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>
21 #include <linux/genhd.h>
22 #include <linux/highmem.h>
23 #include <linux/memcontrol.h>
25 #include <linux/mutex.h>
26 #include <linux/pmem.h>
27 #include <linux/sched.h>
28 #include <linux/uio.h>
29 #include <linux/vmstat.h>
31 int dax_clear_blocks(struct inode
*inode
, sector_t block
, long size
)
33 struct block_device
*bdev
= inode
->i_sb
->s_bdev
;
34 sector_t sector
= block
<< (inode
->i_blkbits
- 9);
42 count
= bdev_direct_access(bdev
, sector
, &addr
, &pfn
, size
);
47 unsigned pgsz
= PAGE_SIZE
- offset_in_page(addr
);
50 clear_pmem((void __pmem
*)addr
, pgsz
);
63 EXPORT_SYMBOL_GPL(dax_clear_blocks
);
65 static long dax_get_addr(struct buffer_head
*bh
, void **addr
, unsigned blkbits
)
68 sector_t sector
= bh
->b_blocknr
<< (blkbits
- 9);
69 return bdev_direct_access(bh
->b_bdev
, sector
, addr
, &pfn
, bh
->b_size
);
72 /* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
73 static void dax_new_buf(void *addr
, unsigned size
, unsigned first
, loff_t pos
,
76 loff_t final
= end
- pos
+ first
; /* The final byte of the buffer */
79 clear_pmem((void __pmem
*)addr
, first
);
81 clear_pmem((void __pmem
*)addr
+ final
, size
- final
);
84 static bool buffer_written(struct buffer_head
*bh
)
86 return buffer_mapped(bh
) && !buffer_unwritten(bh
);
90 * When ext4 encounters a hole, it returns without modifying the buffer_head
91 * which means that we can't trust b_size. To cope with this, we set b_state
92 * to 0 before calling get_block and, if any bit is set, we know we can trust
93 * b_size. Unfortunate, really, since ext4 knows precisely how long a hole is
94 * and would save us time calling get_block repeatedly.
96 static bool buffer_size_valid(struct buffer_head
*bh
)
98 return bh
->b_state
!= 0;
101 static ssize_t
dax_io(struct inode
*inode
, struct iov_iter
*iter
,
102 loff_t start
, loff_t end
, get_block_t get_block
,
103 struct buffer_head
*bh
)
108 loff_t bh_max
= start
;
111 bool need_wmb
= false;
113 if (iov_iter_rw(iter
) != WRITE
)
114 end
= min(end
, i_size_read(inode
));
119 unsigned blkbits
= inode
->i_blkbits
;
120 sector_t block
= pos
>> blkbits
;
121 unsigned first
= pos
- (block
<< blkbits
);
125 bh
->b_size
= PAGE_ALIGN(end
- pos
);
127 retval
= get_block(inode
, block
, bh
,
128 iov_iter_rw(iter
) == WRITE
);
131 if (!buffer_size_valid(bh
))
132 bh
->b_size
= 1 << blkbits
;
133 bh_max
= pos
- first
+ bh
->b_size
;
135 unsigned done
= bh
->b_size
-
136 (bh_max
- (pos
- first
));
137 bh
->b_blocknr
+= done
>> blkbits
;
141 hole
= iov_iter_rw(iter
) != WRITE
&& !buffer_written(bh
);
144 size
= bh
->b_size
- first
;
146 retval
= dax_get_addr(bh
, &addr
, blkbits
);
149 if (buffer_unwritten(bh
) || buffer_new(bh
)) {
150 dax_new_buf(addr
, retval
, first
, pos
,
155 size
= retval
- first
;
157 max
= min(pos
+ size
, end
);
160 if (iov_iter_rw(iter
) == WRITE
) {
161 len
= copy_from_iter_pmem((void __pmem
*)addr
,
165 len
= copy_to_iter(addr
, max
- pos
, iter
);
167 len
= iov_iter_zero(max
- pos
, iter
);
179 return (pos
== start
) ? retval
: pos
- start
;
183 * dax_do_io - Perform I/O to a DAX file
184 * @iocb: The control block for this I/O
185 * @inode: The file which the I/O is directed at
186 * @iter: The addresses to do I/O from or to
187 * @pos: The file offset where the I/O starts
188 * @get_block: The filesystem method used to translate file offsets to blocks
189 * @end_io: A filesystem callback for I/O completion
192 * This function uses the same locking scheme as do_blockdev_direct_IO:
193 * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
194 * caller for writes. For reads, we take and release the i_mutex ourselves.
195 * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
196 * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
199 ssize_t
dax_do_io(struct kiocb
*iocb
, struct inode
*inode
,
200 struct iov_iter
*iter
, loff_t pos
, get_block_t get_block
,
201 dio_iodone_t end_io
, int flags
)
203 struct buffer_head bh
;
204 ssize_t retval
= -EINVAL
;
205 loff_t end
= pos
+ iov_iter_count(iter
);
207 memset(&bh
, 0, sizeof(bh
));
209 if ((flags
& DIO_LOCKING
) && iov_iter_rw(iter
) == READ
) {
210 struct address_space
*mapping
= inode
->i_mapping
;
211 mutex_lock(&inode
->i_mutex
);
212 retval
= filemap_write_and_wait_range(mapping
, pos
, end
- 1);
214 mutex_unlock(&inode
->i_mutex
);
219 /* Protects against truncate */
220 if (!(flags
& DIO_SKIP_DIO_COUNT
))
221 inode_dio_begin(inode
);
223 retval
= dax_io(inode
, iter
, pos
, end
, get_block
, &bh
);
225 if ((flags
& DIO_LOCKING
) && iov_iter_rw(iter
) == READ
)
226 mutex_unlock(&inode
->i_mutex
);
228 if ((retval
> 0) && end_io
)
229 end_io(iocb
, pos
, retval
, bh
.b_private
);
231 if (!(flags
& DIO_SKIP_DIO_COUNT
))
232 inode_dio_end(inode
);
236 EXPORT_SYMBOL_GPL(dax_do_io
);
239 * The user has performed a load from a hole in the file. Allocating
240 * a new page in the file would cause excessive storage usage for
241 * workloads with sparse files. We allocate a page cache page instead.
242 * We'll kick it out of the page cache if it's ever written to,
243 * otherwise it will simply fall out of the page cache under memory
244 * pressure without ever having been dirtied.
246 static int dax_load_hole(struct address_space
*mapping
, struct page
*page
,
247 struct vm_fault
*vmf
)
250 struct inode
*inode
= mapping
->host
;
252 page
= find_or_create_page(mapping
, vmf
->pgoff
,
253 GFP_KERNEL
| __GFP_ZERO
);
256 /* Recheck i_size under page lock to avoid truncate race */
257 size
= (i_size_read(inode
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
258 if (vmf
->pgoff
>= size
) {
260 page_cache_release(page
);
261 return VM_FAULT_SIGBUS
;
265 return VM_FAULT_LOCKED
;
268 static int copy_user_bh(struct page
*to
, struct buffer_head
*bh
,
269 unsigned blkbits
, unsigned long vaddr
)
272 if (dax_get_addr(bh
, &vfrom
, blkbits
) < 0)
274 vto
= kmap_atomic(to
);
275 copy_user_page(vto
, vfrom
, vaddr
, to
);
280 static int dax_insert_mapping(struct inode
*inode
, struct buffer_head
*bh
,
281 struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
283 struct address_space
*mapping
= inode
->i_mapping
;
284 sector_t sector
= bh
->b_blocknr
<< (inode
->i_blkbits
- 9);
285 unsigned long vaddr
= (unsigned long)vmf
->virtual_address
;
291 i_mmap_lock_read(mapping
);
294 * Check truncate didn't happen while we were allocating a block.
295 * If it did, this block may or may not be still allocated to the
296 * file. We can't tell the filesystem to free it because we can't
297 * take i_mutex here. In the worst case, the file still has blocks
298 * allocated past the end of the file.
300 size
= (i_size_read(inode
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
301 if (unlikely(vmf
->pgoff
>= size
)) {
306 error
= bdev_direct_access(bh
->b_bdev
, sector
, &addr
, &pfn
, bh
->b_size
);
309 if (error
< PAGE_SIZE
) {
314 if (buffer_unwritten(bh
) || buffer_new(bh
)) {
315 clear_pmem((void __pmem
*)addr
, PAGE_SIZE
);
319 error
= vm_insert_mixed(vma
, vaddr
, pfn
);
322 i_mmap_unlock_read(mapping
);
328 * __dax_fault - handle a page fault on a DAX file
329 * @vma: The virtual memory area where the fault occurred
330 * @vmf: The description of the fault
331 * @get_block: The filesystem method used to translate file offsets to blocks
333 * When a page fault occurs, filesystems may call this helper in their
334 * fault handler for DAX files. __dax_fault() assumes the caller has done all
335 * the necessary locking for the page fault to proceed successfully.
337 int __dax_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
,
338 get_block_t get_block
, dax_iodone_t complete_unwritten
)
340 struct file
*file
= vma
->vm_file
;
341 struct address_space
*mapping
= file
->f_mapping
;
342 struct inode
*inode
= mapping
->host
;
344 struct buffer_head bh
;
345 unsigned long vaddr
= (unsigned long)vmf
->virtual_address
;
346 unsigned blkbits
= inode
->i_blkbits
;
352 size
= (i_size_read(inode
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
353 if (vmf
->pgoff
>= size
)
354 return VM_FAULT_SIGBUS
;
356 memset(&bh
, 0, sizeof(bh
));
357 block
= (sector_t
)vmf
->pgoff
<< (PAGE_SHIFT
- blkbits
);
358 bh
.b_size
= PAGE_SIZE
;
361 page
= find_get_page(mapping
, vmf
->pgoff
);
363 if (!lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
)) {
364 page_cache_release(page
);
365 return VM_FAULT_RETRY
;
367 if (unlikely(page
->mapping
!= mapping
)) {
369 page_cache_release(page
);
372 size
= (i_size_read(inode
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
373 if (unlikely(vmf
->pgoff
>= size
)) {
375 * We have a struct page covering a hole in the file
376 * from a read fault and we've raced with a truncate
383 error
= get_block(inode
, block
, &bh
, 0);
384 if (!error
&& (bh
.b_size
< PAGE_SIZE
))
385 error
= -EIO
; /* fs corruption? */
389 if (!buffer_mapped(&bh
) && !buffer_unwritten(&bh
) && !vmf
->cow_page
) {
390 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
391 error
= get_block(inode
, block
, &bh
, 1);
392 count_vm_event(PGMAJFAULT
);
393 mem_cgroup_count_vm_event(vma
->vm_mm
, PGMAJFAULT
);
394 major
= VM_FAULT_MAJOR
;
395 if (!error
&& (bh
.b_size
< PAGE_SIZE
))
400 return dax_load_hole(mapping
, page
, vmf
);
405 struct page
*new_page
= vmf
->cow_page
;
406 if (buffer_written(&bh
))
407 error
= copy_user_bh(new_page
, &bh
, blkbits
, vaddr
);
409 clear_user_highpage(new_page
, vaddr
);
414 i_mmap_lock_read(mapping
);
415 /* Check we didn't race with truncate */
416 size
= (i_size_read(inode
) + PAGE_SIZE
- 1) >>
418 if (vmf
->pgoff
>= size
) {
419 i_mmap_unlock_read(mapping
);
424 return VM_FAULT_LOCKED
;
427 /* Check we didn't race with a read fault installing a new page */
429 page
= find_lock_page(mapping
, vmf
->pgoff
);
432 unmap_mapping_range(mapping
, vmf
->pgoff
<< PAGE_SHIFT
,
434 delete_from_page_cache(page
);
436 page_cache_release(page
);
440 * If we successfully insert the new mapping over an unwritten extent,
441 * we need to ensure we convert the unwritten extent. If there is an
442 * error inserting the mapping, the filesystem needs to leave it as
443 * unwritten to prevent exposure of the stale underlying data to
444 * userspace, but we still need to call the completion function so
445 * the private resources on the mapping buffer can be released. We
446 * indicate what the callback should do via the uptodate variable, same
447 * as for normal BH based IO completions.
449 error
= dax_insert_mapping(inode
, &bh
, vma
, vmf
);
450 if (buffer_unwritten(&bh
))
451 complete_unwritten(&bh
, !error
);
454 if (error
== -ENOMEM
)
455 return VM_FAULT_OOM
| major
;
456 /* -EBUSY is fine, somebody else faulted on the same PTE */
457 if ((error
< 0) && (error
!= -EBUSY
))
458 return VM_FAULT_SIGBUS
| major
;
459 return VM_FAULT_NOPAGE
| major
;
464 page_cache_release(page
);
468 EXPORT_SYMBOL(__dax_fault
);
471 * dax_fault - handle a page fault on a DAX file
472 * @vma: The virtual memory area where the fault occurred
473 * @vmf: The description of the fault
474 * @get_block: The filesystem method used to translate file offsets to blocks
476 * When a page fault occurs, filesystems may call this helper in their
477 * fault handler for DAX files.
479 int dax_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
,
480 get_block_t get_block
, dax_iodone_t complete_unwritten
)
483 struct super_block
*sb
= file_inode(vma
->vm_file
)->i_sb
;
485 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
486 sb_start_pagefault(sb
);
487 file_update_time(vma
->vm_file
);
489 result
= __dax_fault(vma
, vmf
, get_block
, complete_unwritten
);
490 if (vmf
->flags
& FAULT_FLAG_WRITE
)
491 sb_end_pagefault(sb
);
495 EXPORT_SYMBOL_GPL(dax_fault
);
498 * dax_pfn_mkwrite - handle first write to DAX page
499 * @vma: The virtual memory area where the fault occurred
500 * @vmf: The description of the fault
503 int dax_pfn_mkwrite(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
505 struct super_block
*sb
= file_inode(vma
->vm_file
)->i_sb
;
507 sb_start_pagefault(sb
);
508 file_update_time(vma
->vm_file
);
509 sb_end_pagefault(sb
);
510 return VM_FAULT_NOPAGE
;
512 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite
);
515 * dax_zero_page_range - zero a range within a page of a DAX file
516 * @inode: The file being truncated
517 * @from: The file offset that is being truncated to
518 * @length: The number of bytes to zero
519 * @get_block: The filesystem method used to translate file offsets to blocks
521 * This function can be called by a filesystem when it is zeroing part of a
522 * page in a DAX file. This is intended for hole-punch operations. If
523 * you are truncating a file, the helper function dax_truncate_page() may be
526 * We work in terms of PAGE_CACHE_SIZE here for commonality with
527 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
528 * took care of disposing of the unnecessary blocks. Even if the filesystem
529 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
530 * since the file might be mmapped.
532 int dax_zero_page_range(struct inode
*inode
, loff_t from
, unsigned length
,
533 get_block_t get_block
)
535 struct buffer_head bh
;
536 pgoff_t index
= from
>> PAGE_CACHE_SHIFT
;
537 unsigned offset
= from
& (PAGE_CACHE_SIZE
-1);
540 /* Block boundary? Nothing to do */
543 BUG_ON((offset
+ length
) > PAGE_CACHE_SIZE
);
545 memset(&bh
, 0, sizeof(bh
));
546 bh
.b_size
= PAGE_CACHE_SIZE
;
547 err
= get_block(inode
, index
, &bh
, 0);
550 if (buffer_written(&bh
)) {
552 err
= dax_get_addr(&bh
, &addr
, inode
->i_blkbits
);
555 clear_pmem((void __pmem
*)addr
+ offset
, length
);
561 EXPORT_SYMBOL_GPL(dax_zero_page_range
);
564 * dax_truncate_page - handle a partial page being truncated in a DAX file
565 * @inode: The file being truncated
566 * @from: The file offset that is being truncated to
567 * @get_block: The filesystem method used to translate file offsets to blocks
569 * Similar to block_truncate_page(), this function can be called by a
570 * filesystem when it is truncating a DAX file to handle the partial page.
572 * We work in terms of PAGE_CACHE_SIZE here for commonality with
573 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
574 * took care of disposing of the unnecessary blocks. Even if the filesystem
575 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
576 * since the file might be mmapped.
578 int dax_truncate_page(struct inode
*inode
, loff_t from
, get_block_t get_block
)
580 unsigned length
= PAGE_CACHE_ALIGN(from
) - from
;
581 return dax_zero_page_range(inode
, from
, length
, get_block
);
583 EXPORT_SYMBOL_GPL(dax_truncate_page
);