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