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
41 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
42
43 static 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 }
51 fs_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
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, unsigned long flags)
75 {
76 return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
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 *entry)
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 *entry)
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_PAGE;
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 smaller order than the entry
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 behaviour
135 * @WAKE_ALL: wake all waiters in the waitqueue
136 * @WAKE_NEXT: wake only the first waiter in the waitqueue
137 */
138 enum dax_wake_mode {
139 WAKE_ALL,
140 WAKE_NEXT,
141 };
142
143 static 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
163 static 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 */
181 static 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 */
209 static 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 */
243 static 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
264 static 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 */
276 static 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 */
292 static 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
298 static 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
310 static 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
323 static 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 */
332 static 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
346 static 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 */
356 static 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
379 static 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
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_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 */
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 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
470 void 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 */
490 dax_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
533 void 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 */
573 static 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
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(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
645 out_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;
654 fallback:
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 */
677 struct 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 }
737 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
738
739 struct page *dax_layout_busy_page(struct address_space *mapping)
740 {
741 return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
742 }
743 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
744
745 static 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;
764 out:
765 put_unlocked_entry(&xas, entry, WAKE_ALL);
766 xas_unlock_irq(&xas);
767 return ret;
768 }
769
770 static 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 */
801 int 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 */
819 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
820 pgoff_t index)
821 {
822 return __dax_invalidate_entry(mapping, index, false);
823 }
824
825 static 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
830 static 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 */
855 static 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 */
869 static 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
926 static 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 */
1025 int 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 }
1064 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1065
1066 static 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
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 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 */
1119 static 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 }
1173 out:
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 */
1186 static 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
1202 static 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
1247 fallback:
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
1254 static 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
1261 static 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 pos = iter->pos;
1266 loff_t length = iomap_length(iter);
1267 int id = 0;
1268 s64 ret = 0;
1269 void *daddr = NULL, *saddr = NULL;
1270
1271 /* don't bother with blocks that are not shared to start with */
1272 if (!(iomap->flags & IOMAP_F_SHARED))
1273 return length;
1274
1275 id = dax_read_lock();
1276 ret = dax_iomap_direct_access(iomap, pos, length, &daddr, NULL);
1277 if (ret < 0)
1278 goto out_unlock;
1279
1280 /* zero the distance if srcmap is HOLE or UNWRITTEN */
1281 if (srcmap->flags & IOMAP_F_SHARED || srcmap->type == IOMAP_UNWRITTEN) {
1282 memset(daddr, 0, length);
1283 dax_flush(iomap->dax_dev, daddr, length);
1284 ret = length;
1285 goto out_unlock;
1286 }
1287
1288 ret = dax_iomap_direct_access(srcmap, pos, length, &saddr, NULL);
1289 if (ret < 0)
1290 goto out_unlock;
1291
1292 if (copy_mc_to_kernel(daddr, saddr, length) == 0)
1293 ret = length;
1294 else
1295 ret = -EIO;
1296
1297 out_unlock:
1298 dax_read_unlock(id);
1299 return dax_mem2blk_err(ret);
1300 }
1301
1302 int dax_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1303 const struct iomap_ops *ops)
1304 {
1305 struct iomap_iter iter = {
1306 .inode = inode,
1307 .pos = pos,
1308 .flags = IOMAP_WRITE | IOMAP_UNSHARE | IOMAP_DAX,
1309 };
1310 loff_t size = i_size_read(inode);
1311 int ret;
1312
1313 if (pos < 0 || pos >= size)
1314 return 0;
1315
1316 iter.len = min(len, size - pos);
1317 while ((ret = iomap_iter(&iter, ops)) > 0)
1318 iter.processed = dax_unshare_iter(&iter);
1319 return ret;
1320 }
1321 EXPORT_SYMBOL_GPL(dax_file_unshare);
1322
1323 static int dax_memzero(struct iomap_iter *iter, loff_t pos, size_t size)
1324 {
1325 const struct iomap *iomap = &iter->iomap;
1326 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1327 unsigned offset = offset_in_page(pos);
1328 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1329 void *kaddr;
1330 long ret;
1331
1332 ret = dax_direct_access(iomap->dax_dev, pgoff, 1, DAX_ACCESS, &kaddr,
1333 NULL);
1334 if (ret < 0)
1335 return dax_mem2blk_err(ret);
1336
1337 memset(kaddr + offset, 0, size);
1338 if (iomap->flags & IOMAP_F_SHARED)
1339 ret = dax_iomap_copy_around(pos, size, PAGE_SIZE, srcmap,
1340 kaddr);
1341 else
1342 dax_flush(iomap->dax_dev, kaddr + offset, size);
1343 return ret;
1344 }
1345
1346 static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero)
1347 {
1348 const struct iomap *iomap = &iter->iomap;
1349 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1350 loff_t pos = iter->pos;
1351 u64 length = iomap_length(iter);
1352 s64 written = 0;
1353
1354 /* already zeroed? we're done. */
1355 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1356 return length;
1357
1358 /*
1359 * invalidate the pages whose sharing state is to be changed
1360 * because of CoW.
1361 */
1362 if (iomap->flags & IOMAP_F_SHARED)
1363 invalidate_inode_pages2_range(iter->inode->i_mapping,
1364 pos >> PAGE_SHIFT,
1365 (pos + length - 1) >> PAGE_SHIFT);
1366
1367 do {
1368 unsigned offset = offset_in_page(pos);
1369 unsigned size = min_t(u64, PAGE_SIZE - offset, length);
1370 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1371 long rc;
1372 int id;
1373
1374 id = dax_read_lock();
1375 if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE)
1376 rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1377 else
1378 rc = dax_memzero(iter, pos, size);
1379 dax_read_unlock(id);
1380
1381 if (rc < 0)
1382 return rc;
1383 pos += size;
1384 length -= size;
1385 written += size;
1386 } while (length > 0);
1387
1388 if (did_zero)
1389 *did_zero = true;
1390 return written;
1391 }
1392
1393 int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1394 const struct iomap_ops *ops)
1395 {
1396 struct iomap_iter iter = {
1397 .inode = inode,
1398 .pos = pos,
1399 .len = len,
1400 .flags = IOMAP_DAX | IOMAP_ZERO,
1401 };
1402 int ret;
1403
1404 while ((ret = iomap_iter(&iter, ops)) > 0)
1405 iter.processed = dax_zero_iter(&iter, did_zero);
1406 return ret;
1407 }
1408 EXPORT_SYMBOL_GPL(dax_zero_range);
1409
1410 int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1411 const struct iomap_ops *ops)
1412 {
1413 unsigned int blocksize = i_blocksize(inode);
1414 unsigned int off = pos & (blocksize - 1);
1415
1416 /* Block boundary? Nothing to do */
1417 if (!off)
1418 return 0;
1419 return dax_zero_range(inode, pos, blocksize - off, did_zero, ops);
1420 }
1421 EXPORT_SYMBOL_GPL(dax_truncate_page);
1422
1423 static loff_t dax_iomap_iter(const struct iomap_iter *iomi,
1424 struct iov_iter *iter)
1425 {
1426 const struct iomap *iomap = &iomi->iomap;
1427 const struct iomap *srcmap = iomap_iter_srcmap(iomi);
1428 loff_t length = iomap_length(iomi);
1429 loff_t pos = iomi->pos;
1430 struct dax_device *dax_dev = iomap->dax_dev;
1431 loff_t end = pos + length, done = 0;
1432 bool write = iov_iter_rw(iter) == WRITE;
1433 bool cow = write && iomap->flags & IOMAP_F_SHARED;
1434 ssize_t ret = 0;
1435 size_t xfer;
1436 int id;
1437
1438 if (!write) {
1439 end = min(end, i_size_read(iomi->inode));
1440 if (pos >= end)
1441 return 0;
1442
1443 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1444 return iov_iter_zero(min(length, end - pos), iter);
1445 }
1446
1447 /*
1448 * In DAX mode, enforce either pure overwrites of written extents, or
1449 * writes to unwritten extents as part of a copy-on-write operation.
1450 */
1451 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED &&
1452 !(iomap->flags & IOMAP_F_SHARED)))
1453 return -EIO;
1454
1455 /*
1456 * Write can allocate block for an area which has a hole page mapped
1457 * into page tables. We have to tear down these mappings so that data
1458 * written by write(2) is visible in mmap.
1459 */
1460 if (iomap->flags & IOMAP_F_NEW || cow) {
1461 /*
1462 * Filesystem allows CoW on non-shared extents. The src extents
1463 * may have been mmapped with dirty mark before. To be able to
1464 * invalidate its dax entries, we need to clear the dirty mark
1465 * in advance.
1466 */
1467 if (cow)
1468 __dax_clear_dirty_range(iomi->inode->i_mapping,
1469 pos >> PAGE_SHIFT,
1470 (end - 1) >> PAGE_SHIFT);
1471 invalidate_inode_pages2_range(iomi->inode->i_mapping,
1472 pos >> PAGE_SHIFT,
1473 (end - 1) >> PAGE_SHIFT);
1474 }
1475
1476 id = dax_read_lock();
1477 while (pos < end) {
1478 unsigned offset = pos & (PAGE_SIZE - 1);
1479 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1480 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1481 ssize_t map_len;
1482 bool recovery = false;
1483 void *kaddr;
1484
1485 if (fatal_signal_pending(current)) {
1486 ret = -EINTR;
1487 break;
1488 }
1489
1490 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1491 DAX_ACCESS, &kaddr, NULL);
1492 if (map_len == -EHWPOISON && iov_iter_rw(iter) == WRITE) {
1493 map_len = dax_direct_access(dax_dev, pgoff,
1494 PHYS_PFN(size), DAX_RECOVERY_WRITE,
1495 &kaddr, NULL);
1496 if (map_len > 0)
1497 recovery = true;
1498 }
1499 if (map_len < 0) {
1500 ret = dax_mem2blk_err(map_len);
1501 break;
1502 }
1503
1504 if (cow) {
1505 ret = dax_iomap_copy_around(pos, length, PAGE_SIZE,
1506 srcmap, kaddr);
1507 if (ret)
1508 break;
1509 }
1510
1511 map_len = PFN_PHYS(map_len);
1512 kaddr += offset;
1513 map_len -= offset;
1514 if (map_len > end - pos)
1515 map_len = end - pos;
1516
1517 if (recovery)
1518 xfer = dax_recovery_write(dax_dev, pgoff, kaddr,
1519 map_len, iter);
1520 else if (write)
1521 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1522 map_len, iter);
1523 else
1524 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1525 map_len, iter);
1526
1527 pos += xfer;
1528 length -= xfer;
1529 done += xfer;
1530
1531 if (xfer == 0)
1532 ret = -EFAULT;
1533 if (xfer < map_len)
1534 break;
1535 }
1536 dax_read_unlock(id);
1537
1538 return done ? done : ret;
1539 }
1540
1541 /**
1542 * dax_iomap_rw - Perform I/O to a DAX file
1543 * @iocb: The control block for this I/O
1544 * @iter: The addresses to do I/O from or to
1545 * @ops: iomap ops passed from the file system
1546 *
1547 * This function performs read and write operations to directly mapped
1548 * persistent memory. The callers needs to take care of read/write exclusion
1549 * and evicting any page cache pages in the region under I/O.
1550 */
1551 ssize_t
1552 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1553 const struct iomap_ops *ops)
1554 {
1555 struct iomap_iter iomi = {
1556 .inode = iocb->ki_filp->f_mapping->host,
1557 .pos = iocb->ki_pos,
1558 .len = iov_iter_count(iter),
1559 .flags = IOMAP_DAX,
1560 };
1561 loff_t done = 0;
1562 int ret;
1563
1564 if (!iomi.len)
1565 return 0;
1566
1567 if (iov_iter_rw(iter) == WRITE) {
1568 lockdep_assert_held_write(&iomi.inode->i_rwsem);
1569 iomi.flags |= IOMAP_WRITE;
1570 } else {
1571 lockdep_assert_held(&iomi.inode->i_rwsem);
1572 }
1573
1574 if (iocb->ki_flags & IOCB_NOWAIT)
1575 iomi.flags |= IOMAP_NOWAIT;
1576
1577 while ((ret = iomap_iter(&iomi, ops)) > 0)
1578 iomi.processed = dax_iomap_iter(&iomi, iter);
1579
1580 done = iomi.pos - iocb->ki_pos;
1581 iocb->ki_pos = iomi.pos;
1582 return done ? done : ret;
1583 }
1584 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1585
1586 static vm_fault_t dax_fault_return(int error)
1587 {
1588 if (error == 0)
1589 return VM_FAULT_NOPAGE;
1590 return vmf_error(error);
1591 }
1592
1593 /*
1594 * When handling a synchronous page fault and the inode need a fsync, we can
1595 * insert the PTE/PMD into page tables only after that fsync happened. Skip
1596 * insertion for now and return the pfn so that caller can insert it after the
1597 * fsync is done.
1598 */
1599 static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn)
1600 {
1601 if (WARN_ON_ONCE(!pfnp))
1602 return VM_FAULT_SIGBUS;
1603 *pfnp = pfn;
1604 return VM_FAULT_NEEDDSYNC;
1605 }
1606
1607 static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf,
1608 const struct iomap_iter *iter)
1609 {
1610 vm_fault_t ret;
1611 int error = 0;
1612
1613 switch (iter->iomap.type) {
1614 case IOMAP_HOLE:
1615 case IOMAP_UNWRITTEN:
1616 clear_user_highpage(vmf->cow_page, vmf->address);
1617 break;
1618 case IOMAP_MAPPED:
1619 error = copy_cow_page_dax(vmf, iter);
1620 break;
1621 default:
1622 WARN_ON_ONCE(1);
1623 error = -EIO;
1624 break;
1625 }
1626
1627 if (error)
1628 return dax_fault_return(error);
1629
1630 __SetPageUptodate(vmf->cow_page);
1631 ret = finish_fault(vmf);
1632 if (!ret)
1633 return VM_FAULT_DONE_COW;
1634 return ret;
1635 }
1636
1637 /**
1638 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault.
1639 * @vmf: vm fault instance
1640 * @iter: iomap iter
1641 * @pfnp: pfn to be returned
1642 * @xas: the dax mapping tree of a file
1643 * @entry: an unlocked dax entry to be inserted
1644 * @pmd: distinguish whether it is a pmd fault
1645 */
1646 static vm_fault_t dax_fault_iter(struct vm_fault *vmf,
1647 const struct iomap_iter *iter, pfn_t *pfnp,
1648 struct xa_state *xas, void **entry, bool pmd)
1649 {
1650 const struct iomap *iomap = &iter->iomap;
1651 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1652 size_t size = pmd ? PMD_SIZE : PAGE_SIZE;
1653 loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT;
1654 bool write = iter->flags & IOMAP_WRITE;
1655 unsigned long entry_flags = pmd ? DAX_PMD : 0;
1656 int err = 0;
1657 pfn_t pfn;
1658 void *kaddr;
1659
1660 if (!pmd && vmf->cow_page)
1661 return dax_fault_cow_page(vmf, iter);
1662
1663 /* if we are reading UNWRITTEN and HOLE, return a hole. */
1664 if (!write &&
1665 (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) {
1666 if (!pmd)
1667 return dax_load_hole(xas, vmf, iter, entry);
1668 return dax_pmd_load_hole(xas, vmf, iter, entry);
1669 }
1670
1671 if (iomap->type != IOMAP_MAPPED && !(iomap->flags & IOMAP_F_SHARED)) {
1672 WARN_ON_ONCE(1);
1673 return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS;
1674 }
1675
1676 err = dax_iomap_direct_access(iomap, pos, size, &kaddr, &pfn);
1677 if (err)
1678 return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err);
1679
1680 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, entry_flags);
1681
1682 if (write && iomap->flags & IOMAP_F_SHARED) {
1683 err = dax_iomap_copy_around(pos, size, size, srcmap, kaddr);
1684 if (err)
1685 return dax_fault_return(err);
1686 }
1687
1688 if (dax_fault_is_synchronous(iter, vmf->vma))
1689 return dax_fault_synchronous_pfnp(pfnp, pfn);
1690
1691 /* insert PMD pfn */
1692 if (pmd)
1693 return vmf_insert_pfn_pmd(vmf, pfn, write);
1694
1695 /* insert PTE pfn */
1696 if (write)
1697 return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1698 return vmf_insert_mixed(vmf->vma, vmf->address, pfn);
1699 }
1700
1701 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1702 int *iomap_errp, const struct iomap_ops *ops)
1703 {
1704 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1705 XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1706 struct iomap_iter iter = {
1707 .inode = mapping->host,
1708 .pos = (loff_t)vmf->pgoff << PAGE_SHIFT,
1709 .len = PAGE_SIZE,
1710 .flags = IOMAP_DAX | IOMAP_FAULT,
1711 };
1712 vm_fault_t ret = 0;
1713 void *entry;
1714 int error;
1715
1716 trace_dax_pte_fault(iter.inode, vmf, ret);
1717 /*
1718 * Check whether offset isn't beyond end of file now. Caller is supposed
1719 * to hold locks serializing us with truncate / punch hole so this is
1720 * a reliable test.
1721 */
1722 if (iter.pos >= i_size_read(iter.inode)) {
1723 ret = VM_FAULT_SIGBUS;
1724 goto out;
1725 }
1726
1727 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1728 iter.flags |= IOMAP_WRITE;
1729
1730 entry = grab_mapping_entry(&xas, mapping, 0);
1731 if (xa_is_internal(entry)) {
1732 ret = xa_to_internal(entry);
1733 goto out;
1734 }
1735
1736 /*
1737 * It is possible, particularly with mixed reads & writes to private
1738 * mappings, that we have raced with a PMD fault that overlaps with
1739 * the PTE we need to set up. If so just return and the fault will be
1740 * retried.
1741 */
1742 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1743 ret = VM_FAULT_NOPAGE;
1744 goto unlock_entry;
1745 }
1746
1747 while ((error = iomap_iter(&iter, ops)) > 0) {
1748 if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) {
1749 iter.processed = -EIO; /* fs corruption? */
1750 continue;
1751 }
1752
1753 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false);
1754 if (ret != VM_FAULT_SIGBUS &&
1755 (iter.iomap.flags & IOMAP_F_NEW)) {
1756 count_vm_event(PGMAJFAULT);
1757 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
1758 ret |= VM_FAULT_MAJOR;
1759 }
1760
1761 if (!(ret & VM_FAULT_ERROR))
1762 iter.processed = PAGE_SIZE;
1763 }
1764
1765 if (iomap_errp)
1766 *iomap_errp = error;
1767 if (!ret && error)
1768 ret = dax_fault_return(error);
1769
1770 unlock_entry:
1771 dax_unlock_entry(&xas, entry);
1772 out:
1773 trace_dax_pte_fault_done(iter.inode, vmf, ret);
1774 return ret;
1775 }
1776
1777 #ifdef CONFIG_FS_DAX_PMD
1778 static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas,
1779 pgoff_t max_pgoff)
1780 {
1781 unsigned long pmd_addr = vmf->address & PMD_MASK;
1782 bool write = vmf->flags & FAULT_FLAG_WRITE;
1783
1784 /*
1785 * Make sure that the faulting address's PMD offset (color) matches
1786 * the PMD offset from the start of the file. This is necessary so
1787 * that a PMD range in the page table overlaps exactly with a PMD
1788 * range in the page cache.
1789 */
1790 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1791 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1792 return true;
1793
1794 /* Fall back to PTEs if we're going to COW */
1795 if (write && !(vmf->vma->vm_flags & VM_SHARED))
1796 return true;
1797
1798 /* If the PMD would extend outside the VMA */
1799 if (pmd_addr < vmf->vma->vm_start)
1800 return true;
1801 if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
1802 return true;
1803
1804 /* If the PMD would extend beyond the file size */
1805 if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff)
1806 return true;
1807
1808 return false;
1809 }
1810
1811 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1812 const struct iomap_ops *ops)
1813 {
1814 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1815 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1816 struct iomap_iter iter = {
1817 .inode = mapping->host,
1818 .len = PMD_SIZE,
1819 .flags = IOMAP_DAX | IOMAP_FAULT,
1820 };
1821 vm_fault_t ret = VM_FAULT_FALLBACK;
1822 pgoff_t max_pgoff;
1823 void *entry;
1824
1825 if (vmf->flags & FAULT_FLAG_WRITE)
1826 iter.flags |= IOMAP_WRITE;
1827
1828 /*
1829 * Check whether offset isn't beyond end of file now. Caller is
1830 * supposed to hold locks serializing us with truncate / punch hole so
1831 * this is a reliable test.
1832 */
1833 max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE);
1834
1835 trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0);
1836
1837 if (xas.xa_index >= max_pgoff) {
1838 ret = VM_FAULT_SIGBUS;
1839 goto out;
1840 }
1841
1842 if (dax_fault_check_fallback(vmf, &xas, max_pgoff))
1843 goto fallback;
1844
1845 /*
1846 * grab_mapping_entry() will make sure we get an empty PMD entry,
1847 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1848 * entry is already in the array, for instance), it will return
1849 * VM_FAULT_FALLBACK.
1850 */
1851 entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1852 if (xa_is_internal(entry)) {
1853 ret = xa_to_internal(entry);
1854 goto fallback;
1855 }
1856
1857 /*
1858 * It is possible, particularly with mixed reads & writes to private
1859 * mappings, that we have raced with a PTE fault that overlaps with
1860 * the PMD we need to set up. If so just return and the fault will be
1861 * retried.
1862 */
1863 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1864 !pmd_devmap(*vmf->pmd)) {
1865 ret = 0;
1866 goto unlock_entry;
1867 }
1868
1869 iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1870 while (iomap_iter(&iter, ops) > 0) {
1871 if (iomap_length(&iter) < PMD_SIZE)
1872 continue; /* actually breaks out of the loop */
1873
1874 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true);
1875 if (ret != VM_FAULT_FALLBACK)
1876 iter.processed = PMD_SIZE;
1877 }
1878
1879 unlock_entry:
1880 dax_unlock_entry(&xas, entry);
1881 fallback:
1882 if (ret == VM_FAULT_FALLBACK) {
1883 split_huge_pmd(vmf->vma, vmf->pmd, vmf->address);
1884 count_vm_event(THP_FAULT_FALLBACK);
1885 }
1886 out:
1887 trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret);
1888 return ret;
1889 }
1890 #else
1891 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1892 const struct iomap_ops *ops)
1893 {
1894 return VM_FAULT_FALLBACK;
1895 }
1896 #endif /* CONFIG_FS_DAX_PMD */
1897
1898 /**
1899 * dax_iomap_fault - handle a page fault on a DAX file
1900 * @vmf: The description of the fault
1901 * @order: Order of the page to fault in
1902 * @pfnp: PFN to insert for synchronous faults if fsync is required
1903 * @iomap_errp: Storage for detailed error code in case of error
1904 * @ops: Iomap ops passed from the file system
1905 *
1906 * When a page fault occurs, filesystems may call this helper in
1907 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1908 * has done all the necessary locking for page fault to proceed
1909 * successfully.
1910 */
1911 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, unsigned int order,
1912 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1913 {
1914 if (order == 0)
1915 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1916 else if (order == PMD_ORDER)
1917 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1918 else
1919 return VM_FAULT_FALLBACK;
1920 }
1921 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1922
1923 /*
1924 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1925 * @vmf: The description of the fault
1926 * @pfn: PFN to insert
1927 * @order: Order of entry to insert.
1928 *
1929 * This function inserts a writeable PTE or PMD entry into the page tables
1930 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1931 */
1932 static vm_fault_t
1933 dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1934 {
1935 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1936 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1937 void *entry;
1938 vm_fault_t ret;
1939
1940 xas_lock_irq(&xas);
1941 entry = get_unlocked_entry(&xas, order);
1942 /* Did we race with someone splitting entry or so? */
1943 if (!entry || dax_is_conflict(entry) ||
1944 (order == 0 && !dax_is_pte_entry(entry))) {
1945 put_unlocked_entry(&xas, entry, WAKE_NEXT);
1946 xas_unlock_irq(&xas);
1947 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1948 VM_FAULT_NOPAGE);
1949 return VM_FAULT_NOPAGE;
1950 }
1951 xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1952 dax_lock_entry(&xas, entry);
1953 xas_unlock_irq(&xas);
1954 if (order == 0)
1955 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1956 #ifdef CONFIG_FS_DAX_PMD
1957 else if (order == PMD_ORDER)
1958 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1959 #endif
1960 else
1961 ret = VM_FAULT_FALLBACK;
1962 dax_unlock_entry(&xas, entry);
1963 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1964 return ret;
1965 }
1966
1967 /**
1968 * dax_finish_sync_fault - finish synchronous page fault
1969 * @vmf: The description of the fault
1970 * @order: Order of entry to be inserted
1971 * @pfn: PFN to insert
1972 *
1973 * This function ensures that the file range touched by the page fault is
1974 * stored persistently on the media and handles inserting of appropriate page
1975 * table entry.
1976 */
1977 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, unsigned int order,
1978 pfn_t pfn)
1979 {
1980 int err;
1981 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1982 size_t len = PAGE_SIZE << order;
1983
1984 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1985 if (err)
1986 return VM_FAULT_SIGBUS;
1987 return dax_insert_pfn_mkwrite(vmf, pfn, order);
1988 }
1989 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
1990
1991 static loff_t dax_range_compare_iter(struct iomap_iter *it_src,
1992 struct iomap_iter *it_dest, u64 len, bool *same)
1993 {
1994 const struct iomap *smap = &it_src->iomap;
1995 const struct iomap *dmap = &it_dest->iomap;
1996 loff_t pos1 = it_src->pos, pos2 = it_dest->pos;
1997 void *saddr, *daddr;
1998 int id, ret;
1999
2000 len = min(len, min(smap->length, dmap->length));
2001
2002 if (smap->type == IOMAP_HOLE && dmap->type == IOMAP_HOLE) {
2003 *same = true;
2004 return len;
2005 }
2006
2007 if (smap->type == IOMAP_HOLE || dmap->type == IOMAP_HOLE) {
2008 *same = false;
2009 return 0;
2010 }
2011
2012 id = dax_read_lock();
2013 ret = dax_iomap_direct_access(smap, pos1, ALIGN(pos1 + len, PAGE_SIZE),
2014 &saddr, NULL);
2015 if (ret < 0)
2016 goto out_unlock;
2017
2018 ret = dax_iomap_direct_access(dmap, pos2, ALIGN(pos2 + len, PAGE_SIZE),
2019 &daddr, NULL);
2020 if (ret < 0)
2021 goto out_unlock;
2022
2023 *same = !memcmp(saddr, daddr, len);
2024 if (!*same)
2025 len = 0;
2026 dax_read_unlock(id);
2027 return len;
2028
2029 out_unlock:
2030 dax_read_unlock(id);
2031 return -EIO;
2032 }
2033
2034 int dax_dedupe_file_range_compare(struct inode *src, loff_t srcoff,
2035 struct inode *dst, loff_t dstoff, loff_t len, bool *same,
2036 const struct iomap_ops *ops)
2037 {
2038 struct iomap_iter src_iter = {
2039 .inode = src,
2040 .pos = srcoff,
2041 .len = len,
2042 .flags = IOMAP_DAX,
2043 };
2044 struct iomap_iter dst_iter = {
2045 .inode = dst,
2046 .pos = dstoff,
2047 .len = len,
2048 .flags = IOMAP_DAX,
2049 };
2050 int ret, compared = 0;
2051
2052 while ((ret = iomap_iter(&src_iter, ops)) > 0 &&
2053 (ret = iomap_iter(&dst_iter, ops)) > 0) {
2054 compared = dax_range_compare_iter(&src_iter, &dst_iter,
2055 min(src_iter.len, dst_iter.len), same);
2056 if (compared < 0)
2057 return ret;
2058 src_iter.processed = dst_iter.processed = compared;
2059 }
2060 return ret;
2061 }
2062
2063 int dax_remap_file_range_prep(struct file *file_in, loff_t pos_in,
2064 struct file *file_out, loff_t pos_out,
2065 loff_t *len, unsigned int remap_flags,
2066 const struct iomap_ops *ops)
2067 {
2068 return __generic_remap_file_range_prep(file_in, pos_in, file_out,
2069 pos_out, len, remap_flags, ops);
2070 }
2071 EXPORT_SYMBOL_GPL(dax_remap_file_range_prep);
2072
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.