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/pmem.h>
28 #include <linux/sched.h>
29 #include <linux/uio.h>
30 #include <linux/vmstat.h>
33 * dax_clear_blocks() is called from within transaction context from XFS,
34 * and hence this means the stack from this point must follow GFP_NOFS
35 * semantics for all operations.
37 int dax_clear_blocks(struct inode *inode, sector_t block, long size)
39 struct block_device *bdev = inode->i_sb->s_bdev;
40 sector_t sector = block << (inode->i_blkbits - 9);
48 count = bdev_direct_access(bdev, sector, &addr, &pfn, size);
53 unsigned pgsz = PAGE_SIZE - offset_in_page(addr);
56 clear_pmem(addr, pgsz);
69 EXPORT_SYMBOL_GPL(dax_clear_blocks);
71 static long dax_get_addr(struct buffer_head *bh, void __pmem **addr,
75 sector_t sector = bh->b_blocknr << (blkbits - 9);
76 return bdev_direct_access(bh->b_bdev, sector, addr, &pfn, bh->b_size);
79 /* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
80 static void dax_new_buf(void __pmem *addr, unsigned size, unsigned first,
81 loff_t pos, loff_t end)
83 loff_t final = end - pos + first; /* The final byte of the buffer */
86 clear_pmem(addr, first);
88 clear_pmem(addr + final, size - final);
91 static bool buffer_written(struct buffer_head *bh)
93 return buffer_mapped(bh) && !buffer_unwritten(bh);
97 * When ext4 encounters a hole, it returns without modifying the buffer_head
98 * which means that we can't trust b_size. To cope with this, we set b_state
99 * to 0 before calling get_block and, if any bit is set, we know we can trust
100 * b_size. Unfortunate, really, since ext4 knows precisely how long a hole is
101 * and would save us time calling get_block repeatedly.
103 static bool buffer_size_valid(struct buffer_head *bh)
105 return bh->b_state != 0;
108 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
109 loff_t start, loff_t end, get_block_t get_block,
110 struct buffer_head *bh)
115 loff_t bh_max = start;
118 bool need_wmb = false;
120 if (iov_iter_rw(iter) != WRITE)
121 end = min(end, i_size_read(inode));
126 unsigned blkbits = inode->i_blkbits;
127 long page = pos >> PAGE_SHIFT;
128 sector_t block = page << (PAGE_SHIFT - blkbits);
129 unsigned first = pos - (block << blkbits);
133 bh->b_size = PAGE_ALIGN(end - pos);
135 retval = get_block(inode, block, bh,
136 iov_iter_rw(iter) == WRITE);
139 if (!buffer_size_valid(bh))
140 bh->b_size = 1 << blkbits;
141 bh_max = pos - first + bh->b_size;
143 unsigned done = bh->b_size -
144 (bh_max - (pos - first));
145 bh->b_blocknr += done >> blkbits;
149 hole = iov_iter_rw(iter) != WRITE && !buffer_written(bh);
152 size = bh->b_size - first;
154 retval = dax_get_addr(bh, &addr, blkbits);
157 if (buffer_unwritten(bh) || buffer_new(bh)) {
158 dax_new_buf(addr, retval, first, pos,
163 size = retval - first;
165 max = min(pos + size, end);
168 if (iov_iter_rw(iter) == WRITE) {
169 len = copy_from_iter_pmem(addr, max - pos, iter);
172 len = copy_to_iter((void __force *)addr, max - pos,
175 len = iov_iter_zero(max - pos, iter);
189 return (pos == start) ? retval : pos - start;
193 * dax_do_io - Perform I/O to a DAX file
194 * @iocb: The control block for this I/O
195 * @inode: The file which the I/O is directed at
196 * @iter: The addresses to do I/O from or to
197 * @pos: The file offset where the I/O starts
198 * @get_block: The filesystem method used to translate file offsets to blocks
199 * @end_io: A filesystem callback for I/O completion
202 * This function uses the same locking scheme as do_blockdev_direct_IO:
203 * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
204 * caller for writes. For reads, we take and release the i_mutex ourselves.
205 * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
206 * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
209 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
210 struct iov_iter *iter, loff_t pos, get_block_t get_block,
211 dio_iodone_t end_io, int flags)
213 struct buffer_head bh;
214 ssize_t retval = -EINVAL;
215 loff_t end = pos + iov_iter_count(iter);
217 memset(&bh, 0, sizeof(bh));
219 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) {
220 struct address_space *mapping = inode->i_mapping;
221 mutex_lock(&inode->i_mutex);
222 retval = filemap_write_and_wait_range(mapping, pos, end - 1);
224 mutex_unlock(&inode->i_mutex);
229 /* Protects against truncate */
230 if (!(flags & DIO_SKIP_DIO_COUNT))
231 inode_dio_begin(inode);
233 retval = dax_io(inode, iter, pos, end, get_block, &bh);
235 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
236 mutex_unlock(&inode->i_mutex);
238 if ((retval > 0) && end_io)
239 end_io(iocb, pos, retval, bh.b_private);
241 if (!(flags & DIO_SKIP_DIO_COUNT))
242 inode_dio_end(inode);
246 EXPORT_SYMBOL_GPL(dax_do_io);
249 * The user has performed a load from a hole in the file. Allocating
250 * a new page in the file would cause excessive storage usage for
251 * workloads with sparse files. We allocate a page cache page instead.
252 * We'll kick it out of the page cache if it's ever written to,
253 * otherwise it will simply fall out of the page cache under memory
254 * pressure without ever having been dirtied.
256 static int dax_load_hole(struct address_space *mapping, struct page *page,
257 struct vm_fault *vmf)
260 struct inode *inode = mapping->host;
262 page = find_or_create_page(mapping, vmf->pgoff,
263 GFP_KERNEL | __GFP_ZERO);
266 /* Recheck i_size under page lock to avoid truncate race */
267 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
268 if (vmf->pgoff >= size) {
270 page_cache_release(page);
271 return VM_FAULT_SIGBUS;
275 return VM_FAULT_LOCKED;
278 static int copy_user_bh(struct page *to, struct buffer_head *bh,
279 unsigned blkbits, unsigned long vaddr)
284 if (dax_get_addr(bh, &vfrom, blkbits) < 0)
286 vto = kmap_atomic(to);
287 copy_user_page(vto, (void __force *)vfrom, vaddr, to);
292 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
293 struct vm_area_struct *vma, struct vm_fault *vmf)
295 struct address_space *mapping = inode->i_mapping;
296 sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
297 unsigned long vaddr = (unsigned long)vmf->virtual_address;
303 i_mmap_lock_read(mapping);
306 * Check truncate didn't happen while we were allocating a block.
307 * If it did, this block may or may not be still allocated to the
308 * file. We can't tell the filesystem to free it because we can't
309 * take i_mutex here. In the worst case, the file still has blocks
310 * allocated past the end of the file.
312 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
313 if (unlikely(vmf->pgoff >= size)) {
318 error = bdev_direct_access(bh->b_bdev, sector, &addr, &pfn, bh->b_size);
321 if (error < PAGE_SIZE) {
326 if (buffer_unwritten(bh) || buffer_new(bh)) {
327 clear_pmem(addr, PAGE_SIZE);
331 error = vm_insert_mixed(vma, vaddr, pfn);
334 i_mmap_unlock_read(mapping);
340 * __dax_fault - handle a page fault on a DAX file
341 * @vma: The virtual memory area where the fault occurred
342 * @vmf: The description of the fault
343 * @get_block: The filesystem method used to translate file offsets to blocks
344 * @complete_unwritten: The filesystem method used to convert unwritten blocks
345 * to written so the data written to them is exposed. This is required for
346 * required by write faults for filesystems that will return unwritten
347 * extent mappings from @get_block, but it is optional for reads as
348 * dax_insert_mapping() will always zero unwritten blocks. If the fs does
349 * not support unwritten extents, the it should pass NULL.
351 * When a page fault occurs, filesystems may call this helper in their
352 * fault handler for DAX files. __dax_fault() assumes the caller has done all
353 * the necessary locking for the page fault to proceed successfully.
355 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
356 get_block_t get_block, dax_iodone_t complete_unwritten)
358 struct file *file = vma->vm_file;
359 struct address_space *mapping = file->f_mapping;
360 struct inode *inode = mapping->host;
362 struct buffer_head bh;
363 unsigned long vaddr = (unsigned long)vmf->virtual_address;
364 unsigned blkbits = inode->i_blkbits;
370 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
371 if (vmf->pgoff >= size)
372 return VM_FAULT_SIGBUS;
374 memset(&bh, 0, sizeof(bh));
375 block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
376 bh.b_size = PAGE_SIZE;
379 page = find_get_page(mapping, vmf->pgoff);
381 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
382 page_cache_release(page);
383 return VM_FAULT_RETRY;
385 if (unlikely(page->mapping != mapping)) {
387 page_cache_release(page);
390 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
391 if (unlikely(vmf->pgoff >= size)) {
393 * We have a struct page covering a hole in the file
394 * from a read fault and we've raced with a truncate
401 error = get_block(inode, block, &bh, 0);
402 if (!error && (bh.b_size < PAGE_SIZE))
403 error = -EIO; /* fs corruption? */
407 if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
408 if (vmf->flags & FAULT_FLAG_WRITE) {
409 error = get_block(inode, block, &bh, 1);
410 count_vm_event(PGMAJFAULT);
411 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
412 major = VM_FAULT_MAJOR;
413 if (!error && (bh.b_size < PAGE_SIZE))
418 return dax_load_hole(mapping, page, vmf);
423 struct page *new_page = vmf->cow_page;
424 if (buffer_written(&bh))
425 error = copy_user_bh(new_page, &bh, blkbits, vaddr);
427 clear_user_highpage(new_page, vaddr);
432 i_mmap_lock_read(mapping);
433 /* Check we didn't race with truncate */
434 size = (i_size_read(inode) + PAGE_SIZE - 1) >>
436 if (vmf->pgoff >= size) {
437 i_mmap_unlock_read(mapping);
442 return VM_FAULT_LOCKED;
445 /* Check we didn't race with a read fault installing a new page */
447 page = find_lock_page(mapping, vmf->pgoff);
450 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
452 delete_from_page_cache(page);
454 page_cache_release(page);
458 * If we successfully insert the new mapping over an unwritten extent,
459 * we need to ensure we convert the unwritten extent. If there is an
460 * error inserting the mapping, the filesystem needs to leave it as
461 * unwritten to prevent exposure of the stale underlying data to
462 * userspace, but we still need to call the completion function so
463 * the private resources on the mapping buffer can be released. We
464 * indicate what the callback should do via the uptodate variable, same
465 * as for normal BH based IO completions.
467 error = dax_insert_mapping(inode, &bh, vma, vmf);
468 if (buffer_unwritten(&bh)) {
469 if (complete_unwritten)
470 complete_unwritten(&bh, !error);
472 WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
476 if (error == -ENOMEM)
477 return VM_FAULT_OOM | major;
478 /* -EBUSY is fine, somebody else faulted on the same PTE */
479 if ((error < 0) && (error != -EBUSY))
480 return VM_FAULT_SIGBUS | major;
481 return VM_FAULT_NOPAGE | major;
486 page_cache_release(page);
490 EXPORT_SYMBOL(__dax_fault);
493 * dax_fault - handle a page fault on a DAX file
494 * @vma: The virtual memory area where the fault occurred
495 * @vmf: The description of the fault
496 * @get_block: The filesystem method used to translate file offsets to blocks
498 * When a page fault occurs, filesystems may call this helper in their
499 * fault handler for DAX files.
501 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
502 get_block_t get_block, dax_iodone_t complete_unwritten)
505 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
507 if (vmf->flags & FAULT_FLAG_WRITE) {
508 sb_start_pagefault(sb);
509 file_update_time(vma->vm_file);
511 result = __dax_fault(vma, vmf, get_block, complete_unwritten);
512 if (vmf->flags & FAULT_FLAG_WRITE)
513 sb_end_pagefault(sb);
517 EXPORT_SYMBOL_GPL(dax_fault);
519 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
521 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
522 * more often than one might expect in the below function.
524 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
526 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
527 pmd_t *pmd, unsigned int flags, get_block_t get_block,
528 dax_iodone_t complete_unwritten)
530 struct file *file = vma->vm_file;
531 struct address_space *mapping = file->f_mapping;
532 struct inode *inode = mapping->host;
533 struct buffer_head bh;
534 unsigned blkbits = inode->i_blkbits;
535 unsigned long pmd_addr = address & PMD_MASK;
536 bool write = flags & FAULT_FLAG_WRITE;
540 sector_t block, sector;
544 /* dax pmd mappings are broken wrt gup and fork */
545 if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
546 return VM_FAULT_FALLBACK;
548 /* Fall back to PTEs if we're going to COW */
549 if (write && !(vma->vm_flags & VM_SHARED))
550 return VM_FAULT_FALLBACK;
551 /* If the PMD would extend outside the VMA */
552 if (pmd_addr < vma->vm_start)
553 return VM_FAULT_FALLBACK;
554 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
555 return VM_FAULT_FALLBACK;
557 pgoff = linear_page_index(vma, pmd_addr);
558 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
560 return VM_FAULT_SIGBUS;
561 /* If the PMD would cover blocks out of the file */
562 if ((pgoff | PG_PMD_COLOUR) >= size)
563 return VM_FAULT_FALLBACK;
565 memset(&bh, 0, sizeof(bh));
566 block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
568 bh.b_size = PMD_SIZE;
569 length = get_block(inode, block, &bh, write);
571 return VM_FAULT_SIGBUS;
572 i_mmap_lock_read(mapping);
575 * If the filesystem isn't willing to tell us the length of a hole,
576 * just fall back to PTEs. Calling get_block 512 times in a loop
579 if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE)
583 * If we allocated new storage, make sure no process has any
584 * zero pages covering this hole
586 if (buffer_new(&bh)) {
587 i_mmap_unlock_read(mapping);
588 unmap_mapping_range(mapping, pgoff << PAGE_SHIFT, PMD_SIZE, 0);
589 i_mmap_lock_read(mapping);
593 * If a truncate happened while we were allocating blocks, we may
594 * leave blocks allocated to the file that are beyond EOF. We can't
595 * take i_mutex here, so just leave them hanging; they'll be freed
596 * when the file is deleted.
598 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
600 result = VM_FAULT_SIGBUS;
603 if ((pgoff | PG_PMD_COLOUR) >= size)
606 if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
609 struct page *zero_page = get_huge_zero_page();
611 if (unlikely(!zero_page))
614 ptl = pmd_lock(vma->vm_mm, pmd);
615 if (!pmd_none(*pmd)) {
620 entry = mk_pmd(zero_page, vma->vm_page_prot);
621 entry = pmd_mkhuge(entry);
622 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
623 result = VM_FAULT_NOPAGE;
626 sector = bh.b_blocknr << (blkbits - 9);
627 length = bdev_direct_access(bh.b_bdev, sector, &kaddr, &pfn,
630 result = VM_FAULT_SIGBUS;
633 if ((length < PMD_SIZE) || (pfn & PG_PMD_COLOUR))
637 * TODO: teach vmf_insert_pfn_pmd() to support
638 * 'pte_special' for pmds
643 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
645 for (i = 0; i < PTRS_PER_PMD; i++)
646 clear_pmem(kaddr + i * PAGE_SIZE, PAGE_SIZE);
648 count_vm_event(PGMAJFAULT);
649 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
650 result |= VM_FAULT_MAJOR;
653 result |= vmf_insert_pfn_pmd(vma, address, pmd, pfn, write);
657 i_mmap_unlock_read(mapping);
659 if (buffer_unwritten(&bh))
660 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
665 count_vm_event(THP_FAULT_FALLBACK);
666 result = VM_FAULT_FALLBACK;
669 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
672 * dax_pmd_fault - handle a PMD fault on a DAX file
673 * @vma: The virtual memory area where the fault occurred
674 * @vmf: The description of the fault
675 * @get_block: The filesystem method used to translate file offsets to blocks
677 * When a page fault occurs, filesystems may call this helper in their
678 * pmd_fault handler for DAX files.
680 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
681 pmd_t *pmd, unsigned int flags, get_block_t get_block,
682 dax_iodone_t complete_unwritten)
685 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
687 if (flags & FAULT_FLAG_WRITE) {
688 sb_start_pagefault(sb);
689 file_update_time(vma->vm_file);
691 result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
693 if (flags & FAULT_FLAG_WRITE)
694 sb_end_pagefault(sb);
698 EXPORT_SYMBOL_GPL(dax_pmd_fault);
699 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
702 * dax_pfn_mkwrite - handle first write to DAX page
703 * @vma: The virtual memory area where the fault occurred
704 * @vmf: The description of the fault
707 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
709 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
711 sb_start_pagefault(sb);
712 file_update_time(vma->vm_file);
713 sb_end_pagefault(sb);
714 return VM_FAULT_NOPAGE;
716 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
719 * dax_zero_page_range - zero a range within a page of a DAX file
720 * @inode: The file being truncated
721 * @from: The file offset that is being truncated to
722 * @length: The number of bytes to zero
723 * @get_block: The filesystem method used to translate file offsets to blocks
725 * This function can be called by a filesystem when it is zeroing part of a
726 * page in a DAX file. This is intended for hole-punch operations. If
727 * you are truncating a file, the helper function dax_truncate_page() may be
730 * We work in terms of PAGE_CACHE_SIZE here for commonality with
731 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
732 * took care of disposing of the unnecessary blocks. Even if the filesystem
733 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
734 * since the file might be mmapped.
736 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
737 get_block_t get_block)
739 struct buffer_head bh;
740 pgoff_t index = from >> PAGE_CACHE_SHIFT;
741 unsigned offset = from & (PAGE_CACHE_SIZE-1);
744 /* Block boundary? Nothing to do */
747 BUG_ON((offset + length) > PAGE_CACHE_SIZE);
749 memset(&bh, 0, sizeof(bh));
750 bh.b_size = PAGE_CACHE_SIZE;
751 err = get_block(inode, index, &bh, 0);
754 if (buffer_written(&bh)) {
756 err = dax_get_addr(&bh, &addr, inode->i_blkbits);
759 clear_pmem(addr + offset, length);
765 EXPORT_SYMBOL_GPL(dax_zero_page_range);
768 * dax_truncate_page - handle a partial page being truncated in a DAX file
769 * @inode: The file being truncated
770 * @from: The file offset that is being truncated to
771 * @get_block: The filesystem method used to translate file offsets to blocks
773 * Similar to block_truncate_page(), this function can be called by a
774 * filesystem when it is truncating a DAX file to handle the partial page.
776 * We work in terms of PAGE_CACHE_SIZE here for commonality with
777 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
778 * took care of disposing of the unnecessary blocks. Even if the filesystem
779 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
780 * since the file might be mmapped.
782 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
784 unsigned length = PAGE_CACHE_ALIGN(from) - from;
785 return dax_zero_page_range(inode, from, length, get_block);
787 EXPORT_SYMBOL_GPL(dax_truncate_page);