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>
34 #include <linux/mmu_notifier.h>
35 #include <linux/iomap.h>
38 /* We choose 4096 entries - same as per-zone page wait tables */
39 #define DAX_WAIT_TABLE_BITS 12
40 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
42 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
44 static int __init init_dax_wait_table(void)
48 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
49 init_waitqueue_head(wait_table + i);
52 fs_initcall(init_dax_wait_table);
54 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
56 struct request_queue *q = bdev->bd_queue;
59 dax->addr = ERR_PTR(-EIO);
60 if (blk_queue_enter(q, true) != 0)
63 rc = bdev_direct_access(bdev, dax);
65 dax->addr = ERR_PTR(rc);
72 static void dax_unmap_atomic(struct block_device *bdev,
73 const struct blk_dax_ctl *dax)
75 if (IS_ERR(dax->addr))
77 blk_queue_exit(bdev->bd_queue);
80 static int dax_is_pmd_entry(void *entry)
82 return (unsigned long)entry & RADIX_DAX_PMD;
85 static int dax_is_pte_entry(void *entry)
87 return !((unsigned long)entry & RADIX_DAX_PMD);
90 static int dax_is_zero_entry(void *entry)
92 return (unsigned long)entry & RADIX_DAX_HZP;
95 static int dax_is_empty_entry(void *entry)
97 return (unsigned long)entry & RADIX_DAX_EMPTY;
100 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
102 struct page *page = alloc_pages(GFP_KERNEL, 0);
103 struct blk_dax_ctl dax = {
105 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
110 return ERR_PTR(-ENOMEM);
112 rc = dax_map_atomic(bdev, &dax);
115 memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
116 dax_unmap_atomic(bdev, &dax);
121 * DAX radix tree locking
123 struct exceptional_entry_key {
124 struct address_space *mapping;
128 struct wait_exceptional_entry_queue {
130 struct exceptional_entry_key key;
133 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
134 pgoff_t index, void *entry, struct exceptional_entry_key *key)
139 * If 'entry' is a PMD, align the 'index' that we use for the wait
140 * queue to the start of that PMD. This ensures that all offsets in
141 * the range covered by the PMD map to the same bit lock.
143 if (dax_is_pmd_entry(entry))
144 index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
146 key->mapping = mapping;
147 key->entry_start = index;
149 hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
150 return wait_table + hash;
153 static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
154 int sync, void *keyp)
156 struct exceptional_entry_key *key = keyp;
157 struct wait_exceptional_entry_queue *ewait =
158 container_of(wait, struct wait_exceptional_entry_queue, wait);
160 if (key->mapping != ewait->key.mapping ||
161 key->entry_start != ewait->key.entry_start)
163 return autoremove_wake_function(wait, mode, sync, NULL);
167 * Check whether the given slot is locked. The function must be called with
168 * mapping->tree_lock held
170 static inline int slot_locked(struct address_space *mapping, void **slot)
172 unsigned long entry = (unsigned long)
173 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
174 return entry & RADIX_DAX_ENTRY_LOCK;
178 * Mark the given slot is locked. The function must be called with
179 * mapping->tree_lock held
181 static inline void *lock_slot(struct address_space *mapping, void **slot)
183 unsigned long entry = (unsigned long)
184 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
186 entry |= RADIX_DAX_ENTRY_LOCK;
187 radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
188 return (void *)entry;
192 * Mark the given slot is unlocked. The function must be called with
193 * mapping->tree_lock held
195 static inline void *unlock_slot(struct address_space *mapping, void **slot)
197 unsigned long entry = (unsigned long)
198 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
200 entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
201 radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
202 return (void *)entry;
206 * Lookup entry in radix tree, wait for it to become unlocked if it is
207 * exceptional entry and return it. The caller must call
208 * put_unlocked_mapping_entry() when he decided not to lock the entry or
209 * put_locked_mapping_entry() when he locked the entry and now wants to
212 * The function must be called with mapping->tree_lock held.
214 static void *get_unlocked_mapping_entry(struct address_space *mapping,
215 pgoff_t index, void ***slotp)
218 struct wait_exceptional_entry_queue ewait;
219 wait_queue_head_t *wq;
221 init_wait(&ewait.wait);
222 ewait.wait.func = wake_exceptional_entry_func;
225 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
227 if (!entry || !radix_tree_exceptional_entry(entry) ||
228 !slot_locked(mapping, slot)) {
234 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
235 prepare_to_wait_exclusive(wq, &ewait.wait,
236 TASK_UNINTERRUPTIBLE);
237 spin_unlock_irq(&mapping->tree_lock);
239 finish_wait(wq, &ewait.wait);
240 spin_lock_irq(&mapping->tree_lock);
244 static void dax_unlock_mapping_entry(struct address_space *mapping,
249 spin_lock_irq(&mapping->tree_lock);
250 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
251 if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
252 !slot_locked(mapping, slot))) {
253 spin_unlock_irq(&mapping->tree_lock);
256 unlock_slot(mapping, slot);
257 spin_unlock_irq(&mapping->tree_lock);
258 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
261 static void put_locked_mapping_entry(struct address_space *mapping,
262 pgoff_t index, void *entry)
264 if (!radix_tree_exceptional_entry(entry)) {
268 dax_unlock_mapping_entry(mapping, index);
273 * Called when we are done with radix tree entry we looked up via
274 * get_unlocked_mapping_entry() and which we didn't lock in the end.
276 static void put_unlocked_mapping_entry(struct address_space *mapping,
277 pgoff_t index, void *entry)
279 if (!radix_tree_exceptional_entry(entry))
282 /* We have to wake up next waiter for the radix tree entry lock */
283 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
287 * Find radix tree entry at given index. If it points to a page, return with
288 * the page locked. If it points to the exceptional entry, return with the
289 * radix tree entry locked. If the radix tree doesn't contain given index,
290 * create empty exceptional entry for the index and return with it locked.
292 * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
293 * either return that locked entry or will return an error. This error will
294 * happen if there are any 4k entries (either zero pages or DAX entries)
295 * within the 2MiB range that we are requesting.
297 * We always favor 4k entries over 2MiB entries. There isn't a flow where we
298 * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB
299 * insertion will fail if it finds any 4k entries already in the tree, and a
300 * 4k insertion will cause an existing 2MiB entry to be unmapped and
301 * downgraded to 4k entries. This happens for both 2MiB huge zero pages as
302 * well as 2MiB empty entries.
304 * The exception to this downgrade path is for 2MiB DAX PMD entries that have
305 * real storage backing them. We will leave these real 2MiB DAX entries in
306 * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
308 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
309 * persistent memory the benefit is doubtful. We can add that later if we can
312 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
313 unsigned long size_flag)
315 bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
319 spin_lock_irq(&mapping->tree_lock);
320 entry = get_unlocked_mapping_entry(mapping, index, &slot);
323 if (size_flag & RADIX_DAX_PMD) {
324 if (!radix_tree_exceptional_entry(entry) ||
325 dax_is_pte_entry(entry)) {
326 put_unlocked_mapping_entry(mapping, index,
328 entry = ERR_PTR(-EEXIST);
331 } else { /* trying to grab a PTE entry */
332 if (radix_tree_exceptional_entry(entry) &&
333 dax_is_pmd_entry(entry) &&
334 (dax_is_zero_entry(entry) ||
335 dax_is_empty_entry(entry))) {
336 pmd_downgrade = true;
341 /* No entry for given index? Make sure radix tree is big enough. */
342 if (!entry || pmd_downgrade) {
347 * Make sure 'entry' remains valid while we drop
348 * mapping->tree_lock.
350 entry = lock_slot(mapping, slot);
353 spin_unlock_irq(&mapping->tree_lock);
355 * Besides huge zero pages the only other thing that gets
356 * downgraded are empty entries which don't need to be
359 if (pmd_downgrade && dax_is_zero_entry(entry))
360 unmap_mapping_range(mapping,
361 (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
363 err = radix_tree_preload(
364 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
367 put_locked_mapping_entry(mapping, index, entry);
370 spin_lock_irq(&mapping->tree_lock);
373 radix_tree_delete(&mapping->page_tree, index);
374 mapping->nrexceptional--;
375 dax_wake_mapping_entry_waiter(mapping, index, entry,
379 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
381 err = __radix_tree_insert(&mapping->page_tree, index,
382 dax_radix_order(entry), entry);
383 radix_tree_preload_end();
385 spin_unlock_irq(&mapping->tree_lock);
387 * Someone already created the entry? This is a
388 * normal failure when inserting PMDs in a range
389 * that already contains PTEs. In that case we want
390 * to return -EEXIST immediately.
392 if (err == -EEXIST && !(size_flag & RADIX_DAX_PMD))
395 * Our insertion of a DAX PMD entry failed, most
396 * likely because it collided with a PTE sized entry
397 * at a different index in the PMD range. We haven't
398 * inserted anything into the radix tree and have no
403 /* Good, we have inserted empty locked entry into the tree. */
404 mapping->nrexceptional++;
405 spin_unlock_irq(&mapping->tree_lock);
408 /* Normal page in radix tree? */
409 if (!radix_tree_exceptional_entry(entry)) {
410 struct page *page = entry;
413 spin_unlock_irq(&mapping->tree_lock);
415 /* Page got truncated? Retry... */
416 if (unlikely(page->mapping != mapping)) {
423 entry = lock_slot(mapping, slot);
425 spin_unlock_irq(&mapping->tree_lock);
430 * We do not necessarily hold the mapping->tree_lock when we call this
431 * function so it is possible that 'entry' is no longer a valid item in the
432 * radix tree. This is okay because all we really need to do is to find the
433 * correct waitqueue where tasks might be waiting for that old 'entry' and
436 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
437 pgoff_t index, void *entry, bool wake_all)
439 struct exceptional_entry_key key;
440 wait_queue_head_t *wq;
442 wq = dax_entry_waitqueue(mapping, index, entry, &key);
445 * Checking for locked entry and prepare_to_wait_exclusive() happens
446 * under mapping->tree_lock, ditto for entry handling in our callers.
447 * So at this point all tasks that could have seen our entry locked
448 * must be in the waitqueue and the following check will see them.
450 if (waitqueue_active(wq))
451 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
454 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
455 pgoff_t index, bool trunc)
459 struct radix_tree_root *page_tree = &mapping->page_tree;
461 spin_lock_irq(&mapping->tree_lock);
462 entry = get_unlocked_mapping_entry(mapping, index, NULL);
463 if (!entry || !radix_tree_exceptional_entry(entry))
466 (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
467 radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)))
469 radix_tree_delete(page_tree, index);
470 mapping->nrexceptional--;
473 put_unlocked_mapping_entry(mapping, index, entry);
474 spin_unlock_irq(&mapping->tree_lock);
478 * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
479 * entry to get unlocked before deleting it.
481 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
483 int ret = __dax_invalidate_mapping_entry(mapping, index, true);
486 * This gets called from truncate / punch_hole path. As such, the caller
487 * must hold locks protecting against concurrent modifications of the
488 * radix tree (usually fs-private i_mmap_sem for writing). Since the
489 * caller has seen exceptional entry for this index, we better find it
490 * at that index as well...
497 * Invalidate exceptional DAX entry if easily possible. This handles DAX
498 * entries for invalidate_inode_pages() so we evict the entry only if we can
499 * do so without blocking.
501 int dax_invalidate_mapping_entry(struct address_space *mapping, pgoff_t index)
505 struct radix_tree_root *page_tree = &mapping->page_tree;
507 spin_lock_irq(&mapping->tree_lock);
508 entry = __radix_tree_lookup(page_tree, index, NULL, &slot);
509 if (!entry || !radix_tree_exceptional_entry(entry) ||
510 slot_locked(mapping, slot))
512 if (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
513 radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
515 radix_tree_delete(page_tree, index);
516 mapping->nrexceptional--;
519 spin_unlock_irq(&mapping->tree_lock);
521 dax_wake_mapping_entry_waiter(mapping, index, entry, true);
526 * Invalidate exceptional DAX entry if it is clean.
528 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
531 return __dax_invalidate_mapping_entry(mapping, index, false);
535 * The user has performed a load from a hole in the file. Allocating
536 * a new page in the file would cause excessive storage usage for
537 * workloads with sparse files. We allocate a page cache page instead.
538 * We'll kick it out of the page cache if it's ever written to,
539 * otherwise it will simply fall out of the page cache under memory
540 * pressure without ever having been dirtied.
542 static int dax_load_hole(struct address_space *mapping, void *entry,
543 struct vm_fault *vmf)
547 /* Hole page already exists? Return it... */
548 if (!radix_tree_exceptional_entry(entry)) {
550 return VM_FAULT_LOCKED;
553 /* This will replace locked radix tree entry with a hole page */
554 page = find_or_create_page(mapping, vmf->pgoff,
555 vmf->gfp_mask | __GFP_ZERO);
559 return VM_FAULT_LOCKED;
562 static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size,
563 struct page *to, unsigned long vaddr)
565 struct blk_dax_ctl dax = {
571 if (dax_map_atomic(bdev, &dax) < 0)
572 return PTR_ERR(dax.addr);
573 vto = kmap_atomic(to);
574 copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
576 dax_unmap_atomic(bdev, &dax);
581 * By this point grab_mapping_entry() has ensured that we have a locked entry
582 * of the appropriate size so we don't have to worry about downgrading PMDs to
583 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
584 * already in the tree, we will skip the insertion and just dirty the PMD as
587 static void *dax_insert_mapping_entry(struct address_space *mapping,
588 struct vm_fault *vmf,
589 void *entry, sector_t sector,
592 struct radix_tree_root *page_tree = &mapping->page_tree;
594 bool hole_fill = false;
596 pgoff_t index = vmf->pgoff;
598 if (vmf->flags & FAULT_FLAG_WRITE)
599 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
601 /* Replacing hole page with block mapping? */
602 if (!radix_tree_exceptional_entry(entry)) {
605 * Unmap the page now before we remove it from page cache below.
606 * The page is locked so it cannot be faulted in again.
608 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
610 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
612 return ERR_PTR(error);
613 } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) {
614 /* replacing huge zero page with PMD block mapping */
615 unmap_mapping_range(mapping,
616 (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
619 spin_lock_irq(&mapping->tree_lock);
620 new_entry = dax_radix_locked_entry(sector, flags);
623 __delete_from_page_cache(entry, NULL);
624 /* Drop pagecache reference */
626 error = __radix_tree_insert(page_tree, index,
627 dax_radix_order(new_entry), new_entry);
629 new_entry = ERR_PTR(error);
632 mapping->nrexceptional++;
633 } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
635 * Only swap our new entry into the radix tree if the current
636 * entry is a zero page or an empty entry. If a normal PTE or
637 * PMD entry is already in the tree, we leave it alone. This
638 * means that if we are trying to insert a PTE and the
639 * existing entry is a PMD, we will just leave the PMD in the
640 * tree and dirty it if necessary.
642 struct radix_tree_node *node;
646 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
647 WARN_ON_ONCE(ret != entry);
648 __radix_tree_replace(page_tree, node, slot,
649 new_entry, NULL, NULL);
651 if (vmf->flags & FAULT_FLAG_WRITE)
652 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
654 spin_unlock_irq(&mapping->tree_lock);
656 radix_tree_preload_end();
658 * We don't need hole page anymore, it has been replaced with
659 * locked radix tree entry now.
661 if (mapping->a_ops->freepage)
662 mapping->a_ops->freepage(entry);
669 static inline unsigned long
670 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
672 unsigned long address;
674 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
675 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
679 /* Walk all mappings of a given index of a file and writeprotect them */
680 static void dax_mapping_entry_mkclean(struct address_space *mapping,
681 pgoff_t index, unsigned long pfn)
683 struct vm_area_struct *vma;
689 i_mmap_lock_read(mapping);
690 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
691 unsigned long address;
695 if (!(vma->vm_flags & VM_SHARED))
698 address = pgoff_address(index, vma);
700 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
702 if (pfn != pte_pfn(*ptep))
704 if (!pte_dirty(*ptep) && !pte_write(*ptep))
707 flush_cache_page(vma, address, pfn);
708 pte = ptep_clear_flush(vma, address, ptep);
709 pte = pte_wrprotect(pte);
710 pte = pte_mkclean(pte);
711 set_pte_at(vma->vm_mm, address, ptep, pte);
714 pte_unmap_unlock(ptep, ptl);
717 mmu_notifier_invalidate_page(vma->vm_mm, address);
719 i_mmap_unlock_read(mapping);
722 static int dax_writeback_one(struct block_device *bdev,
723 struct address_space *mapping, pgoff_t index, void *entry)
725 struct radix_tree_root *page_tree = &mapping->page_tree;
726 struct blk_dax_ctl dax;
727 void *entry2, **slot;
731 * A page got tagged dirty in DAX mapping? Something is seriously
734 if (WARN_ON(!radix_tree_exceptional_entry(entry)))
737 spin_lock_irq(&mapping->tree_lock);
738 entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
739 /* Entry got punched out / reallocated? */
740 if (!entry2 || !radix_tree_exceptional_entry(entry2))
743 * Entry got reallocated elsewhere? No need to writeback. We have to
744 * compare sectors as we must not bail out due to difference in lockbit
747 if (dax_radix_sector(entry2) != dax_radix_sector(entry))
749 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
750 dax_is_zero_entry(entry))) {
755 /* Another fsync thread may have already written back this entry */
756 if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
758 /* Lock the entry to serialize with page faults */
759 entry = lock_slot(mapping, slot);
761 * We can clear the tag now but we have to be careful so that concurrent
762 * dax_writeback_one() calls for the same index cannot finish before we
763 * actually flush the caches. This is achieved as the calls will look
764 * at the entry only under tree_lock and once they do that they will
765 * see the entry locked and wait for it to unlock.
767 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
768 spin_unlock_irq(&mapping->tree_lock);
771 * Even if dax_writeback_mapping_range() was given a wbc->range_start
772 * in the middle of a PMD, the 'index' we are given will be aligned to
773 * the start index of the PMD, as will the sector we pull from
774 * 'entry'. This allows us to flush for PMD_SIZE and not have to
775 * worry about partial PMD writebacks.
777 dax.sector = dax_radix_sector(entry);
778 dax.size = PAGE_SIZE << dax_radix_order(entry);
781 * We cannot hold tree_lock while calling dax_map_atomic() because it
782 * eventually calls cond_resched().
784 ret = dax_map_atomic(bdev, &dax);
786 put_locked_mapping_entry(mapping, index, entry);
790 if (WARN_ON_ONCE(ret < dax.size)) {
795 dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(dax.pfn));
796 wb_cache_pmem(dax.addr, dax.size);
798 * After we have flushed the cache, we can clear the dirty tag. There
799 * cannot be new dirty data in the pfn after the flush has completed as
800 * the pfn mappings are writeprotected and fault waits for mapping
803 spin_lock_irq(&mapping->tree_lock);
804 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
805 spin_unlock_irq(&mapping->tree_lock);
807 dax_unmap_atomic(bdev, &dax);
808 put_locked_mapping_entry(mapping, index, entry);
812 put_unlocked_mapping_entry(mapping, index, entry2);
813 spin_unlock_irq(&mapping->tree_lock);
818 * Flush the mapping to the persistent domain within the byte range of [start,
819 * end]. This is required by data integrity operations to ensure file data is
820 * on persistent storage prior to completion of the operation.
822 int dax_writeback_mapping_range(struct address_space *mapping,
823 struct block_device *bdev, struct writeback_control *wbc)
825 struct inode *inode = mapping->host;
826 pgoff_t start_index, end_index;
827 pgoff_t indices[PAGEVEC_SIZE];
832 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
835 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
838 start_index = wbc->range_start >> PAGE_SHIFT;
839 end_index = wbc->range_end >> PAGE_SHIFT;
841 tag_pages_for_writeback(mapping, start_index, end_index);
843 pagevec_init(&pvec, 0);
845 pvec.nr = find_get_entries_tag(mapping, start_index,
846 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
847 pvec.pages, indices);
852 for (i = 0; i < pvec.nr; i++) {
853 if (indices[i] > end_index) {
858 ret = dax_writeback_one(bdev, mapping, indices[i],
866 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
868 static int dax_insert_mapping(struct address_space *mapping,
869 struct block_device *bdev, sector_t sector, size_t size,
870 void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf)
872 unsigned long vaddr = vmf->address;
873 struct blk_dax_ctl dax = {
878 void *entry = *entryp;
880 if (dax_map_atomic(bdev, &dax) < 0)
881 return PTR_ERR(dax.addr);
882 dax_unmap_atomic(bdev, &dax);
884 ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector, 0);
889 return vm_insert_mixed(vma, vaddr, dax.pfn);
893 * dax_pfn_mkwrite - handle first write to DAX page
894 * @vma: The virtual memory area where the fault occurred
895 * @vmf: The description of the fault
897 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
899 struct file *file = vma->vm_file;
900 struct address_space *mapping = file->f_mapping;
902 pgoff_t index = vmf->pgoff;
904 spin_lock_irq(&mapping->tree_lock);
905 entry = get_unlocked_mapping_entry(mapping, index, &slot);
906 if (!entry || !radix_tree_exceptional_entry(entry)) {
908 put_unlocked_mapping_entry(mapping, index, entry);
909 spin_unlock_irq(&mapping->tree_lock);
910 return VM_FAULT_NOPAGE;
912 radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
913 entry = lock_slot(mapping, slot);
914 spin_unlock_irq(&mapping->tree_lock);
916 * If we race with somebody updating the PTE and finish_mkwrite_fault()
917 * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry
918 * the fault in either case.
920 finish_mkwrite_fault(vmf);
921 put_locked_mapping_entry(mapping, index, entry);
922 return VM_FAULT_NOPAGE;
924 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
926 static bool dax_range_is_aligned(struct block_device *bdev,
927 unsigned int offset, unsigned int length)
929 unsigned short sector_size = bdev_logical_block_size(bdev);
931 if (!IS_ALIGNED(offset, sector_size))
933 if (!IS_ALIGNED(length, sector_size))
939 int __dax_zero_page_range(struct block_device *bdev, sector_t sector,
940 unsigned int offset, unsigned int length)
942 struct blk_dax_ctl dax = {
947 if (dax_range_is_aligned(bdev, offset, length)) {
948 sector_t start_sector = dax.sector + (offset >> 9);
950 return blkdev_issue_zeroout(bdev, start_sector,
951 length >> 9, GFP_NOFS, true);
953 if (dax_map_atomic(bdev, &dax) < 0)
954 return PTR_ERR(dax.addr);
955 clear_pmem(dax.addr + offset, length);
956 dax_unmap_atomic(bdev, &dax);
960 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
962 #ifdef CONFIG_FS_IOMAP
963 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
965 return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
969 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
972 struct iov_iter *iter = data;
973 loff_t end = pos + length, done = 0;
976 if (iov_iter_rw(iter) == READ) {
977 end = min(end, i_size_read(inode));
981 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
982 return iov_iter_zero(min(length, end - pos), iter);
985 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
989 unsigned offset = pos & (PAGE_SIZE - 1);
990 struct blk_dax_ctl dax = { 0 };
993 dax.sector = dax_iomap_sector(iomap, pos);
994 dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK;
995 map_len = dax_map_atomic(iomap->bdev, &dax);
1003 if (map_len > end - pos)
1004 map_len = end - pos;
1006 if (iov_iter_rw(iter) == WRITE)
1007 map_len = copy_from_iter_pmem(dax.addr, map_len, iter);
1009 map_len = copy_to_iter(dax.addr, map_len, iter);
1010 dax_unmap_atomic(iomap->bdev, &dax);
1012 ret = map_len ? map_len : -EFAULT;
1021 return done ? done : ret;
1025 * dax_iomap_rw - Perform I/O to a DAX file
1026 * @iocb: The control block for this I/O
1027 * @iter: The addresses to do I/O from or to
1028 * @ops: iomap ops passed from the file system
1030 * This function performs read and write operations to directly mapped
1031 * persistent memory. The callers needs to take care of read/write exclusion
1032 * and evicting any page cache pages in the region under I/O.
1035 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1036 struct iomap_ops *ops)
1038 struct address_space *mapping = iocb->ki_filp->f_mapping;
1039 struct inode *inode = mapping->host;
1040 loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1043 if (iov_iter_rw(iter) == WRITE)
1044 flags |= IOMAP_WRITE;
1047 * Yes, even DAX files can have page cache attached to them: A zeroed
1048 * page is inserted into the pagecache when we have to serve a write
1049 * fault on a hole. It should never be dirtied and can simply be
1050 * dropped from the pagecache once we get real data for the page.
1052 * XXX: This is racy against mmap, and there's nothing we can do about
1053 * it. We'll eventually need to shift this down even further so that
1054 * we can check if we allocated blocks over a hole first.
1056 if (mapping->nrpages) {
1057 ret = invalidate_inode_pages2_range(mapping,
1059 (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT);
1063 while (iov_iter_count(iter)) {
1064 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1065 iter, dax_iomap_actor);
1072 iocb->ki_pos += done;
1073 return done ? done : ret;
1075 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1078 * dax_iomap_fault - handle a page fault on a DAX file
1079 * @vma: The virtual memory area where the fault occurred
1080 * @vmf: The description of the fault
1081 * @ops: iomap ops passed from the file system
1083 * When a page fault occurs, filesystems may call this helper in their fault
1084 * or mkwrite handler for DAX files. Assumes the caller has done all the
1085 * necessary locking for the page fault to proceed successfully.
1087 int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
1088 struct iomap_ops *ops)
1090 struct address_space *mapping = vma->vm_file->f_mapping;
1091 struct inode *inode = mapping->host;
1092 unsigned long vaddr = vmf->address;
1093 loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1095 struct iomap iomap = { 0 };
1096 unsigned flags = IOMAP_FAULT;
1097 int error, major = 0;
1102 * Check whether offset isn't beyond end of file now. Caller is supposed
1103 * to hold locks serializing us with truncate / punch hole so this is
1106 if (pos >= i_size_read(inode))
1107 return VM_FAULT_SIGBUS;
1109 entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1110 if (IS_ERR(entry)) {
1111 error = PTR_ERR(entry);
1115 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1116 flags |= IOMAP_WRITE;
1119 * Note that we don't bother to use iomap_apply here: DAX required
1120 * the file system block size to be equal the page size, which means
1121 * that we never have to deal with more than a single extent here.
1123 error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1126 if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1127 error = -EIO; /* fs corruption? */
1131 sector = dax_iomap_sector(&iomap, pos);
1133 if (vmf->cow_page) {
1134 switch (iomap.type) {
1136 case IOMAP_UNWRITTEN:
1137 clear_user_highpage(vmf->cow_page, vaddr);
1140 error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE,
1141 vmf->cow_page, vaddr);
1152 __SetPageUptodate(vmf->cow_page);
1153 vmf_ret = finish_fault(vmf);
1155 vmf_ret = VM_FAULT_DONE_COW;
1159 switch (iomap.type) {
1161 if (iomap.flags & IOMAP_F_NEW) {
1162 count_vm_event(PGMAJFAULT);
1163 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1164 major = VM_FAULT_MAJOR;
1166 error = dax_insert_mapping(mapping, iomap.bdev, sector,
1167 PAGE_SIZE, &entry, vma, vmf);
1169 case IOMAP_UNWRITTEN:
1171 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
1172 vmf_ret = dax_load_hole(mapping, entry, vmf);
1183 if (ops->iomap_end) {
1184 if (error || (vmf_ret & VM_FAULT_ERROR)) {
1185 /* keep previous error */
1186 ops->iomap_end(inode, pos, PAGE_SIZE, 0, flags,
1189 error = ops->iomap_end(inode, pos, PAGE_SIZE,
1190 PAGE_SIZE, flags, &iomap);
1194 if (vmf_ret != VM_FAULT_LOCKED || error)
1195 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1197 if (error == -ENOMEM)
1198 return VM_FAULT_OOM | major;
1199 /* -EBUSY is fine, somebody else faulted on the same PTE */
1200 if (error < 0 && error != -EBUSY)
1201 return VM_FAULT_SIGBUS | major;
1203 WARN_ON_ONCE(error); /* -EBUSY from ops->iomap_end? */
1206 return VM_FAULT_NOPAGE | major;
1208 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1210 #ifdef CONFIG_FS_DAX_PMD
1212 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
1213 * more often than one might expect in the below functions.
1215 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
1217 static int dax_pmd_insert_mapping(struct vm_area_struct *vma, pmd_t *pmd,
1218 struct vm_fault *vmf, unsigned long address,
1219 struct iomap *iomap, loff_t pos, bool write, void **entryp)
1221 struct address_space *mapping = vma->vm_file->f_mapping;
1222 struct block_device *bdev = iomap->bdev;
1223 struct blk_dax_ctl dax = {
1224 .sector = dax_iomap_sector(iomap, pos),
1227 long length = dax_map_atomic(bdev, &dax);
1230 if (length < 0) /* dax_map_atomic() failed */
1231 return VM_FAULT_FALLBACK;
1232 if (length < PMD_SIZE)
1233 goto unmap_fallback;
1234 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR)
1235 goto unmap_fallback;
1236 if (!pfn_t_devmap(dax.pfn))
1237 goto unmap_fallback;
1239 dax_unmap_atomic(bdev, &dax);
1241 ret = dax_insert_mapping_entry(mapping, vmf, *entryp, dax.sector,
1244 return VM_FAULT_FALLBACK;
1247 return vmf_insert_pfn_pmd(vma, address, pmd, dax.pfn, write);
1250 dax_unmap_atomic(bdev, &dax);
1251 return VM_FAULT_FALLBACK;
1254 static int dax_pmd_load_hole(struct vm_area_struct *vma, pmd_t *pmd,
1255 struct vm_fault *vmf, unsigned long address,
1256 struct iomap *iomap, void **entryp)
1258 struct address_space *mapping = vma->vm_file->f_mapping;
1259 unsigned long pmd_addr = address & PMD_MASK;
1260 struct page *zero_page;
1265 zero_page = mm_get_huge_zero_page(vma->vm_mm);
1267 if (unlikely(!zero_page))
1268 return VM_FAULT_FALLBACK;
1270 ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0,
1271 RADIX_DAX_PMD | RADIX_DAX_HZP);
1273 return VM_FAULT_FALLBACK;
1276 ptl = pmd_lock(vma->vm_mm, pmd);
1277 if (!pmd_none(*pmd)) {
1279 return VM_FAULT_FALLBACK;
1282 pmd_entry = mk_pmd(zero_page, vma->vm_page_prot);
1283 pmd_entry = pmd_mkhuge(pmd_entry);
1284 set_pmd_at(vma->vm_mm, pmd_addr, pmd, pmd_entry);
1286 return VM_FAULT_NOPAGE;
1289 int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1290 pmd_t *pmd, unsigned int flags, struct iomap_ops *ops)
1292 struct address_space *mapping = vma->vm_file->f_mapping;
1293 unsigned long pmd_addr = address & PMD_MASK;
1294 bool write = flags & FAULT_FLAG_WRITE;
1295 unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1296 struct inode *inode = mapping->host;
1297 int result = VM_FAULT_FALLBACK;
1298 struct iomap iomap = { 0 };
1299 pgoff_t max_pgoff, pgoff;
1300 struct vm_fault vmf;
1305 /* Fall back to PTEs if we're going to COW */
1306 if (write && !(vma->vm_flags & VM_SHARED))
1309 /* If the PMD would extend outside the VMA */
1310 if (pmd_addr < vma->vm_start)
1312 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1316 * Check whether offset isn't beyond end of file now. Caller is
1317 * supposed to hold locks serializing us with truncate / punch hole so
1318 * this is a reliable test.
1320 pgoff = linear_page_index(vma, pmd_addr);
1321 max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1323 if (pgoff > max_pgoff)
1324 return VM_FAULT_SIGBUS;
1326 /* If the PMD would extend beyond the file size */
1327 if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1331 * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1332 * PMD or a HZP entry. If it can't (because a 4k page is already in
1333 * the tree, for instance), it will return -EEXIST and we just fall
1334 * back to 4k entries.
1336 entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1341 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1342 * setting up a mapping, so really we're using iomap_begin() as a way
1343 * to look up our filesystem block.
1345 pos = (loff_t)pgoff << PAGE_SHIFT;
1346 error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1349 if (iomap.offset + iomap.length < pos + PMD_SIZE)
1354 vmf.gfp_mask = mapping_gfp_mask(mapping) | __GFP_IO;
1356 switch (iomap.type) {
1358 result = dax_pmd_insert_mapping(vma, pmd, &vmf, address,
1359 &iomap, pos, write, &entry);
1361 case IOMAP_UNWRITTEN:
1363 if (WARN_ON_ONCE(write))
1365 result = dax_pmd_load_hole(vma, pmd, &vmf, address, &iomap,
1374 if (ops->iomap_end) {
1375 if (result == VM_FAULT_FALLBACK) {
1376 ops->iomap_end(inode, pos, PMD_SIZE, 0, iomap_flags,
1379 error = ops->iomap_end(inode, pos, PMD_SIZE, PMD_SIZE,
1380 iomap_flags, &iomap);
1382 result = VM_FAULT_FALLBACK;
1386 put_locked_mapping_entry(mapping, pgoff, entry);
1388 if (result == VM_FAULT_FALLBACK) {
1389 split_huge_pmd(vma, pmd, address);
1390 count_vm_event(THP_FAULT_FALLBACK);
1394 EXPORT_SYMBOL_GPL(dax_iomap_pmd_fault);
1395 #endif /* CONFIG_FS_DAX_PMD */
1396 #endif /* CONFIG_FS_IOMAP */