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