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/sched.h>
27 #include <linux/uio.h>
28 #include <linux/vmstat.h>
30 int dax_clear_blocks(struct inode *inode, sector_t block, long size)
32 struct block_device *bdev = inode->i_sb->s_bdev;
33 sector_t sector = block << (inode->i_blkbits - 9);
41 count = bdev_direct_access(bdev, sector, &addr, &pfn, size);
46 unsigned pgsz = PAGE_SIZE - offset_in_page(addr);
50 memset(addr, 0, pgsz);
64 EXPORT_SYMBOL_GPL(dax_clear_blocks);
66 static long dax_get_addr(struct buffer_head *bh, void **addr, unsigned blkbits)
69 sector_t sector = bh->b_blocknr << (blkbits - 9);
70 return bdev_direct_access(bh->b_bdev, sector, addr, &pfn, bh->b_size);
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 memset(addr, 0, first);
81 memset(addr + final, 0, 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;
112 if (iov_iter_rw(iter) != WRITE)
113 end = min(end, i_size_read(inode));
118 unsigned blkbits = inode->i_blkbits;
119 sector_t block = pos >> blkbits;
120 unsigned first = pos - (block << blkbits);
124 bh->b_size = PAGE_ALIGN(end - pos);
126 retval = get_block(inode, block, bh,
127 iov_iter_rw(iter) == WRITE);
130 if (!buffer_size_valid(bh))
131 bh->b_size = 1 << blkbits;
132 bh_max = pos - first + bh->b_size;
134 unsigned done = bh->b_size -
135 (bh_max - (pos - first));
136 bh->b_blocknr += done >> blkbits;
140 hole = iov_iter_rw(iter) != WRITE && !buffer_written(bh);
143 size = bh->b_size - first;
145 retval = dax_get_addr(bh, &addr, blkbits);
148 if (buffer_unwritten(bh) || buffer_new(bh))
149 dax_new_buf(addr, retval, first, pos,
152 size = retval - first;
154 max = min(pos + size, end);
157 if (iov_iter_rw(iter) == WRITE)
158 len = copy_from_iter_nocache(addr, max - pos, iter);
160 len = copy_to_iter(addr, max - pos, iter);
162 len = iov_iter_zero(max - pos, iter);
171 return (pos == start) ? retval : pos - start;
175 * dax_do_io - Perform I/O to a DAX file
176 * @iocb: The control block for this I/O
177 * @inode: The file which the I/O is directed at
178 * @iter: The addresses to do I/O from or to
179 * @pos: The file offset where the I/O starts
180 * @get_block: The filesystem method used to translate file offsets to blocks
181 * @end_io: A filesystem callback for I/O completion
184 * This function uses the same locking scheme as do_blockdev_direct_IO:
185 * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
186 * caller for writes. For reads, we take and release the i_mutex ourselves.
187 * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
188 * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
191 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
192 struct iov_iter *iter, loff_t pos, get_block_t get_block,
193 dio_iodone_t end_io, int flags)
195 struct buffer_head bh;
196 ssize_t retval = -EINVAL;
197 loff_t end = pos + iov_iter_count(iter);
199 memset(&bh, 0, sizeof(bh));
201 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) {
202 struct address_space *mapping = inode->i_mapping;
203 mutex_lock(&inode->i_mutex);
204 retval = filemap_write_and_wait_range(mapping, pos, end - 1);
206 mutex_unlock(&inode->i_mutex);
211 /* Protects against truncate */
212 if (!(flags & DIO_SKIP_DIO_COUNT))
213 inode_dio_begin(inode);
215 retval = dax_io(inode, iter, pos, end, get_block, &bh);
217 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
218 mutex_unlock(&inode->i_mutex);
220 if ((retval > 0) && end_io)
221 end_io(iocb, pos, retval, bh.b_private);
223 if (!(flags & DIO_SKIP_DIO_COUNT))
224 inode_dio_end(inode);
228 EXPORT_SYMBOL_GPL(dax_do_io);
231 * The user has performed a load from a hole in the file. Allocating
232 * a new page in the file would cause excessive storage usage for
233 * workloads with sparse files. We allocate a page cache page instead.
234 * We'll kick it out of the page cache if it's ever written to,
235 * otherwise it will simply fall out of the page cache under memory
236 * pressure without ever having been dirtied.
238 static int dax_load_hole(struct address_space *mapping, struct page *page,
239 struct vm_fault *vmf)
242 struct inode *inode = mapping->host;
244 page = find_or_create_page(mapping, vmf->pgoff,
245 GFP_KERNEL | __GFP_ZERO);
248 /* Recheck i_size under page lock to avoid truncate race */
249 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
250 if (vmf->pgoff >= size) {
252 page_cache_release(page);
253 return VM_FAULT_SIGBUS;
257 return VM_FAULT_LOCKED;
260 static int copy_user_bh(struct page *to, struct buffer_head *bh,
261 unsigned blkbits, unsigned long vaddr)
264 if (dax_get_addr(bh, &vfrom, blkbits) < 0)
266 vto = kmap_atomic(to);
267 copy_user_page(vto, vfrom, vaddr, to);
272 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
273 struct vm_area_struct *vma, struct vm_fault *vmf)
275 sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
276 unsigned long vaddr = (unsigned long)vmf->virtual_address;
283 * Check truncate didn't happen while we were allocating a block.
284 * If it did, this block may or may not be still allocated to the
285 * file. We can't tell the filesystem to free it because we can't
286 * take i_mutex here. In the worst case, the file still has blocks
287 * allocated past the end of the file.
289 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
290 if (unlikely(vmf->pgoff >= size)) {
295 error = bdev_direct_access(bh->b_bdev, sector, &addr, &pfn, bh->b_size);
298 if (error < PAGE_SIZE) {
303 if (buffer_unwritten(bh) || buffer_new(bh))
306 error = vm_insert_mixed(vma, vaddr, pfn);
313 * __dax_fault - handle a page fault on a DAX file
314 * @vma: The virtual memory area where the fault occurred
315 * @vmf: The description of the fault
316 * @get_block: The filesystem method used to translate file offsets to blocks
317 * @complete_unwritten: The filesystem method used to convert unwritten blocks
318 * to written so the data written to them is exposed. This is required for
319 * required by write faults for filesystems that will return unwritten
320 * extent mappings from @get_block, but it is optional for reads as
321 * dax_insert_mapping() will always zero unwritten blocks. If the fs does
322 * not support unwritten extents, the it should pass NULL.
324 * When a page fault occurs, filesystems may call this helper in their
325 * fault handler for DAX files. __dax_fault() assumes the caller has done all
326 * the necessary locking for the page fault to proceed successfully.
328 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
329 get_block_t get_block, dax_iodone_t complete_unwritten)
331 struct file *file = vma->vm_file;
332 struct address_space *mapping = file->f_mapping;
333 struct inode *inode = mapping->host;
335 struct buffer_head bh;
336 unsigned long vaddr = (unsigned long)vmf->virtual_address;
337 unsigned blkbits = inode->i_blkbits;
343 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
344 if (vmf->pgoff >= size)
345 return VM_FAULT_SIGBUS;
347 memset(&bh, 0, sizeof(bh));
348 block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
349 bh.b_size = PAGE_SIZE;
352 page = find_get_page(mapping, vmf->pgoff);
354 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
355 page_cache_release(page);
356 return VM_FAULT_RETRY;
358 if (unlikely(page->mapping != mapping)) {
360 page_cache_release(page);
363 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
364 if (unlikely(vmf->pgoff >= size)) {
366 * We have a struct page covering a hole in the file
367 * from a read fault and we've raced with a truncate
373 i_mmap_lock_write(mapping);
376 error = get_block(inode, block, &bh, 0);
377 if (!error && (bh.b_size < PAGE_SIZE))
378 error = -EIO; /* fs corruption? */
382 if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
383 if (vmf->flags & FAULT_FLAG_WRITE) {
384 error = get_block(inode, block, &bh, 1);
385 count_vm_event(PGMAJFAULT);
386 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
387 major = VM_FAULT_MAJOR;
388 if (!error && (bh.b_size < PAGE_SIZE))
393 i_mmap_unlock_write(mapping);
394 return dax_load_hole(mapping, page, vmf);
399 struct page *new_page = vmf->cow_page;
400 if (buffer_written(&bh))
401 error = copy_user_bh(new_page, &bh, blkbits, vaddr);
403 clear_user_highpage(new_page, vaddr);
408 /* Check we didn't race with truncate */
409 size = (i_size_read(inode) + PAGE_SIZE - 1) >>
411 if (vmf->pgoff >= size) {
416 return VM_FAULT_LOCKED;
419 /* Check we didn't race with a read fault installing a new page */
421 page = find_lock_page(mapping, vmf->pgoff);
424 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
426 delete_from_page_cache(page);
428 page_cache_release(page);
432 * If we successfully insert the new mapping over an unwritten extent,
433 * we need to ensure we convert the unwritten extent. If there is an
434 * error inserting the mapping, the filesystem needs to leave it as
435 * unwritten to prevent exposure of the stale underlying data to
436 * userspace, but we still need to call the completion function so
437 * the private resources on the mapping buffer can be released. We
438 * indicate what the callback should do via the uptodate variable, same
439 * as for normal BH based IO completions.
441 error = dax_insert_mapping(inode, &bh, vma, vmf);
442 if (buffer_unwritten(&bh)) {
443 if (complete_unwritten)
444 complete_unwritten(&bh, !error);
446 WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
450 i_mmap_unlock_write(mapping);
452 if (error == -ENOMEM)
453 return VM_FAULT_OOM | major;
454 /* -EBUSY is fine, somebody else faulted on the same PTE */
455 if ((error < 0) && (error != -EBUSY))
456 return VM_FAULT_SIGBUS | major;
457 return VM_FAULT_NOPAGE | major;
462 page_cache_release(page);
464 i_mmap_unlock_write(mapping);
469 EXPORT_SYMBOL(__dax_fault);
472 * dax_fault - handle a page fault on a DAX file
473 * @vma: The virtual memory area where the fault occurred
474 * @vmf: The description of the fault
475 * @get_block: The filesystem method used to translate file offsets to blocks
477 * When a page fault occurs, filesystems may call this helper in their
478 * fault handler for DAX files.
480 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
481 get_block_t get_block, dax_iodone_t complete_unwritten)
484 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
486 if (vmf->flags & FAULT_FLAG_WRITE) {
487 sb_start_pagefault(sb);
488 file_update_time(vma->vm_file);
490 result = __dax_fault(vma, vmf, get_block, complete_unwritten);
491 if (vmf->flags & FAULT_FLAG_WRITE)
492 sb_end_pagefault(sb);
496 EXPORT_SYMBOL_GPL(dax_fault);
498 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
500 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
501 * more often than one might expect in the below function.
503 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
505 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
506 pmd_t *pmd, unsigned int flags, get_block_t get_block,
507 dax_iodone_t complete_unwritten)
509 struct file *file = vma->vm_file;
510 struct address_space *mapping = file->f_mapping;
511 struct inode *inode = mapping->host;
512 struct buffer_head bh;
513 unsigned blkbits = inode->i_blkbits;
514 unsigned long pmd_addr = address & PMD_MASK;
515 bool write = flags & FAULT_FLAG_WRITE;
519 sector_t block, sector;
523 /* Fall back to PTEs if we're going to COW */
524 if (write && !(vma->vm_flags & VM_SHARED))
525 return VM_FAULT_FALLBACK;
526 /* If the PMD would extend outside the VMA */
527 if (pmd_addr < vma->vm_start)
528 return VM_FAULT_FALLBACK;
529 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
530 return VM_FAULT_FALLBACK;
532 pgoff = linear_page_index(vma, pmd_addr);
533 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
535 return VM_FAULT_SIGBUS;
536 /* If the PMD would cover blocks out of the file */
537 if ((pgoff | PG_PMD_COLOUR) >= size)
538 return VM_FAULT_FALLBACK;
540 memset(&bh, 0, sizeof(bh));
541 block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
543 bh.b_size = PMD_SIZE;
544 i_mmap_lock_write(mapping);
545 length = get_block(inode, block, &bh, write);
547 return VM_FAULT_SIGBUS;
550 * If the filesystem isn't willing to tell us the length of a hole,
551 * just fall back to PTEs. Calling get_block 512 times in a loop
554 if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE)
557 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
559 for (i = 0; i < PTRS_PER_PMD; i++)
560 clear_page(kaddr + i * PAGE_SIZE);
561 count_vm_event(PGMAJFAULT);
562 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
563 result |= VM_FAULT_MAJOR;
567 * If we allocated new storage, make sure no process has any
568 * zero pages covering this hole
570 if (buffer_new(&bh)) {
571 i_mmap_unlock_write(mapping);
572 unmap_mapping_range(mapping, pgoff << PAGE_SHIFT, PMD_SIZE, 0);
573 i_mmap_lock_write(mapping);
577 * If a truncate happened while we were allocating blocks, we may
578 * leave blocks allocated to the file that are beyond EOF. We can't
579 * take i_mutex here, so just leave them hanging; they'll be freed
580 * when the file is deleted.
582 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
584 result = VM_FAULT_SIGBUS;
587 if ((pgoff | PG_PMD_COLOUR) >= size)
590 if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
593 struct page *zero_page = get_huge_zero_page();
595 if (unlikely(!zero_page))
598 ptl = pmd_lock(vma->vm_mm, pmd);
599 if (!pmd_none(*pmd)) {
604 entry = mk_pmd(zero_page, vma->vm_page_prot);
605 entry = pmd_mkhuge(entry);
606 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
607 result = VM_FAULT_NOPAGE;
610 sector = bh.b_blocknr << (blkbits - 9);
611 length = bdev_direct_access(bh.b_bdev, sector, &kaddr, &pfn,
614 result = VM_FAULT_SIGBUS;
617 if ((length < PMD_SIZE) || (pfn & PG_PMD_COLOUR))
620 result |= vmf_insert_pfn_pmd(vma, address, pmd, pfn, write);
624 if (buffer_unwritten(&bh))
625 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
627 i_mmap_unlock_write(mapping);
632 count_vm_event(THP_FAULT_FALLBACK);
633 result = VM_FAULT_FALLBACK;
636 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
639 * dax_pmd_fault - handle a PMD fault on a DAX file
640 * @vma: The virtual memory area where the fault occurred
641 * @vmf: The description of the fault
642 * @get_block: The filesystem method used to translate file offsets to blocks
644 * When a page fault occurs, filesystems may call this helper in their
645 * pmd_fault handler for DAX files.
647 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
648 pmd_t *pmd, unsigned int flags, get_block_t get_block,
649 dax_iodone_t complete_unwritten)
652 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
654 if (flags & FAULT_FLAG_WRITE) {
655 sb_start_pagefault(sb);
656 file_update_time(vma->vm_file);
658 result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
660 if (flags & FAULT_FLAG_WRITE)
661 sb_end_pagefault(sb);
665 EXPORT_SYMBOL_GPL(dax_pmd_fault);
666 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
669 * dax_pfn_mkwrite - handle first write to DAX page
670 * @vma: The virtual memory area where the fault occurred
671 * @vmf: The description of the fault
674 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
676 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
678 sb_start_pagefault(sb);
679 file_update_time(vma->vm_file);
680 sb_end_pagefault(sb);
681 return VM_FAULT_NOPAGE;
683 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
686 * dax_zero_page_range - zero a range within a page of a DAX file
687 * @inode: The file being truncated
688 * @from: The file offset that is being truncated to
689 * @length: The number of bytes to zero
690 * @get_block: The filesystem method used to translate file offsets to blocks
692 * This function can be called by a filesystem when it is zeroing part of a
693 * page in a DAX file. This is intended for hole-punch operations. If
694 * you are truncating a file, the helper function dax_truncate_page() may be
697 * We work in terms of PAGE_CACHE_SIZE here for commonality with
698 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
699 * took care of disposing of the unnecessary blocks. Even if the filesystem
700 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
701 * since the file might be mmapped.
703 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
704 get_block_t get_block)
706 struct buffer_head bh;
707 pgoff_t index = from >> PAGE_CACHE_SHIFT;
708 unsigned offset = from & (PAGE_CACHE_SIZE-1);
711 /* Block boundary? Nothing to do */
714 BUG_ON((offset + length) > PAGE_CACHE_SIZE);
716 memset(&bh, 0, sizeof(bh));
717 bh.b_size = PAGE_CACHE_SIZE;
718 err = get_block(inode, index, &bh, 0);
721 if (buffer_written(&bh)) {
723 err = dax_get_addr(&bh, &addr, inode->i_blkbits);
726 memset(addr + offset, 0, length);
731 EXPORT_SYMBOL_GPL(dax_zero_page_range);
734 * dax_truncate_page - handle a partial page being truncated in a DAX file
735 * @inode: The file being truncated
736 * @from: The file offset that is being truncated to
737 * @get_block: The filesystem method used to translate file offsets to blocks
739 * Similar to block_truncate_page(), this function can be called by a
740 * filesystem when it is truncating a DAX file to handle the partial page.
742 * We work in terms of PAGE_CACHE_SIZE here for commonality with
743 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
744 * took care of disposing of the unnecessary blocks. Even if the filesystem
745 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
746 * since the file might be mmapped.
748 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
750 unsigned length = PAGE_CACHE_ALIGN(from) - from;
751 return dax_zero_page_range(inode, from, length, get_block);
753 EXPORT_SYMBOL_GPL(dax_truncate_page);