Loading...
1/*
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>
6 *
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.
10 *
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
14 * more details.
15 */
16
17#include <linux/atomic.h>
18#include <linux/blkdev.h>
19#include <linux/buffer_head.h>
20#include <linux/dax.h>
21#include <linux/fs.h>
22#include <linux/genhd.h>
23#include <linux/highmem.h>
24#include <linux/memcontrol.h>
25#include <linux/mm.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>
36#include "internal.h"
37
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)
41
42static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
43
44static int __init init_dax_wait_table(void)
45{
46 int i;
47
48 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
49 init_waitqueue_head(wait_table + i);
50 return 0;
51}
52fs_initcall(init_dax_wait_table);
53
54static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
55{
56 struct request_queue *q = bdev->bd_queue;
57 long rc = -EIO;
58
59 dax->addr = ERR_PTR(-EIO);
60 if (blk_queue_enter(q, true) != 0)
61 return rc;
62
63 rc = bdev_direct_access(bdev, dax);
64 if (rc < 0) {
65 dax->addr = ERR_PTR(rc);
66 blk_queue_exit(q);
67 return rc;
68 }
69 return rc;
70}
71
72static void dax_unmap_atomic(struct block_device *bdev,
73 const struct blk_dax_ctl *dax)
74{
75 if (IS_ERR(dax->addr))
76 return;
77 blk_queue_exit(bdev->bd_queue);
78}
79
80static int dax_is_pmd_entry(void *entry)
81{
82 return (unsigned long)entry & RADIX_DAX_PMD;
83}
84
85static int dax_is_pte_entry(void *entry)
86{
87 return !((unsigned long)entry & RADIX_DAX_PMD);
88}
89
90static int dax_is_zero_entry(void *entry)
91{
92 return (unsigned long)entry & RADIX_DAX_HZP;
93}
94
95static int dax_is_empty_entry(void *entry)
96{
97 return (unsigned long)entry & RADIX_DAX_EMPTY;
98}
99
100struct page *read_dax_sector(struct block_device *bdev, sector_t n)
101{
102 struct page *page = alloc_pages(GFP_KERNEL, 0);
103 struct blk_dax_ctl dax = {
104 .size = PAGE_SIZE,
105 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
106 };
107 long rc;
108
109 if (!page)
110 return ERR_PTR(-ENOMEM);
111
112 rc = dax_map_atomic(bdev, &dax);
113 if (rc < 0)
114 return ERR_PTR(rc);
115 memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
116 dax_unmap_atomic(bdev, &dax);
117 return page;
118}
119
120/*
121 * DAX radix tree locking
122 */
123struct exceptional_entry_key {
124 struct address_space *mapping;
125 pgoff_t entry_start;
126};
127
128struct wait_exceptional_entry_queue {
129 wait_queue_t wait;
130 struct exceptional_entry_key key;
131};
132
133static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
134 pgoff_t index, void *entry, struct exceptional_entry_key *key)
135{
136 unsigned long hash;
137
138 /*
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.
142 */
143 if (dax_is_pmd_entry(entry))
144 index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
145
146 key->mapping = mapping;
147 key->entry_start = index;
148
149 hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
150 return wait_table + hash;
151}
152
153static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
154 int sync, void *keyp)
155{
156 struct exceptional_entry_key *key = keyp;
157 struct wait_exceptional_entry_queue *ewait =
158 container_of(wait, struct wait_exceptional_entry_queue, wait);
159
160 if (key->mapping != ewait->key.mapping ||
161 key->entry_start != ewait->key.entry_start)
162 return 0;
163 return autoremove_wake_function(wait, mode, sync, NULL);
164}
165
166/*
167 * Check whether the given slot is locked. The function must be called with
168 * mapping->tree_lock held
169 */
170static inline int slot_locked(struct address_space *mapping, void **slot)
171{
172 unsigned long entry = (unsigned long)
173 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
174 return entry & RADIX_DAX_ENTRY_LOCK;
175}
176
177/*
178 * Mark the given slot is locked. The function must be called with
179 * mapping->tree_lock held
180 */
181static inline void *lock_slot(struct address_space *mapping, void **slot)
182{
183 unsigned long entry = (unsigned long)
184 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
185
186 entry |= RADIX_DAX_ENTRY_LOCK;
187 radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
188 return (void *)entry;
189}
190
191/*
192 * Mark the given slot is unlocked. The function must be called with
193 * mapping->tree_lock held
194 */
195static inline void *unlock_slot(struct address_space *mapping, void **slot)
196{
197 unsigned long entry = (unsigned long)
198 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
199
200 entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
201 radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
202 return (void *)entry;
203}
204
205/*
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
210 * unlock it.
211 *
212 * The function must be called with mapping->tree_lock held.
213 */
214static void *get_unlocked_mapping_entry(struct address_space *mapping,
215 pgoff_t index, void ***slotp)
216{
217 void *entry, **slot;
218 struct wait_exceptional_entry_queue ewait;
219 wait_queue_head_t *wq;
220
221 init_wait(&ewait.wait);
222 ewait.wait.func = wake_exceptional_entry_func;
223
224 for (;;) {
225 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
226 &slot);
227 if (!entry || !radix_tree_exceptional_entry(entry) ||
228 !slot_locked(mapping, slot)) {
229 if (slotp)
230 *slotp = slot;
231 return entry;
232 }
233
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);
238 schedule();
239 finish_wait(wq, &ewait.wait);
240 spin_lock_irq(&mapping->tree_lock);
241 }
242}
243
244static void dax_unlock_mapping_entry(struct address_space *mapping,
245 pgoff_t index)
246{
247 void *entry, **slot;
248
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);
254 return;
255 }
256 unlock_slot(mapping, slot);
257 spin_unlock_irq(&mapping->tree_lock);
258 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
259}
260
261static void put_locked_mapping_entry(struct address_space *mapping,
262 pgoff_t index, void *entry)
263{
264 if (!radix_tree_exceptional_entry(entry)) {
265 unlock_page(entry);
266 put_page(entry);
267 } else {
268 dax_unlock_mapping_entry(mapping, index);
269 }
270}
271
272/*
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.
275 */
276static void put_unlocked_mapping_entry(struct address_space *mapping,
277 pgoff_t index, void *entry)
278{
279 if (!radix_tree_exceptional_entry(entry))
280 return;
281
282 /* We have to wake up next waiter for the radix tree entry lock */
283 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
284}
285
286/*
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.
291 *
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.
296 *
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.
303 *
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.
307 *
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
310 * show it helps.
311 */
312static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
313 unsigned long size_flag)
314{
315 bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
316 void *entry, **slot;
317
318restart:
319 spin_lock_irq(&mapping->tree_lock);
320 entry = get_unlocked_mapping_entry(mapping, index, &slot);
321
322 if (entry) {
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,
327 entry);
328 entry = ERR_PTR(-EEXIST);
329 goto out_unlock;
330 }
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;
337 }
338 }
339 }
340
341 /* No entry for given index? Make sure radix tree is big enough. */
342 if (!entry || pmd_downgrade) {
343 int err;
344
345 if (pmd_downgrade) {
346 /*
347 * Make sure 'entry' remains valid while we drop
348 * mapping->tree_lock.
349 */
350 entry = lock_slot(mapping, slot);
351 }
352
353 spin_unlock_irq(&mapping->tree_lock);
354 /*
355 * Besides huge zero pages the only other thing that gets
356 * downgraded are empty entries which don't need to be
357 * unmapped.
358 */
359 if (pmd_downgrade && dax_is_zero_entry(entry))
360 unmap_mapping_range(mapping,
361 (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
362
363 err = radix_tree_preload(
364 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
365 if (err) {
366 if (pmd_downgrade)
367 put_locked_mapping_entry(mapping, index, entry);
368 return ERR_PTR(err);
369 }
370 spin_lock_irq(&mapping->tree_lock);
371
372 if (!entry) {
373 /*
374 * We needed to drop the page_tree lock while calling
375 * radix_tree_preload() and we didn't have an entry to
376 * lock. See if another thread inserted an entry at
377 * our index during this time.
378 */
379 entry = __radix_tree_lookup(&mapping->page_tree, index,
380 NULL, &slot);
381 if (entry) {
382 radix_tree_preload_end();
383 spin_unlock_irq(&mapping->tree_lock);
384 goto restart;
385 }
386 }
387
388 if (pmd_downgrade) {
389 radix_tree_delete(&mapping->page_tree, index);
390 mapping->nrexceptional--;
391 dax_wake_mapping_entry_waiter(mapping, index, entry,
392 true);
393 }
394
395 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
396
397 err = __radix_tree_insert(&mapping->page_tree, index,
398 dax_radix_order(entry), entry);
399 radix_tree_preload_end();
400 if (err) {
401 spin_unlock_irq(&mapping->tree_lock);
402 /*
403 * Our insertion of a DAX entry failed, most likely
404 * because we were inserting a PMD entry and it
405 * collided with a PTE sized entry at a different
406 * index in the PMD range. We haven't inserted
407 * anything into the radix tree and have no waiters to
408 * wake.
409 */
410 return ERR_PTR(err);
411 }
412 /* Good, we have inserted empty locked entry into the tree. */
413 mapping->nrexceptional++;
414 spin_unlock_irq(&mapping->tree_lock);
415 return entry;
416 }
417 /* Normal page in radix tree? */
418 if (!radix_tree_exceptional_entry(entry)) {
419 struct page *page = entry;
420
421 get_page(page);
422 spin_unlock_irq(&mapping->tree_lock);
423 lock_page(page);
424 /* Page got truncated? Retry... */
425 if (unlikely(page->mapping != mapping)) {
426 unlock_page(page);
427 put_page(page);
428 goto restart;
429 }
430 return page;
431 }
432 entry = lock_slot(mapping, slot);
433 out_unlock:
434 spin_unlock_irq(&mapping->tree_lock);
435 return entry;
436}
437
438/*
439 * We do not necessarily hold the mapping->tree_lock when we call this
440 * function so it is possible that 'entry' is no longer a valid item in the
441 * radix tree. This is okay because all we really need to do is to find the
442 * correct waitqueue where tasks might be waiting for that old 'entry' and
443 * wake them.
444 */
445void dax_wake_mapping_entry_waiter(struct address_space *mapping,
446 pgoff_t index, void *entry, bool wake_all)
447{
448 struct exceptional_entry_key key;
449 wait_queue_head_t *wq;
450
451 wq = dax_entry_waitqueue(mapping, index, entry, &key);
452
453 /*
454 * Checking for locked entry and prepare_to_wait_exclusive() happens
455 * under mapping->tree_lock, ditto for entry handling in our callers.
456 * So at this point all tasks that could have seen our entry locked
457 * must be in the waitqueue and the following check will see them.
458 */
459 if (waitqueue_active(wq))
460 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
461}
462
463static int __dax_invalidate_mapping_entry(struct address_space *mapping,
464 pgoff_t index, bool trunc)
465{
466 int ret = 0;
467 void *entry;
468 struct radix_tree_root *page_tree = &mapping->page_tree;
469
470 spin_lock_irq(&mapping->tree_lock);
471 entry = get_unlocked_mapping_entry(mapping, index, NULL);
472 if (!entry || !radix_tree_exceptional_entry(entry))
473 goto out;
474 if (!trunc &&
475 (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
476 radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)))
477 goto out;
478 radix_tree_delete(page_tree, index);
479 mapping->nrexceptional--;
480 ret = 1;
481out:
482 put_unlocked_mapping_entry(mapping, index, entry);
483 spin_unlock_irq(&mapping->tree_lock);
484 return ret;
485}
486/*
487 * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
488 * entry to get unlocked before deleting it.
489 */
490int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
491{
492 int ret = __dax_invalidate_mapping_entry(mapping, index, true);
493
494 /*
495 * This gets called from truncate / punch_hole path. As such, the caller
496 * must hold locks protecting against concurrent modifications of the
497 * radix tree (usually fs-private i_mmap_sem for writing). Since the
498 * caller has seen exceptional entry for this index, we better find it
499 * at that index as well...
500 */
501 WARN_ON_ONCE(!ret);
502 return ret;
503}
504
505/*
506 * Invalidate exceptional DAX entry if easily possible. This handles DAX
507 * entries for invalidate_inode_pages() so we evict the entry only if we can
508 * do so without blocking.
509 */
510int dax_invalidate_mapping_entry(struct address_space *mapping, pgoff_t index)
511{
512 int ret = 0;
513 void *entry, **slot;
514 struct radix_tree_root *page_tree = &mapping->page_tree;
515
516 spin_lock_irq(&mapping->tree_lock);
517 entry = __radix_tree_lookup(page_tree, index, NULL, &slot);
518 if (!entry || !radix_tree_exceptional_entry(entry) ||
519 slot_locked(mapping, slot))
520 goto out;
521 if (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
522 radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
523 goto out;
524 radix_tree_delete(page_tree, index);
525 mapping->nrexceptional--;
526 ret = 1;
527out:
528 spin_unlock_irq(&mapping->tree_lock);
529 if (ret)
530 dax_wake_mapping_entry_waiter(mapping, index, entry, true);
531 return ret;
532}
533
534/*
535 * Invalidate exceptional DAX entry if it is clean.
536 */
537int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
538 pgoff_t index)
539{
540 return __dax_invalidate_mapping_entry(mapping, index, false);
541}
542
543/*
544 * The user has performed a load from a hole in the file. Allocating
545 * a new page in the file would cause excessive storage usage for
546 * workloads with sparse files. We allocate a page cache page instead.
547 * We'll kick it out of the page cache if it's ever written to,
548 * otherwise it will simply fall out of the page cache under memory
549 * pressure without ever having been dirtied.
550 */
551static int dax_load_hole(struct address_space *mapping, void **entry,
552 struct vm_fault *vmf)
553{
554 struct page *page;
555 int ret;
556
557 /* Hole page already exists? Return it... */
558 if (!radix_tree_exceptional_entry(*entry)) {
559 page = *entry;
560 goto out;
561 }
562
563 /* This will replace locked radix tree entry with a hole page */
564 page = find_or_create_page(mapping, vmf->pgoff,
565 vmf->gfp_mask | __GFP_ZERO);
566 if (!page)
567 return VM_FAULT_OOM;
568 out:
569 vmf->page = page;
570 ret = finish_fault(vmf);
571 vmf->page = NULL;
572 *entry = page;
573 if (!ret) {
574 /* Grab reference for PTE that is now referencing the page */
575 get_page(page);
576 return VM_FAULT_NOPAGE;
577 }
578 return ret;
579}
580
581static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size,
582 struct page *to, unsigned long vaddr)
583{
584 struct blk_dax_ctl dax = {
585 .sector = sector,
586 .size = size,
587 };
588 void *vto;
589
590 if (dax_map_atomic(bdev, &dax) < 0)
591 return PTR_ERR(dax.addr);
592 vto = kmap_atomic(to);
593 copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
594 kunmap_atomic(vto);
595 dax_unmap_atomic(bdev, &dax);
596 return 0;
597}
598
599/*
600 * By this point grab_mapping_entry() has ensured that we have a locked entry
601 * of the appropriate size so we don't have to worry about downgrading PMDs to
602 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
603 * already in the tree, we will skip the insertion and just dirty the PMD as
604 * appropriate.
605 */
606static void *dax_insert_mapping_entry(struct address_space *mapping,
607 struct vm_fault *vmf,
608 void *entry, sector_t sector,
609 unsigned long flags)
610{
611 struct radix_tree_root *page_tree = &mapping->page_tree;
612 int error = 0;
613 bool hole_fill = false;
614 void *new_entry;
615 pgoff_t index = vmf->pgoff;
616
617 if (vmf->flags & FAULT_FLAG_WRITE)
618 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
619
620 /* Replacing hole page with block mapping? */
621 if (!radix_tree_exceptional_entry(entry)) {
622 hole_fill = true;
623 /*
624 * Unmap the page now before we remove it from page cache below.
625 * The page is locked so it cannot be faulted in again.
626 */
627 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
628 PAGE_SIZE, 0);
629 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
630 if (error)
631 return ERR_PTR(error);
632 } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) {
633 /* replacing huge zero page with PMD block mapping */
634 unmap_mapping_range(mapping,
635 (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
636 }
637
638 spin_lock_irq(&mapping->tree_lock);
639 new_entry = dax_radix_locked_entry(sector, flags);
640
641 if (hole_fill) {
642 __delete_from_page_cache(entry, NULL);
643 /* Drop pagecache reference */
644 put_page(entry);
645 error = __radix_tree_insert(page_tree, index,
646 dax_radix_order(new_entry), new_entry);
647 if (error) {
648 new_entry = ERR_PTR(error);
649 goto unlock;
650 }
651 mapping->nrexceptional++;
652 } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
653 /*
654 * Only swap our new entry into the radix tree if the current
655 * entry is a zero page or an empty entry. If a normal PTE or
656 * PMD entry is already in the tree, we leave it alone. This
657 * means that if we are trying to insert a PTE and the
658 * existing entry is a PMD, we will just leave the PMD in the
659 * tree and dirty it if necessary.
660 */
661 struct radix_tree_node *node;
662 void **slot;
663 void *ret;
664
665 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
666 WARN_ON_ONCE(ret != entry);
667 __radix_tree_replace(page_tree, node, slot,
668 new_entry, NULL, NULL);
669 }
670 if (vmf->flags & FAULT_FLAG_WRITE)
671 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
672 unlock:
673 spin_unlock_irq(&mapping->tree_lock);
674 if (hole_fill) {
675 radix_tree_preload_end();
676 /*
677 * We don't need hole page anymore, it has been replaced with
678 * locked radix tree entry now.
679 */
680 if (mapping->a_ops->freepage)
681 mapping->a_ops->freepage(entry);
682 unlock_page(entry);
683 put_page(entry);
684 }
685 return new_entry;
686}
687
688static inline unsigned long
689pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
690{
691 unsigned long address;
692
693 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
694 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
695 return address;
696}
697
698/* Walk all mappings of a given index of a file and writeprotect them */
699static void dax_mapping_entry_mkclean(struct address_space *mapping,
700 pgoff_t index, unsigned long pfn)
701{
702 struct vm_area_struct *vma;
703 pte_t pte, *ptep = NULL;
704 pmd_t *pmdp = NULL;
705 spinlock_t *ptl;
706 bool changed;
707
708 i_mmap_lock_read(mapping);
709 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
710 unsigned long address;
711
712 cond_resched();
713
714 if (!(vma->vm_flags & VM_SHARED))
715 continue;
716
717 address = pgoff_address(index, vma);
718 changed = false;
719 if (follow_pte_pmd(vma->vm_mm, address, &ptep, &pmdp, &ptl))
720 continue;
721
722 if (pmdp) {
723#ifdef CONFIG_FS_DAX_PMD
724 pmd_t pmd;
725
726 if (pfn != pmd_pfn(*pmdp))
727 goto unlock_pmd;
728 if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
729 goto unlock_pmd;
730
731 flush_cache_page(vma, address, pfn);
732 pmd = pmdp_huge_clear_flush(vma, address, pmdp);
733 pmd = pmd_wrprotect(pmd);
734 pmd = pmd_mkclean(pmd);
735 set_pmd_at(vma->vm_mm, address, pmdp, pmd);
736 changed = true;
737unlock_pmd:
738 spin_unlock(ptl);
739#endif
740 } else {
741 if (pfn != pte_pfn(*ptep))
742 goto unlock_pte;
743 if (!pte_dirty(*ptep) && !pte_write(*ptep))
744 goto unlock_pte;
745
746 flush_cache_page(vma, address, pfn);
747 pte = ptep_clear_flush(vma, address, ptep);
748 pte = pte_wrprotect(pte);
749 pte = pte_mkclean(pte);
750 set_pte_at(vma->vm_mm, address, ptep, pte);
751 changed = true;
752unlock_pte:
753 pte_unmap_unlock(ptep, ptl);
754 }
755
756 if (changed)
757 mmu_notifier_invalidate_page(vma->vm_mm, address);
758 }
759 i_mmap_unlock_read(mapping);
760}
761
762static int dax_writeback_one(struct block_device *bdev,
763 struct address_space *mapping, pgoff_t index, void *entry)
764{
765 struct radix_tree_root *page_tree = &mapping->page_tree;
766 struct blk_dax_ctl dax;
767 void *entry2, **slot;
768 int ret = 0;
769
770 /*
771 * A page got tagged dirty in DAX mapping? Something is seriously
772 * wrong.
773 */
774 if (WARN_ON(!radix_tree_exceptional_entry(entry)))
775 return -EIO;
776
777 spin_lock_irq(&mapping->tree_lock);
778 entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
779 /* Entry got punched out / reallocated? */
780 if (!entry2 || !radix_tree_exceptional_entry(entry2))
781 goto put_unlocked;
782 /*
783 * Entry got reallocated elsewhere? No need to writeback. We have to
784 * compare sectors as we must not bail out due to difference in lockbit
785 * or entry type.
786 */
787 if (dax_radix_sector(entry2) != dax_radix_sector(entry))
788 goto put_unlocked;
789 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
790 dax_is_zero_entry(entry))) {
791 ret = -EIO;
792 goto put_unlocked;
793 }
794
795 /* Another fsync thread may have already written back this entry */
796 if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
797 goto put_unlocked;
798 /* Lock the entry to serialize with page faults */
799 entry = lock_slot(mapping, slot);
800 /*
801 * We can clear the tag now but we have to be careful so that concurrent
802 * dax_writeback_one() calls for the same index cannot finish before we
803 * actually flush the caches. This is achieved as the calls will look
804 * at the entry only under tree_lock and once they do that they will
805 * see the entry locked and wait for it to unlock.
806 */
807 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
808 spin_unlock_irq(&mapping->tree_lock);
809
810 /*
811 * Even if dax_writeback_mapping_range() was given a wbc->range_start
812 * in the middle of a PMD, the 'index' we are given will be aligned to
813 * the start index of the PMD, as will the sector we pull from
814 * 'entry'. This allows us to flush for PMD_SIZE and not have to
815 * worry about partial PMD writebacks.
816 */
817 dax.sector = dax_radix_sector(entry);
818 dax.size = PAGE_SIZE << dax_radix_order(entry);
819
820 /*
821 * We cannot hold tree_lock while calling dax_map_atomic() because it
822 * eventually calls cond_resched().
823 */
824 ret = dax_map_atomic(bdev, &dax);
825 if (ret < 0) {
826 put_locked_mapping_entry(mapping, index, entry);
827 return ret;
828 }
829
830 if (WARN_ON_ONCE(ret < dax.size)) {
831 ret = -EIO;
832 goto unmap;
833 }
834
835 dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(dax.pfn));
836 wb_cache_pmem(dax.addr, dax.size);
837 /*
838 * After we have flushed the cache, we can clear the dirty tag. There
839 * cannot be new dirty data in the pfn after the flush has completed as
840 * the pfn mappings are writeprotected and fault waits for mapping
841 * entry lock.
842 */
843 spin_lock_irq(&mapping->tree_lock);
844 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
845 spin_unlock_irq(&mapping->tree_lock);
846 unmap:
847 dax_unmap_atomic(bdev, &dax);
848 put_locked_mapping_entry(mapping, index, entry);
849 return ret;
850
851 put_unlocked:
852 put_unlocked_mapping_entry(mapping, index, entry2);
853 spin_unlock_irq(&mapping->tree_lock);
854 return ret;
855}
856
857/*
858 * Flush the mapping to the persistent domain within the byte range of [start,
859 * end]. This is required by data integrity operations to ensure file data is
860 * on persistent storage prior to completion of the operation.
861 */
862int dax_writeback_mapping_range(struct address_space *mapping,
863 struct block_device *bdev, struct writeback_control *wbc)
864{
865 struct inode *inode = mapping->host;
866 pgoff_t start_index, end_index;
867 pgoff_t indices[PAGEVEC_SIZE];
868 struct pagevec pvec;
869 bool done = false;
870 int i, ret = 0;
871
872 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
873 return -EIO;
874
875 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
876 return 0;
877
878 start_index = wbc->range_start >> PAGE_SHIFT;
879 end_index = wbc->range_end >> PAGE_SHIFT;
880
881 tag_pages_for_writeback(mapping, start_index, end_index);
882
883 pagevec_init(&pvec, 0);
884 while (!done) {
885 pvec.nr = find_get_entries_tag(mapping, start_index,
886 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
887 pvec.pages, indices);
888
889 if (pvec.nr == 0)
890 break;
891
892 for (i = 0; i < pvec.nr; i++) {
893 if (indices[i] > end_index) {
894 done = true;
895 break;
896 }
897
898 ret = dax_writeback_one(bdev, mapping, indices[i],
899 pvec.pages[i]);
900 if (ret < 0)
901 return ret;
902 }
903 }
904 return 0;
905}
906EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
907
908static int dax_insert_mapping(struct address_space *mapping,
909 struct block_device *bdev, sector_t sector, size_t size,
910 void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf)
911{
912 unsigned long vaddr = vmf->address;
913 struct blk_dax_ctl dax = {
914 .sector = sector,
915 .size = size,
916 };
917 void *ret;
918 void *entry = *entryp;
919
920 if (dax_map_atomic(bdev, &dax) < 0)
921 return PTR_ERR(dax.addr);
922 dax_unmap_atomic(bdev, &dax);
923
924 ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector, 0);
925 if (IS_ERR(ret))
926 return PTR_ERR(ret);
927 *entryp = ret;
928
929 return vm_insert_mixed(vma, vaddr, dax.pfn);
930}
931
932/**
933 * dax_pfn_mkwrite - handle first write to DAX page
934 * @vma: The virtual memory area where the fault occurred
935 * @vmf: The description of the fault
936 */
937int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
938{
939 struct file *file = vma->vm_file;
940 struct address_space *mapping = file->f_mapping;
941 void *entry, **slot;
942 pgoff_t index = vmf->pgoff;
943
944 spin_lock_irq(&mapping->tree_lock);
945 entry = get_unlocked_mapping_entry(mapping, index, &slot);
946 if (!entry || !radix_tree_exceptional_entry(entry)) {
947 if (entry)
948 put_unlocked_mapping_entry(mapping, index, entry);
949 spin_unlock_irq(&mapping->tree_lock);
950 return VM_FAULT_NOPAGE;
951 }
952 radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
953 entry = lock_slot(mapping, slot);
954 spin_unlock_irq(&mapping->tree_lock);
955 /*
956 * If we race with somebody updating the PTE and finish_mkwrite_fault()
957 * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry
958 * the fault in either case.
959 */
960 finish_mkwrite_fault(vmf);
961 put_locked_mapping_entry(mapping, index, entry);
962 return VM_FAULT_NOPAGE;
963}
964EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
965
966static bool dax_range_is_aligned(struct block_device *bdev,
967 unsigned int offset, unsigned int length)
968{
969 unsigned short sector_size = bdev_logical_block_size(bdev);
970
971 if (!IS_ALIGNED(offset, sector_size))
972 return false;
973 if (!IS_ALIGNED(length, sector_size))
974 return false;
975
976 return true;
977}
978
979int __dax_zero_page_range(struct block_device *bdev, sector_t sector,
980 unsigned int offset, unsigned int length)
981{
982 struct blk_dax_ctl dax = {
983 .sector = sector,
984 .size = PAGE_SIZE,
985 };
986
987 if (dax_range_is_aligned(bdev, offset, length)) {
988 sector_t start_sector = dax.sector + (offset >> 9);
989
990 return blkdev_issue_zeroout(bdev, start_sector,
991 length >> 9, GFP_NOFS, true);
992 } else {
993 if (dax_map_atomic(bdev, &dax) < 0)
994 return PTR_ERR(dax.addr);
995 clear_pmem(dax.addr + offset, length);
996 dax_unmap_atomic(bdev, &dax);
997 }
998 return 0;
999}
1000EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1001
1002static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
1003{
1004 return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
1005}
1006
1007static loff_t
1008dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1009 struct iomap *iomap)
1010{
1011 struct iov_iter *iter = data;
1012 loff_t end = pos + length, done = 0;
1013 ssize_t ret = 0;
1014
1015 if (iov_iter_rw(iter) == READ) {
1016 end = min(end, i_size_read(inode));
1017 if (pos >= end)
1018 return 0;
1019
1020 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1021 return iov_iter_zero(min(length, end - pos), iter);
1022 }
1023
1024 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1025 return -EIO;
1026
1027 /*
1028 * Write can allocate block for an area which has a hole page mapped
1029 * into page tables. We have to tear down these mappings so that data
1030 * written by write(2) is visible in mmap.
1031 */
1032 if ((iomap->flags & IOMAP_F_NEW) && inode->i_mapping->nrpages) {
1033 invalidate_inode_pages2_range(inode->i_mapping,
1034 pos >> PAGE_SHIFT,
1035 (end - 1) >> PAGE_SHIFT);
1036 }
1037
1038 while (pos < end) {
1039 unsigned offset = pos & (PAGE_SIZE - 1);
1040 struct blk_dax_ctl dax = { 0 };
1041 ssize_t map_len;
1042
1043 if (fatal_signal_pending(current)) {
1044 ret = -EINTR;
1045 break;
1046 }
1047
1048 dax.sector = dax_iomap_sector(iomap, pos);
1049 dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK;
1050 map_len = dax_map_atomic(iomap->bdev, &dax);
1051 if (map_len < 0) {
1052 ret = map_len;
1053 break;
1054 }
1055
1056 dax.addr += offset;
1057 map_len -= offset;
1058 if (map_len > end - pos)
1059 map_len = end - pos;
1060
1061 if (iov_iter_rw(iter) == WRITE)
1062 map_len = copy_from_iter_pmem(dax.addr, map_len, iter);
1063 else
1064 map_len = copy_to_iter(dax.addr, map_len, iter);
1065 dax_unmap_atomic(iomap->bdev, &dax);
1066 if (map_len <= 0) {
1067 ret = map_len ? map_len : -EFAULT;
1068 break;
1069 }
1070
1071 pos += map_len;
1072 length -= map_len;
1073 done += map_len;
1074 }
1075
1076 return done ? done : ret;
1077}
1078
1079/**
1080 * dax_iomap_rw - Perform I/O to a DAX file
1081 * @iocb: The control block for this I/O
1082 * @iter: The addresses to do I/O from or to
1083 * @ops: iomap ops passed from the file system
1084 *
1085 * This function performs read and write operations to directly mapped
1086 * persistent memory. The callers needs to take care of read/write exclusion
1087 * and evicting any page cache pages in the region under I/O.
1088 */
1089ssize_t
1090dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1091 struct iomap_ops *ops)
1092{
1093 struct address_space *mapping = iocb->ki_filp->f_mapping;
1094 struct inode *inode = mapping->host;
1095 loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1096 unsigned flags = 0;
1097
1098 if (iov_iter_rw(iter) == WRITE)
1099 flags |= IOMAP_WRITE;
1100
1101 while (iov_iter_count(iter)) {
1102 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1103 iter, dax_iomap_actor);
1104 if (ret <= 0)
1105 break;
1106 pos += ret;
1107 done += ret;
1108 }
1109
1110 iocb->ki_pos += done;
1111 return done ? done : ret;
1112}
1113EXPORT_SYMBOL_GPL(dax_iomap_rw);
1114
1115static int dax_fault_return(int error)
1116{
1117 if (error == 0)
1118 return VM_FAULT_NOPAGE;
1119 if (error == -ENOMEM)
1120 return VM_FAULT_OOM;
1121 return VM_FAULT_SIGBUS;
1122}
1123
1124/**
1125 * dax_iomap_fault - handle a page fault on a DAX file
1126 * @vma: The virtual memory area where the fault occurred
1127 * @vmf: The description of the fault
1128 * @ops: iomap ops passed from the file system
1129 *
1130 * When a page fault occurs, filesystems may call this helper in their fault
1131 * or mkwrite handler for DAX files. Assumes the caller has done all the
1132 * necessary locking for the page fault to proceed successfully.
1133 */
1134int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
1135 struct iomap_ops *ops)
1136{
1137 struct address_space *mapping = vma->vm_file->f_mapping;
1138 struct inode *inode = mapping->host;
1139 unsigned long vaddr = vmf->address;
1140 loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1141 sector_t sector;
1142 struct iomap iomap = { 0 };
1143 unsigned flags = IOMAP_FAULT;
1144 int error, major = 0;
1145 int vmf_ret = 0;
1146 void *entry;
1147
1148 /*
1149 * Check whether offset isn't beyond end of file now. Caller is supposed
1150 * to hold locks serializing us with truncate / punch hole so this is
1151 * a reliable test.
1152 */
1153 if (pos >= i_size_read(inode))
1154 return VM_FAULT_SIGBUS;
1155
1156 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1157 flags |= IOMAP_WRITE;
1158
1159 /*
1160 * Note that we don't bother to use iomap_apply here: DAX required
1161 * the file system block size to be equal the page size, which means
1162 * that we never have to deal with more than a single extent here.
1163 */
1164 error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1165 if (error)
1166 return dax_fault_return(error);
1167 if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1168 vmf_ret = dax_fault_return(-EIO); /* fs corruption? */
1169 goto finish_iomap;
1170 }
1171
1172 entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1173 if (IS_ERR(entry)) {
1174 vmf_ret = dax_fault_return(PTR_ERR(entry));
1175 goto finish_iomap;
1176 }
1177
1178 sector = dax_iomap_sector(&iomap, pos);
1179
1180 if (vmf->cow_page) {
1181 switch (iomap.type) {
1182 case IOMAP_HOLE:
1183 case IOMAP_UNWRITTEN:
1184 clear_user_highpage(vmf->cow_page, vaddr);
1185 break;
1186 case IOMAP_MAPPED:
1187 error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE,
1188 vmf->cow_page, vaddr);
1189 break;
1190 default:
1191 WARN_ON_ONCE(1);
1192 error = -EIO;
1193 break;
1194 }
1195
1196 if (error)
1197 goto error_unlock_entry;
1198
1199 __SetPageUptodate(vmf->cow_page);
1200 vmf_ret = finish_fault(vmf);
1201 if (!vmf_ret)
1202 vmf_ret = VM_FAULT_DONE_COW;
1203 goto unlock_entry;
1204 }
1205
1206 switch (iomap.type) {
1207 case IOMAP_MAPPED:
1208 if (iomap.flags & IOMAP_F_NEW) {
1209 count_vm_event(PGMAJFAULT);
1210 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1211 major = VM_FAULT_MAJOR;
1212 }
1213 error = dax_insert_mapping(mapping, iomap.bdev, sector,
1214 PAGE_SIZE, &entry, vma, vmf);
1215 /* -EBUSY is fine, somebody else faulted on the same PTE */
1216 if (error == -EBUSY)
1217 error = 0;
1218 break;
1219 case IOMAP_UNWRITTEN:
1220 case IOMAP_HOLE:
1221 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
1222 vmf_ret = dax_load_hole(mapping, &entry, vmf);
1223 goto unlock_entry;
1224 }
1225 /*FALLTHRU*/
1226 default:
1227 WARN_ON_ONCE(1);
1228 error = -EIO;
1229 break;
1230 }
1231
1232 error_unlock_entry:
1233 vmf_ret = dax_fault_return(error) | major;
1234 unlock_entry:
1235 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1236 finish_iomap:
1237 if (ops->iomap_end) {
1238 int copied = PAGE_SIZE;
1239
1240 if (vmf_ret & VM_FAULT_ERROR)
1241 copied = 0;
1242 /*
1243 * The fault is done by now and there's no way back (other
1244 * thread may be already happily using PTE we have installed).
1245 * Just ignore error from ->iomap_end since we cannot do much
1246 * with it.
1247 */
1248 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1249 }
1250 return vmf_ret;
1251}
1252EXPORT_SYMBOL_GPL(dax_iomap_fault);
1253
1254#ifdef CONFIG_FS_DAX_PMD
1255/*
1256 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
1257 * more often than one might expect in the below functions.
1258 */
1259#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
1260
1261static int dax_pmd_insert_mapping(struct vm_area_struct *vma, pmd_t *pmd,
1262 struct vm_fault *vmf, unsigned long address,
1263 struct iomap *iomap, loff_t pos, bool write, void **entryp)
1264{
1265 struct address_space *mapping = vma->vm_file->f_mapping;
1266 struct block_device *bdev = iomap->bdev;
1267 struct blk_dax_ctl dax = {
1268 .sector = dax_iomap_sector(iomap, pos),
1269 .size = PMD_SIZE,
1270 };
1271 long length = dax_map_atomic(bdev, &dax);
1272 void *ret;
1273
1274 if (length < 0) /* dax_map_atomic() failed */
1275 return VM_FAULT_FALLBACK;
1276 if (length < PMD_SIZE)
1277 goto unmap_fallback;
1278 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR)
1279 goto unmap_fallback;
1280 if (!pfn_t_devmap(dax.pfn))
1281 goto unmap_fallback;
1282
1283 dax_unmap_atomic(bdev, &dax);
1284
1285 ret = dax_insert_mapping_entry(mapping, vmf, *entryp, dax.sector,
1286 RADIX_DAX_PMD);
1287 if (IS_ERR(ret))
1288 return VM_FAULT_FALLBACK;
1289 *entryp = ret;
1290
1291 return vmf_insert_pfn_pmd(vma, address, pmd, dax.pfn, write);
1292
1293 unmap_fallback:
1294 dax_unmap_atomic(bdev, &dax);
1295 return VM_FAULT_FALLBACK;
1296}
1297
1298static int dax_pmd_load_hole(struct vm_area_struct *vma, pmd_t *pmd,
1299 struct vm_fault *vmf, unsigned long address,
1300 struct iomap *iomap, void **entryp)
1301{
1302 struct address_space *mapping = vma->vm_file->f_mapping;
1303 unsigned long pmd_addr = address & PMD_MASK;
1304 struct page *zero_page;
1305 spinlock_t *ptl;
1306 pmd_t pmd_entry;
1307 void *ret;
1308
1309 zero_page = mm_get_huge_zero_page(vma->vm_mm);
1310
1311 if (unlikely(!zero_page))
1312 return VM_FAULT_FALLBACK;
1313
1314 ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0,
1315 RADIX_DAX_PMD | RADIX_DAX_HZP);
1316 if (IS_ERR(ret))
1317 return VM_FAULT_FALLBACK;
1318 *entryp = ret;
1319
1320 ptl = pmd_lock(vma->vm_mm, pmd);
1321 if (!pmd_none(*pmd)) {
1322 spin_unlock(ptl);
1323 return VM_FAULT_FALLBACK;
1324 }
1325
1326 pmd_entry = mk_pmd(zero_page, vma->vm_page_prot);
1327 pmd_entry = pmd_mkhuge(pmd_entry);
1328 set_pmd_at(vma->vm_mm, pmd_addr, pmd, pmd_entry);
1329 spin_unlock(ptl);
1330 return VM_FAULT_NOPAGE;
1331}
1332
1333int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1334 pmd_t *pmd, unsigned int flags, struct iomap_ops *ops)
1335{
1336 struct address_space *mapping = vma->vm_file->f_mapping;
1337 unsigned long pmd_addr = address & PMD_MASK;
1338 bool write = flags & FAULT_FLAG_WRITE;
1339 unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1340 struct inode *inode = mapping->host;
1341 int result = VM_FAULT_FALLBACK;
1342 struct iomap iomap = { 0 };
1343 pgoff_t max_pgoff, pgoff;
1344 struct vm_fault vmf;
1345 void *entry;
1346 loff_t pos;
1347 int error;
1348
1349 /* Fall back to PTEs if we're going to COW */
1350 if (write && !(vma->vm_flags & VM_SHARED))
1351 goto fallback;
1352
1353 /* If the PMD would extend outside the VMA */
1354 if (pmd_addr < vma->vm_start)
1355 goto fallback;
1356 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1357 goto fallback;
1358
1359 /*
1360 * Check whether offset isn't beyond end of file now. Caller is
1361 * supposed to hold locks serializing us with truncate / punch hole so
1362 * this is a reliable test.
1363 */
1364 pgoff = linear_page_index(vma, pmd_addr);
1365 max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1366
1367 if (pgoff > max_pgoff)
1368 return VM_FAULT_SIGBUS;
1369
1370 /* If the PMD would extend beyond the file size */
1371 if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1372 goto fallback;
1373
1374 /*
1375 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1376 * setting up a mapping, so really we're using iomap_begin() as a way
1377 * to look up our filesystem block.
1378 */
1379 pos = (loff_t)pgoff << PAGE_SHIFT;
1380 error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1381 if (error)
1382 goto fallback;
1383
1384 if (iomap.offset + iomap.length < pos + PMD_SIZE)
1385 goto finish_iomap;
1386
1387 /*
1388 * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1389 * PMD or a HZP entry. If it can't (because a 4k page is already in
1390 * the tree, for instance), it will return -EEXIST and we just fall
1391 * back to 4k entries.
1392 */
1393 entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1394 if (IS_ERR(entry))
1395 goto finish_iomap;
1396
1397 vmf.pgoff = pgoff;
1398 vmf.flags = flags;
1399 vmf.gfp_mask = mapping_gfp_mask(mapping) | __GFP_IO;
1400
1401 switch (iomap.type) {
1402 case IOMAP_MAPPED:
1403 result = dax_pmd_insert_mapping(vma, pmd, &vmf, address,
1404 &iomap, pos, write, &entry);
1405 break;
1406 case IOMAP_UNWRITTEN:
1407 case IOMAP_HOLE:
1408 if (WARN_ON_ONCE(write))
1409 goto unlock_entry;
1410 result = dax_pmd_load_hole(vma, pmd, &vmf, address, &iomap,
1411 &entry);
1412 break;
1413 default:
1414 WARN_ON_ONCE(1);
1415 break;
1416 }
1417
1418 unlock_entry:
1419 put_locked_mapping_entry(mapping, pgoff, entry);
1420 finish_iomap:
1421 if (ops->iomap_end) {
1422 int copied = PMD_SIZE;
1423
1424 if (result == VM_FAULT_FALLBACK)
1425 copied = 0;
1426 /*
1427 * The fault is done by now and there's no way back (other
1428 * thread may be already happily using PMD we have installed).
1429 * Just ignore error from ->iomap_end since we cannot do much
1430 * with it.
1431 */
1432 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1433 &iomap);
1434 }
1435 fallback:
1436 if (result == VM_FAULT_FALLBACK) {
1437 split_huge_pmd(vma, pmd, address);
1438 count_vm_event(THP_FAULT_FALLBACK);
1439 }
1440 return result;
1441}
1442EXPORT_SYMBOL_GPL(dax_iomap_pmd_fault);
1443#endif /* CONFIG_FS_DAX_PMD */
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * fs/dax.c - Direct Access filesystem code
4 * Copyright (c) 2013-2014 Intel Corporation
5 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
6 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
7 */
8
9#include <linux/atomic.h>
10#include <linux/blkdev.h>
11#include <linux/buffer_head.h>
12#include <linux/dax.h>
13#include <linux/fs.h>
14#include <linux/highmem.h>
15#include <linux/memcontrol.h>
16#include <linux/mm.h>
17#include <linux/mutex.h>
18#include <linux/pagevec.h>
19#include <linux/sched.h>
20#include <linux/sched/signal.h>
21#include <linux/uio.h>
22#include <linux/vmstat.h>
23#include <linux/pfn_t.h>
24#include <linux/sizes.h>
25#include <linux/mmu_notifier.h>
26#include <linux/iomap.h>
27#include <linux/rmap.h>
28#include <asm/pgalloc.h>
29
30#define CREATE_TRACE_POINTS
31#include <trace/events/fs_dax.h>
32
33/* We choose 4096 entries - same as per-zone page wait tables */
34#define DAX_WAIT_TABLE_BITS 12
35#define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
36
37/* The 'colour' (ie low bits) within a PMD of a page offset. */
38#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
39#define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
40
41static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
42
43static int __init init_dax_wait_table(void)
44{
45 int i;
46
47 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
48 init_waitqueue_head(wait_table + i);
49 return 0;
50}
51fs_initcall(init_dax_wait_table);
52
53/*
54 * DAX pagecache entries use XArray value entries so they can't be mistaken
55 * for pages. We use one bit for locking, one bit for the entry size (PMD)
56 * and two more to tell us if the entry is a zero page or an empty entry that
57 * is just used for locking. In total four special bits.
58 *
59 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
60 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
61 * block allocation.
62 */
63#define DAX_SHIFT (4)
64#define DAX_LOCKED (1UL << 0)
65#define DAX_PMD (1UL << 1)
66#define DAX_ZERO_PAGE (1UL << 2)
67#define DAX_EMPTY (1UL << 3)
68
69static unsigned long dax_to_pfn(void *entry)
70{
71 return xa_to_value(entry) >> DAX_SHIFT;
72}
73
74static void *dax_make_entry(pfn_t pfn, unsigned long flags)
75{
76 return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
77}
78
79static bool dax_is_locked(void *entry)
80{
81 return xa_to_value(entry) & DAX_LOCKED;
82}
83
84static unsigned int dax_entry_order(void *entry)
85{
86 if (xa_to_value(entry) & DAX_PMD)
87 return PMD_ORDER;
88 return 0;
89}
90
91static unsigned long dax_is_pmd_entry(void *entry)
92{
93 return xa_to_value(entry) & DAX_PMD;
94}
95
96static bool dax_is_pte_entry(void *entry)
97{
98 return !(xa_to_value(entry) & DAX_PMD);
99}
100
101static int dax_is_zero_entry(void *entry)
102{
103 return xa_to_value(entry) & DAX_ZERO_PAGE;
104}
105
106static int dax_is_empty_entry(void *entry)
107{
108 return xa_to_value(entry) & DAX_EMPTY;
109}
110
111/*
112 * true if the entry that was found is of a smaller order than the entry
113 * we were looking for
114 */
115static bool dax_is_conflict(void *entry)
116{
117 return entry == XA_RETRY_ENTRY;
118}
119
120/*
121 * DAX page cache entry locking
122 */
123struct exceptional_entry_key {
124 struct xarray *xa;
125 pgoff_t entry_start;
126};
127
128struct wait_exceptional_entry_queue {
129 wait_queue_entry_t wait;
130 struct exceptional_entry_key key;
131};
132
133/**
134 * enum dax_wake_mode: waitqueue wakeup behaviour
135 * @WAKE_ALL: wake all waiters in the waitqueue
136 * @WAKE_NEXT: wake only the first waiter in the waitqueue
137 */
138enum dax_wake_mode {
139 WAKE_ALL,
140 WAKE_NEXT,
141};
142
143static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
144 void *entry, struct exceptional_entry_key *key)
145{
146 unsigned long hash;
147 unsigned long index = xas->xa_index;
148
149 /*
150 * If 'entry' is a PMD, align the 'index' that we use for the wait
151 * queue to the start of that PMD. This ensures that all offsets in
152 * the range covered by the PMD map to the same bit lock.
153 */
154 if (dax_is_pmd_entry(entry))
155 index &= ~PG_PMD_COLOUR;
156 key->xa = xas->xa;
157 key->entry_start = index;
158
159 hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
160 return wait_table + hash;
161}
162
163static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
164 unsigned int mode, int sync, void *keyp)
165{
166 struct exceptional_entry_key *key = keyp;
167 struct wait_exceptional_entry_queue *ewait =
168 container_of(wait, struct wait_exceptional_entry_queue, wait);
169
170 if (key->xa != ewait->key.xa ||
171 key->entry_start != ewait->key.entry_start)
172 return 0;
173 return autoremove_wake_function(wait, mode, sync, NULL);
174}
175
176/*
177 * @entry may no longer be the entry at the index in the mapping.
178 * The important information it's conveying is whether the entry at
179 * this index used to be a PMD entry.
180 */
181static void dax_wake_entry(struct xa_state *xas, void *entry,
182 enum dax_wake_mode mode)
183{
184 struct exceptional_entry_key key;
185 wait_queue_head_t *wq;
186
187 wq = dax_entry_waitqueue(xas, entry, &key);
188
189 /*
190 * Checking for locked entry and prepare_to_wait_exclusive() happens
191 * under the i_pages lock, ditto for entry handling in our callers.
192 * So at this point all tasks that could have seen our entry locked
193 * must be in the waitqueue and the following check will see them.
194 */
195 if (waitqueue_active(wq))
196 __wake_up(wq, TASK_NORMAL, mode == WAKE_ALL ? 0 : 1, &key);
197}
198
199/*
200 * Look up entry in page cache, wait for it to become unlocked if it
201 * is a DAX entry and return it. The caller must subsequently call
202 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
203 * if it did. The entry returned may have a larger order than @order.
204 * If @order is larger than the order of the entry found in i_pages, this
205 * function returns a dax_is_conflict entry.
206 *
207 * Must be called with the i_pages lock held.
208 */
209static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
210{
211 void *entry;
212 struct wait_exceptional_entry_queue ewait;
213 wait_queue_head_t *wq;
214
215 init_wait(&ewait.wait);
216 ewait.wait.func = wake_exceptional_entry_func;
217
218 for (;;) {
219 entry = xas_find_conflict(xas);
220 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
221 return entry;
222 if (dax_entry_order(entry) < order)
223 return XA_RETRY_ENTRY;
224 if (!dax_is_locked(entry))
225 return entry;
226
227 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
228 prepare_to_wait_exclusive(wq, &ewait.wait,
229 TASK_UNINTERRUPTIBLE);
230 xas_unlock_irq(xas);
231 xas_reset(xas);
232 schedule();
233 finish_wait(wq, &ewait.wait);
234 xas_lock_irq(xas);
235 }
236}
237
238/*
239 * The only thing keeping the address space around is the i_pages lock
240 * (it's cycled in clear_inode() after removing the entries from i_pages)
241 * After we call xas_unlock_irq(), we cannot touch xas->xa.
242 */
243static void wait_entry_unlocked(struct xa_state *xas, void *entry)
244{
245 struct wait_exceptional_entry_queue ewait;
246 wait_queue_head_t *wq;
247
248 init_wait(&ewait.wait);
249 ewait.wait.func = wake_exceptional_entry_func;
250
251 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
252 /*
253 * Unlike get_unlocked_entry() there is no guarantee that this
254 * path ever successfully retrieves an unlocked entry before an
255 * inode dies. Perform a non-exclusive wait in case this path
256 * never successfully performs its own wake up.
257 */
258 prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
259 xas_unlock_irq(xas);
260 schedule();
261 finish_wait(wq, &ewait.wait);
262}
263
264static void put_unlocked_entry(struct xa_state *xas, void *entry,
265 enum dax_wake_mode mode)
266{
267 if (entry && !dax_is_conflict(entry))
268 dax_wake_entry(xas, entry, mode);
269}
270
271/*
272 * We used the xa_state to get the entry, but then we locked the entry and
273 * dropped the xa_lock, so we know the xa_state is stale and must be reset
274 * before use.
275 */
276static void dax_unlock_entry(struct xa_state *xas, void *entry)
277{
278 void *old;
279
280 BUG_ON(dax_is_locked(entry));
281 xas_reset(xas);
282 xas_lock_irq(xas);
283 old = xas_store(xas, entry);
284 xas_unlock_irq(xas);
285 BUG_ON(!dax_is_locked(old));
286 dax_wake_entry(xas, entry, WAKE_NEXT);
287}
288
289/*
290 * Return: The entry stored at this location before it was locked.
291 */
292static void *dax_lock_entry(struct xa_state *xas, void *entry)
293{
294 unsigned long v = xa_to_value(entry);
295 return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
296}
297
298static unsigned long dax_entry_size(void *entry)
299{
300 if (dax_is_zero_entry(entry))
301 return 0;
302 else if (dax_is_empty_entry(entry))
303 return 0;
304 else if (dax_is_pmd_entry(entry))
305 return PMD_SIZE;
306 else
307 return PAGE_SIZE;
308}
309
310static unsigned long dax_end_pfn(void *entry)
311{
312 return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
313}
314
315/*
316 * Iterate through all mapped pfns represented by an entry, i.e. skip
317 * 'empty' and 'zero' entries.
318 */
319#define for_each_mapped_pfn(entry, pfn) \
320 for (pfn = dax_to_pfn(entry); \
321 pfn < dax_end_pfn(entry); pfn++)
322
323static inline bool dax_page_is_shared(struct page *page)
324{
325 return page->mapping == PAGE_MAPPING_DAX_SHARED;
326}
327
328/*
329 * Set the page->mapping with PAGE_MAPPING_DAX_SHARED flag, increase the
330 * refcount.
331 */
332static inline void dax_page_share_get(struct page *page)
333{
334 if (page->mapping != PAGE_MAPPING_DAX_SHARED) {
335 /*
336 * Reset the index if the page was already mapped
337 * regularly before.
338 */
339 if (page->mapping)
340 page->share = 1;
341 page->mapping = PAGE_MAPPING_DAX_SHARED;
342 }
343 page->share++;
344}
345
346static inline unsigned long dax_page_share_put(struct page *page)
347{
348 return --page->share;
349}
350
351/*
352 * When it is called in dax_insert_entry(), the shared flag will indicate that
353 * whether this entry is shared by multiple files. If so, set the page->mapping
354 * PAGE_MAPPING_DAX_SHARED, and use page->share as refcount.
355 */
356static void dax_associate_entry(void *entry, struct address_space *mapping,
357 struct vm_area_struct *vma, unsigned long address, bool shared)
358{
359 unsigned long size = dax_entry_size(entry), pfn, index;
360 int i = 0;
361
362 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
363 return;
364
365 index = linear_page_index(vma, address & ~(size - 1));
366 for_each_mapped_pfn(entry, pfn) {
367 struct page *page = pfn_to_page(pfn);
368
369 if (shared) {
370 dax_page_share_get(page);
371 } else {
372 WARN_ON_ONCE(page->mapping);
373 page->mapping = mapping;
374 page->index = index + i++;
375 }
376 }
377}
378
379static void dax_disassociate_entry(void *entry, struct address_space *mapping,
380 bool trunc)
381{
382 unsigned long pfn;
383
384 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
385 return;
386
387 for_each_mapped_pfn(entry, pfn) {
388 struct page *page = pfn_to_page(pfn);
389
390 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
391 if (dax_page_is_shared(page)) {
392 /* keep the shared flag if this page is still shared */
393 if (dax_page_share_put(page) > 0)
394 continue;
395 } else
396 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
397 page->mapping = NULL;
398 page->index = 0;
399 }
400}
401
402static struct page *dax_busy_page(void *entry)
403{
404 unsigned long pfn;
405
406 for_each_mapped_pfn(entry, pfn) {
407 struct page *page = pfn_to_page(pfn);
408
409 if (page_ref_count(page) > 1)
410 return page;
411 }
412 return NULL;
413}
414
415/**
416 * dax_lock_folio - Lock the DAX entry corresponding to a folio
417 * @folio: The folio whose entry we want to lock
418 *
419 * Context: Process context.
420 * Return: A cookie to pass to dax_unlock_folio() or 0 if the entry could
421 * not be locked.
422 */
423dax_entry_t dax_lock_folio(struct folio *folio)
424{
425 XA_STATE(xas, NULL, 0);
426 void *entry;
427
428 /* Ensure folio->mapping isn't freed while we look at it */
429 rcu_read_lock();
430 for (;;) {
431 struct address_space *mapping = READ_ONCE(folio->mapping);
432
433 entry = NULL;
434 if (!mapping || !dax_mapping(mapping))
435 break;
436
437 /*
438 * In the device-dax case there's no need to lock, a
439 * struct dev_pagemap pin is sufficient to keep the
440 * inode alive, and we assume we have dev_pagemap pin
441 * otherwise we would not have a valid pfn_to_page()
442 * translation.
443 */
444 entry = (void *)~0UL;
445 if (S_ISCHR(mapping->host->i_mode))
446 break;
447
448 xas.xa = &mapping->i_pages;
449 xas_lock_irq(&xas);
450 if (mapping != folio->mapping) {
451 xas_unlock_irq(&xas);
452 continue;
453 }
454 xas_set(&xas, folio->index);
455 entry = xas_load(&xas);
456 if (dax_is_locked(entry)) {
457 rcu_read_unlock();
458 wait_entry_unlocked(&xas, entry);
459 rcu_read_lock();
460 continue;
461 }
462 dax_lock_entry(&xas, entry);
463 xas_unlock_irq(&xas);
464 break;
465 }
466 rcu_read_unlock();
467 return (dax_entry_t)entry;
468}
469
470void dax_unlock_folio(struct folio *folio, dax_entry_t cookie)
471{
472 struct address_space *mapping = folio->mapping;
473 XA_STATE(xas, &mapping->i_pages, folio->index);
474
475 if (S_ISCHR(mapping->host->i_mode))
476 return;
477
478 dax_unlock_entry(&xas, (void *)cookie);
479}
480
481/*
482 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping
483 * @mapping: the file's mapping whose entry we want to lock
484 * @index: the offset within this file
485 * @page: output the dax page corresponding to this dax entry
486 *
487 * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry
488 * could not be locked.
489 */
490dax_entry_t dax_lock_mapping_entry(struct address_space *mapping, pgoff_t index,
491 struct page **page)
492{
493 XA_STATE(xas, NULL, 0);
494 void *entry;
495
496 rcu_read_lock();
497 for (;;) {
498 entry = NULL;
499 if (!dax_mapping(mapping))
500 break;
501
502 xas.xa = &mapping->i_pages;
503 xas_lock_irq(&xas);
504 xas_set(&xas, index);
505 entry = xas_load(&xas);
506 if (dax_is_locked(entry)) {
507 rcu_read_unlock();
508 wait_entry_unlocked(&xas, entry);
509 rcu_read_lock();
510 continue;
511 }
512 if (!entry ||
513 dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
514 /*
515 * Because we are looking for entry from file's mapping
516 * and index, so the entry may not be inserted for now,
517 * or even a zero/empty entry. We don't think this is
518 * an error case. So, return a special value and do
519 * not output @page.
520 */
521 entry = (void *)~0UL;
522 } else {
523 *page = pfn_to_page(dax_to_pfn(entry));
524 dax_lock_entry(&xas, entry);
525 }
526 xas_unlock_irq(&xas);
527 break;
528 }
529 rcu_read_unlock();
530 return (dax_entry_t)entry;
531}
532
533void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index,
534 dax_entry_t cookie)
535{
536 XA_STATE(xas, &mapping->i_pages, index);
537
538 if (cookie == ~0UL)
539 return;
540
541 dax_unlock_entry(&xas, (void *)cookie);
542}
543
544/*
545 * Find page cache entry at given index. If it is a DAX entry, return it
546 * with the entry locked. If the page cache doesn't contain an entry at
547 * that index, add a locked empty entry.
548 *
549 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
550 * either return that locked entry or will return VM_FAULT_FALLBACK.
551 * This will happen if there are any PTE entries within the PMD range
552 * that we are requesting.
553 *
554 * We always favor PTE entries over PMD entries. There isn't a flow where we
555 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
556 * insertion will fail if it finds any PTE entries already in the tree, and a
557 * PTE insertion will cause an existing PMD entry to be unmapped and
558 * downgraded to PTE entries. This happens for both PMD zero pages as
559 * well as PMD empty entries.
560 *
561 * The exception to this downgrade path is for PMD entries that have
562 * real storage backing them. We will leave these real PMD entries in
563 * the tree, and PTE writes will simply dirty the entire PMD entry.
564 *
565 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
566 * persistent memory the benefit is doubtful. We can add that later if we can
567 * show it helps.
568 *
569 * On error, this function does not return an ERR_PTR. Instead it returns
570 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
571 * overlap with xarray value entries.
572 */
573static void *grab_mapping_entry(struct xa_state *xas,
574 struct address_space *mapping, unsigned int order)
575{
576 unsigned long index = xas->xa_index;
577 bool pmd_downgrade; /* splitting PMD entry into PTE entries? */
578 void *entry;
579
580retry:
581 pmd_downgrade = false;
582 xas_lock_irq(xas);
583 entry = get_unlocked_entry(xas, order);
584
585 if (entry) {
586 if (dax_is_conflict(entry))
587 goto fallback;
588 if (!xa_is_value(entry)) {
589 xas_set_err(xas, -EIO);
590 goto out_unlock;
591 }
592
593 if (order == 0) {
594 if (dax_is_pmd_entry(entry) &&
595 (dax_is_zero_entry(entry) ||
596 dax_is_empty_entry(entry))) {
597 pmd_downgrade = true;
598 }
599 }
600 }
601
602 if (pmd_downgrade) {
603 /*
604 * Make sure 'entry' remains valid while we drop
605 * the i_pages lock.
606 */
607 dax_lock_entry(xas, entry);
608
609 /*
610 * Besides huge zero pages the only other thing that gets
611 * downgraded are empty entries which don't need to be
612 * unmapped.
613 */
614 if (dax_is_zero_entry(entry)) {
615 xas_unlock_irq(xas);
616 unmap_mapping_pages(mapping,
617 xas->xa_index & ~PG_PMD_COLOUR,
618 PG_PMD_NR, false);
619 xas_reset(xas);
620 xas_lock_irq(xas);
621 }
622
623 dax_disassociate_entry(entry, mapping, false);
624 xas_store(xas, NULL); /* undo the PMD join */
625 dax_wake_entry(xas, entry, WAKE_ALL);
626 mapping->nrpages -= PG_PMD_NR;
627 entry = NULL;
628 xas_set(xas, index);
629 }
630
631 if (entry) {
632 dax_lock_entry(xas, entry);
633 } else {
634 unsigned long flags = DAX_EMPTY;
635
636 if (order > 0)
637 flags |= DAX_PMD;
638 entry = dax_make_entry(pfn_to_pfn_t(0), flags);
639 dax_lock_entry(xas, entry);
640 if (xas_error(xas))
641 goto out_unlock;
642 mapping->nrpages += 1UL << order;
643 }
644
645out_unlock:
646 xas_unlock_irq(xas);
647 if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
648 goto retry;
649 if (xas->xa_node == XA_ERROR(-ENOMEM))
650 return xa_mk_internal(VM_FAULT_OOM);
651 if (xas_error(xas))
652 return xa_mk_internal(VM_FAULT_SIGBUS);
653 return entry;
654fallback:
655 xas_unlock_irq(xas);
656 return xa_mk_internal(VM_FAULT_FALLBACK);
657}
658
659/**
660 * dax_layout_busy_page_range - find first pinned page in @mapping
661 * @mapping: address space to scan for a page with ref count > 1
662 * @start: Starting offset. Page containing 'start' is included.
663 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
664 * pages from 'start' till the end of file are included.
665 *
666 * DAX requires ZONE_DEVICE mapped pages. These pages are never
667 * 'onlined' to the page allocator so they are considered idle when
668 * page->count == 1. A filesystem uses this interface to determine if
669 * any page in the mapping is busy, i.e. for DMA, or other
670 * get_user_pages() usages.
671 *
672 * It is expected that the filesystem is holding locks to block the
673 * establishment of new mappings in this address_space. I.e. it expects
674 * to be able to run unmap_mapping_range() and subsequently not race
675 * mapping_mapped() becoming true.
676 */
677struct page *dax_layout_busy_page_range(struct address_space *mapping,
678 loff_t start, loff_t end)
679{
680 void *entry;
681 unsigned int scanned = 0;
682 struct page *page = NULL;
683 pgoff_t start_idx = start >> PAGE_SHIFT;
684 pgoff_t end_idx;
685 XA_STATE(xas, &mapping->i_pages, start_idx);
686
687 /*
688 * In the 'limited' case get_user_pages() for dax is disabled.
689 */
690 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
691 return NULL;
692
693 if (!dax_mapping(mapping) || !mapping_mapped(mapping))
694 return NULL;
695
696 /* If end == LLONG_MAX, all pages from start to till end of file */
697 if (end == LLONG_MAX)
698 end_idx = ULONG_MAX;
699 else
700 end_idx = end >> PAGE_SHIFT;
701 /*
702 * If we race get_user_pages_fast() here either we'll see the
703 * elevated page count in the iteration and wait, or
704 * get_user_pages_fast() will see that the page it took a reference
705 * against is no longer mapped in the page tables and bail to the
706 * get_user_pages() slow path. The slow path is protected by
707 * pte_lock() and pmd_lock(). New references are not taken without
708 * holding those locks, and unmap_mapping_pages() will not zero the
709 * pte or pmd without holding the respective lock, so we are
710 * guaranteed to either see new references or prevent new
711 * references from being established.
712 */
713 unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0);
714
715 xas_lock_irq(&xas);
716 xas_for_each(&xas, entry, end_idx) {
717 if (WARN_ON_ONCE(!xa_is_value(entry)))
718 continue;
719 if (unlikely(dax_is_locked(entry)))
720 entry = get_unlocked_entry(&xas, 0);
721 if (entry)
722 page = dax_busy_page(entry);
723 put_unlocked_entry(&xas, entry, WAKE_NEXT);
724 if (page)
725 break;
726 if (++scanned % XA_CHECK_SCHED)
727 continue;
728
729 xas_pause(&xas);
730 xas_unlock_irq(&xas);
731 cond_resched();
732 xas_lock_irq(&xas);
733 }
734 xas_unlock_irq(&xas);
735 return page;
736}
737EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
738
739struct page *dax_layout_busy_page(struct address_space *mapping)
740{
741 return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
742}
743EXPORT_SYMBOL_GPL(dax_layout_busy_page);
744
745static int __dax_invalidate_entry(struct address_space *mapping,
746 pgoff_t index, bool trunc)
747{
748 XA_STATE(xas, &mapping->i_pages, index);
749 int ret = 0;
750 void *entry;
751
752 xas_lock_irq(&xas);
753 entry = get_unlocked_entry(&xas, 0);
754 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
755 goto out;
756 if (!trunc &&
757 (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
758 xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
759 goto out;
760 dax_disassociate_entry(entry, mapping, trunc);
761 xas_store(&xas, NULL);
762 mapping->nrpages -= 1UL << dax_entry_order(entry);
763 ret = 1;
764out:
765 put_unlocked_entry(&xas, entry, WAKE_ALL);
766 xas_unlock_irq(&xas);
767 return ret;
768}
769
770static int __dax_clear_dirty_range(struct address_space *mapping,
771 pgoff_t start, pgoff_t end)
772{
773 XA_STATE(xas, &mapping->i_pages, start);
774 unsigned int scanned = 0;
775 void *entry;
776
777 xas_lock_irq(&xas);
778 xas_for_each(&xas, entry, end) {
779 entry = get_unlocked_entry(&xas, 0);
780 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
781 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
782 put_unlocked_entry(&xas, entry, WAKE_NEXT);
783
784 if (++scanned % XA_CHECK_SCHED)
785 continue;
786
787 xas_pause(&xas);
788 xas_unlock_irq(&xas);
789 cond_resched();
790 xas_lock_irq(&xas);
791 }
792 xas_unlock_irq(&xas);
793
794 return 0;
795}
796
797/*
798 * Delete DAX entry at @index from @mapping. Wait for it
799 * to be unlocked before deleting it.
800 */
801int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
802{
803 int ret = __dax_invalidate_entry(mapping, index, true);
804
805 /*
806 * This gets called from truncate / punch_hole path. As such, the caller
807 * must hold locks protecting against concurrent modifications of the
808 * page cache (usually fs-private i_mmap_sem for writing). Since the
809 * caller has seen a DAX entry for this index, we better find it
810 * at that index as well...
811 */
812 WARN_ON_ONCE(!ret);
813 return ret;
814}
815
816/*
817 * Invalidate DAX entry if it is clean.
818 */
819int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
820 pgoff_t index)
821{
822 return __dax_invalidate_entry(mapping, index, false);
823}
824
825static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos)
826{
827 return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset);
828}
829
830static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter)
831{
832 pgoff_t pgoff = dax_iomap_pgoff(&iter->iomap, iter->pos);
833 void *vto, *kaddr;
834 long rc;
835 int id;
836
837 id = dax_read_lock();
838 rc = dax_direct_access(iter->iomap.dax_dev, pgoff, 1, DAX_ACCESS,
839 &kaddr, NULL);
840 if (rc < 0) {
841 dax_read_unlock(id);
842 return rc;
843 }
844 vto = kmap_atomic(vmf->cow_page);
845 copy_user_page(vto, kaddr, vmf->address, vmf->cow_page);
846 kunmap_atomic(vto);
847 dax_read_unlock(id);
848 return 0;
849}
850
851/*
852 * MAP_SYNC on a dax mapping guarantees dirty metadata is
853 * flushed on write-faults (non-cow), but not read-faults.
854 */
855static bool dax_fault_is_synchronous(const struct iomap_iter *iter,
856 struct vm_area_struct *vma)
857{
858 return (iter->flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) &&
859 (iter->iomap.flags & IOMAP_F_DIRTY);
860}
861
862/*
863 * By this point grab_mapping_entry() has ensured that we have a locked entry
864 * of the appropriate size so we don't have to worry about downgrading PMDs to
865 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
866 * already in the tree, we will skip the insertion and just dirty the PMD as
867 * appropriate.
868 */
869static void *dax_insert_entry(struct xa_state *xas, struct vm_fault *vmf,
870 const struct iomap_iter *iter, void *entry, pfn_t pfn,
871 unsigned long flags)
872{
873 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
874 void *new_entry = dax_make_entry(pfn, flags);
875 bool write = iter->flags & IOMAP_WRITE;
876 bool dirty = write && !dax_fault_is_synchronous(iter, vmf->vma);
877 bool shared = iter->iomap.flags & IOMAP_F_SHARED;
878
879 if (dirty)
880 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
881
882 if (shared || (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE))) {
883 unsigned long index = xas->xa_index;
884 /* we are replacing a zero page with block mapping */
885 if (dax_is_pmd_entry(entry))
886 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
887 PG_PMD_NR, false);
888 else /* pte entry */
889 unmap_mapping_pages(mapping, index, 1, false);
890 }
891
892 xas_reset(xas);
893 xas_lock_irq(xas);
894 if (shared || dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
895 void *old;
896
897 dax_disassociate_entry(entry, mapping, false);
898 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address,
899 shared);
900 /*
901 * Only swap our new entry into the page cache if the current
902 * entry is a zero page or an empty entry. If a normal PTE or
903 * PMD entry is already in the cache, we leave it alone. This
904 * means that if we are trying to insert a PTE and the
905 * existing entry is a PMD, we will just leave the PMD in the
906 * tree and dirty it if necessary.
907 */
908 old = dax_lock_entry(xas, new_entry);
909 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
910 DAX_LOCKED));
911 entry = new_entry;
912 } else {
913 xas_load(xas); /* Walk the xa_state */
914 }
915
916 if (dirty)
917 xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
918
919 if (write && shared)
920 xas_set_mark(xas, PAGECACHE_TAG_TOWRITE);
921
922 xas_unlock_irq(xas);
923 return entry;
924}
925
926static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
927 struct address_space *mapping, void *entry)
928{
929 unsigned long pfn, index, count, end;
930 long ret = 0;
931 struct vm_area_struct *vma;
932
933 /*
934 * A page got tagged dirty in DAX mapping? Something is seriously
935 * wrong.
936 */
937 if (WARN_ON(!xa_is_value(entry)))
938 return -EIO;
939
940 if (unlikely(dax_is_locked(entry))) {
941 void *old_entry = entry;
942
943 entry = get_unlocked_entry(xas, 0);
944
945 /* Entry got punched out / reallocated? */
946 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
947 goto put_unlocked;
948 /*
949 * Entry got reallocated elsewhere? No need to writeback.
950 * We have to compare pfns as we must not bail out due to
951 * difference in lockbit or entry type.
952 */
953 if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
954 goto put_unlocked;
955 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
956 dax_is_zero_entry(entry))) {
957 ret = -EIO;
958 goto put_unlocked;
959 }
960
961 /* Another fsync thread may have already done this entry */
962 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
963 goto put_unlocked;
964 }
965
966 /* Lock the entry to serialize with page faults */
967 dax_lock_entry(xas, entry);
968
969 /*
970 * We can clear the tag now but we have to be careful so that concurrent
971 * dax_writeback_one() calls for the same index cannot finish before we
972 * actually flush the caches. This is achieved as the calls will look
973 * at the entry only under the i_pages lock and once they do that
974 * they will see the entry locked and wait for it to unlock.
975 */
976 xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
977 xas_unlock_irq(xas);
978
979 /*
980 * If dax_writeback_mapping_range() was given a wbc->range_start
981 * in the middle of a PMD, the 'index' we use needs to be
982 * aligned to the start of the PMD.
983 * This allows us to flush for PMD_SIZE and not have to worry about
984 * partial PMD writebacks.
985 */
986 pfn = dax_to_pfn(entry);
987 count = 1UL << dax_entry_order(entry);
988 index = xas->xa_index & ~(count - 1);
989 end = index + count - 1;
990
991 /* Walk all mappings of a given index of a file and writeprotect them */
992 i_mmap_lock_read(mapping);
993 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, end) {
994 pfn_mkclean_range(pfn, count, index, vma);
995 cond_resched();
996 }
997 i_mmap_unlock_read(mapping);
998
999 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
1000 /*
1001 * After we have flushed the cache, we can clear the dirty tag. There
1002 * cannot be new dirty data in the pfn after the flush has completed as
1003 * the pfn mappings are writeprotected and fault waits for mapping
1004 * entry lock.
1005 */
1006 xas_reset(xas);
1007 xas_lock_irq(xas);
1008 xas_store(xas, entry);
1009 xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
1010 dax_wake_entry(xas, entry, WAKE_NEXT);
1011
1012 trace_dax_writeback_one(mapping->host, index, count);
1013 return ret;
1014
1015 put_unlocked:
1016 put_unlocked_entry(xas, entry, WAKE_NEXT);
1017 return ret;
1018}
1019
1020/*
1021 * Flush the mapping to the persistent domain within the byte range of [start,
1022 * end]. This is required by data integrity operations to ensure file data is
1023 * on persistent storage prior to completion of the operation.
1024 */
1025int dax_writeback_mapping_range(struct address_space *mapping,
1026 struct dax_device *dax_dev, struct writeback_control *wbc)
1027{
1028 XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
1029 struct inode *inode = mapping->host;
1030 pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
1031 void *entry;
1032 int ret = 0;
1033 unsigned int scanned = 0;
1034
1035 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
1036 return -EIO;
1037
1038 if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL)
1039 return 0;
1040
1041 trace_dax_writeback_range(inode, xas.xa_index, end_index);
1042
1043 tag_pages_for_writeback(mapping, xas.xa_index, end_index);
1044
1045 xas_lock_irq(&xas);
1046 xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
1047 ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
1048 if (ret < 0) {
1049 mapping_set_error(mapping, ret);
1050 break;
1051 }
1052 if (++scanned % XA_CHECK_SCHED)
1053 continue;
1054
1055 xas_pause(&xas);
1056 xas_unlock_irq(&xas);
1057 cond_resched();
1058 xas_lock_irq(&xas);
1059 }
1060 xas_unlock_irq(&xas);
1061 trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
1062 return ret;
1063}
1064EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1065
1066static int dax_iomap_direct_access(const struct iomap *iomap, loff_t pos,
1067 size_t size, void **kaddr, pfn_t *pfnp)
1068{
1069 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1070 int id, rc = 0;
1071 long length;
1072
1073 id = dax_read_lock();
1074 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1075 DAX_ACCESS, kaddr, pfnp);
1076 if (length < 0) {
1077 rc = length;
1078 goto out;
1079 }
1080 if (!pfnp)
1081 goto out_check_addr;
1082 rc = -EINVAL;
1083 if (PFN_PHYS(length) < size)
1084 goto out;
1085 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1086 goto out;
1087 /* For larger pages we need devmap */
1088 if (length > 1 && !pfn_t_devmap(*pfnp))
1089 goto out;
1090 rc = 0;
1091
1092out_check_addr:
1093 if (!kaddr)
1094 goto out;
1095 if (!*kaddr)
1096 rc = -EFAULT;
1097out:
1098 dax_read_unlock(id);
1099 return rc;
1100}
1101
1102/**
1103 * dax_iomap_copy_around - Prepare for an unaligned write to a shared/cow page
1104 * by copying the data before and after the range to be written.
1105 * @pos: address to do copy from.
1106 * @length: size of copy operation.
1107 * @align_size: aligned w.r.t align_size (either PMD_SIZE or PAGE_SIZE)
1108 * @srcmap: iomap srcmap
1109 * @daddr: destination address to copy to.
1110 *
1111 * This can be called from two places. Either during DAX write fault (page
1112 * aligned), to copy the length size data to daddr. Or, while doing normal DAX
1113 * write operation, dax_iomap_iter() might call this to do the copy of either
1114 * start or end unaligned address. In the latter case the rest of the copy of
1115 * aligned ranges is taken care by dax_iomap_iter() itself.
1116 * If the srcmap contains invalid data, such as HOLE and UNWRITTEN, zero the
1117 * area to make sure no old data remains.
1118 */
1119static int dax_iomap_copy_around(loff_t pos, uint64_t length, size_t align_size,
1120 const struct iomap *srcmap, void *daddr)
1121{
1122 loff_t head_off = pos & (align_size - 1);
1123 size_t size = ALIGN(head_off + length, align_size);
1124 loff_t end = pos + length;
1125 loff_t pg_end = round_up(end, align_size);
1126 /* copy_all is usually in page fault case */
1127 bool copy_all = head_off == 0 && end == pg_end;
1128 /* zero the edges if srcmap is a HOLE or IOMAP_UNWRITTEN */
1129 bool zero_edge = srcmap->flags & IOMAP_F_SHARED ||
1130 srcmap->type == IOMAP_UNWRITTEN;
1131 void *saddr = NULL;
1132 int ret = 0;
1133
1134 if (!zero_edge) {
1135 ret = dax_iomap_direct_access(srcmap, pos, size, &saddr, NULL);
1136 if (ret)
1137 return dax_mem2blk_err(ret);
1138 }
1139
1140 if (copy_all) {
1141 if (zero_edge)
1142 memset(daddr, 0, size);
1143 else
1144 ret = copy_mc_to_kernel(daddr, saddr, length);
1145 goto out;
1146 }
1147
1148 /* Copy the head part of the range */
1149 if (head_off) {
1150 if (zero_edge)
1151 memset(daddr, 0, head_off);
1152 else {
1153 ret = copy_mc_to_kernel(daddr, saddr, head_off);
1154 if (ret)
1155 return -EIO;
1156 }
1157 }
1158
1159 /* Copy the tail part of the range */
1160 if (end < pg_end) {
1161 loff_t tail_off = head_off + length;
1162 loff_t tail_len = pg_end - end;
1163
1164 if (zero_edge)
1165 memset(daddr + tail_off, 0, tail_len);
1166 else {
1167 ret = copy_mc_to_kernel(daddr + tail_off,
1168 saddr + tail_off, tail_len);
1169 if (ret)
1170 return -EIO;
1171 }
1172 }
1173out:
1174 if (zero_edge)
1175 dax_flush(srcmap->dax_dev, daddr, size);
1176 return ret ? -EIO : 0;
1177}
1178
1179/*
1180 * The user has performed a load from a hole in the file. Allocating a new
1181 * page in the file would cause excessive storage usage for workloads with
1182 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1183 * If this page is ever written to we will re-fault and change the mapping to
1184 * point to real DAX storage instead.
1185 */
1186static vm_fault_t dax_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1187 const struct iomap_iter *iter, void **entry)
1188{
1189 struct inode *inode = iter->inode;
1190 unsigned long vaddr = vmf->address;
1191 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1192 vm_fault_t ret;
1193
1194 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, DAX_ZERO_PAGE);
1195
1196 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1197 trace_dax_load_hole(inode, vmf, ret);
1198 return ret;
1199}
1200
1201#ifdef CONFIG_FS_DAX_PMD
1202static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1203 const struct iomap_iter *iter, void **entry)
1204{
1205 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1206 unsigned long pmd_addr = vmf->address & PMD_MASK;
1207 struct vm_area_struct *vma = vmf->vma;
1208 struct inode *inode = mapping->host;
1209 pgtable_t pgtable = NULL;
1210 struct folio *zero_folio;
1211 spinlock_t *ptl;
1212 pmd_t pmd_entry;
1213 pfn_t pfn;
1214
1215 zero_folio = mm_get_huge_zero_folio(vmf->vma->vm_mm);
1216
1217 if (unlikely(!zero_folio))
1218 goto fallback;
1219
1220 pfn = page_to_pfn_t(&zero_folio->page);
1221 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn,
1222 DAX_PMD | DAX_ZERO_PAGE);
1223
1224 if (arch_needs_pgtable_deposit()) {
1225 pgtable = pte_alloc_one(vma->vm_mm);
1226 if (!pgtable)
1227 return VM_FAULT_OOM;
1228 }
1229
1230 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1231 if (!pmd_none(*(vmf->pmd))) {
1232 spin_unlock(ptl);
1233 goto fallback;
1234 }
1235
1236 if (pgtable) {
1237 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1238 mm_inc_nr_ptes(vma->vm_mm);
1239 }
1240 pmd_entry = mk_pmd(&zero_folio->page, vmf->vma->vm_page_prot);
1241 pmd_entry = pmd_mkhuge(pmd_entry);
1242 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1243 spin_unlock(ptl);
1244 trace_dax_pmd_load_hole(inode, vmf, zero_folio, *entry);
1245 return VM_FAULT_NOPAGE;
1246
1247fallback:
1248 if (pgtable)
1249 pte_free(vma->vm_mm, pgtable);
1250 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_folio, *entry);
1251 return VM_FAULT_FALLBACK;
1252}
1253#else
1254static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1255 const struct iomap_iter *iter, void **entry)
1256{
1257 return VM_FAULT_FALLBACK;
1258}
1259#endif /* CONFIG_FS_DAX_PMD */
1260
1261static s64 dax_unshare_iter(struct iomap_iter *iter)
1262{
1263 struct iomap *iomap = &iter->iomap;
1264 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1265 loff_t copy_pos = iter->pos;
1266 u64 copy_len = iomap_length(iter);
1267 u32 mod;
1268 int id = 0;
1269 s64 ret = 0;
1270 void *daddr = NULL, *saddr = NULL;
1271
1272 if (!iomap_want_unshare_iter(iter))
1273 return iomap_length(iter);
1274
1275 /*
1276 * Extend the file range to be aligned to fsblock/pagesize, because
1277 * we need to copy entire blocks, not just the byte range specified.
1278 * Invalidate the mapping because we're about to CoW.
1279 */
1280 mod = offset_in_page(copy_pos);
1281 if (mod) {
1282 copy_len += mod;
1283 copy_pos -= mod;
1284 }
1285
1286 mod = offset_in_page(copy_pos + copy_len);
1287 if (mod)
1288 copy_len += PAGE_SIZE - mod;
1289
1290 invalidate_inode_pages2_range(iter->inode->i_mapping,
1291 copy_pos >> PAGE_SHIFT,
1292 (copy_pos + copy_len - 1) >> PAGE_SHIFT);
1293
1294 id = dax_read_lock();
1295 ret = dax_iomap_direct_access(iomap, copy_pos, copy_len, &daddr, NULL);
1296 if (ret < 0)
1297 goto out_unlock;
1298
1299 ret = dax_iomap_direct_access(srcmap, copy_pos, copy_len, &saddr, NULL);
1300 if (ret < 0)
1301 goto out_unlock;
1302
1303 if (copy_mc_to_kernel(daddr, saddr, copy_len) == 0)
1304 ret = iomap_length(iter);
1305 else
1306 ret = -EIO;
1307
1308out_unlock:
1309 dax_read_unlock(id);
1310 return dax_mem2blk_err(ret);
1311}
1312
1313int dax_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1314 const struct iomap_ops *ops)
1315{
1316 struct iomap_iter iter = {
1317 .inode = inode,
1318 .pos = pos,
1319 .flags = IOMAP_WRITE | IOMAP_UNSHARE | IOMAP_DAX,
1320 };
1321 loff_t size = i_size_read(inode);
1322 int ret;
1323
1324 if (pos < 0 || pos >= size)
1325 return 0;
1326
1327 iter.len = min(len, size - pos);
1328 while ((ret = iomap_iter(&iter, ops)) > 0)
1329 iter.processed = dax_unshare_iter(&iter);
1330 return ret;
1331}
1332EXPORT_SYMBOL_GPL(dax_file_unshare);
1333
1334static int dax_memzero(struct iomap_iter *iter, loff_t pos, size_t size)
1335{
1336 const struct iomap *iomap = &iter->iomap;
1337 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1338 unsigned offset = offset_in_page(pos);
1339 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1340 void *kaddr;
1341 long ret;
1342
1343 ret = dax_direct_access(iomap->dax_dev, pgoff, 1, DAX_ACCESS, &kaddr,
1344 NULL);
1345 if (ret < 0)
1346 return dax_mem2blk_err(ret);
1347
1348 memset(kaddr + offset, 0, size);
1349 if (iomap->flags & IOMAP_F_SHARED)
1350 ret = dax_iomap_copy_around(pos, size, PAGE_SIZE, srcmap,
1351 kaddr);
1352 else
1353 dax_flush(iomap->dax_dev, kaddr + offset, size);
1354 return ret;
1355}
1356
1357static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero)
1358{
1359 const struct iomap *iomap = &iter->iomap;
1360 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1361 loff_t pos = iter->pos;
1362 u64 length = iomap_length(iter);
1363 s64 written = 0;
1364
1365 /* already zeroed? we're done. */
1366 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1367 return length;
1368
1369 /*
1370 * invalidate the pages whose sharing state is to be changed
1371 * because of CoW.
1372 */
1373 if (iomap->flags & IOMAP_F_SHARED)
1374 invalidate_inode_pages2_range(iter->inode->i_mapping,
1375 pos >> PAGE_SHIFT,
1376 (pos + length - 1) >> PAGE_SHIFT);
1377
1378 do {
1379 unsigned offset = offset_in_page(pos);
1380 unsigned size = min_t(u64, PAGE_SIZE - offset, length);
1381 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1382 long rc;
1383 int id;
1384
1385 id = dax_read_lock();
1386 if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE)
1387 rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1388 else
1389 rc = dax_memzero(iter, pos, size);
1390 dax_read_unlock(id);
1391
1392 if (rc < 0)
1393 return rc;
1394 pos += size;
1395 length -= size;
1396 written += size;
1397 } while (length > 0);
1398
1399 if (did_zero)
1400 *did_zero = true;
1401 return written;
1402}
1403
1404int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1405 const struct iomap_ops *ops)
1406{
1407 struct iomap_iter iter = {
1408 .inode = inode,
1409 .pos = pos,
1410 .len = len,
1411 .flags = IOMAP_DAX | IOMAP_ZERO,
1412 };
1413 int ret;
1414
1415 while ((ret = iomap_iter(&iter, ops)) > 0)
1416 iter.processed = dax_zero_iter(&iter, did_zero);
1417 return ret;
1418}
1419EXPORT_SYMBOL_GPL(dax_zero_range);
1420
1421int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1422 const struct iomap_ops *ops)
1423{
1424 unsigned int blocksize = i_blocksize(inode);
1425 unsigned int off = pos & (blocksize - 1);
1426
1427 /* Block boundary? Nothing to do */
1428 if (!off)
1429 return 0;
1430 return dax_zero_range(inode, pos, blocksize - off, did_zero, ops);
1431}
1432EXPORT_SYMBOL_GPL(dax_truncate_page);
1433
1434static loff_t dax_iomap_iter(const struct iomap_iter *iomi,
1435 struct iov_iter *iter)
1436{
1437 const struct iomap *iomap = &iomi->iomap;
1438 const struct iomap *srcmap = iomap_iter_srcmap(iomi);
1439 loff_t length = iomap_length(iomi);
1440 loff_t pos = iomi->pos;
1441 struct dax_device *dax_dev = iomap->dax_dev;
1442 loff_t end = pos + length, done = 0;
1443 bool write = iov_iter_rw(iter) == WRITE;
1444 bool cow = write && iomap->flags & IOMAP_F_SHARED;
1445 ssize_t ret = 0;
1446 size_t xfer;
1447 int id;
1448
1449 if (!write) {
1450 end = min(end, i_size_read(iomi->inode));
1451 if (pos >= end)
1452 return 0;
1453
1454 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1455 return iov_iter_zero(min(length, end - pos), iter);
1456 }
1457
1458 /*
1459 * In DAX mode, enforce either pure overwrites of written extents, or
1460 * writes to unwritten extents as part of a copy-on-write operation.
1461 */
1462 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED &&
1463 !(iomap->flags & IOMAP_F_SHARED)))
1464 return -EIO;
1465
1466 /*
1467 * Write can allocate block for an area which has a hole page mapped
1468 * into page tables. We have to tear down these mappings so that data
1469 * written by write(2) is visible in mmap.
1470 */
1471 if (iomap->flags & IOMAP_F_NEW || cow) {
1472 /*
1473 * Filesystem allows CoW on non-shared extents. The src extents
1474 * may have been mmapped with dirty mark before. To be able to
1475 * invalidate its dax entries, we need to clear the dirty mark
1476 * in advance.
1477 */
1478 if (cow)
1479 __dax_clear_dirty_range(iomi->inode->i_mapping,
1480 pos >> PAGE_SHIFT,
1481 (end - 1) >> PAGE_SHIFT);
1482 invalidate_inode_pages2_range(iomi->inode->i_mapping,
1483 pos >> PAGE_SHIFT,
1484 (end - 1) >> PAGE_SHIFT);
1485 }
1486
1487 id = dax_read_lock();
1488 while (pos < end) {
1489 unsigned offset = pos & (PAGE_SIZE - 1);
1490 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1491 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1492 ssize_t map_len;
1493 bool recovery = false;
1494 void *kaddr;
1495
1496 if (fatal_signal_pending(current)) {
1497 ret = -EINTR;
1498 break;
1499 }
1500
1501 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1502 DAX_ACCESS, &kaddr, NULL);
1503 if (map_len == -EHWPOISON && iov_iter_rw(iter) == WRITE) {
1504 map_len = dax_direct_access(dax_dev, pgoff,
1505 PHYS_PFN(size), DAX_RECOVERY_WRITE,
1506 &kaddr, NULL);
1507 if (map_len > 0)
1508 recovery = true;
1509 }
1510 if (map_len < 0) {
1511 ret = dax_mem2blk_err(map_len);
1512 break;
1513 }
1514
1515 if (cow) {
1516 ret = dax_iomap_copy_around(pos, length, PAGE_SIZE,
1517 srcmap, kaddr);
1518 if (ret)
1519 break;
1520 }
1521
1522 map_len = PFN_PHYS(map_len);
1523 kaddr += offset;
1524 map_len -= offset;
1525 if (map_len > end - pos)
1526 map_len = end - pos;
1527
1528 if (recovery)
1529 xfer = dax_recovery_write(dax_dev, pgoff, kaddr,
1530 map_len, iter);
1531 else if (write)
1532 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1533 map_len, iter);
1534 else
1535 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1536 map_len, iter);
1537
1538 pos += xfer;
1539 length -= xfer;
1540 done += xfer;
1541
1542 if (xfer == 0)
1543 ret = -EFAULT;
1544 if (xfer < map_len)
1545 break;
1546 }
1547 dax_read_unlock(id);
1548
1549 return done ? done : ret;
1550}
1551
1552/**
1553 * dax_iomap_rw - Perform I/O to a DAX file
1554 * @iocb: The control block for this I/O
1555 * @iter: The addresses to do I/O from or to
1556 * @ops: iomap ops passed from the file system
1557 *
1558 * This function performs read and write operations to directly mapped
1559 * persistent memory. The callers needs to take care of read/write exclusion
1560 * and evicting any page cache pages in the region under I/O.
1561 */
1562ssize_t
1563dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1564 const struct iomap_ops *ops)
1565{
1566 struct iomap_iter iomi = {
1567 .inode = iocb->ki_filp->f_mapping->host,
1568 .pos = iocb->ki_pos,
1569 .len = iov_iter_count(iter),
1570 .flags = IOMAP_DAX,
1571 };
1572 loff_t done = 0;
1573 int ret;
1574
1575 if (!iomi.len)
1576 return 0;
1577
1578 if (iov_iter_rw(iter) == WRITE) {
1579 lockdep_assert_held_write(&iomi.inode->i_rwsem);
1580 iomi.flags |= IOMAP_WRITE;
1581 } else {
1582 lockdep_assert_held(&iomi.inode->i_rwsem);
1583 }
1584
1585 if (iocb->ki_flags & IOCB_NOWAIT)
1586 iomi.flags |= IOMAP_NOWAIT;
1587
1588 while ((ret = iomap_iter(&iomi, ops)) > 0)
1589 iomi.processed = dax_iomap_iter(&iomi, iter);
1590
1591 done = iomi.pos - iocb->ki_pos;
1592 iocb->ki_pos = iomi.pos;
1593 return done ? done : ret;
1594}
1595EXPORT_SYMBOL_GPL(dax_iomap_rw);
1596
1597static vm_fault_t dax_fault_return(int error)
1598{
1599 if (error == 0)
1600 return VM_FAULT_NOPAGE;
1601 return vmf_error(error);
1602}
1603
1604/*
1605 * When handling a synchronous page fault and the inode need a fsync, we can
1606 * insert the PTE/PMD into page tables only after that fsync happened. Skip
1607 * insertion for now and return the pfn so that caller can insert it after the
1608 * fsync is done.
1609 */
1610static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn)
1611{
1612 if (WARN_ON_ONCE(!pfnp))
1613 return VM_FAULT_SIGBUS;
1614 *pfnp = pfn;
1615 return VM_FAULT_NEEDDSYNC;
1616}
1617
1618static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf,
1619 const struct iomap_iter *iter)
1620{
1621 vm_fault_t ret;
1622 int error = 0;
1623
1624 switch (iter->iomap.type) {
1625 case IOMAP_HOLE:
1626 case IOMAP_UNWRITTEN:
1627 clear_user_highpage(vmf->cow_page, vmf->address);
1628 break;
1629 case IOMAP_MAPPED:
1630 error = copy_cow_page_dax(vmf, iter);
1631 break;
1632 default:
1633 WARN_ON_ONCE(1);
1634 error = -EIO;
1635 break;
1636 }
1637
1638 if (error)
1639 return dax_fault_return(error);
1640
1641 __SetPageUptodate(vmf->cow_page);
1642 ret = finish_fault(vmf);
1643 if (!ret)
1644 return VM_FAULT_DONE_COW;
1645 return ret;
1646}
1647
1648/**
1649 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault.
1650 * @vmf: vm fault instance
1651 * @iter: iomap iter
1652 * @pfnp: pfn to be returned
1653 * @xas: the dax mapping tree of a file
1654 * @entry: an unlocked dax entry to be inserted
1655 * @pmd: distinguish whether it is a pmd fault
1656 */
1657static vm_fault_t dax_fault_iter(struct vm_fault *vmf,
1658 const struct iomap_iter *iter, pfn_t *pfnp,
1659 struct xa_state *xas, void **entry, bool pmd)
1660{
1661 const struct iomap *iomap = &iter->iomap;
1662 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1663 size_t size = pmd ? PMD_SIZE : PAGE_SIZE;
1664 loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT;
1665 bool write = iter->flags & IOMAP_WRITE;
1666 unsigned long entry_flags = pmd ? DAX_PMD : 0;
1667 int err = 0;
1668 pfn_t pfn;
1669 void *kaddr;
1670
1671 if (!pmd && vmf->cow_page)
1672 return dax_fault_cow_page(vmf, iter);
1673
1674 /* if we are reading UNWRITTEN and HOLE, return a hole. */
1675 if (!write &&
1676 (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) {
1677 if (!pmd)
1678 return dax_load_hole(xas, vmf, iter, entry);
1679 return dax_pmd_load_hole(xas, vmf, iter, entry);
1680 }
1681
1682 if (iomap->type != IOMAP_MAPPED && !(iomap->flags & IOMAP_F_SHARED)) {
1683 WARN_ON_ONCE(1);
1684 return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS;
1685 }
1686
1687 err = dax_iomap_direct_access(iomap, pos, size, &kaddr, &pfn);
1688 if (err)
1689 return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err);
1690
1691 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, entry_flags);
1692
1693 if (write && iomap->flags & IOMAP_F_SHARED) {
1694 err = dax_iomap_copy_around(pos, size, size, srcmap, kaddr);
1695 if (err)
1696 return dax_fault_return(err);
1697 }
1698
1699 if (dax_fault_is_synchronous(iter, vmf->vma))
1700 return dax_fault_synchronous_pfnp(pfnp, pfn);
1701
1702 /* insert PMD pfn */
1703 if (pmd)
1704 return vmf_insert_pfn_pmd(vmf, pfn, write);
1705
1706 /* insert PTE pfn */
1707 if (write)
1708 return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1709 return vmf_insert_mixed(vmf->vma, vmf->address, pfn);
1710}
1711
1712static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1713 int *iomap_errp, const struct iomap_ops *ops)
1714{
1715 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1716 XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1717 struct iomap_iter iter = {
1718 .inode = mapping->host,
1719 .pos = (loff_t)vmf->pgoff << PAGE_SHIFT,
1720 .len = PAGE_SIZE,
1721 .flags = IOMAP_DAX | IOMAP_FAULT,
1722 };
1723 vm_fault_t ret = 0;
1724 void *entry;
1725 int error;
1726
1727 trace_dax_pte_fault(iter.inode, vmf, ret);
1728 /*
1729 * Check whether offset isn't beyond end of file now. Caller is supposed
1730 * to hold locks serializing us with truncate / punch hole so this is
1731 * a reliable test.
1732 */
1733 if (iter.pos >= i_size_read(iter.inode)) {
1734 ret = VM_FAULT_SIGBUS;
1735 goto out;
1736 }
1737
1738 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1739 iter.flags |= IOMAP_WRITE;
1740
1741 entry = grab_mapping_entry(&xas, mapping, 0);
1742 if (xa_is_internal(entry)) {
1743 ret = xa_to_internal(entry);
1744 goto out;
1745 }
1746
1747 /*
1748 * It is possible, particularly with mixed reads & writes to private
1749 * mappings, that we have raced with a PMD fault that overlaps with
1750 * the PTE we need to set up. If so just return and the fault will be
1751 * retried.
1752 */
1753 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1754 ret = VM_FAULT_NOPAGE;
1755 goto unlock_entry;
1756 }
1757
1758 while ((error = iomap_iter(&iter, ops)) > 0) {
1759 if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) {
1760 iter.processed = -EIO; /* fs corruption? */
1761 continue;
1762 }
1763
1764 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false);
1765 if (ret != VM_FAULT_SIGBUS &&
1766 (iter.iomap.flags & IOMAP_F_NEW)) {
1767 count_vm_event(PGMAJFAULT);
1768 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
1769 ret |= VM_FAULT_MAJOR;
1770 }
1771
1772 if (!(ret & VM_FAULT_ERROR))
1773 iter.processed = PAGE_SIZE;
1774 }
1775
1776 if (iomap_errp)
1777 *iomap_errp = error;
1778 if (!ret && error)
1779 ret = dax_fault_return(error);
1780
1781unlock_entry:
1782 dax_unlock_entry(&xas, entry);
1783out:
1784 trace_dax_pte_fault_done(iter.inode, vmf, ret);
1785 return ret;
1786}
1787
1788#ifdef CONFIG_FS_DAX_PMD
1789static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas,
1790 pgoff_t max_pgoff)
1791{
1792 unsigned long pmd_addr = vmf->address & PMD_MASK;
1793 bool write = vmf->flags & FAULT_FLAG_WRITE;
1794
1795 /*
1796 * Make sure that the faulting address's PMD offset (color) matches
1797 * the PMD offset from the start of the file. This is necessary so
1798 * that a PMD range in the page table overlaps exactly with a PMD
1799 * range in the page cache.
1800 */
1801 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1802 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1803 return true;
1804
1805 /* Fall back to PTEs if we're going to COW */
1806 if (write && !(vmf->vma->vm_flags & VM_SHARED))
1807 return true;
1808
1809 /* If the PMD would extend outside the VMA */
1810 if (pmd_addr < vmf->vma->vm_start)
1811 return true;
1812 if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
1813 return true;
1814
1815 /* If the PMD would extend beyond the file size */
1816 if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff)
1817 return true;
1818
1819 return false;
1820}
1821
1822static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1823 const struct iomap_ops *ops)
1824{
1825 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1826 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1827 struct iomap_iter iter = {
1828 .inode = mapping->host,
1829 .len = PMD_SIZE,
1830 .flags = IOMAP_DAX | IOMAP_FAULT,
1831 };
1832 vm_fault_t ret = VM_FAULT_FALLBACK;
1833 pgoff_t max_pgoff;
1834 void *entry;
1835
1836 if (vmf->flags & FAULT_FLAG_WRITE)
1837 iter.flags |= IOMAP_WRITE;
1838
1839 /*
1840 * Check whether offset isn't beyond end of file now. Caller is
1841 * supposed to hold locks serializing us with truncate / punch hole so
1842 * this is a reliable test.
1843 */
1844 max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE);
1845
1846 trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0);
1847
1848 if (xas.xa_index >= max_pgoff) {
1849 ret = VM_FAULT_SIGBUS;
1850 goto out;
1851 }
1852
1853 if (dax_fault_check_fallback(vmf, &xas, max_pgoff))
1854 goto fallback;
1855
1856 /*
1857 * grab_mapping_entry() will make sure we get an empty PMD entry,
1858 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1859 * entry is already in the array, for instance), it will return
1860 * VM_FAULT_FALLBACK.
1861 */
1862 entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1863 if (xa_is_internal(entry)) {
1864 ret = xa_to_internal(entry);
1865 goto fallback;
1866 }
1867
1868 /*
1869 * It is possible, particularly with mixed reads & writes to private
1870 * mappings, that we have raced with a PTE fault that overlaps with
1871 * the PMD we need to set up. If so just return and the fault will be
1872 * retried.
1873 */
1874 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1875 !pmd_devmap(*vmf->pmd)) {
1876 ret = 0;
1877 goto unlock_entry;
1878 }
1879
1880 iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1881 while (iomap_iter(&iter, ops) > 0) {
1882 if (iomap_length(&iter) < PMD_SIZE)
1883 continue; /* actually breaks out of the loop */
1884
1885 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true);
1886 if (ret != VM_FAULT_FALLBACK)
1887 iter.processed = PMD_SIZE;
1888 }
1889
1890unlock_entry:
1891 dax_unlock_entry(&xas, entry);
1892fallback:
1893 if (ret == VM_FAULT_FALLBACK) {
1894 split_huge_pmd(vmf->vma, vmf->pmd, vmf->address);
1895 count_vm_event(THP_FAULT_FALLBACK);
1896 }
1897out:
1898 trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret);
1899 return ret;
1900}
1901#else
1902static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1903 const struct iomap_ops *ops)
1904{
1905 return VM_FAULT_FALLBACK;
1906}
1907#endif /* CONFIG_FS_DAX_PMD */
1908
1909/**
1910 * dax_iomap_fault - handle a page fault on a DAX file
1911 * @vmf: The description of the fault
1912 * @order: Order of the page to fault in
1913 * @pfnp: PFN to insert for synchronous faults if fsync is required
1914 * @iomap_errp: Storage for detailed error code in case of error
1915 * @ops: Iomap ops passed from the file system
1916 *
1917 * When a page fault occurs, filesystems may call this helper in
1918 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1919 * has done all the necessary locking for page fault to proceed
1920 * successfully.
1921 */
1922vm_fault_t dax_iomap_fault(struct vm_fault *vmf, unsigned int order,
1923 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1924{
1925 if (order == 0)
1926 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1927 else if (order == PMD_ORDER)
1928 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1929 else
1930 return VM_FAULT_FALLBACK;
1931}
1932EXPORT_SYMBOL_GPL(dax_iomap_fault);
1933
1934/*
1935 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1936 * @vmf: The description of the fault
1937 * @pfn: PFN to insert
1938 * @order: Order of entry to insert.
1939 *
1940 * This function inserts a writeable PTE or PMD entry into the page tables
1941 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1942 */
1943static vm_fault_t
1944dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1945{
1946 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1947 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1948 void *entry;
1949 vm_fault_t ret;
1950
1951 xas_lock_irq(&xas);
1952 entry = get_unlocked_entry(&xas, order);
1953 /* Did we race with someone splitting entry or so? */
1954 if (!entry || dax_is_conflict(entry) ||
1955 (order == 0 && !dax_is_pte_entry(entry))) {
1956 put_unlocked_entry(&xas, entry, WAKE_NEXT);
1957 xas_unlock_irq(&xas);
1958 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1959 VM_FAULT_NOPAGE);
1960 return VM_FAULT_NOPAGE;
1961 }
1962 xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1963 dax_lock_entry(&xas, entry);
1964 xas_unlock_irq(&xas);
1965 if (order == 0)
1966 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1967#ifdef CONFIG_FS_DAX_PMD
1968 else if (order == PMD_ORDER)
1969 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1970#endif
1971 else
1972 ret = VM_FAULT_FALLBACK;
1973 dax_unlock_entry(&xas, entry);
1974 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1975 return ret;
1976}
1977
1978/**
1979 * dax_finish_sync_fault - finish synchronous page fault
1980 * @vmf: The description of the fault
1981 * @order: Order of entry to be inserted
1982 * @pfn: PFN to insert
1983 *
1984 * This function ensures that the file range touched by the page fault is
1985 * stored persistently on the media and handles inserting of appropriate page
1986 * table entry.
1987 */
1988vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, unsigned int order,
1989 pfn_t pfn)
1990{
1991 int err;
1992 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1993 size_t len = PAGE_SIZE << order;
1994
1995 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1996 if (err)
1997 return VM_FAULT_SIGBUS;
1998 return dax_insert_pfn_mkwrite(vmf, pfn, order);
1999}
2000EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
2001
2002static loff_t dax_range_compare_iter(struct iomap_iter *it_src,
2003 struct iomap_iter *it_dest, u64 len, bool *same)
2004{
2005 const struct iomap *smap = &it_src->iomap;
2006 const struct iomap *dmap = &it_dest->iomap;
2007 loff_t pos1 = it_src->pos, pos2 = it_dest->pos;
2008 void *saddr, *daddr;
2009 int id, ret;
2010
2011 len = min(len, min(smap->length, dmap->length));
2012
2013 if (smap->type == IOMAP_HOLE && dmap->type == IOMAP_HOLE) {
2014 *same = true;
2015 return len;
2016 }
2017
2018 if (smap->type == IOMAP_HOLE || dmap->type == IOMAP_HOLE) {
2019 *same = false;
2020 return 0;
2021 }
2022
2023 id = dax_read_lock();
2024 ret = dax_iomap_direct_access(smap, pos1, ALIGN(pos1 + len, PAGE_SIZE),
2025 &saddr, NULL);
2026 if (ret < 0)
2027 goto out_unlock;
2028
2029 ret = dax_iomap_direct_access(dmap, pos2, ALIGN(pos2 + len, PAGE_SIZE),
2030 &daddr, NULL);
2031 if (ret < 0)
2032 goto out_unlock;
2033
2034 *same = !memcmp(saddr, daddr, len);
2035 if (!*same)
2036 len = 0;
2037 dax_read_unlock(id);
2038 return len;
2039
2040out_unlock:
2041 dax_read_unlock(id);
2042 return -EIO;
2043}
2044
2045int dax_dedupe_file_range_compare(struct inode *src, loff_t srcoff,
2046 struct inode *dst, loff_t dstoff, loff_t len, bool *same,
2047 const struct iomap_ops *ops)
2048{
2049 struct iomap_iter src_iter = {
2050 .inode = src,
2051 .pos = srcoff,
2052 .len = len,
2053 .flags = IOMAP_DAX,
2054 };
2055 struct iomap_iter dst_iter = {
2056 .inode = dst,
2057 .pos = dstoff,
2058 .len = len,
2059 .flags = IOMAP_DAX,
2060 };
2061 int ret, compared = 0;
2062
2063 while ((ret = iomap_iter(&src_iter, ops)) > 0 &&
2064 (ret = iomap_iter(&dst_iter, ops)) > 0) {
2065 compared = dax_range_compare_iter(&src_iter, &dst_iter,
2066 min(src_iter.len, dst_iter.len), same);
2067 if (compared < 0)
2068 return ret;
2069 src_iter.processed = dst_iter.processed = compared;
2070 }
2071 return ret;
2072}
2073
2074int dax_remap_file_range_prep(struct file *file_in, loff_t pos_in,
2075 struct file *file_out, loff_t pos_out,
2076 loff_t *len, unsigned int remap_flags,
2077 const struct iomap_ops *ops)
2078{
2079 return __generic_remap_file_range_prep(file_in, pos_in, file_out,
2080 pos_out, len, remap_flags, ops);
2081}
2082EXPORT_SYMBOL_GPL(dax_remap_file_range_prep);