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>
35 #define RADIX_DAX_MASK 0xf
36 #define RADIX_DAX_SHIFT 4
37 #define RADIX_DAX_PTE (0x4 | RADIX_TREE_EXCEPTIONAL_ENTRY)
38 #define RADIX_DAX_PMD (0x8 | RADIX_TREE_EXCEPTIONAL_ENTRY)
39 #define RADIX_DAX_TYPE(entry) ((unsigned long)entry & RADIX_DAX_MASK)
40 #define RADIX_DAX_SECTOR(entry) (((unsigned long)entry >> RADIX_DAX_SHIFT))
41 #define RADIX_DAX_ENTRY(sector, pmd) ((void *)((unsigned long)sector << \
42 RADIX_DAX_SHIFT | (pmd ? RADIX_DAX_PMD : RADIX_DAX_PTE)))
44 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
46 struct request_queue *q = bdev->bd_queue;
49 dax->addr = (void __pmem *) ERR_PTR(-EIO);
50 if (blk_queue_enter(q, true) != 0)
53 rc = bdev_direct_access(bdev, dax);
55 dax->addr = (void __pmem *) ERR_PTR(rc);
62 static void dax_unmap_atomic(struct block_device *bdev,
63 const struct blk_dax_ctl *dax)
65 if (IS_ERR(dax->addr))
67 blk_queue_exit(bdev->bd_queue);
70 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
72 struct page *page = alloc_pages(GFP_KERNEL, 0);
73 struct blk_dax_ctl dax = {
75 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
80 return ERR_PTR(-ENOMEM);
82 rc = dax_map_atomic(bdev, &dax);
85 memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
86 dax_unmap_atomic(bdev, &dax);
91 * dax_clear_sectors() is called from within transaction context from XFS,
92 * and hence this means the stack from this point must follow GFP_NOFS
93 * semantics for all operations.
95 int dax_clear_sectors(struct block_device *bdev, sector_t _sector, long _size)
97 struct blk_dax_ctl dax = {
106 count = dax_map_atomic(bdev, &dax);
109 sz = min_t(long, count, SZ_128K);
110 clear_pmem(dax.addr, sz);
112 dax.sector += sz / 512;
113 dax_unmap_atomic(bdev, &dax);
120 EXPORT_SYMBOL_GPL(dax_clear_sectors);
122 static bool buffer_written(struct buffer_head *bh)
124 return buffer_mapped(bh) && !buffer_unwritten(bh);
128 * When ext4 encounters a hole, it returns without modifying the buffer_head
129 * which means that we can't trust b_size. To cope with this, we set b_state
130 * to 0 before calling get_block and, if any bit is set, we know we can trust
131 * b_size. Unfortunate, really, since ext4 knows precisely how long a hole is
132 * and would save us time calling get_block repeatedly.
134 static bool buffer_size_valid(struct buffer_head *bh)
136 return bh->b_state != 0;
140 static sector_t to_sector(const struct buffer_head *bh,
141 const struct inode *inode)
143 sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
148 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
149 loff_t start, loff_t end, get_block_t get_block,
150 struct buffer_head *bh)
152 loff_t pos = start, max = start, bh_max = start;
153 bool hole = false, need_wmb = false;
154 struct block_device *bdev = NULL;
155 int rw = iov_iter_rw(iter), rc;
157 struct blk_dax_ctl dax = {
158 .addr = (void __pmem *) ERR_PTR(-EIO),
160 unsigned blkbits = inode->i_blkbits;
161 sector_t file_blks = (i_size_read(inode) + (1 << blkbits) - 1)
165 end = min(end, i_size_read(inode));
170 long page = pos >> PAGE_SHIFT;
171 sector_t block = page << (PAGE_SHIFT - blkbits);
172 unsigned first = pos - (block << blkbits);
176 bh->b_size = PAGE_ALIGN(end - pos);
178 rc = get_block(inode, block, bh, rw == WRITE);
181 if (!buffer_size_valid(bh))
182 bh->b_size = 1 << blkbits;
183 bh_max = pos - first + bh->b_size;
186 * We allow uninitialized buffers for writes
187 * beyond EOF as those cannot race with faults
190 (buffer_new(bh) && block < file_blks) ||
191 (rw == WRITE && buffer_unwritten(bh)));
193 unsigned done = bh->b_size -
194 (bh_max - (pos - first));
195 bh->b_blocknr += done >> blkbits;
199 hole = rw == READ && !buffer_written(bh);
201 size = bh->b_size - first;
203 dax_unmap_atomic(bdev, &dax);
204 dax.sector = to_sector(bh, inode);
205 dax.size = bh->b_size;
206 map_len = dax_map_atomic(bdev, &dax);
212 size = map_len - first;
214 max = min(pos + size, end);
217 if (iov_iter_rw(iter) == WRITE) {
218 len = copy_from_iter_pmem(dax.addr, max - pos, iter);
221 len = copy_to_iter((void __force *) dax.addr, max - pos,
224 len = iov_iter_zero(max - pos, iter);
232 if (!IS_ERR(dax.addr))
238 dax_unmap_atomic(bdev, &dax);
240 return (pos == start) ? rc : pos - start;
244 * dax_do_io - Perform I/O to a DAX file
245 * @iocb: The control block for this I/O
246 * @inode: The file which the I/O is directed at
247 * @iter: The addresses to do I/O from or to
248 * @pos: The file offset where the I/O starts
249 * @get_block: The filesystem method used to translate file offsets to blocks
250 * @end_io: A filesystem callback for I/O completion
253 * This function uses the same locking scheme as do_blockdev_direct_IO:
254 * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
255 * caller for writes. For reads, we take and release the i_mutex ourselves.
256 * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
257 * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
260 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
261 struct iov_iter *iter, loff_t pos, get_block_t get_block,
262 dio_iodone_t end_io, int flags)
264 struct buffer_head bh;
265 ssize_t retval = -EINVAL;
266 loff_t end = pos + iov_iter_count(iter);
268 memset(&bh, 0, sizeof(bh));
269 bh.b_bdev = inode->i_sb->s_bdev;
271 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
274 /* Protects against truncate */
275 if (!(flags & DIO_SKIP_DIO_COUNT))
276 inode_dio_begin(inode);
278 retval = dax_io(inode, iter, pos, end, get_block, &bh);
280 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
286 err = end_io(iocb, pos, retval, bh.b_private);
291 if (!(flags & DIO_SKIP_DIO_COUNT))
292 inode_dio_end(inode);
295 EXPORT_SYMBOL_GPL(dax_do_io);
298 * The user has performed a load from a hole in the file. Allocating
299 * a new page in the file would cause excessive storage usage for
300 * workloads with sparse files. We allocate a page cache page instead.
301 * We'll kick it out of the page cache if it's ever written to,
302 * otherwise it will simply fall out of the page cache under memory
303 * pressure without ever having been dirtied.
305 static int dax_load_hole(struct address_space *mapping, struct page *page,
306 struct vm_fault *vmf)
309 struct inode *inode = mapping->host;
311 page = find_or_create_page(mapping, vmf->pgoff,
312 GFP_KERNEL | __GFP_ZERO);
315 /* Recheck i_size under page lock to avoid truncate race */
316 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
317 if (vmf->pgoff >= size) {
320 return VM_FAULT_SIGBUS;
324 return VM_FAULT_LOCKED;
327 static int copy_user_bh(struct page *to, struct inode *inode,
328 struct buffer_head *bh, unsigned long vaddr)
330 struct blk_dax_ctl dax = {
331 .sector = to_sector(bh, inode),
334 struct block_device *bdev = bh->b_bdev;
337 if (dax_map_atomic(bdev, &dax) < 0)
338 return PTR_ERR(dax.addr);
339 vto = kmap_atomic(to);
340 copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
342 dax_unmap_atomic(bdev, &dax);
347 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
349 static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
350 sector_t sector, bool pmd_entry, bool dirty)
352 struct radix_tree_root *page_tree = &mapping->page_tree;
353 pgoff_t pmd_index = DAX_PMD_INDEX(index);
357 WARN_ON_ONCE(pmd_entry && !dirty);
359 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
361 spin_lock_irq(&mapping->tree_lock);
363 entry = radix_tree_lookup(page_tree, pmd_index);
364 if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
369 entry = radix_tree_lookup(page_tree, index);
371 type = RADIX_DAX_TYPE(entry);
372 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
373 type != RADIX_DAX_PMD)) {
378 if (!pmd_entry || type == RADIX_DAX_PMD)
382 * We only insert dirty PMD entries into the radix tree. This
383 * means we don't need to worry about removing a dirty PTE
384 * entry and inserting a clean PMD entry, thus reducing the
385 * range we would flush with a follow-up fsync/msync call.
387 radix_tree_delete(&mapping->page_tree, index);
388 mapping->nrexceptional--;
391 if (sector == NO_SECTOR) {
393 * This can happen during correct operation if our pfn_mkwrite
394 * fault raced against a hole punch operation. If this
395 * happens the pte that was hole punched will have been
396 * unmapped and the radix tree entry will have been removed by
397 * the time we are called, but the call will still happen. We
398 * will return all the way up to wp_pfn_shared(), where the
399 * pte_same() check will fail, eventually causing page fault
400 * to be retried by the CPU.
405 error = radix_tree_insert(page_tree, index,
406 RADIX_DAX_ENTRY(sector, pmd_entry));
410 mapping->nrexceptional++;
413 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
415 spin_unlock_irq(&mapping->tree_lock);
419 static int dax_writeback_one(struct block_device *bdev,
420 struct address_space *mapping, pgoff_t index, void *entry)
422 struct radix_tree_root *page_tree = &mapping->page_tree;
423 int type = RADIX_DAX_TYPE(entry);
424 struct radix_tree_node *node;
425 struct blk_dax_ctl dax;
429 spin_lock_irq(&mapping->tree_lock);
431 * Regular page slots are stabilized by the page lock even
432 * without the tree itself locked. These unlocked entries
433 * need verification under the tree lock.
435 if (!__radix_tree_lookup(page_tree, index, &node, &slot))
440 /* another fsync thread may have already written back this entry */
441 if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
444 if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
449 dax.sector = RADIX_DAX_SECTOR(entry);
450 dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
451 spin_unlock_irq(&mapping->tree_lock);
454 * We cannot hold tree_lock while calling dax_map_atomic() because it
455 * eventually calls cond_resched().
457 ret = dax_map_atomic(bdev, &dax);
461 if (WARN_ON_ONCE(ret < dax.size)) {
466 wb_cache_pmem(dax.addr, dax.size);
468 spin_lock_irq(&mapping->tree_lock);
469 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
470 spin_unlock_irq(&mapping->tree_lock);
472 dax_unmap_atomic(bdev, &dax);
476 spin_unlock_irq(&mapping->tree_lock);
481 * Flush the mapping to the persistent domain within the byte range of [start,
482 * end]. This is required by data integrity operations to ensure file data is
483 * on persistent storage prior to completion of the operation.
485 int dax_writeback_mapping_range(struct address_space *mapping,
486 struct block_device *bdev, struct writeback_control *wbc)
488 struct inode *inode = mapping->host;
489 pgoff_t start_index, end_index, pmd_index;
490 pgoff_t indices[PAGEVEC_SIZE];
496 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
499 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
502 start_index = wbc->range_start >> PAGE_SHIFT;
503 end_index = wbc->range_end >> PAGE_SHIFT;
504 pmd_index = DAX_PMD_INDEX(start_index);
507 entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
510 /* see if the start of our range is covered by a PMD entry */
511 if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
512 start_index = pmd_index;
514 tag_pages_for_writeback(mapping, start_index, end_index);
516 pagevec_init(&pvec, 0);
518 pvec.nr = find_get_entries_tag(mapping, start_index,
519 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
520 pvec.pages, indices);
525 for (i = 0; i < pvec.nr; i++) {
526 if (indices[i] > end_index) {
531 ret = dax_writeback_one(bdev, mapping, indices[i],
540 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
542 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
543 struct vm_area_struct *vma, struct vm_fault *vmf)
545 unsigned long vaddr = (unsigned long)vmf->virtual_address;
546 struct address_space *mapping = inode->i_mapping;
547 struct block_device *bdev = bh->b_bdev;
548 struct blk_dax_ctl dax = {
549 .sector = to_sector(bh, inode),
555 i_mmap_lock_read(mapping);
558 * Check truncate didn't happen while we were allocating a block.
559 * If it did, this block may or may not be still allocated to the
560 * file. We can't tell the filesystem to free it because we can't
561 * take i_mutex here. In the worst case, the file still has blocks
562 * allocated past the end of the file.
564 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
565 if (unlikely(vmf->pgoff >= size)) {
570 if (dax_map_atomic(bdev, &dax) < 0) {
571 error = PTR_ERR(dax.addr);
574 dax_unmap_atomic(bdev, &dax);
576 error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
577 vmf->flags & FAULT_FLAG_WRITE);
581 error = vm_insert_mixed(vma, vaddr, dax.pfn);
584 i_mmap_unlock_read(mapping);
590 * __dax_fault - handle a page fault on a DAX file
591 * @vma: The virtual memory area where the fault occurred
592 * @vmf: The description of the fault
593 * @get_block: The filesystem method used to translate file offsets to blocks
595 * When a page fault occurs, filesystems may call this helper in their
596 * fault handler for DAX files. __dax_fault() assumes the caller has done all
597 * the necessary locking for the page fault to proceed successfully.
599 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
600 get_block_t get_block)
602 struct file *file = vma->vm_file;
603 struct address_space *mapping = file->f_mapping;
604 struct inode *inode = mapping->host;
606 struct buffer_head bh;
607 unsigned long vaddr = (unsigned long)vmf->virtual_address;
608 unsigned blkbits = inode->i_blkbits;
614 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
615 if (vmf->pgoff >= size)
616 return VM_FAULT_SIGBUS;
618 memset(&bh, 0, sizeof(bh));
619 block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
620 bh.b_bdev = inode->i_sb->s_bdev;
621 bh.b_size = PAGE_SIZE;
624 page = find_get_page(mapping, vmf->pgoff);
626 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
628 return VM_FAULT_RETRY;
630 if (unlikely(page->mapping != mapping)) {
635 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
636 if (unlikely(vmf->pgoff >= size)) {
638 * We have a struct page covering a hole in the file
639 * from a read fault and we've raced with a truncate
646 error = get_block(inode, block, &bh, 0);
647 if (!error && (bh.b_size < PAGE_SIZE))
648 error = -EIO; /* fs corruption? */
652 if (!buffer_mapped(&bh) && !vmf->cow_page) {
653 if (vmf->flags & FAULT_FLAG_WRITE) {
654 error = get_block(inode, block, &bh, 1);
655 count_vm_event(PGMAJFAULT);
656 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
657 major = VM_FAULT_MAJOR;
658 if (!error && (bh.b_size < PAGE_SIZE))
663 return dax_load_hole(mapping, page, vmf);
668 struct page *new_page = vmf->cow_page;
669 if (buffer_written(&bh))
670 error = copy_user_bh(new_page, inode, &bh, vaddr);
672 clear_user_highpage(new_page, vaddr);
677 i_mmap_lock_read(mapping);
678 /* Check we didn't race with truncate */
679 size = (i_size_read(inode) + PAGE_SIZE - 1) >>
681 if (vmf->pgoff >= size) {
682 i_mmap_unlock_read(mapping);
687 return VM_FAULT_LOCKED;
690 /* Check we didn't race with a read fault installing a new page */
692 page = find_lock_page(mapping, vmf->pgoff);
695 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
697 delete_from_page_cache(page);
703 /* Filesystem should not return unwritten buffers to us! */
704 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
705 error = dax_insert_mapping(inode, &bh, vma, vmf);
708 if (error == -ENOMEM)
709 return VM_FAULT_OOM | major;
710 /* -EBUSY is fine, somebody else faulted on the same PTE */
711 if ((error < 0) && (error != -EBUSY))
712 return VM_FAULT_SIGBUS | major;
713 return VM_FAULT_NOPAGE | major;
722 EXPORT_SYMBOL(__dax_fault);
725 * dax_fault - handle a page fault on a DAX file
726 * @vma: The virtual memory area where the fault occurred
727 * @vmf: The description of the fault
728 * @get_block: The filesystem method used to translate file offsets to blocks
730 * When a page fault occurs, filesystems may call this helper in their
731 * fault handler for DAX files.
733 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
734 get_block_t get_block)
737 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
739 if (vmf->flags & FAULT_FLAG_WRITE) {
740 sb_start_pagefault(sb);
741 file_update_time(vma->vm_file);
743 result = __dax_fault(vma, vmf, get_block);
744 if (vmf->flags & FAULT_FLAG_WRITE)
745 sb_end_pagefault(sb);
749 EXPORT_SYMBOL_GPL(dax_fault);
751 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
753 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
754 * more often than one might expect in the below function.
756 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
758 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
759 const char *reason, const char *fn)
762 char bname[BDEVNAME_SIZE];
763 bdevname(bh->b_bdev, bname);
764 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
765 "length %zd fallback: %s\n", fn, current->comm,
766 address, bname, bh->b_state, (u64)bh->b_blocknr,
769 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
770 current->comm, address, reason);
774 #define dax_pmd_dbg(bh, address, reason) __dax_dbg(bh, address, reason, "dax_pmd")
776 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
777 pmd_t *pmd, unsigned int flags, get_block_t get_block)
779 struct file *file = vma->vm_file;
780 struct address_space *mapping = file->f_mapping;
781 struct inode *inode = mapping->host;
782 struct buffer_head bh;
783 unsigned blkbits = inode->i_blkbits;
784 unsigned long pmd_addr = address & PMD_MASK;
785 bool write = flags & FAULT_FLAG_WRITE;
786 struct block_device *bdev;
789 int error, result = 0;
792 /* dax pmd mappings require pfn_t_devmap() */
793 if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
794 return VM_FAULT_FALLBACK;
796 /* Fall back to PTEs if we're going to COW */
797 if (write && !(vma->vm_flags & VM_SHARED)) {
798 split_huge_pmd(vma, pmd, address);
799 dax_pmd_dbg(NULL, address, "cow write");
800 return VM_FAULT_FALLBACK;
802 /* If the PMD would extend outside the VMA */
803 if (pmd_addr < vma->vm_start) {
804 dax_pmd_dbg(NULL, address, "vma start unaligned");
805 return VM_FAULT_FALLBACK;
807 if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
808 dax_pmd_dbg(NULL, address, "vma end unaligned");
809 return VM_FAULT_FALLBACK;
812 pgoff = linear_page_index(vma, pmd_addr);
813 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
815 return VM_FAULT_SIGBUS;
816 /* If the PMD would cover blocks out of the file */
817 if ((pgoff | PG_PMD_COLOUR) >= size) {
818 dax_pmd_dbg(NULL, address,
819 "offset + huge page size > file size");
820 return VM_FAULT_FALLBACK;
823 memset(&bh, 0, sizeof(bh));
824 bh.b_bdev = inode->i_sb->s_bdev;
825 block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
827 bh.b_size = PMD_SIZE;
829 if (get_block(inode, block, &bh, 0) != 0)
830 return VM_FAULT_SIGBUS;
832 if (!buffer_mapped(&bh) && write) {
833 if (get_block(inode, block, &bh, 1) != 0)
834 return VM_FAULT_SIGBUS;
836 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
842 * If the filesystem isn't willing to tell us the length of a hole,
843 * just fall back to PTEs. Calling get_block 512 times in a loop
846 if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
847 dax_pmd_dbg(&bh, address, "allocated block too small");
848 return VM_FAULT_FALLBACK;
852 * If we allocated new storage, make sure no process has any
853 * zero pages covering this hole
856 loff_t lstart = pgoff << PAGE_SHIFT;
857 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
859 truncate_pagecache_range(inode, lstart, lend);
862 i_mmap_lock_read(mapping);
865 * If a truncate happened while we were allocating blocks, we may
866 * leave blocks allocated to the file that are beyond EOF. We can't
867 * take i_mutex here, so just leave them hanging; they'll be freed
868 * when the file is deleted.
870 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
872 result = VM_FAULT_SIGBUS;
875 if ((pgoff | PG_PMD_COLOUR) >= size) {
876 dax_pmd_dbg(&bh, address,
877 "offset + huge page size > file size");
881 if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
884 struct page *zero_page = get_huge_zero_page();
886 if (unlikely(!zero_page)) {
887 dax_pmd_dbg(&bh, address, "no zero page");
891 ptl = pmd_lock(vma->vm_mm, pmd);
892 if (!pmd_none(*pmd)) {
894 dax_pmd_dbg(&bh, address, "pmd already present");
898 dev_dbg(part_to_dev(bdev->bd_part),
899 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
900 __func__, current->comm, address,
901 (unsigned long long) to_sector(&bh, inode));
903 entry = mk_pmd(zero_page, vma->vm_page_prot);
904 entry = pmd_mkhuge(entry);
905 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
906 result = VM_FAULT_NOPAGE;
909 struct blk_dax_ctl dax = {
910 .sector = to_sector(&bh, inode),
913 long length = dax_map_atomic(bdev, &dax);
916 result = VM_FAULT_SIGBUS;
919 if (length < PMD_SIZE) {
920 dax_pmd_dbg(&bh, address, "dax-length too small");
921 dax_unmap_atomic(bdev, &dax);
924 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
925 dax_pmd_dbg(&bh, address, "pfn unaligned");
926 dax_unmap_atomic(bdev, &dax);
930 if (!pfn_t_devmap(dax.pfn)) {
931 dax_unmap_atomic(bdev, &dax);
932 dax_pmd_dbg(&bh, address, "pfn not in memmap");
935 dax_unmap_atomic(bdev, &dax);
938 * For PTE faults we insert a radix tree entry for reads, and
939 * leave it clean. Then on the first write we dirty the radix
940 * tree entry via the dax_pfn_mkwrite() path. This sequence
941 * allows the dax_pfn_mkwrite() call to be simpler and avoid a
942 * call into get_block() to translate the pgoff to a sector in
943 * order to be able to create a new radix tree entry.
945 * The PMD path doesn't have an equivalent to
946 * dax_pfn_mkwrite(), though, so for a read followed by a
947 * write we traverse all the way through __dax_pmd_fault()
948 * twice. This means we can just skip inserting a radix tree
949 * entry completely on the initial read and just wait until
950 * the write to insert a dirty entry.
953 error = dax_radix_entry(mapping, pgoff, dax.sector,
956 dax_pmd_dbg(&bh, address,
957 "PMD radix insertion failed");
962 dev_dbg(part_to_dev(bdev->bd_part),
963 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
964 __func__, current->comm, address,
965 pfn_t_to_pfn(dax.pfn),
966 (unsigned long long) dax.sector);
967 result |= vmf_insert_pfn_pmd(vma, address, pmd,
972 i_mmap_unlock_read(mapping);
977 count_vm_event(THP_FAULT_FALLBACK);
978 result = VM_FAULT_FALLBACK;
981 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
984 * dax_pmd_fault - handle a PMD fault on a DAX file
985 * @vma: The virtual memory area where the fault occurred
986 * @vmf: The description of the fault
987 * @get_block: The filesystem method used to translate file offsets to blocks
989 * When a page fault occurs, filesystems may call this helper in their
990 * pmd_fault handler for DAX files.
992 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
993 pmd_t *pmd, unsigned int flags, get_block_t get_block)
996 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
998 if (flags & FAULT_FLAG_WRITE) {
999 sb_start_pagefault(sb);
1000 file_update_time(vma->vm_file);
1002 result = __dax_pmd_fault(vma, address, pmd, flags, get_block);
1003 if (flags & FAULT_FLAG_WRITE)
1004 sb_end_pagefault(sb);
1008 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1009 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1012 * dax_pfn_mkwrite - handle first write to DAX page
1013 * @vma: The virtual memory area where the fault occurred
1014 * @vmf: The description of the fault
1016 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1018 struct file *file = vma->vm_file;
1022 * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1023 * RADIX_DAX_PTE entry already exists in the radix tree from a
1024 * previous call to __dax_fault(). We just want to look up that PTE
1025 * entry using vmf->pgoff and make sure the dirty tag is set. This
1026 * saves us from having to make a call to get_block() here to look
1029 error = dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false,
1032 if (error == -ENOMEM)
1033 return VM_FAULT_OOM;
1035 return VM_FAULT_SIGBUS;
1036 return VM_FAULT_NOPAGE;
1038 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1041 * dax_zero_page_range - zero a range within a page of a DAX file
1042 * @inode: The file being truncated
1043 * @from: The file offset that is being truncated to
1044 * @length: The number of bytes to zero
1045 * @get_block: The filesystem method used to translate file offsets to blocks
1047 * This function can be called by a filesystem when it is zeroing part of a
1048 * page in a DAX file. This is intended for hole-punch operations. If
1049 * you are truncating a file, the helper function dax_truncate_page() may be
1052 * We work in terms of PAGE_SIZE here for commonality with
1053 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1054 * took care of disposing of the unnecessary blocks. Even if the filesystem
1055 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1056 * since the file might be mmapped.
1058 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1059 get_block_t get_block)
1061 struct buffer_head bh;
1062 pgoff_t index = from >> PAGE_SHIFT;
1063 unsigned offset = from & (PAGE_SIZE-1);
1066 /* Block boundary? Nothing to do */
1069 BUG_ON((offset + length) > PAGE_SIZE);
1071 memset(&bh, 0, sizeof(bh));
1072 bh.b_bdev = inode->i_sb->s_bdev;
1073 bh.b_size = PAGE_SIZE;
1074 err = get_block(inode, index, &bh, 0);
1077 if (buffer_written(&bh)) {
1078 struct block_device *bdev = bh.b_bdev;
1079 struct blk_dax_ctl dax = {
1080 .sector = to_sector(&bh, inode),
1084 if (dax_map_atomic(bdev, &dax) < 0)
1085 return PTR_ERR(dax.addr);
1086 clear_pmem(dax.addr + offset, length);
1088 dax_unmap_atomic(bdev, &dax);
1093 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1096 * dax_truncate_page - handle a partial page being truncated in a DAX file
1097 * @inode: The file being truncated
1098 * @from: The file offset that is being truncated to
1099 * @get_block: The filesystem method used to translate file offsets to blocks
1101 * Similar to block_truncate_page(), this function can be called by a
1102 * filesystem when it is truncating a DAX file to handle the partial page.
1104 * We work in terms of PAGE_SIZE here for commonality with
1105 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1106 * took care of disposing of the unnecessary blocks. Even if the filesystem
1107 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1108 * since the file might be mmapped.
1110 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1112 unsigned length = PAGE_ALIGN(from) - from;
1113 return dax_zero_page_range(inode, from, length, get_block);
1115 EXPORT_SYMBOL_GPL(dax_truncate_page);