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>
36 * We use lowest available bit in exceptional entry for locking, other two
37 * bits to determine entry type. In total 3 special bits.
39 #define RADIX_DAX_SHIFT (RADIX_TREE_EXCEPTIONAL_SHIFT + 3)
40 #define RADIX_DAX_PTE (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1))
41 #define RADIX_DAX_PMD (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2))
42 #define RADIX_DAX_TYPE_MASK (RADIX_DAX_PTE | RADIX_DAX_PMD)
43 #define RADIX_DAX_TYPE(entry) ((unsigned long)entry & RADIX_DAX_TYPE_MASK)
44 #define RADIX_DAX_SECTOR(entry) (((unsigned long)entry >> RADIX_DAX_SHIFT))
45 #define RADIX_DAX_ENTRY(sector, pmd) ((void *)((unsigned long)sector << \
46 RADIX_DAX_SHIFT | (pmd ? RADIX_DAX_PMD : RADIX_DAX_PTE) | \
47 RADIX_TREE_EXCEPTIONAL_ENTRY))
49 /* We choose 4096 entries - same as per-zone page wait tables */
50 #define DAX_WAIT_TABLE_BITS 12
51 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
53 wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
55 static int __init init_dax_wait_table(void)
59 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
60 init_waitqueue_head(wait_table + i);
63 fs_initcall(init_dax_wait_table);
65 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
68 unsigned long hash = hash_long((unsigned long)mapping ^ index,
70 return wait_table + hash;
73 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
75 struct request_queue *q = bdev->bd_queue;
78 dax->addr = (void __pmem *) ERR_PTR(-EIO);
79 if (blk_queue_enter(q, true) != 0)
82 rc = bdev_direct_access(bdev, dax);
84 dax->addr = (void __pmem *) ERR_PTR(rc);
91 static void dax_unmap_atomic(struct block_device *bdev,
92 const struct blk_dax_ctl *dax)
94 if (IS_ERR(dax->addr))
96 blk_queue_exit(bdev->bd_queue);
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);
119 static bool buffer_written(struct buffer_head *bh)
121 return buffer_mapped(bh) && !buffer_unwritten(bh);
125 * When ext4 encounters a hole, it returns without modifying the buffer_head
126 * which means that we can't trust b_size. To cope with this, we set b_state
127 * to 0 before calling get_block and, if any bit is set, we know we can trust
128 * b_size. Unfortunate, really, since ext4 knows precisely how long a hole is
129 * and would save us time calling get_block repeatedly.
131 static bool buffer_size_valid(struct buffer_head *bh)
133 return bh->b_state != 0;
137 static sector_t to_sector(const struct buffer_head *bh,
138 const struct inode *inode)
140 sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
145 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
146 loff_t start, loff_t end, get_block_t get_block,
147 struct buffer_head *bh)
149 loff_t pos = start, max = start, bh_max = start;
151 struct block_device *bdev = NULL;
152 int rw = iov_iter_rw(iter), rc;
154 struct blk_dax_ctl dax = {
155 .addr = (void __pmem *) ERR_PTR(-EIO),
157 unsigned blkbits = inode->i_blkbits;
158 sector_t file_blks = (i_size_read(inode) + (1 << blkbits) - 1)
162 end = min(end, i_size_read(inode));
167 long page = pos >> PAGE_SHIFT;
168 sector_t block = page << (PAGE_SHIFT - blkbits);
169 unsigned first = pos - (block << blkbits);
173 bh->b_size = PAGE_ALIGN(end - pos);
175 rc = get_block(inode, block, bh, rw == WRITE);
178 if (!buffer_size_valid(bh))
179 bh->b_size = 1 << blkbits;
180 bh_max = pos - first + bh->b_size;
183 * We allow uninitialized buffers for writes
184 * beyond EOF as those cannot race with faults
187 (buffer_new(bh) && block < file_blks) ||
188 (rw == WRITE && buffer_unwritten(bh)));
190 unsigned done = bh->b_size -
191 (bh_max - (pos - first));
192 bh->b_blocknr += done >> blkbits;
196 hole = rw == READ && !buffer_written(bh);
198 size = bh->b_size - first;
200 dax_unmap_atomic(bdev, &dax);
201 dax.sector = to_sector(bh, inode);
202 dax.size = bh->b_size;
203 map_len = dax_map_atomic(bdev, &dax);
209 size = map_len - first;
211 max = min(pos + size, end);
214 if (iov_iter_rw(iter) == WRITE) {
215 len = copy_from_iter_pmem(dax.addr, max - pos, iter);
217 len = copy_to_iter((void __force *) dax.addr, max - pos,
220 len = iov_iter_zero(max - pos, iter);
228 if (!IS_ERR(dax.addr))
232 dax_unmap_atomic(bdev, &dax);
234 return (pos == start) ? rc : pos - start;
238 * dax_do_io - Perform I/O to a DAX file
239 * @iocb: The control block for this I/O
240 * @inode: The file which the I/O is directed at
241 * @iter: The addresses to do I/O from or to
242 * @get_block: The filesystem method used to translate file offsets to blocks
243 * @end_io: A filesystem callback for I/O completion
246 * This function uses the same locking scheme as do_blockdev_direct_IO:
247 * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
248 * caller for writes. For reads, we take and release the i_mutex ourselves.
249 * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
250 * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
253 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
254 struct iov_iter *iter, get_block_t get_block,
255 dio_iodone_t end_io, int flags)
257 struct buffer_head bh;
258 ssize_t retval = -EINVAL;
259 loff_t pos = iocb->ki_pos;
260 loff_t end = pos + iov_iter_count(iter);
262 memset(&bh, 0, sizeof(bh));
263 bh.b_bdev = inode->i_sb->s_bdev;
265 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
268 /* Protects against truncate */
269 if (!(flags & DIO_SKIP_DIO_COUNT))
270 inode_dio_begin(inode);
272 retval = dax_io(inode, iter, pos, end, get_block, &bh);
274 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
280 err = end_io(iocb, pos, retval, bh.b_private);
285 if (!(flags & DIO_SKIP_DIO_COUNT))
286 inode_dio_end(inode);
289 EXPORT_SYMBOL_GPL(dax_do_io);
292 * DAX radix tree locking
294 struct exceptional_entry_key {
295 struct address_space *mapping;
299 struct wait_exceptional_entry_queue {
301 struct exceptional_entry_key key;
304 static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
305 int sync, void *keyp)
307 struct exceptional_entry_key *key = keyp;
308 struct wait_exceptional_entry_queue *ewait =
309 container_of(wait, struct wait_exceptional_entry_queue, wait);
311 if (key->mapping != ewait->key.mapping ||
312 key->index != ewait->key.index)
314 return autoremove_wake_function(wait, mode, sync, NULL);
318 * Check whether the given slot is locked. The function must be called with
319 * mapping->tree_lock held
321 static inline int slot_locked(struct address_space *mapping, void **slot)
323 unsigned long entry = (unsigned long)
324 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
325 return entry & RADIX_DAX_ENTRY_LOCK;
329 * Mark the given slot is locked. The function must be called with
330 * mapping->tree_lock held
332 static inline void *lock_slot(struct address_space *mapping, void **slot)
334 unsigned long entry = (unsigned long)
335 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
337 entry |= RADIX_DAX_ENTRY_LOCK;
338 radix_tree_replace_slot(slot, (void *)entry);
339 return (void *)entry;
343 * Mark the given slot is unlocked. The function must be called with
344 * mapping->tree_lock held
346 static inline void *unlock_slot(struct address_space *mapping, void **slot)
348 unsigned long entry = (unsigned long)
349 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
351 entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
352 radix_tree_replace_slot(slot, (void *)entry);
353 return (void *)entry;
357 * Lookup entry in radix tree, wait for it to become unlocked if it is
358 * exceptional entry and return it. The caller must call
359 * put_unlocked_mapping_entry() when he decided not to lock the entry or
360 * put_locked_mapping_entry() when he locked the entry and now wants to
363 * The function must be called with mapping->tree_lock held.
365 static void *get_unlocked_mapping_entry(struct address_space *mapping,
366 pgoff_t index, void ***slotp)
369 struct wait_exceptional_entry_queue ewait;
370 wait_queue_head_t *wq = dax_entry_waitqueue(mapping, index);
372 init_wait(&ewait.wait);
373 ewait.wait.func = wake_exceptional_entry_func;
374 ewait.key.mapping = mapping;
375 ewait.key.index = index;
378 ret = __radix_tree_lookup(&mapping->page_tree, index, NULL,
380 if (!ret || !radix_tree_exceptional_entry(ret) ||
381 !slot_locked(mapping, slot)) {
386 prepare_to_wait_exclusive(wq, &ewait.wait,
387 TASK_UNINTERRUPTIBLE);
388 spin_unlock_irq(&mapping->tree_lock);
390 finish_wait(wq, &ewait.wait);
391 spin_lock_irq(&mapping->tree_lock);
396 * Find radix tree entry at given index. If it points to a page, return with
397 * the page locked. If it points to the exceptional entry, return with the
398 * radix tree entry locked. If the radix tree doesn't contain given index,
399 * create empty exceptional entry for the index and return with it locked.
401 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
402 * persistent memory the benefit is doubtful. We can add that later if we can
405 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index)
410 spin_lock_irq(&mapping->tree_lock);
411 ret = get_unlocked_mapping_entry(mapping, index, &slot);
412 /* No entry for given index? Make sure radix tree is big enough. */
416 spin_unlock_irq(&mapping->tree_lock);
417 err = radix_tree_preload(
418 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
421 ret = (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY |
422 RADIX_DAX_ENTRY_LOCK);
423 spin_lock_irq(&mapping->tree_lock);
424 err = radix_tree_insert(&mapping->page_tree, index, ret);
425 radix_tree_preload_end();
427 spin_unlock_irq(&mapping->tree_lock);
428 /* Someone already created the entry? */
433 /* Good, we have inserted empty locked entry into the tree. */
434 mapping->nrexceptional++;
435 spin_unlock_irq(&mapping->tree_lock);
438 /* Normal page in radix tree? */
439 if (!radix_tree_exceptional_entry(ret)) {
440 struct page *page = ret;
443 spin_unlock_irq(&mapping->tree_lock);
445 /* Page got truncated? Retry... */
446 if (unlikely(page->mapping != mapping)) {
453 ret = lock_slot(mapping, slot);
454 spin_unlock_irq(&mapping->tree_lock);
458 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
459 pgoff_t index, bool wake_all)
461 wait_queue_head_t *wq = dax_entry_waitqueue(mapping, index);
464 * Checking for locked entry and prepare_to_wait_exclusive() happens
465 * under mapping->tree_lock, ditto for entry handling in our callers.
466 * So at this point all tasks that could have seen our entry locked
467 * must be in the waitqueue and the following check will see them.
469 if (waitqueue_active(wq)) {
470 struct exceptional_entry_key key;
472 key.mapping = mapping;
474 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
478 void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index)
482 spin_lock_irq(&mapping->tree_lock);
483 ret = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
484 if (WARN_ON_ONCE(!ret || !radix_tree_exceptional_entry(ret) ||
485 !slot_locked(mapping, slot))) {
486 spin_unlock_irq(&mapping->tree_lock);
489 unlock_slot(mapping, slot);
490 spin_unlock_irq(&mapping->tree_lock);
491 dax_wake_mapping_entry_waiter(mapping, index, false);
494 static void put_locked_mapping_entry(struct address_space *mapping,
495 pgoff_t index, void *entry)
497 if (!radix_tree_exceptional_entry(entry)) {
501 dax_unlock_mapping_entry(mapping, index);
506 * Called when we are done with radix tree entry we looked up via
507 * get_unlocked_mapping_entry() and which we didn't lock in the end.
509 static void put_unlocked_mapping_entry(struct address_space *mapping,
510 pgoff_t index, void *entry)
512 if (!radix_tree_exceptional_entry(entry))
515 /* We have to wake up next waiter for the radix tree entry lock */
516 dax_wake_mapping_entry_waiter(mapping, index, false);
520 * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
521 * entry to get unlocked before deleting it.
523 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
527 spin_lock_irq(&mapping->tree_lock);
528 entry = get_unlocked_mapping_entry(mapping, index, NULL);
530 * This gets called from truncate / punch_hole path. As such, the caller
531 * must hold locks protecting against concurrent modifications of the
532 * radix tree (usually fs-private i_mmap_sem for writing). Since the
533 * caller has seen exceptional entry for this index, we better find it
534 * at that index as well...
536 if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry))) {
537 spin_unlock_irq(&mapping->tree_lock);
540 radix_tree_delete(&mapping->page_tree, index);
541 mapping->nrexceptional--;
542 spin_unlock_irq(&mapping->tree_lock);
543 dax_wake_mapping_entry_waiter(mapping, index, true);
549 * The user has performed a load from a hole in the file. Allocating
550 * a new page in the file would cause excessive storage usage for
551 * workloads with sparse files. We allocate a page cache page instead.
552 * We'll kick it out of the page cache if it's ever written to,
553 * otherwise it will simply fall out of the page cache under memory
554 * pressure without ever having been dirtied.
556 static int dax_load_hole(struct address_space *mapping, void *entry,
557 struct vm_fault *vmf)
561 /* Hole page already exists? Return it... */
562 if (!radix_tree_exceptional_entry(entry)) {
564 return VM_FAULT_LOCKED;
567 /* This will replace locked radix tree entry with a hole page */
568 page = find_or_create_page(mapping, vmf->pgoff,
569 vmf->gfp_mask | __GFP_ZERO);
571 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
575 return VM_FAULT_LOCKED;
578 static int copy_user_bh(struct page *to, struct inode *inode,
579 struct buffer_head *bh, unsigned long vaddr)
581 struct blk_dax_ctl dax = {
582 .sector = to_sector(bh, inode),
585 struct block_device *bdev = bh->b_bdev;
588 if (dax_map_atomic(bdev, &dax) < 0)
589 return PTR_ERR(dax.addr);
590 vto = kmap_atomic(to);
591 copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
593 dax_unmap_atomic(bdev, &dax);
597 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
599 static void *dax_insert_mapping_entry(struct address_space *mapping,
600 struct vm_fault *vmf,
601 void *entry, sector_t sector)
603 struct radix_tree_root *page_tree = &mapping->page_tree;
605 bool hole_fill = false;
607 pgoff_t index = vmf->pgoff;
609 if (vmf->flags & FAULT_FLAG_WRITE)
610 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
612 /* Replacing hole page with block mapping? */
613 if (!radix_tree_exceptional_entry(entry)) {
616 * Unmap the page now before we remove it from page cache below.
617 * The page is locked so it cannot be faulted in again.
619 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
621 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
623 return ERR_PTR(error);
626 spin_lock_irq(&mapping->tree_lock);
627 new_entry = (void *)((unsigned long)RADIX_DAX_ENTRY(sector, false) |
628 RADIX_DAX_ENTRY_LOCK);
630 __delete_from_page_cache(entry, NULL);
631 /* Drop pagecache reference */
633 error = radix_tree_insert(page_tree, index, new_entry);
635 new_entry = ERR_PTR(error);
638 mapping->nrexceptional++;
643 ret = __radix_tree_lookup(page_tree, index, NULL, &slot);
644 WARN_ON_ONCE(ret != entry);
645 radix_tree_replace_slot(slot, new_entry);
647 if (vmf->flags & FAULT_FLAG_WRITE)
648 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
650 spin_unlock_irq(&mapping->tree_lock);
652 radix_tree_preload_end();
654 * We don't need hole page anymore, it has been replaced with
655 * locked radix tree entry now.
657 if (mapping->a_ops->freepage)
658 mapping->a_ops->freepage(entry);
665 static int dax_writeback_one(struct block_device *bdev,
666 struct address_space *mapping, pgoff_t index, void *entry)
668 struct radix_tree_root *page_tree = &mapping->page_tree;
669 int type = RADIX_DAX_TYPE(entry);
670 struct radix_tree_node *node;
671 struct blk_dax_ctl dax;
675 spin_lock_irq(&mapping->tree_lock);
677 * Regular page slots are stabilized by the page lock even
678 * without the tree itself locked. These unlocked entries
679 * need verification under the tree lock.
681 if (!__radix_tree_lookup(page_tree, index, &node, &slot))
686 /* another fsync thread may have already written back this entry */
687 if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
690 if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
695 dax.sector = RADIX_DAX_SECTOR(entry);
696 dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
697 spin_unlock_irq(&mapping->tree_lock);
700 * We cannot hold tree_lock while calling dax_map_atomic() because it
701 * eventually calls cond_resched().
703 ret = dax_map_atomic(bdev, &dax);
707 if (WARN_ON_ONCE(ret < dax.size)) {
712 wb_cache_pmem(dax.addr, dax.size);
714 spin_lock_irq(&mapping->tree_lock);
715 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
716 spin_unlock_irq(&mapping->tree_lock);
718 dax_unmap_atomic(bdev, &dax);
722 spin_unlock_irq(&mapping->tree_lock);
727 * Flush the mapping to the persistent domain within the byte range of [start,
728 * end]. This is required by data integrity operations to ensure file data is
729 * on persistent storage prior to completion of the operation.
731 int dax_writeback_mapping_range(struct address_space *mapping,
732 struct block_device *bdev, struct writeback_control *wbc)
734 struct inode *inode = mapping->host;
735 pgoff_t start_index, end_index, pmd_index;
736 pgoff_t indices[PAGEVEC_SIZE];
742 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
745 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
748 start_index = wbc->range_start >> PAGE_SHIFT;
749 end_index = wbc->range_end >> PAGE_SHIFT;
750 pmd_index = DAX_PMD_INDEX(start_index);
753 entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
756 /* see if the start of our range is covered by a PMD entry */
757 if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
758 start_index = pmd_index;
760 tag_pages_for_writeback(mapping, start_index, end_index);
762 pagevec_init(&pvec, 0);
764 pvec.nr = find_get_entries_tag(mapping, start_index,
765 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
766 pvec.pages, indices);
771 for (i = 0; i < pvec.nr; i++) {
772 if (indices[i] > end_index) {
777 ret = dax_writeback_one(bdev, mapping, indices[i],
785 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
787 static int dax_insert_mapping(struct address_space *mapping,
788 struct buffer_head *bh, void **entryp,
789 struct vm_area_struct *vma, struct vm_fault *vmf)
791 unsigned long vaddr = (unsigned long)vmf->virtual_address;
792 struct block_device *bdev = bh->b_bdev;
793 struct blk_dax_ctl dax = {
794 .sector = to_sector(bh, mapping->host),
798 void *entry = *entryp;
800 if (dax_map_atomic(bdev, &dax) < 0)
801 return PTR_ERR(dax.addr);
802 dax_unmap_atomic(bdev, &dax);
804 ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector);
809 return vm_insert_mixed(vma, vaddr, dax.pfn);
813 * __dax_fault - handle a page fault on a DAX file
814 * @vma: The virtual memory area where the fault occurred
815 * @vmf: The description of the fault
816 * @get_block: The filesystem method used to translate file offsets to blocks
818 * When a page fault occurs, filesystems may call this helper in their
819 * fault handler for DAX files. __dax_fault() assumes the caller has done all
820 * the necessary locking for the page fault to proceed successfully.
822 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
823 get_block_t get_block)
825 struct file *file = vma->vm_file;
826 struct address_space *mapping = file->f_mapping;
827 struct inode *inode = mapping->host;
829 struct buffer_head bh;
830 unsigned long vaddr = (unsigned long)vmf->virtual_address;
831 unsigned blkbits = inode->i_blkbits;
838 * Check whether offset isn't beyond end of file now. Caller is supposed
839 * to hold locks serializing us with truncate / punch hole so this is
842 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
843 if (vmf->pgoff >= size)
844 return VM_FAULT_SIGBUS;
846 memset(&bh, 0, sizeof(bh));
847 block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
848 bh.b_bdev = inode->i_sb->s_bdev;
849 bh.b_size = PAGE_SIZE;
851 entry = grab_mapping_entry(mapping, vmf->pgoff);
853 error = PTR_ERR(entry);
857 error = get_block(inode, block, &bh, 0);
858 if (!error && (bh.b_size < PAGE_SIZE))
859 error = -EIO; /* fs corruption? */
864 struct page *new_page = vmf->cow_page;
865 if (buffer_written(&bh))
866 error = copy_user_bh(new_page, inode, &bh, vaddr);
868 clear_user_highpage(new_page, vaddr);
871 if (!radix_tree_exceptional_entry(entry)) {
873 return VM_FAULT_LOCKED;
876 return VM_FAULT_DAX_LOCKED;
879 if (!buffer_mapped(&bh)) {
880 if (vmf->flags & FAULT_FLAG_WRITE) {
881 error = get_block(inode, block, &bh, 1);
882 count_vm_event(PGMAJFAULT);
883 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
884 major = VM_FAULT_MAJOR;
885 if (!error && (bh.b_size < PAGE_SIZE))
890 return dax_load_hole(mapping, entry, vmf);
894 /* Filesystem should not return unwritten buffers to us! */
895 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
896 error = dax_insert_mapping(mapping, &bh, &entry, vma, vmf);
898 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
900 if (error == -ENOMEM)
901 return VM_FAULT_OOM | major;
902 /* -EBUSY is fine, somebody else faulted on the same PTE */
903 if ((error < 0) && (error != -EBUSY))
904 return VM_FAULT_SIGBUS | major;
905 return VM_FAULT_NOPAGE | major;
907 EXPORT_SYMBOL(__dax_fault);
910 * dax_fault - handle a page fault on a DAX file
911 * @vma: The virtual memory area where the fault occurred
912 * @vmf: The description of the fault
913 * @get_block: The filesystem method used to translate file offsets to blocks
915 * When a page fault occurs, filesystems may call this helper in their
916 * fault handler for DAX files.
918 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
919 get_block_t get_block)
922 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
924 if (vmf->flags & FAULT_FLAG_WRITE) {
925 sb_start_pagefault(sb);
926 file_update_time(vma->vm_file);
928 result = __dax_fault(vma, vmf, get_block);
929 if (vmf->flags & FAULT_FLAG_WRITE)
930 sb_end_pagefault(sb);
934 EXPORT_SYMBOL_GPL(dax_fault);
936 #if defined(CONFIG_TRANSPARENT_HUGEPAGE)
938 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
939 * more often than one might expect in the below function.
941 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
943 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
944 const char *reason, const char *fn)
947 char bname[BDEVNAME_SIZE];
948 bdevname(bh->b_bdev, bname);
949 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
950 "length %zd fallback: %s\n", fn, current->comm,
951 address, bname, bh->b_state, (u64)bh->b_blocknr,
954 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
955 current->comm, address, reason);
959 #define dax_pmd_dbg(bh, address, reason) __dax_dbg(bh, address, reason, "dax_pmd")
961 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
962 pmd_t *pmd, unsigned int flags, get_block_t get_block)
964 struct file *file = vma->vm_file;
965 struct address_space *mapping = file->f_mapping;
966 struct inode *inode = mapping->host;
967 struct buffer_head bh;
968 unsigned blkbits = inode->i_blkbits;
969 unsigned long pmd_addr = address & PMD_MASK;
970 bool write = flags & FAULT_FLAG_WRITE;
971 struct block_device *bdev;
977 /* dax pmd mappings require pfn_t_devmap() */
978 if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
979 return VM_FAULT_FALLBACK;
981 /* Fall back to PTEs if we're going to COW */
982 if (write && !(vma->vm_flags & VM_SHARED)) {
983 split_huge_pmd(vma, pmd, address);
984 dax_pmd_dbg(NULL, address, "cow write");
985 return VM_FAULT_FALLBACK;
987 /* If the PMD would extend outside the VMA */
988 if (pmd_addr < vma->vm_start) {
989 dax_pmd_dbg(NULL, address, "vma start unaligned");
990 return VM_FAULT_FALLBACK;
992 if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
993 dax_pmd_dbg(NULL, address, "vma end unaligned");
994 return VM_FAULT_FALLBACK;
997 pgoff = linear_page_index(vma, pmd_addr);
998 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
1000 return VM_FAULT_SIGBUS;
1001 /* If the PMD would cover blocks out of the file */
1002 if ((pgoff | PG_PMD_COLOUR) >= size) {
1003 dax_pmd_dbg(NULL, address,
1004 "offset + huge page size > file size");
1005 return VM_FAULT_FALLBACK;
1008 memset(&bh, 0, sizeof(bh));
1009 bh.b_bdev = inode->i_sb->s_bdev;
1010 block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
1012 bh.b_size = PMD_SIZE;
1014 if (get_block(inode, block, &bh, 0) != 0)
1015 return VM_FAULT_SIGBUS;
1017 if (!buffer_mapped(&bh) && write) {
1018 if (get_block(inode, block, &bh, 1) != 0)
1019 return VM_FAULT_SIGBUS;
1021 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
1027 * If the filesystem isn't willing to tell us the length of a hole,
1028 * just fall back to PTEs. Calling get_block 512 times in a loop
1031 if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
1032 dax_pmd_dbg(&bh, address, "allocated block too small");
1033 return VM_FAULT_FALLBACK;
1037 * If we allocated new storage, make sure no process has any
1038 * zero pages covering this hole
1041 loff_t lstart = pgoff << PAGE_SHIFT;
1042 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
1044 truncate_pagecache_range(inode, lstart, lend);
1047 if (!write && !buffer_mapped(&bh)) {
1050 struct page *zero_page = get_huge_zero_page();
1052 if (unlikely(!zero_page)) {
1053 dax_pmd_dbg(&bh, address, "no zero page");
1057 ptl = pmd_lock(vma->vm_mm, pmd);
1058 if (!pmd_none(*pmd)) {
1060 dax_pmd_dbg(&bh, address, "pmd already present");
1064 dev_dbg(part_to_dev(bdev->bd_part),
1065 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
1066 __func__, current->comm, address,
1067 (unsigned long long) to_sector(&bh, inode));
1069 entry = mk_pmd(zero_page, vma->vm_page_prot);
1070 entry = pmd_mkhuge(entry);
1071 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
1072 result = VM_FAULT_NOPAGE;
1075 struct blk_dax_ctl dax = {
1076 .sector = to_sector(&bh, inode),
1079 long length = dax_map_atomic(bdev, &dax);
1082 dax_pmd_dbg(&bh, address, "dax-error fallback");
1085 if (length < PMD_SIZE) {
1086 dax_pmd_dbg(&bh, address, "dax-length too small");
1087 dax_unmap_atomic(bdev, &dax);
1090 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
1091 dax_pmd_dbg(&bh, address, "pfn unaligned");
1092 dax_unmap_atomic(bdev, &dax);
1096 if (!pfn_t_devmap(dax.pfn)) {
1097 dax_unmap_atomic(bdev, &dax);
1098 dax_pmd_dbg(&bh, address, "pfn not in memmap");
1101 dax_unmap_atomic(bdev, &dax);
1104 * For PTE faults we insert a radix tree entry for reads, and
1105 * leave it clean. Then on the first write we dirty the radix
1106 * tree entry via the dax_pfn_mkwrite() path. This sequence
1107 * allows the dax_pfn_mkwrite() call to be simpler and avoid a
1108 * call into get_block() to translate the pgoff to a sector in
1109 * order to be able to create a new radix tree entry.
1111 * The PMD path doesn't have an equivalent to
1112 * dax_pfn_mkwrite(), though, so for a read followed by a
1113 * write we traverse all the way through __dax_pmd_fault()
1114 * twice. This means we can just skip inserting a radix tree
1115 * entry completely on the initial read and just wait until
1116 * the write to insert a dirty entry.
1120 * We should insert radix-tree entry and dirty it here.
1121 * For now this is broken...
1125 dev_dbg(part_to_dev(bdev->bd_part),
1126 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
1127 __func__, current->comm, address,
1128 pfn_t_to_pfn(dax.pfn),
1129 (unsigned long long) dax.sector);
1130 result |= vmf_insert_pfn_pmd(vma, address, pmd,
1138 count_vm_event(THP_FAULT_FALLBACK);
1139 result = VM_FAULT_FALLBACK;
1142 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
1145 * dax_pmd_fault - handle a PMD fault on a DAX file
1146 * @vma: The virtual memory area where the fault occurred
1147 * @vmf: The description of the fault
1148 * @get_block: The filesystem method used to translate file offsets to blocks
1150 * When a page fault occurs, filesystems may call this helper in their
1151 * pmd_fault handler for DAX files.
1153 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1154 pmd_t *pmd, unsigned int flags, get_block_t get_block)
1157 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
1159 if (flags & FAULT_FLAG_WRITE) {
1160 sb_start_pagefault(sb);
1161 file_update_time(vma->vm_file);
1163 result = __dax_pmd_fault(vma, address, pmd, flags, get_block);
1164 if (flags & FAULT_FLAG_WRITE)
1165 sb_end_pagefault(sb);
1169 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1170 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1173 * dax_pfn_mkwrite - handle first write to DAX page
1174 * @vma: The virtual memory area where the fault occurred
1175 * @vmf: The description of the fault
1177 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1179 struct file *file = vma->vm_file;
1180 struct address_space *mapping = file->f_mapping;
1182 pgoff_t index = vmf->pgoff;
1184 spin_lock_irq(&mapping->tree_lock);
1185 entry = get_unlocked_mapping_entry(mapping, index, NULL);
1186 if (!entry || !radix_tree_exceptional_entry(entry))
1188 radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
1189 put_unlocked_mapping_entry(mapping, index, entry);
1191 spin_unlock_irq(&mapping->tree_lock);
1192 return VM_FAULT_NOPAGE;
1194 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1196 static bool dax_range_is_aligned(struct block_device *bdev,
1197 unsigned int offset, unsigned int length)
1199 unsigned short sector_size = bdev_logical_block_size(bdev);
1201 if (!IS_ALIGNED(offset, sector_size))
1203 if (!IS_ALIGNED(length, sector_size))
1209 int __dax_zero_page_range(struct block_device *bdev, sector_t sector,
1210 unsigned int offset, unsigned int length)
1212 struct blk_dax_ctl dax = {
1217 if (dax_range_is_aligned(bdev, offset, length)) {
1218 sector_t start_sector = dax.sector + (offset >> 9);
1220 return blkdev_issue_zeroout(bdev, start_sector,
1221 length >> 9, GFP_NOFS, true);
1223 if (dax_map_atomic(bdev, &dax) < 0)
1224 return PTR_ERR(dax.addr);
1225 clear_pmem(dax.addr + offset, length);
1226 dax_unmap_atomic(bdev, &dax);
1230 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1233 * dax_zero_page_range - zero a range within a page of a DAX file
1234 * @inode: The file being truncated
1235 * @from: The file offset that is being truncated to
1236 * @length: The number of bytes to zero
1237 * @get_block: The filesystem method used to translate file offsets to blocks
1239 * This function can be called by a filesystem when it is zeroing part of a
1240 * page in a DAX file. This is intended for hole-punch operations. If
1241 * you are truncating a file, the helper function dax_truncate_page() may be
1244 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1245 get_block_t get_block)
1247 struct buffer_head bh;
1248 pgoff_t index = from >> PAGE_SHIFT;
1249 unsigned offset = from & (PAGE_SIZE-1);
1252 /* Block boundary? Nothing to do */
1255 BUG_ON((offset + length) > PAGE_SIZE);
1257 memset(&bh, 0, sizeof(bh));
1258 bh.b_bdev = inode->i_sb->s_bdev;
1259 bh.b_size = PAGE_SIZE;
1260 err = get_block(inode, index, &bh, 0);
1261 if (err < 0 || !buffer_written(&bh))
1264 return __dax_zero_page_range(bh.b_bdev, to_sector(&bh, inode),
1267 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1270 * dax_truncate_page - handle a partial page being truncated in a DAX file
1271 * @inode: The file being truncated
1272 * @from: The file offset that is being truncated to
1273 * @get_block: The filesystem method used to translate file offsets to blocks
1275 * Similar to block_truncate_page(), this function can be called by a
1276 * filesystem when it is truncating a DAX file to handle the partial page.
1278 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1280 unsigned length = PAGE_ALIGN(from) - from;
1281 return dax_zero_page_range(inode, from, length, get_block);
1283 EXPORT_SYMBOL_GPL(dax_truncate_page);