range(mapping
,
343 (index
<< PAGE_SHIFT
) & PMD_MASK
, PMD_SIZE
, 0);
345 err
= radix_tree_preload(
346 mapping_gfp_mask(mapping
) & ~__GFP_HIGHMEM
);
349 put_locked_mapping_entry(mapping
, index
, entry
);
352 spin_lock_irq(&mapping
->tree_lock
);
355 radix_tree_delete(&mapping
->page_tree
, index
);
356 mapping
->nrexceptional
--;
357 dax_wake_mapping_entry_waiter(mapping
, index
, entry
,
361 entry
= dax_radix_locked_entry(0, size_flag
| RADIX_DAX_EMPTY
);
363 err
= __radix_tree_insert(&mapping
->page_tree
, index
,
364 dax_radix_order(entry
), entry
);
365 radix_tree_preload_end();
367 spin_unlock_irq(&mapping
->tree_lock
);
369 * Someone already created the entry? This is a
370 * normal failure when inserting PMDs in a range
371 * that already contains PTEs. In that case we want
372 * to return -EEXIST immediately.
374 if (err
== -EEXIST
&& !(size_flag
& RADIX_DAX_PMD
))
377 * Our insertion of a DAX PMD entry failed, most
378 * likely because it collided with a PTE sized entry
379 * at a different index in the PMD range. We haven't
380 * inserted anything into the radix tree and have no
385 /* Good, we have inserted empty locked entry into the tree. */
386 mapping
->nrexceptional
++;
387 spin_unlock_irq(&mapping
->tree_lock
);
390 /* Normal page in radix tree? */
391 if (!radix_tree_exceptional_entry(entry
)) {
392 struct page
*page
= entry
;
395 spin_unlock_irq(&mapping
->tree_lock
);
397 /* Page got truncated? Retry... */
398 if (unlikely(page
->mapping
!= mapping
)) {
405 entry
= lock_slot(mapping
, slot
);
407 spin_unlock_irq(&mapping
->tree_lock
);
412 * We do not necessarily hold the mapping->tree_lock when we call this
413 * function so it is possible that 'entry' is no longer a valid item in the
414 * radix tree. This is okay because all we really need to do is to find the
415 * correct waitqueue where tasks might be waiting for that old 'entry' and
418 void dax_wake_mapping_entry_waiter(struct address_space
*mapping
,
419 pgoff_t index
, void *entry
, bool wake_all
)
421 struct exceptional_entry_key key
;
422 wait_queue_head_t
*wq
;
424 wq
= dax_entry_waitqueue(mapping
, index
, entry
, &key
);
427 * Checking for locked entry and prepare_to_wait_exclusive() happens
428 * under mapping->tree_lock, ditto for entry handling in our callers.
429 * So at this point all tasks that could have seen our entry locked
430 * must be in the waitqueue and the following check will see them.
432 if (waitqueue_active(wq
))
433 __wake_up(wq
, TASK_NORMAL
, wake_all
? 0 : 1, &key
);
436 void dax_unlock_mapping_entry(struct address_space
*mapping
, pgoff_t index
)
440 spin_lock_irq(&mapping
->tree_lock
);
441 entry
= __radix_tree_lookup(&mapping
->page_tree
, index
, NULL
, &slot
);
442 if (WARN_ON_ONCE(!entry
|| !radix_tree_exceptional_entry(entry
) ||
443 !slot_locked(mapping
, slot
))) {
444 spin_unlock_irq(&mapping
->tree_lock
);
447 unlock_slot(mapping
, slot
);
448 spin_unlock_irq(&mapping
->tree_lock
);
449 dax_wake_mapping_entry_waiter(mapping
, index
, entry
, false);
453 * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
454 * entry to get unlocked before deleting it.
456 int dax_delete_mapping_entry(struct address_space
*mapping
, pgoff_t index
)
460 spin_lock_irq(&mapping
->tree_lock
);
461 entry
= get_unlocked_mapping_entry(mapping
, index
, NULL
);
463 * This gets called from truncate / punch_hole path. As such, the caller
464 * must hold locks protecting against concurrent modifications of the
465 * radix tree (usually fs-private i_mmap_sem for writing). Since the
466 * caller has seen exceptional entry for this index, we better find it
467 * at that index as well...
469 if (WARN_ON_ONCE(!entry
|| !radix_tree_exceptional_entry(entry
))) {
470 spin_unlock_irq(&mapping
->tree_lock
);
473 radix_tree_delete(&mapping
->page_tree
, index
);
474 mapping
->nrexceptional
--;
475 spin_unlock_irq(&mapping
->tree_lock
);
476 dax_wake_mapping_entry_waiter(mapping
, index
, entry
, true);
482 * The user has performed a load from a hole in the file. Allocating
483 * a new page in the file would cause excessive storage usage for
484 * workloads with sparse files. We allocate a page cache page instead.
485 * We'll kick it out of the page cache if it's ever written to,
486 * otherwise it will simply fall out of the page cache under memory
487 * pressure without ever having been dirtied.
489 static int dax_load_hole(struct address_space
*mapping
, void *entry
,
490 struct vm_fault
*vmf
)
494 /* Hole page already exists? Return it... */
495 if (!radix_tree_exceptional_entry(entry
)) {
497 return VM_FAULT_LOCKED
;
500 /* This will replace locked radix tree entry with a hole page */
501 page
= find_or_create_page(mapping
, vmf
->pgoff
,
502 vmf
->gfp_mask
| __GFP_ZERO
);
504 put_locked_mapping_entry(mapping
, vmf
->pgoff
, entry
);
508 return VM_FAULT_LOCKED
;
511 static int copy_user_dax(struct block_device
*bdev
, sector_t sector
, size_t size
,
512 struct page
*to
, unsigned long vaddr
)
514 struct blk_dax_ctl dax
= {
520 if (dax_map_atomic(bdev
, &dax
) < 0)
521 return PTR_ERR(dax
.addr
);
522 vto
= kmap_atomic(to
);
523 copy_user_page(vto
, (void __force
*)dax
.addr
, vaddr
, to
);
525 dax_unmap_atomic(bdev
, &dax
);
530 * By this point grab_mapping_entry() has ensured that we have a locked entry
531 * of the appropriate size so we don't have to worry about downgrading PMDs to
532 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
533 * already in the tree, we will skip the insertion and just dirty the PMD as
536 static void *dax_insert_mapping_entry(struct address_space
*mapping
,
537 struct vm_fault
*vmf
,
538 void *entry
, sector_t sector
,
541 struct radix_tree_root
*page_tree
= &mapping
->page_tree
;
543 bool hole_fill
= false;
545 pgoff_t index
= vmf
->pgoff
;
547 if (vmf
->flags
& FAULT_FLAG_WRITE
)
548 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
550 /* Replacing hole page with block mapping? */
551 if (!radix_tree_exceptional_entry(entry
)) {
554 * Unmap the page now before we remove it from page cache below.
555 * The page is locked so it cannot be faulted in again.
557 unmap_mapping_range(mapping
, vmf
->pgoff
<< PAGE_SHIFT
,
559 error
= radix_tree_preload(vmf
->gfp_mask
& ~__GFP_HIGHMEM
);
561 return ERR_PTR(error
);
562 } else if (dax_is_zero_entry(entry
) && !(flags
& RADIX_DAX_HZP
)) {
563 /* replacing huge zero page with PMD block mapping */
564 unmap_mapping_range(mapping
,
565 (vmf
->pgoff
<< PAGE_SHIFT
) & PMD_MASK
, PMD_SIZE
, 0);
568 spin_lock_irq(&mapping
->tree_lock
);
569 new_entry
= dax_radix_locked_entry(sector
, flags
);
572 __delete_from_page_cache(entry
, NULL
);
573 /* Drop pagecache reference */
575 error
= __radix_tree_insert(page_tree
, index
,
576 dax_radix_order(new_entry
), new_entry
);
578 new_entry
= ERR_PTR(error
);
581 mapping
->nrexceptional
++;
582 } else if (dax_is_zero_entry(entry
) || dax_is_empty_entry(entry
)) {
584 * Only swap our new entry into the radix tree if the current
585 * entry is a zero page or an empty entry. If a normal PTE or
586 * PMD entry is already in the tree, we leave it alone. This
587 * means that if we are trying to insert a PTE and the
588 * existing entry is a PMD, we will just leave the PMD in the
589 * tree and dirty it if necessary.
594 ret
= __radix_tree_lookup(page_tree
, index
, NULL
, &slot
);
595 WARN_ON_ONCE(ret
!= entry
);
596 radix_tree_replace_slot(slot
, new_entry
);
598 if (vmf
->flags
& FAULT_FLAG_WRITE
)
599 radix_tree_tag_set(page_tree
, index
, PAGECACHE_TAG_DIRTY
);
601 spin_unlock_irq(&mapping
->tree_lock
);
603 radix_tree_preload_end();
605 * We don't need hole page anymore, it has been replaced with
606 * locked radix tree entry now.
608 if (mapping
->a_ops
->freepage
)
609 mapping
->a_ops
->freepage(entry
);
616 static int dax_writeback_one(struct block_device
*bdev
,
617 struct address_space
*mapping
, pgoff_t index
, void *entry
)
619 struct radix_tree_root
*page_tree
= &mapping
->page_tree
;
620 struct radix_tree_node
*node
;
621 struct blk_dax_ctl dax
;
625 spin_lock_irq(&mapping
->tree_lock
);
627 * Regular page slots are stabilized by the page lock even
628 * without the tree itself locked. These unlocked entries
629 * need verification under the tree lock.
631 if (!__radix_tree_lookup(page_tree
, index
, &node
, &slot
))
636 /* another fsync thread may have already written back this entry */
637 if (!radix_tree_tag_get(page_tree
, index
, PAGECACHE_TAG_TOWRITE
))
640 if (WARN_ON_ONCE(dax_is_empty_entry(entry
) ||
641 dax_is_zero_entry(entry
))) {
647 * Even if dax_writeback_mapping_range() was given a wbc->range_start
648 * in the middle of a PMD, the 'index' we are given will be aligned to
649 * the start index of the PMD, as will the sector we pull from
650 * 'entry'. This allows us to flush for PMD_SIZE and not have to
651 * worry about partial PMD writebacks.
653 dax
.sector
= dax_radix_sector(entry
);
654 dax
.size
= PAGE_SIZE
<< dax_radix_order(entry
);
655 spin_unlock_irq(&mapping
->tree_lock
);
658 * We cannot hold tree_lock while calling dax_map_atomic() because it
659 * eventually calls cond_resched().
661 ret
= dax_map_atomic(bdev
, &dax
);
665 if (WARN_ON_ONCE(ret
< dax
.size
)) {
670 wb_cache_pmem(dax
.addr
, dax
.size
);
672 spin_lock_irq(&mapping
->tree_lock
);
673 radix_tree_tag_clear(page_tree
, index
, PAGECACHE_TAG_TOWRITE
);
674 spin_unlock_irq(&mapping
->tree_lock
);
676 dax_unmap_atomic(bdev
, &dax
);
680 spin_unlock_irq(&mapping
->tree_lock
);
685 * Flush the mapping to the persistent domain within the byte range of [start,
686 * end]. This is required by data integrity operations to ensure file data is
687 * on persistent storage prior to completion of the operation.
689 int dax_writeback_mapping_range(struct address_space
*mapping
,
690 struct block_device
*bdev
, struct writeback_control
*wbc
)
692 struct inode
*inode
= mapping
->host
;
693 pgoff_t start_index
, end_index
;
694 pgoff_t indices
[PAGEVEC_SIZE
];
699 if (WARN_ON_ONCE(inode
->i_blkbits
!= PAGE_SHIFT
))
702 if (!mapping
->nrexceptional
|| wbc
->sync_mode
!= WB_SYNC_ALL
)
705 start_index
= wbc
->range_start
>> PAGE_SHIFT
;
706 end_index
= wbc
->range_end
>> PAGE_SHIFT
;
708 tag_pages_for_writeback(mapping
, start_index
, end_index
);
710 pagevec_init(&pvec
, 0);
712 pvec
.nr
= find_get_entries_tag(mapping
, start_index
,
713 PAGECACHE_TAG_TOWRITE
, PAGEVEC_SIZE
,
714 pvec
.pages
, indices
);
719 for (i
= 0; i
< pvec
.nr
; i
++) {
720 if (indices
[i
] > end_index
) {
725 ret
= dax_writeback_one(bdev
, mapping
, indices
[i
],
733 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range
);
735 static int dax_insert_mapping(struct address_space
*mapping
,
736 struct block_device
*bdev
, sector_t sector
, size_t size
,
737 void **entryp
, struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
739 unsigned long vaddr
= (unsigned long)vmf
->virtual_address
;
740 struct blk_dax_ctl dax
= {
745 void *entry
= *entryp
;
747 if (dax_map_atomic(bdev
, &dax
) < 0)
748 return PTR_ERR(dax
.addr
);
749 dax_unmap_atomic(bdev
, &dax
);
751 ret
= dax_insert_mapping_entry(mapping
, vmf
, entry
, dax
.sector
, 0);
756 return vm_insert_mixed(vma
, vaddr
, dax
.pfn
);
760 * dax_pfn_mkwrite - handle first write to DAX page
761 * @vma: The virtual memory area where the fault occurred
762 * @vmf: The description of the fault
764 int dax_pfn_mkwrite(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
766 struct file
*file
= vma
->vm_file
;
767 struct address_space
*mapping
= file
->f_mapping
;
769 pgoff_t index
= vmf
->pgoff
;
771 spin_lock_irq(&mapping
->tree_lock
);
772 entry
= get_unlocked_mapping_entry(mapping
, index
, NULL
);
773 if (!entry
|| !radix_tree_exceptional_entry(entry
))
775 radix_tree_tag_set(&mapping
->page_tree
, index
, PAGECACHE_TAG_DIRTY
);
776 put_unlocked_mapping_entry(mapping
, index
, entry
);
778 spin_unlock_irq(&mapping
->tree_lock
);
779 return VM_FAULT_NOPAGE
;
781 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite
);
783 static bool dax_range_is_aligned(struct block_device
*bdev
,
784 unsigned int offset
, unsigned int length
)
786 unsigned short sector_size
= bdev_logical_block_size(bdev
);
788 if (!IS_ALIGNED(offset
, sector_size
))
790 if (!IS_ALIGNED(length
, sector_size
))
796 int __dax_zero_page_range(struct block_device
*bdev
, sector_t sector
,
797 unsigned int offset
, unsigned int length
)
799 struct blk_dax_ctl dax
= {
804 if (dax_range_is_aligned(bdev
, offset
, length
)) {
805 sector_t start_sector
= dax
.sector
+ (offset
>> 9);
807 return blkdev_issue_zeroout(bdev
, start_sector
,
808 length
>> 9, GFP_NOFS
, true);
810 if (dax_map_atomic(bdev
, &dax
) < 0)
811 return PTR_ERR(dax
.addr
);
812 clear_pmem(dax
.addr
+ offset
, length
);
813 dax_unmap_atomic(bdev
, &dax
);
817 EXPORT_SYMBOL_GPL(__dax_zero_page_range
);
819 #ifdef CONFIG_FS_IOMAP
820 static sector_t
dax_iomap_sector(struct iomap
*iomap
, loff_t pos
)
822 return iomap
->blkno
+ (((pos
& PAGE_MASK
) - iomap
->offset
) >> 9);
826 dax_iomap_actor(struct inode
*inode
, loff_t pos
, loff_t length
, void *data
,
829 struct iov_iter
*iter
= data
;
830 loff_t end
= pos
+ length
, done
= 0;
833 if (iov_iter_rw(iter
) == READ
) {
834 end
= min(end
, i_size_read(inode
));
838 if (iomap
->type
== IOMAP_HOLE
|| iomap
->type
== IOMAP_UNWRITTEN
)
839 return iov_iter_zero(min(length
, end
- pos
), iter
);
842 if (WARN_ON_ONCE(iomap
->type
!= IOMAP_MAPPED
))
846 unsigned offset
= pos
& (PAGE_SIZE
- 1);
847 struct blk_dax_ctl dax
= { 0 };
850 dax
.sector
= dax_iomap_sector(iomap
, pos
);
851 dax
.size
= (length
+ offset
+ PAGE_SIZE
- 1) & PAGE_MASK
;
852 map_len
= dax_map_atomic(iomap
->bdev
, &dax
);
860 if (map_len
> end
- pos
)
863 if (iov_iter_rw(iter
) == WRITE
)
864 map_len
= copy_from_iter_pmem(dax
.addr
, map_len
, iter
);
866 map_len
= copy_to_iter(dax
.addr
, map_len
, iter
);
867 dax_unmap_atomic(iomap
->bdev
, &dax
);
869 ret
= map_len
? map_len
: -EFAULT
;
878 return done
? done
: ret
;
882 * dax_iomap_rw - Perform I/O to a DAX file
883 * @iocb: The control block for this I/O
884 * @iter: The addresses to do I/O from or to
885 * @ops: iomap ops passed from the file system
887 * This function performs read and write operations to directly mapped
888 * persistent memory. The callers needs to take care of read/write exclusion
889 * and evicting any page cache pages in the region under I/O.
892 dax_iomap_rw(struct kiocb
*iocb
, struct iov_iter
*iter
,
893 struct iomap_ops
*ops
)
895 struct address_space
*mapping
= iocb
->ki_filp
->f_mapping
;
896 struct inode
*inode
= mapping
->host
;
897 loff_t pos
= iocb
->ki_pos
, ret
= 0, done
= 0;
900 if (iov_iter_rw(iter
) == WRITE
)
901 flags
|= IOMAP_WRITE
;
904 * Yes, even DAX files can have page cache attached to them: A zeroed
905 * page is inserted into the pagecache when we have to serve a write
906 * fault on a hole. It should never be dirtied and can simply be
907 * dropped from the pagecache once we get real data for the page.
909 * XXX: This is racy against mmap, and there's nothing we can do about
910 * it. We'll eventually need to shift this down even further so that
911 * we can check if we allocated blocks over a hole first.
913 if (mapping
->nrpages
) {
914 ret
= invalidate_inode_pages2_range(mapping
,
916 (pos
+ iov_iter_count(iter
) - 1) >> PAGE_SHIFT
);
920 while (iov_iter_count(iter
)) {
921 ret
= iomap_apply(inode
, pos
, iov_iter_count(iter
), flags
, ops
,
922 iter
, dax_iomap_actor
);
929 iocb
->ki_pos
+= done
;
930 return done
? done
: ret
;
932 EXPORT_SYMBOL_GPL(dax_iomap_rw
);
935 * dax_iomap_fault - handle a page fault on a DAX file
936 * @vma: The virtual memory area where the fault occurred
937 * @vmf: The description of the fault
938 * @ops: iomap ops passed from the file system
940 * When a page fault occurs, filesystems may call this helper in their fault
941 * or mkwrite handler for DAX files. Assumes the caller has done all the
942 * necessary locking for the page fault to proceed successfully.
944 int dax_iomap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
,
945 struct iomap_ops
*ops
)
947 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
948 struct inode
*inode
= mapping
->host
;
949 unsigned long vaddr
= (unsigned long)vmf
->virtual_address
;
950 loff_t pos
= (loff_t
)vmf
->pgoff
<< PAGE_SHIFT
;
952 struct iomap iomap
= { 0 };
953 unsigned flags
= IOMAP_FAULT
;
954 int error
, major
= 0;
955 int locked_status
= 0;
959 * Check whether offset isn't beyond end of file now. Caller is supposed
960 * to hold locks serializing us with truncate / punch hole so this is
963 if (pos
>= i_size_read(inode
))
964 return VM_FAULT_SIGBUS
;
966 entry
= grab_mapping_entry(mapping
, vmf
->pgoff
, 0);
968 error
= PTR_ERR(entry
);
972 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && !vmf
->cow_page
)
973 flags
|= IOMAP_WRITE
;
976 * Note that we don't bother to use iomap_apply here: DAX required
977 * the file system block size to be equal the page size, which means
978 * that we never have to deal with more than a single extent here.
980 error
= ops
->iomap_begin(inode
, pos
, PAGE_SIZE
, flags
, &iomap
);
983 if (WARN_ON_ONCE(iomap
.offset
+ iomap
.length
< pos
+ PAGE_SIZE
)) {
984 error
= -EIO
; /* fs corruption? */
988 sector
= dax_iomap_sector(&iomap
, pos
);
991 switch (iomap
.type
) {
993 case IOMAP_UNWRITTEN
:
994 clear_user_highpage(vmf
->cow_page
, vaddr
);
997 error
= copy_user_dax(iomap
.bdev
, sector
, PAGE_SIZE
,
998 vmf
->cow_page
, vaddr
);
1008 if (!radix_tree_exceptional_entry(entry
)) {
1010 locked_status
= VM_FAULT_LOCKED
;
1013 locked_status
= VM_FAULT_DAX_LOCKED
;
1018 switch (iomap
.type
) {
1020 if (iomap
.flags
& IOMAP_F_NEW
) {
1021 count_vm_event(PGMAJFAULT
);
1022 mem_cgroup_count_vm_event(vma
->vm_mm
, PGMAJFAULT
);
1023 major
= VM_FAULT_MAJOR
;
1025 error
= dax_insert_mapping(mapping
, iomap
.bdev
, sector
,
1026 PAGE_SIZE
, &entry
, vma
, vmf
);
1028 case IOMAP_UNWRITTEN
:
1030 if (!(vmf
->flags
& FAULT_FLAG_WRITE
)) {
1031 locked_status
= dax_load_hole(mapping
, entry
, vmf
);
1042 if (ops
->iomap_end
) {
1044 /* keep previous error */
1045 ops
->iomap_end(inode
, pos
, PAGE_SIZE
, 0, flags
,
1048 error
= ops
->iomap_end(inode
, pos
, PAGE_SIZE
,
1049 PAGE_SIZE
, flags
, &iomap
);
1053 if (!locked_status
|| error
)
1054 put_locked_mapping_entry(mapping
, vmf
->pgoff
, entry
);
1056 if (error
== -ENOMEM
)
1057 return VM_FAULT_OOM
| major
;
1058 /* -EBUSY is fine, somebody else faulted on the same PTE */
1059 if (error
< 0 && error
!= -EBUSY
)
1060 return VM_FAULT_SIGBUS
| major
;
1061 if (locked_status
) {
1062 WARN_ON_ONCE(error
); /* -EBUSY from ops->iomap_end? */
1063 return locked_status
;
1065 return VM_FAULT_NOPAGE
| major
;
1067 EXPORT_SYMBOL_GPL(dax_iomap_fault
);
1069 #ifdef CONFIG_FS_DAX_PMD
1071 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
1072 * more often than one might expect in the below functions.
1074 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
1076 static int dax_pmd_insert_mapping(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1077 struct vm_fault
*vmf
, unsigned long address
,
1078 struct iomap
*iomap
, loff_t pos
, bool write
, void **entryp
)
1080 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1081 struct block_device
*bdev
= iomap
->bdev
;
1082 struct blk_dax_ctl dax
= {
1083 .sector
= dax_iomap_sector(iomap
, pos
),
1086 long length
= dax_map_atomic(bdev
, &dax
);
1089 if (length
< 0) /* dax_map_atomic() failed */
1090 return VM_FAULT_FALLBACK
;
1091 if (length
< PMD_SIZE
)
1092 goto unmap_fallback
;
1093 if (pfn_t_to_pfn(dax
.pfn
) & PG_PMD_COLOUR
)
1094 goto unmap_fallback
;
1095 if (!pfn_t_devmap(dax
.pfn
))
1096 goto unmap_fallback
;
1098 dax_unmap_atomic(bdev
, &dax
);
1100 ret
= dax_insert_mapping_entry(mapping
, vmf
, *entryp
, dax
.sector
,
1103 return VM_FAULT_FALLBACK
;
1106 return vmf_insert_pfn_pmd(vma
, address
, pmd
, dax
.pfn
, write
);
1109 dax_unmap_atomic(bdev
, &dax
);
1110 return VM_FAULT_FALLBACK
;
1113 static int dax_pmd_load_hole(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1114 struct vm_fault
*vmf
, unsigned long address
,
1115 struct iomap
*iomap
, void **entryp
)
1117 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1118 unsigned long pmd_addr
= address
& PMD_MASK
;
1119 struct page
*zero_page
;
1124 zero_page
= mm_get_huge_zero_page(vma
->vm_mm
);
1126 if (unlikely(!zero_page
))
1127 return VM_FAULT_FALLBACK
;
1129 ret
= dax_insert_mapping_entry(mapping
, vmf
, *entryp
, 0,
1130 RADIX_DAX_PMD
| RADIX_DAX_HZP
);
1132 return VM_FAULT_FALLBACK
;
1135 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1136 if (!pmd_none(*pmd
)) {
1138 return VM_FAULT_FALLBACK
;
1141 pmd_entry
= mk_pmd(zero_page
, vma
->vm_page_prot
);
1142 pmd_entry
= pmd_mkhuge(pmd_entry
);
1143 set_pmd_at(vma
->vm_mm
, pmd_addr
, pmd
, pmd_entry
);
1145 return VM_FAULT_NOPAGE
;
1148 int dax_iomap_pmd_fault(struct vm_area_struct
*vma
, unsigned long address
,
1149 pmd_t
*pmd
, unsigned int flags
, struct iomap_ops
*ops
)
1151 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1152 unsigned long pmd_addr
= address
& PMD_MASK
;
1153 bool write
= flags
& FAULT_FLAG_WRITE
;
1154 unsigned int iomap_flags
= (write
? IOMAP_WRITE
: 0) | IOMAP_FAULT
;
1155 struct inode
*inode
= mapping
->host
;
1156 int result
= VM_FAULT_FALLBACK
;
1157 struct iomap iomap
= { 0 };
1158 pgoff_t max_pgoff
, pgoff
;
1159 struct vm_fault vmf
;
1164 /* Fall back to PTEs if we're going to COW */
1165 if (write
&& !(vma
->vm_flags
& VM_SHARED
))
1168 /* If the PMD would extend outside the VMA */
1169 if (pmd_addr
< vma
->vm_start
)
1171 if ((pmd_addr
+ PMD_SIZE
) > vma
->vm_end
)
1175 * Check whether offset isn't beyond end of file now. Caller is
1176 * supposed to hold locks serializing us with truncate / punch hole so
1177 * this is a reliable test.
1179 pgoff
= linear_page_index(vma
, pmd_addr
);
1180 max_pgoff
= (i_size_read(inode
) - 1) >> PAGE_SHIFT
;
1182 if (pgoff
> max_pgoff
)
1183 return VM_FAULT_SIGBUS
;
1185 /* If the PMD would extend beyond the file size */
1186 if ((pgoff
| PG_PMD_COLOUR
) > max_pgoff
)
1190 * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1191 * PMD or a HZP entry. If it can't (because a 4k page is already in
1192 * the tree, for instance), it will return -EEXIST and we just fall
1193 * back to 4k entries.
1195 entry
= grab_mapping_entry(mapping
, pgoff
, RADIX_DAX_PMD
);
1200 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1201 * setting up a mapping, so really we're using iomap_begin() as a way
1202 * to look up our filesystem block.
1204 pos
= (loff_t
)pgoff
<< PAGE_SHIFT
;
1205 error
= ops
->iomap_begin(inode
, pos
, PMD_SIZE
, iomap_flags
, &iomap
);
1208 if (iomap
.offset
+ iomap
.length
< pos
+ PMD_SIZE
)
1213 vmf
.gfp_mask
= mapping_gfp_mask(mapping
) | __GFP_IO
;
1215 switch (iomap
.type
) {
1217 result
= dax_pmd_insert_mapping(vma
, pmd
, &vmf
, address
,
1218 &iomap
, pos
, write
, &entry
);
1220 case IOMAP_UNWRITTEN
:
1222 if (WARN_ON_ONCE(write
))
1224 result
= dax_pmd_load_hole(vma
, pmd
, &vmf
, address
, &iomap
,
1233 if (ops
->iomap_end
) {
1234 if (result
== VM_FAULT_FALLBACK
) {
1235 ops
->iomap_end(inode
, pos
, PMD_SIZE
, 0, iomap_flags
,
1238 error
= ops
->iomap_end(inode
, pos
, PMD_SIZE
, PMD_SIZE
,
1239 iomap_flags
, &iomap
);
1241 result
= VM_FAULT_FALLBACK
;
1245 put_locked_mapping_entry(mapping
, pgoff
, entry
);
1247 if (result
== VM_FAULT_FALLBACK
) {
1248 split_huge_pmd(vma
, pmd
, address
);
1249 count_vm_event(THP_FAULT_FALLBACK
);
1253 EXPORT_SYMBOL_GPL(dax_iomap_pmd_fault
);
1254 #endif /* CONFIG_FS_DAX_PMD */
1255 #endif /* CONFIG_FS_IOMAP */