1 // SPDX-License-Identifier: GPL-2.0-only 1
2 /*
3 * fs/dax.c - Direct Access filesystem code
4 * Copyright (c) 2013-2014 Intel Corporation
5 * Author: Matthew Wilcox <matthew.r.wilcox@in
6 * Author: Ross Zwisler <ross.zwisler@linux.in
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 p
34 #define DAX_WAIT_TABLE_BITS 12
35 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_
36
37 /* The 'colour' (ie low bits) within a PMD of
38 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHI
39 #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIF
40
41 static wait_queue_head_t wait_table[DAX_WAIT_T
42
43 static int __init init_dax_wait_table(void)
44 {
45 int i;
46
47 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES
48 init_waitqueue_head(wait_table
49 return 0;
50 }
51 fs_initcall(init_dax_wait_table);
52
53 /*
54 * DAX pagecache entries use XArray value entr
55 * for pages. We use one bit for locking, one
56 * and two more to tell us if the entry is a z
57 * is just used for locking. In total four sp
58 *
59 * If the PMD bit isn't set the entry has size
60 * and EMPTY bits aren't set the entry is a no
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
69 static unsigned long dax_to_pfn(void *entry)
70 {
71 return xa_to_value(entry) >> DAX_SHIFT
72 }
73
74 static void *dax_make_entry(pfn_t pfn, unsigne
75 {
76 return xa_mk_value(flags | (pfn_t_to_p
77 }
78
79 static bool dax_is_locked(void *entry)
80 {
81 return xa_to_value(entry) & DAX_LOCKED
82 }
83
84 static unsigned int dax_entry_order(void *entr
85 {
86 if (xa_to_value(entry) & DAX_PMD)
87 return PMD_ORDER;
88 return 0;
89 }
90
91 static unsigned long dax_is_pmd_entry(void *en
92 {
93 return xa_to_value(entry) & DAX_PMD;
94 }
95
96 static bool dax_is_pte_entry(void *entry)
97 {
98 return !(xa_to_value(entry) & DAX_PMD)
99 }
100
101 static int dax_is_zero_entry(void *entry)
102 {
103 return xa_to_value(entry) & DAX_ZERO_P
104 }
105
106 static 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 sm
113 * we were looking for
114 */
115 static bool dax_is_conflict(void *entry)
116 {
117 return entry == XA_RETRY_ENTRY;
118 }
119
120 /*
121 * DAX page cache entry locking
122 */
123 struct exceptional_entry_key {
124 struct xarray *xa;
125 pgoff_t entry_start;
126 };
127
128 struct 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 behavi
135 * @WAKE_ALL: wake all waiters in the waitqueu
136 * @WAKE_NEXT: wake only the first waiter in t
137 */
138 enum dax_wake_mode {
139 WAKE_ALL,
140 WAKE_NEXT,
141 };
142
143 static wait_queue_head_t *dax_entry_waitqueue(
144 void *entry, struct exceptiona
145 {
146 unsigned long hash;
147 unsigned long index = xas->xa_index;
148
149 /*
150 * If 'entry' is a PMD, align the 'ind
151 * queue to the start of that PMD. Th
152 * the range covered by the PMD map to
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->x
160 return wait_table + hash;
161 }
162
163 static int wake_exceptional_entry_func(wait_qu
164 unsigned int mode, int sync, v
165 {
166 struct exceptional_entry_key *key = ke
167 struct wait_exceptional_entry_queue *e
168 container_of(wait, struct wait
169
170 if (key->xa != ewait->key.xa ||
171 key->entry_start != ewait->key.ent
172 return 0;
173 return autoremove_wake_function(wait,
174 }
175
176 /*
177 * @entry may no longer be the entry at the in
178 * The important information it's conveying is
179 * this index used to be a PMD entry.
180 */
181 static void dax_wake_entry(struct xa_state *xa
182 enum dax_wake_mode
183 {
184 struct exceptional_entry_key key;
185 wait_queue_head_t *wq;
186
187 wq = dax_entry_waitqueue(xas, entry, &
188
189 /*
190 * Checking for locked entry and prepa
191 * under the i_pages lock, ditto for e
192 * So at this point all tasks that cou
193 * must be in the waitqueue and the fo
194 */
195 if (waitqueue_active(wq))
196 __wake_up(wq, TASK_NORMAL, mod
197 }
198
199 /*
200 * Look up entry in page cache, wait for it to
201 * is a DAX entry and return it. The caller m
202 * put_unlocked_entry() if it did not lock the
203 * if it did. The entry returned may have a l
204 * If @order is larger than the order of the e
205 * function returns a dax_is_conflict entry.
206 *
207 * Must be called with the i_pages lock held.
208 */
209 static void *get_unlocked_entry(struct xa_stat
210 {
211 void *entry;
212 struct wait_exceptional_entry_queue ew
213 wait_queue_head_t *wq;
214
215 init_wait(&ewait.wait);
216 ewait.wait.func = wake_exceptional_ent
217
218 for (;;) {
219 entry = xas_find_conflict(xas)
220 if (!entry || WARN_ON_ONCE(!xa
221 return entry;
222 if (dax_entry_order(entry) < o
223 return XA_RETRY_ENTRY;
224 if (!dax_is_locked(entry))
225 return entry;
226
227 wq = dax_entry_waitqueue(xas,
228 prepare_to_wait_exclusive(wq,
229 TASK
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 ar
240 * (it's cycled in clear_inode() after removin
241 * After we call xas_unlock_irq(), we cannot t
242 */
243 static void wait_entry_unlocked(struct xa_stat
244 {
245 struct wait_exceptional_entry_queue ew
246 wait_queue_head_t *wq;
247
248 init_wait(&ewait.wait);
249 ewait.wait.func = wake_exceptional_ent
250
251 wq = dax_entry_waitqueue(xas, entry, &
252 /*
253 * Unlike get_unlocked_entry() there i
254 * path ever successfully retrieves an
255 * inode dies. Perform a non-exclusive
256 * never successfully performs its own
257 */
258 prepare_to_wait(wq, &ewait.wait, TASK_
259 xas_unlock_irq(xas);
260 schedule();
261 finish_wait(wq, &ewait.wait);
262 }
263
264 static void put_unlocked_entry(struct xa_state
265 enum dax_wake_m
266 {
267 if (entry && !dax_is_conflict(entry))
268 dax_wake_entry(xas, entry, mod
269 }
270
271 /*
272 * We used the xa_state to get the entry, but
273 * dropped the xa_lock, so we know the xa_stat
274 * before use.
275 */
276 static void dax_unlock_entry(struct xa_state *
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 b
291 */
292 static void *dax_lock_entry(struct xa_state *x
293 {
294 unsigned long v = xa_to_value(entry);
295 return xas_store(xas, xa_mk_value(v |
296 }
297
298 static unsigned long dax_entry_size(void *entr
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
310 static unsigned long dax_end_pfn(void *entry)
311 {
312 return dax_to_pfn(entry) + dax_entry_s
313 }
314
315 /*
316 * Iterate through all mapped pfns represented
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(entr
322
323 static inline bool dax_page_is_shared(struct p
324 {
325 return page->mapping == PAGE_MAPPING_D
326 }
327
328 /*
329 * Set the page->mapping with PAGE_MAPPING_DAX
330 * refcount.
331 */
332 static inline void dax_page_share_get(struct p
333 {
334 if (page->mapping != PAGE_MAPPING_DAX_
335 /*
336 * Reset the index if the page
337 * regularly before.
338 */
339 if (page->mapping)
340 page->share = 1;
341 page->mapping = PAGE_MAPPING_D
342 }
343 page->share++;
344 }
345
346 static inline unsigned long dax_page_share_put
347 {
348 return --page->share;
349 }
350
351 /*
352 * When it is called in dax_insert_entry(), th
353 * whether this entry is shared by multiple fi
354 * PAGE_MAPPING_DAX_SHARED, and use page->shar
355 */
356 static void dax_associate_entry(void *entry, s
357 struct vm_area_struct *vma, un
358 {
359 unsigned long size = dax_entry_size(en
360 int i = 0;
361
362 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
363 return;
364
365 index = linear_page_index(vma, address
366 for_each_mapped_pfn(entry, pfn) {
367 struct page *page = pfn_to_pag
368
369 if (shared) {
370 dax_page_share_get(pag
371 } else {
372 WARN_ON_ONCE(page->map
373 page->mapping = mappin
374 page->index = index +
375 }
376 }
377 }
378
379 static void dax_disassociate_entry(void *entry
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_pag
389
390 WARN_ON_ONCE(trunc && page_ref
391 if (dax_page_is_shared(page))
392 /* keep the shared fla
393 if (dax_page_share_put
394 continue;
395 } else
396 WARN_ON_ONCE(page->map
397 page->mapping = NULL;
398 page->index = 0;
399 }
400 }
401
402 static 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_pag
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 corresp
417 * @folio: The folio whose entry we want to lo
418 *
419 * Context: Process context.
420 * Return: A cookie to pass to dax_unlock_foli
421 * not be locked.
422 */
423 dax_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 w
429 rcu_read_lock();
430 for (;;) {
431 struct address_space *mapping
432
433 entry = NULL;
434 if (!mapping || !dax_mapping(m
435 break;
436
437 /*
438 * In the device-dax case ther
439 * struct dev_pagemap pin is s
440 * inode alive, and we assume
441 * otherwise we would not have
442 * translation.
443 */
444 entry = (void *)~0UL;
445 if (S_ISCHR(mapping->host->i_m
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(&x
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
470 void dax_unlock_folio(struct folio *folio, dax
471 {
472 struct address_space *mapping = folio-
473 XA_STATE(xas, &mapping->i_pages, folio
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
483 * @mapping: the file's mapping whose entry we
484 * @index: the offset within this file
485 * @page: output the dax page corresponding to
486 *
487 * Return: A cookie to pass to dax_unlock_mapp
488 * could not be locked.
489 */
490 dax_entry_t dax_lock_mapping_entry(struct addr
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(&x
509 rcu_read_lock();
510 continue;
511 }
512 if (!entry ||
513 dax_is_zero_entry(entry) |
514 /*
515 * Because we are look
516 * and index, so the e
517 * or even a zero/empt
518 * an error case. So,
519 * not output @page.
520 */
521 entry = (void *)~0UL;
522 } else {
523 *page = pfn_to_page(da
524 dax_lock_entry(&xas, e
525 }
526 xas_unlock_irq(&xas);
527 break;
528 }
529 rcu_read_unlock();
530 return (dax_entry_t)entry;
531 }
532
533 void dax_unlock_mapping_entry(struct address_s
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
546 * with the entry locked. If the page cache do
547 * that index, add a locked empty entry.
548 *
549 * When requesting an entry with size DAX_PMD,
550 * either return that locked entry or will ret
551 * This will happen if there are any PTE entri
552 * that we are requesting.
553 *
554 * We always favor PTE entries over PMD entrie
555 * evict PTE entries in order to 'upgrade' the
556 * insertion will fail if it finds any PTE ent
557 * PTE insertion will cause an existing PMD en
558 * downgraded to PTE entries. This happens fo
559 * well as PMD empty entries.
560 *
561 * The exception to this downgrade path is for
562 * real storage backing them. We will leave t
563 * the tree, and PTE writes will simply dirty
564 *
565 * Note: Unlike filemap_fault() we don't honor
566 * persistent memory the benefit is doubtful.
567 * show it helps.
568 *
569 * On error, this function does not return an
570 * a VM_FAULT code, encoded as an xarray inter
571 * overlap with xarray value entries.
572 */
573 static void *grab_mapping_entry(struct xa_stat
574 struct address_space *mapping,
575 {
576 unsigned long index = xas->xa_index;
577 bool pmd_downgrade; /* splitting P
578 void *entry;
579
580 retry:
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(e
595 (dax_is_zero_entry
596 dax_is_empty_entr
597 pmd_downgrade
598 }
599 }
600 }
601
602 if (pmd_downgrade) {
603 /*
604 * Make sure 'entry' remains v
605 * the i_pages lock.
606 */
607 dax_lock_entry(xas, entry);
608
609 /*
610 * Besides huge zero pages the
611 * downgraded are empty entrie
612 * unmapped.
613 */
614 if (dax_is_zero_entry(entry))
615 xas_unlock_irq(xas);
616 unmap_mapping_pages(ma
617 xas->x
618 PG_PMD
619 xas_reset(xas);
620 xas_lock_irq(xas);
621 }
622
623 dax_disassociate_entry(entry,
624 xas_store(xas, NULL); /* und
625 dax_wake_entry(xas, entry, WAK
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_EMPT
635
636 if (order > 0)
637 flags |= DAX_PMD;
638 entry = dax_make_entry(pfn_to_
639 dax_lock_entry(xas, entry);
640 if (xas_error(xas))
641 goto out_unlock;
642 mapping->nrpages += 1UL << ord
643 }
644
645 out_unlock:
646 xas_unlock_irq(xas);
647 if (xas_nomem(xas, mapping_gfp_mask(ma
648 goto retry;
649 if (xas->xa_node == XA_ERROR(-ENOMEM))
650 return xa_mk_internal(VM_FAULT
651 if (xas_error(xas))
652 return xa_mk_internal(VM_FAULT
653 return entry;
654 fallback:
655 xas_unlock_irq(xas);
656 return xa_mk_internal(VM_FAULT_FALLBAC
657 }
658
659 /**
660 * dax_layout_busy_page_range - find first pin
661 * @mapping: address space to scan for a page
662 * @start: Starting offset. Page containing 's
663 * @end: End offset. Page containing 'end' is
664 * pages from 'start' till the end of fi
665 *
666 * DAX requires ZONE_DEVICE mapped pages. Thes
667 * 'onlined' to the page allocator so they are
668 * page->count == 1. A filesystem uses this in
669 * any page in the mapping is busy, i.e. for D
670 * get_user_pages() usages.
671 *
672 * It is expected that the filesystem is holdi
673 * establishment of new mappings in this addre
674 * to be able to run unmap_mapping_range() and
675 * mapping_mapped() becoming true.
676 */
677 struct page *dax_layout_busy_page_range(struct
678 loff_t
679 {
680 void *entry;
681 unsigned int scanned = 0;
682 struct page *page = NULL;
683 pgoff_t start_idx = start >> PAGE_SHIF
684 pgoff_t end_idx;
685 XA_STATE(xas, &mapping->i_pages, start
686
687 /*
688 * In the 'limited' case get_user_page
689 */
690 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
691 return NULL;
692
693 if (!dax_mapping(mapping) || !mapping_
694 return NULL;
695
696 /* If end == LLONG_MAX, all pages from
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() he
703 * elevated page count in the iteratio
704 * get_user_pages_fast() will see that
705 * against is no longer mapped in the
706 * get_user_pages() slow path. The sl
707 * pte_lock() and pmd_lock(). New refe
708 * holding those locks, and unmap_mapp
709 * pte or pmd without holding the resp
710 * guaranteed to either see new refere
711 * references from being established.
712 */
713 unmap_mapping_pages(mapping, start_idx
714
715 xas_lock_irq(&xas);
716 xas_for_each(&xas, entry, end_idx) {
717 if (WARN_ON_ONCE(!xa_is_value(
718 continue;
719 if (unlikely(dax_is_locked(ent
720 entry = get_unlocked_e
721 if (entry)
722 page = dax_busy_page(e
723 put_unlocked_entry(&xas, entry
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 }
737 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
738
739 struct page *dax_layout_busy_page(struct addre
740 {
741 return dax_layout_busy_page_range(mapp
742 }
743 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
744
745 static int __dax_invalidate_entry(struct addre
746 pgof
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_valu
755 goto out;
756 if (!trunc &&
757 (xas_get_mark(&xas, PAGECACHE_TAG_
758 xas_get_mark(&xas, PAGECACHE_TAG_
759 goto out;
760 dax_disassociate_entry(entry, mapping,
761 xas_store(&xas, NULL);
762 mapping->nrpages -= 1UL << dax_entry_o
763 ret = 1;
764 out:
765 put_unlocked_entry(&xas, entry, WAKE_A
766 xas_unlock_irq(&xas);
767 return ret;
768 }
769
770 static int __dax_clear_dirty_range(struct addr
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(&xa
780 xas_clear_mark(&xas, PAGECACHE
781 xas_clear_mark(&xas, PAGECACHE
782 put_unlocked_entry(&xas, entry
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.
799 * to be unlocked before deleting it.
800 */
801 int dax_delete_mapping_entry(struct address_sp
802 {
803 int ret = __dax_invalidate_entry(mappi
804
805 /*
806 * This gets called from truncate / pu
807 * must hold locks protecting against
808 * page cache (usually fs-private i_mm
809 * caller has seen a DAX entry for thi
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 */
819 int dax_invalidate_mapping_entry_sync(struct a
820 pgoff_t
821 {
822 return __dax_invalidate_entry(mapping,
823 }
824
825 static pgoff_t dax_iomap_pgoff(const struct io
826 {
827 return PHYS_PFN(iomap->addr + (pos & P
828 }
829
830 static int copy_cow_page_dax(struct vm_fault *
831 {
832 pgoff_t pgoff = dax_iomap_pgoff(&iter-
833 void *vto, *kaddr;
834 long rc;
835 int id;
836
837 id = dax_read_lock();
838 rc = dax_direct_access(iter->iomap.dax
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->addres
846 kunmap_atomic(vto);
847 dax_read_unlock(id);
848 return 0;
849 }
850
851 /*
852 * MAP_SYNC on a dax mapping guarantees dirty
853 * flushed on write-faults (non-cow), but not
854 */
855 static bool dax_fault_is_synchronous(const str
856 struct vm_area_struct *vma)
857 {
858 return (iter->flags & IOMAP_WRITE) &&
859 (iter->iomap.flags & IOMAP_F_D
860 }
861
862 /*
863 * By this point grab_mapping_entry() has ensu
864 * of the appropriate size so we don't have to
865 * PTEs. If we happen to be trying to insert
866 * already in the tree, we will skip the inser
867 * appropriate.
868 */
869 static void *dax_insert_entry(struct xa_state
870 const struct iomap_iter *iter,
871 unsigned long flags)
872 {
873 struct address_space *mapping = vmf->v
874 void *new_entry = dax_make_entry(pfn,
875 bool write = iter->flags & IOMAP_WRITE
876 bool dirty = write && !dax_fault_is_sy
877 bool shared = iter->iomap.flags & IOMA
878
879 if (dirty)
880 __mark_inode_dirty(mapping->ho
881
882 if (shared || (dax_is_zero_entry(entry
883 unsigned long index = xas->xa_
884 /* we are replacing a zero pag
885 if (dax_is_pmd_entry(entry))
886 unmap_mapping_pages(ma
887 PG_PMD
888 else /* pte entry */
889 unmap_mapping_pages(ma
890 }
891
892 xas_reset(xas);
893 xas_lock_irq(xas);
894 if (shared || dax_is_zero_entry(entry)
895 void *old;
896
897 dax_disassociate_entry(entry,
898 dax_associate_entry(new_entry,
899 shared);
900 /*
901 * Only swap our new entry int
902 * entry is a zero page or an
903 * PMD entry is already in the
904 * means that if we are trying
905 * existing entry is a PMD, we
906 * tree and dirty it if necess
907 */
908 old = dax_lock_entry(xas, new_
909 WARN_ON_ONCE(old != xa_mk_valu
910 DAX_LO
911 entry = new_entry;
912 } else {
913 xas_load(xas); /* Walk the xa
914 }
915
916 if (dirty)
917 xas_set_mark(xas, PAGECACHE_TA
918
919 if (write && shared)
920 xas_set_mark(xas, PAGECACHE_TA
921
922 xas_unlock_irq(xas);
923 return entry;
924 }
925
926 static int dax_writeback_one(struct xa_state *
927 struct address_space *mapping,
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 mapp
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
944
945 /* Entry got punched out / rea
946 if (!entry || WARN_ON_ONCE(!xa
947 goto put_unlocked;
948 /*
949 * Entry got reallocated elsew
950 * We have to compare pfns as
951 * difference in lockbit or en
952 */
953 if (dax_to_pfn(old_entry) != d
954 goto put_unlocked;
955 if (WARN_ON_ONCE(dax_is_empty_
956 dax_is
957 ret = -EIO;
958 goto put_unlocked;
959 }
960
961 /* Another fsync thread may ha
962 if (!xas_get_mark(xas, PAGECAC
963 goto put_unlocked;
964 }
965
966 /* Lock the entry to serialize with pa
967 dax_lock_entry(xas, entry);
968
969 /*
970 * We can clear the tag now but we hav
971 * dax_writeback_one() calls for the s
972 * actually flush the caches. This is
973 * at the entry only under the i_pages
974 * they will see the entry locked and
975 */
976 xas_clear_mark(xas, PAGECACHE_TAG_TOWR
977 xas_unlock_irq(xas);
978
979 /*
980 * If dax_writeback_mapping_range() wa
981 * in the middle of a PMD, the 'index'
982 * aligned to the start of the PMD.
983 * This allows us to flush for PMD_SIZ
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
992 i_mmap_lock_read(mapping);
993 vma_interval_tree_foreach(vma, &mappin
994 pfn_mkclean_range(pfn, count,
995 cond_resched();
996 }
997 i_mmap_unlock_read(mapping);
998
999 dax_flush(dax_dev, page_address(pfn_to
1000 /*
1001 * After we have flushed the cache, w
1002 * cannot be new dirty data in the pf
1003 * the pfn mappings are writeprotecte
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_DIR
1010 dax_wake_entry(xas, entry, WAKE_NEXT)
1011
1012 trace_dax_writeback_one(mapping->host
1013 return ret;
1014
1015 put_unlocked:
1016 put_unlocked_entry(xas, entry, WAKE_N
1017 return ret;
1018 }
1019
1020 /*
1021 * Flush the mapping to the persistent domain
1022 * end]. This is required by data integrity o
1023 * on persistent storage prior to completion
1024 */
1025 int dax_writeback_mapping_range(struct addres
1026 struct dax_device *dax_dev, s
1027 {
1028 XA_STATE(xas, &mapping->i_pages, wbc-
1029 struct inode *inode = mapping->host;
1030 pgoff_t end_index = wbc->range_end >>
1031 void *entry;
1032 int ret = 0;
1033 unsigned int scanned = 0;
1034
1035 if (WARN_ON_ONCE(inode->i_blkbits !=
1036 return -EIO;
1037
1038 if (mapping_empty(mapping) || wbc->sy
1039 return 0;
1040
1041 trace_dax_writeback_range(inode, xas.
1042
1043 tag_pages_for_writeback(mapping, xas.
1044
1045 xas_lock_irq(&xas);
1046 xas_for_each_marked(&xas, entry, end_
1047 ret = dax_writeback_one(&xas,
1048 if (ret < 0) {
1049 mapping_set_error(map
1050 break;
1051 }
1052 if (++scanned % XA_CHECK_SCHE
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,
1062 return ret;
1063 }
1064 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range
1065
1066 static int dax_iomap_direct_access(const stru
1067 size_t size, void **kaddr, pf
1068 {
1069 pgoff_t pgoff = dax_iomap_pgoff(iomap
1070 int id, rc = 0;
1071 long length;
1072
1073 id = dax_read_lock();
1074 length = dax_direct_access(iomap->dax
1075 DAX_ACCESS
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(s
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
1092 out_check_addr:
1093 if (!kaddr)
1094 goto out;
1095 if (!*kaddr)
1096 rc = -EFAULT;
1097 out:
1098 dax_read_unlock(id);
1099 return rc;
1100 }
1101
1102 /**
1103 * dax_iomap_copy_around - Prepare for an una
1104 * by copying the data before and after the r
1105 * @pos: address to do copy from.
1106 * @length: size of copy operation.
1107 * @align_size: aligned w.r.t align_size (eit
1108 * @srcmap: iomap srcmap
1109 * @daddr: destination address to copy t
1110 *
1111 * This can be called from two places. Either
1112 * aligned), to copy the length size data to
1113 * write operation, dax_iomap_iter() might ca
1114 * start or end unaligned address. In the lat
1115 * aligned ranges is taken care by dax_iomap_
1116 * If the srcmap contains invalid data, such
1117 * area to make sure no old data remains.
1118 */
1119 static int dax_iomap_copy_around(loff_t pos,
1120 const struct iomap *srcmap, v
1121 {
1122 loff_t head_off = pos & (align_size -
1123 size_t size = ALIGN(head_off + length
1124 loff_t end = pos + length;
1125 loff_t pg_end = round_up(end, align_s
1126 /* copy_all is usually in page fault
1127 bool copy_all = head_off == 0 && end
1128 /* zero the edges if srcmap is a HOLE
1129 bool zero_edge = srcmap->flags & IOMA
1130 srcmap->type == IOMA
1131 void *saddr = NULL;
1132 int ret = 0;
1133
1134 if (!zero_edge) {
1135 ret = dax_iomap_direct_access
1136 if (ret)
1137 return dax_mem2blk_er
1138 }
1139
1140 if (copy_all) {
1141 if (zero_edge)
1142 memset(daddr, 0, size
1143 else
1144 ret = copy_mc_to_kern
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
1152 else {
1153 ret = copy_mc_to_kern
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 +
1162 loff_t tail_len = pg_end - en
1163
1164 if (zero_edge)
1165 memset(daddr + tail_o
1166 else {
1167 ret = copy_mc_to_kern
1168
1169 if (ret)
1170 return -EIO;
1171 }
1172 }
1173 out:
1174 if (zero_edge)
1175 dax_flush(srcmap->dax_dev, da
1176 return ret ? -EIO : 0;
1177 }
1178
1179 /*
1180 * The user has performed a load from a hole
1181 * page in the file would cause excessive sto
1182 * sparse files. Instead we insert a read-on
1183 * If this page is ever written to we will re
1184 * point to real DAX storage instead.
1185 */
1186 static vm_fault_t dax_load_hole(struct xa_sta
1187 const struct iomap_iter *iter
1188 {
1189 struct inode *inode = iter->inode;
1190 unsigned long vaddr = vmf->address;
1191 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(
1192 vm_fault_t ret;
1193
1194 *entry = dax_insert_entry(xas, vmf, i
1195
1196 ret = vmf_insert_mixed(vmf->vma, vadd
1197 trace_dax_load_hole(inode, vmf, ret);
1198 return ret;
1199 }
1200
1201 #ifdef CONFIG_FS_DAX_PMD
1202 static vm_fault_t dax_pmd_load_hole(struct xa
1203 const struct iomap_iter *iter
1204 {
1205 struct address_space *mapping = vmf->
1206 unsigned long pmd_addr = vmf->address
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(v
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, i
1222 DAX_PMD | D
1223
1224 if (arch_needs_pgtable_deposit()) {
1225 pgtable = pte_alloc_one(vma->
1226 if (!pgtable)
1227 return VM_FAULT_OOM;
1228 }
1229
1230 ptl = pmd_lock(vmf->vma->vm_mm, vmf->
1231 if (!pmd_none(*(vmf->pmd))) {
1232 spin_unlock(ptl);
1233 goto fallback;
1234 }
1235
1236 if (pgtable) {
1237 pgtable_trans_huge_deposit(vm
1238 mm_inc_nr_ptes(vma->vm_mm);
1239 }
1240 pmd_entry = mk_pmd(&zero_folio->page,
1241 pmd_entry = pmd_mkhuge(pmd_entry);
1242 set_pmd_at(vmf->vma->vm_mm, pmd_addr,
1243 spin_unlock(ptl);
1244 trace_dax_pmd_load_hole(inode, vmf, z
1245 return VM_FAULT_NOPAGE;
1246
1247 fallback:
1248 if (pgtable)
1249 pte_free(vma->vm_mm, pgtable)
1250 trace_dax_pmd_load_hole_fallback(inod
1251 return VM_FAULT_FALLBACK;
1252 }
1253 #else
1254 static vm_fault_t dax_pmd_load_hole(struct xa
1255 const struct iomap_iter *iter
1256 {
1257 return VM_FAULT_FALLBACK;
1258 }
1259 #endif /* CONFIG_FS_DAX_PMD */
1260
1261 static s64 dax_unshare_iter(struct iomap_iter
1262 {
1263 struct iomap *iomap = &iter->iomap;
1264 const struct iomap *srcmap = iomap_it
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 aligne
1277 * we need to copy entire blocks, not
1278 * Invalidate the mapping because we'
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_
1287 if (mod)
1288 copy_len += PAGE_SIZE - mod;
1289
1290 invalidate_inode_pages2_range(iter->i
1291 copy_po
1292 (copy_p
1293
1294 id = dax_read_lock();
1295 ret = dax_iomap_direct_access(iomap,
1296 if (ret < 0)
1297 goto out_unlock;
1298
1299 ret = dax_iomap_direct_access(srcmap,
1300 if (ret < 0)
1301 goto out_unlock;
1302
1303 if (copy_mc_to_kernel(daddr, saddr, c
1304 ret = iomap_length(iter);
1305 else
1306 ret = -EIO;
1307
1308 out_unlock:
1309 dax_read_unlock(id);
1310 return dax_mem2blk_err(ret);
1311 }
1312
1313 int dax_file_unshare(struct inode *inode, lof
1314 const struct iomap_ops *ops)
1315 {
1316 struct iomap_iter iter = {
1317 .inode = inode,
1318 .pos = pos,
1319 .flags = IOMAP_WRITE
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))
1329 iter.processed = dax_unshare_
1330 return ret;
1331 }
1332 EXPORT_SYMBOL_GPL(dax_file_unshare);
1333
1334 static int dax_memzero(struct iomap_iter *ite
1335 {
1336 const struct iomap *iomap = &iter->io
1337 const struct iomap *srcmap = iomap_it
1338 unsigned offset = offset_in_page(pos)
1339 pgoff_t pgoff = dax_iomap_pgoff(iomap
1340 void *kaddr;
1341 long ret;
1342
1343 ret = dax_direct_access(iomap->dax_de
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(p
1351 k
1352 else
1353 dax_flush(iomap->dax_dev, kad
1354 return ret;
1355 }
1356
1357 static s64 dax_zero_iter(struct iomap_iter *i
1358 {
1359 const struct iomap *iomap = &iter->io
1360 const struct iomap *srcmap = iomap_it
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 || src
1367 return length;
1368
1369 /*
1370 * invalidate the pages whose sharing
1371 * because of CoW.
1372 */
1373 if (iomap->flags & IOMAP_F_SHARED)
1374 invalidate_inode_pages2_range
1375
1376
1377
1378 do {
1379 unsigned offset = offset_in_p
1380 unsigned size = min_t(u64, PA
1381 pgoff_t pgoff = dax_iomap_pgo
1382 long rc;
1383 int id;
1384
1385 id = dax_read_lock();
1386 if (IS_ALIGNED(pos, PAGE_SIZE
1387 rc = dax_zero_page_ra
1388 else
1389 rc = dax_memzero(iter
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
1404 int dax_zero_range(struct inode *inode, loff_
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 |
1412 };
1413 int ret;
1414
1415 while ((ret = iomap_iter(&iter, ops))
1416 iter.processed = dax_zero_ite
1417 return ret;
1418 }
1419 EXPORT_SYMBOL_GPL(dax_zero_range);
1420
1421 int dax_truncate_page(struct inode *inode, lo
1422 const struct iomap_ops *ops)
1423 {
1424 unsigned int blocksize = i_blocksize(
1425 unsigned int off = pos & (blocksize -
1426
1427 /* Block boundary? Nothing to do */
1428 if (!off)
1429 return 0;
1430 return dax_zero_range(inode, pos, blo
1431 }
1432 EXPORT_SYMBOL_GPL(dax_truncate_page);
1433
1434 static loff_t dax_iomap_iter(const struct iom
1435 struct iov_iter *iter)
1436 {
1437 const struct iomap *iomap = &iomi->io
1438 const struct iomap *srcmap = iomap_it
1439 loff_t length = iomap_length(iomi);
1440 loff_t pos = iomi->pos;
1441 struct dax_device *dax_dev = iomap->d
1442 loff_t end = pos + length, done = 0;
1443 bool write = iov_iter_rw(iter) == WRI
1444 bool cow = write && iomap->flags & IO
1445 ssize_t ret = 0;
1446 size_t xfer;
1447 int id;
1448
1449 if (!write) {
1450 end = min(end, i_size_read(io
1451 if (pos >= end)
1452 return 0;
1453
1454 if (iomap->type == IOMAP_HOLE
1455 return iov_iter_zero(
1456 }
1457
1458 /*
1459 * In DAX mode, enforce either pure o
1460 * writes to unwritten extents as par
1461 */
1462 if (WARN_ON_ONCE(iomap->type != IOMAP
1463 !(iomap->flags & IOMA
1464 return -EIO;
1465
1466 /*
1467 * Write can allocate block for an ar
1468 * into page tables. We have to tear
1469 * written by write(2) is visible in
1470 */
1471 if (iomap->flags & IOMAP_F_NEW || cow
1472 /*
1473 * Filesystem allows CoW on n
1474 * may have been mmapped with
1475 * invalidate its dax entries
1476 * in advance.
1477 */
1478 if (cow)
1479 __dax_clear_dirty_ran
1480
1481
1482 invalidate_inode_pages2_range
1483
1484
1485 }
1486
1487 id = dax_read_lock();
1488 while (pos < end) {
1489 unsigned offset = pos & (PAGE
1490 const size_t size = ALIGN(len
1491 pgoff_t pgoff = dax_iomap_pgo
1492 ssize_t map_len;
1493 bool recovery = false;
1494 void *kaddr;
1495
1496 if (fatal_signal_pending(curr
1497 ret = -EINTR;
1498 break;
1499 }
1500
1501 map_len = dax_direct_access(d
1502 DAX_ACCESS, &
1503 if (map_len == -EHWPOISON &&
1504 map_len = dax_direct_
1505 PHYS_
1506 &kadd
1507 if (map_len > 0)
1508 recovery = tr
1509 }
1510 if (map_len < 0) {
1511 ret = dax_mem2blk_err
1512 break;
1513 }
1514
1515 if (cow) {
1516 ret = dax_iomap_copy_
1517
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_w
1530 map_l
1531 else if (write)
1532 xfer = dax_copy_from_
1533 map_l
1534 else
1535 xfer = dax_copy_to_it
1536 map_l
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/
1555 * @iter: The addresses to do I/O from
1556 * @ops: iomap ops passed from the fil
1557 *
1558 * This function performs read and write oper
1559 * persistent memory. The callers needs to t
1560 * and evicting any page cache pages in the r
1561 */
1562 ssize_t
1563 dax_iomap_rw(struct kiocb *iocb, struct iov_i
1564 const struct iomap_ops *ops)
1565 {
1566 struct iomap_iter iomi = {
1567 .inode = iocb->ki_fi
1568 .pos = iocb->ki_po
1569 .len = iov_iter_co
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(&io
1580 iomi.flags |= IOMAP_WRITE;
1581 } else {
1582 lockdep_assert_held(&iomi.ino
1583 }
1584
1585 if (iocb->ki_flags & IOCB_NOWAIT)
1586 iomi.flags |= IOMAP_NOWAIT;
1587
1588 while ((ret = iomap_iter(&iomi, ops))
1589 iomi.processed = dax_iomap_it
1590
1591 done = iomi.pos - iocb->ki_pos;
1592 iocb->ki_pos = iomi.pos;
1593 return done ? done : ret;
1594 }
1595 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1596
1597 static 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
1606 * insert the PTE/PMD into page tables only a
1607 * insertion for now and return the pfn so th
1608 * fsync is done.
1609 */
1610 static vm_fault_t dax_fault_synchronous_pfnp(
1611 {
1612 if (WARN_ON_ONCE(!pfnp))
1613 return VM_FAULT_SIGBUS;
1614 *pfnp = pfn;
1615 return VM_FAULT_NEEDDSYNC;
1616 }
1617
1618 static vm_fault_t dax_fault_cow_page(struct v
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_
1628 break;
1629 case IOMAP_MAPPED:
1630 error = copy_cow_page_dax(vmf
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 pf
1650 * @vmf: vm fault instance
1651 * @iter: iomap iter
1652 * @pfnp: pfn to be returned
1653 * @xas: the dax mapping tree of a fil
1654 * @entry: an unlocked dax entry to be i
1655 * @pmd: distinguish whether it is a p
1656 */
1657 static vm_fault_t dax_fault_iter(struct vm_fa
1658 const struct iomap_iter *iter
1659 struct xa_state *xas, void **
1660 {
1661 const struct iomap *iomap = &iter->io
1662 const struct iomap *srcmap = iomap_it
1663 size_t size = pmd ? PMD_SIZE : PAGE_S
1664 loff_t pos = (loff_t)xas->xa_index <<
1665 bool write = iter->flags & IOMAP_WRIT
1666 unsigned long entry_flags = pmd ? DAX
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
1673
1674 /* if we are reading UNWRITTEN and HO
1675 if (!write &&
1676 (iomap->type == IOMAP_UNWRITTEN |
1677 if (!pmd)
1678 return dax_load_hole(
1679 return dax_pmd_load_hole(xas,
1680 }
1681
1682 if (iomap->type != IOMAP_MAPPED && !(
1683 WARN_ON_ONCE(1);
1684 return pmd ? VM_FAULT_FALLBAC
1685 }
1686
1687 err = dax_iomap_direct_access(iomap,
1688 if (err)
1689 return pmd ? VM_FAULT_FALLBAC
1690
1691 *entry = dax_insert_entry(xas, vmf, i
1692
1693 if (write && iomap->flags & IOMAP_F_S
1694 err = dax_iomap_copy_around(p
1695 if (err)
1696 return dax_fault_retu
1697 }
1698
1699 if (dax_fault_is_synchronous(iter, vm
1700 return dax_fault_synchronous_
1701
1702 /* insert PMD pfn */
1703 if (pmd)
1704 return vmf_insert_pfn_pmd(vmf
1705
1706 /* insert PTE pfn */
1707 if (write)
1708 return vmf_insert_mixed_mkwri
1709 return vmf_insert_mixed(vmf->vma, vmf
1710 }
1711
1712 static vm_fault_t dax_iomap_pte_fault(struct
1713 int *iomap_err
1714 {
1715 struct address_space *mapping = vmf->
1716 XA_STATE(xas, &mapping->i_pages, vmf-
1717 struct iomap_iter iter = {
1718 .inode = mapping->ho
1719 .pos = (loff_t)vmf
1720 .len = PAGE_SIZE,
1721 .flags = IOMAP_DAX |
1722 };
1723 vm_fault_t ret = 0;
1724 void *entry;
1725 int error;
1726
1727 trace_dax_pte_fault(iter.inode, vmf,
1728 /*
1729 * Check whether offset isn't beyond
1730 * to hold locks serializing us with
1731 * a reliable test.
1732 */
1733 if (iter.pos >= i_size_read(iter.inod
1734 ret = VM_FAULT_SIGBUS;
1735 goto out;
1736 }
1737
1738 if ((vmf->flags & FAULT_FLAG_WRITE) &
1739 iter.flags |= IOMAP_WRITE;
1740
1741 entry = grab_mapping_entry(&xas, mapp
1742 if (xa_is_internal(entry)) {
1743 ret = xa_to_internal(entry);
1744 goto out;
1745 }
1746
1747 /*
1748 * It is possible, particularly with
1749 * mappings, that we have raced with
1750 * the PTE we need to set up. If so
1751 * retried.
1752 */
1753 if (pmd_trans_huge(*vmf->pmd) || pmd_
1754 ret = VM_FAULT_NOPAGE;
1755 goto unlock_entry;
1756 }
1757
1758 while ((error = iomap_iter(&iter, ops
1759 if (WARN_ON_ONCE(iomap_length
1760 iter.processed = -EIO
1761 continue;
1762 }
1763
1764 ret = dax_fault_iter(vmf, &it
1765 if (ret != VM_FAULT_SIGBUS &&
1766 (iter.iomap.flags & IOMAP
1767 count_vm_event(PGMAJF
1768 count_memcg_event_mm(
1769 ret |= VM_FAULT_MAJOR
1770 }
1771
1772 if (!(ret & VM_FAULT_ERROR))
1773 iter.processed = PAGE
1774 }
1775
1776 if (iomap_errp)
1777 *iomap_errp = error;
1778 if (!ret && error)
1779 ret = dax_fault_return(error)
1780
1781 unlock_entry:
1782 dax_unlock_entry(&xas, entry);
1783 out:
1784 trace_dax_pte_fault_done(iter.inode,
1785 return ret;
1786 }
1787
1788 #ifdef CONFIG_FS_DAX_PMD
1789 static bool dax_fault_check_fallback(struct v
1790 pgoff_t max_pgoff)
1791 {
1792 unsigned long pmd_addr = vmf->address
1793 bool write = vmf->flags & FAULT_FLAG_
1794
1795 /*
1796 * Make sure that the faulting addres
1797 * the PMD offset from the start of t
1798 * that a PMD range in the page table
1799 * range in the page cache.
1800 */
1801 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1802 ((vmf->address >> PAGE_SHIFT) & P
1803 return true;
1804
1805 /* Fall back to PTEs if we're going t
1806 if (write && !(vmf->vma->vm_flags & V
1807 return true;
1808
1809 /* If the PMD would extend outside th
1810 if (pmd_addr < vmf->vma->vm_start)
1811 return true;
1812 if ((pmd_addr + PMD_SIZE) > vmf->vma-
1813 return true;
1814
1815 /* If the PMD would extend beyond the
1816 if ((xas->xa_index | PG_PMD_COLOUR) >
1817 return true;
1818
1819 return false;
1820 }
1821
1822 static vm_fault_t dax_iomap_pmd_fault(struct
1823 const struct i
1824 {
1825 struct address_space *mapping = vmf->
1826 XA_STATE_ORDER(xas, &mapping->i_pages
1827 struct iomap_iter iter = {
1828 .inode = mapping->ho
1829 .len = PMD_SIZE,
1830 .flags = IOMAP_DAX |
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
1841 * supposed to hold locks serializing
1842 * this is a reliable test.
1843 */
1844 max_pgoff = DIV_ROUND_UP(i_size_read(
1845
1846 trace_dax_pmd_fault(iter.inode, vmf,
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, &xa
1854 goto fallback;
1855
1856 /*
1857 * grab_mapping_entry() will make sur
1858 * a zero PMD entry or a DAX PMD. If
1859 * entry is already in the array, for
1860 * VM_FAULT_FALLBACK.
1861 */
1862 entry = grab_mapping_entry(&xas, mapp
1863 if (xa_is_internal(entry)) {
1864 ret = xa_to_internal(entry);
1865 goto fallback;
1866 }
1867
1868 /*
1869 * It is possible, particularly with
1870 * mappings, that we have raced with
1871 * the PMD we need to set up. If so
1872 * retried.
1873 */
1874 if (!pmd_none(*vmf->pmd) && !pmd_tran
1875 !pmd_devmap(*vmf->pmd
1876 ret = 0;
1877 goto unlock_entry;
1878 }
1879
1880 iter.pos = (loff_t)xas.xa_index << PA
1881 while (iomap_iter(&iter, ops) > 0) {
1882 if (iomap_length(&iter) < PMD
1883 continue; /* actually
1884
1885 ret = dax_fault_iter(vmf, &it
1886 if (ret != VM_FAULT_FALLBACK)
1887 iter.processed = PMD_
1888 }
1889
1890 unlock_entry:
1891 dax_unlock_entry(&xas, entry);
1892 fallback:
1893 if (ret == VM_FAULT_FALLBACK) {
1894 split_huge_pmd(vmf->vma, vmf-
1895 count_vm_event(THP_FAULT_FALL
1896 }
1897 out:
1898 trace_dax_pmd_fault_done(iter.inode,
1899 return ret;
1900 }
1901 #else
1902 static vm_fault_t dax_iomap_pmd_fault(struct
1903 const struct i
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
1911 * @vmf: The description of the fault
1912 * @order: Order of the page to fault in
1913 * @pfnp: PFN to insert for synchronous fault
1914 * @iomap_errp: Storage for detailed error co
1915 * @ops: Iomap ops passed from the file syste
1916 *
1917 * When a page fault occurs, filesystems may
1918 * their fault handler for DAX files. dax_iom
1919 * has done all the necessary locking for pag
1920 * successfully.
1921 */
1922 vm_fault_t dax_iomap_fault(struct vm_fault *v
1923 pfn_t *pfnp, int *iomap_e
1924 {
1925 if (order == 0)
1926 return dax_iomap_pte_fault(vm
1927 else if (order == PMD_ORDER)
1928 return dax_iomap_pmd_fault(vm
1929 else
1930 return VM_FAULT_FALLBACK;
1931 }
1932 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1933
1934 /*
1935 * dax_insert_pfn_mkwrite - insert PTE or PMD
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 P
1941 * for an mmaped DAX file. It also marks the
1942 */
1943 static vm_fault_t
1944 dax_insert_pfn_mkwrite(struct vm_fault *vmf,
1945 {
1946 struct address_space *mapping = vmf->
1947 XA_STATE_ORDER(xas, &mapping->i_pages
1948 void *entry;
1949 vm_fault_t ret;
1950
1951 xas_lock_irq(&xas);
1952 entry = get_unlocked_entry(&xas, orde
1953 /* Did we race with someone splitting
1954 if (!entry || dax_is_conflict(entry)
1955 (order == 0 && !dax_is_pte_entry(
1956 put_unlocked_entry(&xas, entr
1957 xas_unlock_irq(&xas);
1958 trace_dax_insert_pfn_mkwrite_
1959
1960 return VM_FAULT_NOPAGE;
1961 }
1962 xas_set_mark(&xas, PAGECACHE_TAG_DIRT
1963 dax_lock_entry(&xas, entry);
1964 xas_unlock_irq(&xas);
1965 if (order == 0)
1966 ret = vmf_insert_mixed_mkwrit
1967 #ifdef CONFIG_FS_DAX_PMD
1968 else if (order == PMD_ORDER)
1969 ret = vmf_insert_pfn_pmd(vmf,
1970 #endif
1971 else
1972 ret = VM_FAULT_FALLBACK;
1973 dax_unlock_entry(&xas, entry);
1974 trace_dax_insert_pfn_mkwrite(mapping-
1975 return ret;
1976 }
1977
1978 /**
1979 * dax_finish_sync_fault - finish synchronous
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
1985 * stored persistently on the media and handl
1986 * table entry.
1987 */
1988 vm_fault_t dax_finish_sync_fault(struct vm_fa
1989 pfn_t pfn)
1990 {
1991 int err;
1992 loff_t start = ((loff_t)vmf->pgoff) <
1993 size_t len = PAGE_SIZE << order;
1994
1995 err = vfs_fsync_range(vmf->vma->vm_fi
1996 if (err)
1997 return VM_FAULT_SIGBUS;
1998 return dax_insert_pfn_mkwrite(vmf, pf
1999 }
2000 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
2001
2002 static loff_t dax_range_compare_iter(struct i
2003 struct iomap_iter *it_dest, u
2004 {
2005 const struct iomap *smap = &it_src->i
2006 const struct iomap *dmap = &it_dest->
2007 loff_t pos1 = it_src->pos, pos2 = it_
2008 void *saddr, *daddr;
2009 int id, ret;
2010
2011 len = min(len, min(smap->length, dmap
2012
2013 if (smap->type == IOMAP_HOLE && dmap-
2014 *same = true;
2015 return len;
2016 }
2017
2018 if (smap->type == IOMAP_HOLE || dmap-
2019 *same = false;
2020 return 0;
2021 }
2022
2023 id = dax_read_lock();
2024 ret = dax_iomap_direct_access(smap, p
2025 &saddr,
2026 if (ret < 0)
2027 goto out_unlock;
2028
2029 ret = dax_iomap_direct_access(dmap, p
2030 &daddr,
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
2040 out_unlock:
2041 dax_read_unlock(id);
2042 return -EIO;
2043 }
2044
2045 int dax_dedupe_file_range_compare(struct inod
2046 struct inode *dst, loff_t dst
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, o
2064 (ret = iomap_iter(&dst_iter, o
2065 compared = dax_range_compare_
2066 min(src_iter.
2067 if (compared < 0)
2068 return ret;
2069 src_iter.processed = dst_iter
2070 }
2071 return ret;
2072 }
2073
2074 int dax_remap_file_range_prep(struct file *fi
2075 struct file *fi
2076 loff_t *len, un
2077 const struct io
2078 {
2079 return __generic_remap_file_range_pre
2080
2081 }
2082 EXPORT_SYMBOL_GPL(dax_remap_file_range_prep);
2083
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.