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/genhd.h>
15#include <linux/highmem.h>
16#include <linux/memcontrol.h>
17#include <linux/mm.h>
18#include <linux/mutex.h>
19#include <linux/pagevec.h>
20#include <linux/sched.h>
21#include <linux/sched/signal.h>
22#include <linux/uio.h>
23#include <linux/vmstat.h>
24#include <linux/pfn_t.h>
25#include <linux/sizes.h>
26#include <linux/mmu_notifier.h>
27#include <linux/iomap.h>
28#include <asm/pgalloc.h>
29
30#define CREATE_TRACE_POINTS
31#include <trace/events/fs_dax.h>
32
33static inline unsigned int pe_order(enum page_entry_size pe_size)
34{
35 if (pe_size == PE_SIZE_PTE)
36 return PAGE_SHIFT - PAGE_SHIFT;
37 if (pe_size == PE_SIZE_PMD)
38 return PMD_SHIFT - PAGE_SHIFT;
39 if (pe_size == PE_SIZE_PUD)
40 return PUD_SHIFT - PAGE_SHIFT;
41 return ~0;
42}
43
44/* We choose 4096 entries - same as per-zone page wait tables */
45#define DAX_WAIT_TABLE_BITS 12
46#define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
47
48/* The 'colour' (ie low bits) within a PMD of a page offset. */
49#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
50#define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
51
52/* The order of a PMD entry */
53#define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT)
54
55static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
56
57static int __init init_dax_wait_table(void)
58{
59 int i;
60
61 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
62 init_waitqueue_head(wait_table + i);
63 return 0;
64}
65fs_initcall(init_dax_wait_table);
66
67/*
68 * DAX pagecache entries use XArray value entries so they can't be mistaken
69 * for pages. We use one bit for locking, one bit for the entry size (PMD)
70 * and two more to tell us if the entry is a zero page or an empty entry that
71 * is just used for locking. In total four special bits.
72 *
73 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
74 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
75 * block allocation.
76 */
77#define DAX_SHIFT (4)
78#define DAX_LOCKED (1UL << 0)
79#define DAX_PMD (1UL << 1)
80#define DAX_ZERO_PAGE (1UL << 2)
81#define DAX_EMPTY (1UL << 3)
82
83static unsigned long dax_to_pfn(void *entry)
84{
85 return xa_to_value(entry) >> DAX_SHIFT;
86}
87
88static void *dax_make_entry(pfn_t pfn, unsigned long flags)
89{
90 return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
91}
92
93static bool dax_is_locked(void *entry)
94{
95 return xa_to_value(entry) & DAX_LOCKED;
96}
97
98static unsigned int dax_entry_order(void *entry)
99{
100 if (xa_to_value(entry) & DAX_PMD)
101 return PMD_ORDER;
102 return 0;
103}
104
105static unsigned long dax_is_pmd_entry(void *entry)
106{
107 return xa_to_value(entry) & DAX_PMD;
108}
109
110static bool dax_is_pte_entry(void *entry)
111{
112 return !(xa_to_value(entry) & DAX_PMD);
113}
114
115static int dax_is_zero_entry(void *entry)
116{
117 return xa_to_value(entry) & DAX_ZERO_PAGE;
118}
119
120static int dax_is_empty_entry(void *entry)
121{
122 return xa_to_value(entry) & DAX_EMPTY;
123}
124
125/*
126 * true if the entry that was found is of a smaller order than the entry
127 * we were looking for
128 */
129static bool dax_is_conflict(void *entry)
130{
131 return entry == XA_RETRY_ENTRY;
132}
133
134/*
135 * DAX page cache entry locking
136 */
137struct exceptional_entry_key {
138 struct xarray *xa;
139 pgoff_t entry_start;
140};
141
142struct wait_exceptional_entry_queue {
143 wait_queue_entry_t wait;
144 struct exceptional_entry_key key;
145};
146
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
337/*
338 * TODO: for reflink+dax we need a way to associate a single page with
339 * multiple address_space instances at different linear_page_index()
340 * offsets.
341 */
342static void dax_associate_entry(void *entry, struct address_space *mapping,
343 struct vm_area_struct *vma, unsigned long address)
344{
345 unsigned long size = dax_entry_size(entry), pfn, index;
346 int i = 0;
347
348 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
349 return;
350
351 index = linear_page_index(vma, address & ~(size - 1));
352 for_each_mapped_pfn(entry, pfn) {
353 struct page *page = pfn_to_page(pfn);
354
355 WARN_ON_ONCE(page->mapping);
356 page->mapping = mapping;
357 page->index = index + i++;
358 }
359}
360
361static void dax_disassociate_entry(void *entry, struct address_space *mapping,
362 bool trunc)
363{
364 unsigned long pfn;
365
366 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
367 return;
368
369 for_each_mapped_pfn(entry, pfn) {
370 struct page *page = pfn_to_page(pfn);
371
372 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
373 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
374 page->mapping = NULL;
375 page->index = 0;
376 }
377}
378
379static struct page *dax_busy_page(void *entry)
380{
381 unsigned long pfn;
382
383 for_each_mapped_pfn(entry, pfn) {
384 struct page *page = pfn_to_page(pfn);
385
386 if (page_ref_count(page) > 1)
387 return page;
388 }
389 return NULL;
390}
391
392/*
393 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page
394 * @page: The page whose entry we want to lock
395 *
396 * Context: Process context.
397 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
398 * not be locked.
399 */
400dax_entry_t dax_lock_page(struct page *page)
401{
402 XA_STATE(xas, NULL, 0);
403 void *entry;
404
405 /* Ensure page->mapping isn't freed while we look at it */
406 rcu_read_lock();
407 for (;;) {
408 struct address_space *mapping = READ_ONCE(page->mapping);
409
410 entry = NULL;
411 if (!mapping || !dax_mapping(mapping))
412 break;
413
414 /*
415 * In the device-dax case there's no need to lock, a
416 * struct dev_pagemap pin is sufficient to keep the
417 * inode alive, and we assume we have dev_pagemap pin
418 * otherwise we would not have a valid pfn_to_page()
419 * translation.
420 */
421 entry = (void *)~0UL;
422 if (S_ISCHR(mapping->host->i_mode))
423 break;
424
425 xas.xa = &mapping->i_pages;
426 xas_lock_irq(&xas);
427 if (mapping != page->mapping) {
428 xas_unlock_irq(&xas);
429 continue;
430 }
431 xas_set(&xas, page->index);
432 entry = xas_load(&xas);
433 if (dax_is_locked(entry)) {
434 rcu_read_unlock();
435 wait_entry_unlocked(&xas, entry);
436 rcu_read_lock();
437 continue;
438 }
439 dax_lock_entry(&xas, entry);
440 xas_unlock_irq(&xas);
441 break;
442 }
443 rcu_read_unlock();
444 return (dax_entry_t)entry;
445}
446
447void dax_unlock_page(struct page *page, dax_entry_t cookie)
448{
449 struct address_space *mapping = page->mapping;
450 XA_STATE(xas, &mapping->i_pages, page->index);
451
452 if (S_ISCHR(mapping->host->i_mode))
453 return;
454
455 dax_unlock_entry(&xas, (void *)cookie);
456}
457
458/*
459 * Find page cache entry at given index. If it is a DAX entry, return it
460 * with the entry locked. If the page cache doesn't contain an entry at
461 * that index, add a locked empty entry.
462 *
463 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
464 * either return that locked entry or will return VM_FAULT_FALLBACK.
465 * This will happen if there are any PTE entries within the PMD range
466 * that we are requesting.
467 *
468 * We always favor PTE entries over PMD entries. There isn't a flow where we
469 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
470 * insertion will fail if it finds any PTE entries already in the tree, and a
471 * PTE insertion will cause an existing PMD entry to be unmapped and
472 * downgraded to PTE entries. This happens for both PMD zero pages as
473 * well as PMD empty entries.
474 *
475 * The exception to this downgrade path is for PMD entries that have
476 * real storage backing them. We will leave these real PMD entries in
477 * the tree, and PTE writes will simply dirty the entire PMD entry.
478 *
479 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
480 * persistent memory the benefit is doubtful. We can add that later if we can
481 * show it helps.
482 *
483 * On error, this function does not return an ERR_PTR. Instead it returns
484 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
485 * overlap with xarray value entries.
486 */
487static void *grab_mapping_entry(struct xa_state *xas,
488 struct address_space *mapping, unsigned int order)
489{
490 unsigned long index = xas->xa_index;
491 bool pmd_downgrade; /* splitting PMD entry into PTE entries? */
492 void *entry;
493
494retry:
495 pmd_downgrade = false;
496 xas_lock_irq(xas);
497 entry = get_unlocked_entry(xas, order);
498
499 if (entry) {
500 if (dax_is_conflict(entry))
501 goto fallback;
502 if (!xa_is_value(entry)) {
503 xas_set_err(xas, -EIO);
504 goto out_unlock;
505 }
506
507 if (order == 0) {
508 if (dax_is_pmd_entry(entry) &&
509 (dax_is_zero_entry(entry) ||
510 dax_is_empty_entry(entry))) {
511 pmd_downgrade = true;
512 }
513 }
514 }
515
516 if (pmd_downgrade) {
517 /*
518 * Make sure 'entry' remains valid while we drop
519 * the i_pages lock.
520 */
521 dax_lock_entry(xas, entry);
522
523 /*
524 * Besides huge zero pages the only other thing that gets
525 * downgraded are empty entries which don't need to be
526 * unmapped.
527 */
528 if (dax_is_zero_entry(entry)) {
529 xas_unlock_irq(xas);
530 unmap_mapping_pages(mapping,
531 xas->xa_index & ~PG_PMD_COLOUR,
532 PG_PMD_NR, false);
533 xas_reset(xas);
534 xas_lock_irq(xas);
535 }
536
537 dax_disassociate_entry(entry, mapping, false);
538 xas_store(xas, NULL); /* undo the PMD join */
539 dax_wake_entry(xas, entry, WAKE_ALL);
540 mapping->nrpages -= PG_PMD_NR;
541 entry = NULL;
542 xas_set(xas, index);
543 }
544
545 if (entry) {
546 dax_lock_entry(xas, entry);
547 } else {
548 unsigned long flags = DAX_EMPTY;
549
550 if (order > 0)
551 flags |= DAX_PMD;
552 entry = dax_make_entry(pfn_to_pfn_t(0), flags);
553 dax_lock_entry(xas, entry);
554 if (xas_error(xas))
555 goto out_unlock;
556 mapping->nrpages += 1UL << order;
557 }
558
559out_unlock:
560 xas_unlock_irq(xas);
561 if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
562 goto retry;
563 if (xas->xa_node == XA_ERROR(-ENOMEM))
564 return xa_mk_internal(VM_FAULT_OOM);
565 if (xas_error(xas))
566 return xa_mk_internal(VM_FAULT_SIGBUS);
567 return entry;
568fallback:
569 xas_unlock_irq(xas);
570 return xa_mk_internal(VM_FAULT_FALLBACK);
571}
572
573/**
574 * dax_layout_busy_page_range - find first pinned page in @mapping
575 * @mapping: address space to scan for a page with ref count > 1
576 * @start: Starting offset. Page containing 'start' is included.
577 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
578 * pages from 'start' till the end of file are included.
579 *
580 * DAX requires ZONE_DEVICE mapped pages. These pages are never
581 * 'onlined' to the page allocator so they are considered idle when
582 * page->count == 1. A filesystem uses this interface to determine if
583 * any page in the mapping is busy, i.e. for DMA, or other
584 * get_user_pages() usages.
585 *
586 * It is expected that the filesystem is holding locks to block the
587 * establishment of new mappings in this address_space. I.e. it expects
588 * to be able to run unmap_mapping_range() and subsequently not race
589 * mapping_mapped() becoming true.
590 */
591struct page *dax_layout_busy_page_range(struct address_space *mapping,
592 loff_t start, loff_t end)
593{
594 void *entry;
595 unsigned int scanned = 0;
596 struct page *page = NULL;
597 pgoff_t start_idx = start >> PAGE_SHIFT;
598 pgoff_t end_idx;
599 XA_STATE(xas, &mapping->i_pages, start_idx);
600
601 /*
602 * In the 'limited' case get_user_pages() for dax is disabled.
603 */
604 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
605 return NULL;
606
607 if (!dax_mapping(mapping) || !mapping_mapped(mapping))
608 return NULL;
609
610 /* If end == LLONG_MAX, all pages from start to till end of file */
611 if (end == LLONG_MAX)
612 end_idx = ULONG_MAX;
613 else
614 end_idx = end >> PAGE_SHIFT;
615 /*
616 * If we race get_user_pages_fast() here either we'll see the
617 * elevated page count in the iteration and wait, or
618 * get_user_pages_fast() will see that the page it took a reference
619 * against is no longer mapped in the page tables and bail to the
620 * get_user_pages() slow path. The slow path is protected by
621 * pte_lock() and pmd_lock(). New references are not taken without
622 * holding those locks, and unmap_mapping_pages() will not zero the
623 * pte or pmd without holding the respective lock, so we are
624 * guaranteed to either see new references or prevent new
625 * references from being established.
626 */
627 unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0);
628
629 xas_lock_irq(&xas);
630 xas_for_each(&xas, entry, end_idx) {
631 if (WARN_ON_ONCE(!xa_is_value(entry)))
632 continue;
633 if (unlikely(dax_is_locked(entry)))
634 entry = get_unlocked_entry(&xas, 0);
635 if (entry)
636 page = dax_busy_page(entry);
637 put_unlocked_entry(&xas, entry, WAKE_NEXT);
638 if (page)
639 break;
640 if (++scanned % XA_CHECK_SCHED)
641 continue;
642
643 xas_pause(&xas);
644 xas_unlock_irq(&xas);
645 cond_resched();
646 xas_lock_irq(&xas);
647 }
648 xas_unlock_irq(&xas);
649 return page;
650}
651EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
652
653struct page *dax_layout_busy_page(struct address_space *mapping)
654{
655 return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
656}
657EXPORT_SYMBOL_GPL(dax_layout_busy_page);
658
659static int __dax_invalidate_entry(struct address_space *mapping,
660 pgoff_t index, bool trunc)
661{
662 XA_STATE(xas, &mapping->i_pages, index);
663 int ret = 0;
664 void *entry;
665
666 xas_lock_irq(&xas);
667 entry = get_unlocked_entry(&xas, 0);
668 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
669 goto out;
670 if (!trunc &&
671 (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
672 xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
673 goto out;
674 dax_disassociate_entry(entry, mapping, trunc);
675 xas_store(&xas, NULL);
676 mapping->nrpages -= 1UL << dax_entry_order(entry);
677 ret = 1;
678out:
679 put_unlocked_entry(&xas, entry, WAKE_ALL);
680 xas_unlock_irq(&xas);
681 return ret;
682}
683
684/*
685 * Delete DAX entry at @index from @mapping. Wait for it
686 * to be unlocked before deleting it.
687 */
688int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
689{
690 int ret = __dax_invalidate_entry(mapping, index, true);
691
692 /*
693 * This gets called from truncate / punch_hole path. As such, the caller
694 * must hold locks protecting against concurrent modifications of the
695 * page cache (usually fs-private i_mmap_sem for writing). Since the
696 * caller has seen a DAX entry for this index, we better find it
697 * at that index as well...
698 */
699 WARN_ON_ONCE(!ret);
700 return ret;
701}
702
703/*
704 * Invalidate DAX entry if it is clean.
705 */
706int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
707 pgoff_t index)
708{
709 return __dax_invalidate_entry(mapping, index, false);
710}
711
712static int copy_cow_page_dax(struct block_device *bdev, struct dax_device *dax_dev,
713 sector_t sector, struct page *to, unsigned long vaddr)
714{
715 void *vto, *kaddr;
716 pgoff_t pgoff;
717 long rc;
718 int id;
719
720 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
721 if (rc)
722 return rc;
723
724 id = dax_read_lock();
725 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, NULL);
726 if (rc < 0) {
727 dax_read_unlock(id);
728 return rc;
729 }
730 vto = kmap_atomic(to);
731 copy_user_page(vto, (void __force *)kaddr, vaddr, to);
732 kunmap_atomic(vto);
733 dax_read_unlock(id);
734 return 0;
735}
736
737/*
738 * By this point grab_mapping_entry() has ensured that we have a locked entry
739 * of the appropriate size so we don't have to worry about downgrading PMDs to
740 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
741 * already in the tree, we will skip the insertion and just dirty the PMD as
742 * appropriate.
743 */
744static void *dax_insert_entry(struct xa_state *xas,
745 struct address_space *mapping, struct vm_fault *vmf,
746 void *entry, pfn_t pfn, unsigned long flags, bool dirty)
747{
748 void *new_entry = dax_make_entry(pfn, flags);
749
750 if (dirty)
751 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
752
753 if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) {
754 unsigned long index = xas->xa_index;
755 /* we are replacing a zero page with block mapping */
756 if (dax_is_pmd_entry(entry))
757 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
758 PG_PMD_NR, false);
759 else /* pte entry */
760 unmap_mapping_pages(mapping, index, 1, false);
761 }
762
763 xas_reset(xas);
764 xas_lock_irq(xas);
765 if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
766 void *old;
767
768 dax_disassociate_entry(entry, mapping, false);
769 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address);
770 /*
771 * Only swap our new entry into the page cache if the current
772 * entry is a zero page or an empty entry. If a normal PTE or
773 * PMD entry is already in the cache, we leave it alone. This
774 * means that if we are trying to insert a PTE and the
775 * existing entry is a PMD, we will just leave the PMD in the
776 * tree and dirty it if necessary.
777 */
778 old = dax_lock_entry(xas, new_entry);
779 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
780 DAX_LOCKED));
781 entry = new_entry;
782 } else {
783 xas_load(xas); /* Walk the xa_state */
784 }
785
786 if (dirty)
787 xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
788
789 xas_unlock_irq(xas);
790 return entry;
791}
792
793static inline
794unsigned long pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
795{
796 unsigned long address;
797
798 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
799 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
800 return address;
801}
802
803/* Walk all mappings of a given index of a file and writeprotect them */
804static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
805 unsigned long pfn)
806{
807 struct vm_area_struct *vma;
808 pte_t pte, *ptep = NULL;
809 pmd_t *pmdp = NULL;
810 spinlock_t *ptl;
811
812 i_mmap_lock_read(mapping);
813 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
814 struct mmu_notifier_range range;
815 unsigned long address;
816
817 cond_resched();
818
819 if (!(vma->vm_flags & VM_SHARED))
820 continue;
821
822 address = pgoff_address(index, vma);
823
824 /*
825 * follow_invalidate_pte() will use the range to call
826 * mmu_notifier_invalidate_range_start() on our behalf before
827 * taking any lock.
828 */
829 if (follow_invalidate_pte(vma->vm_mm, address, &range, &ptep,
830 &pmdp, &ptl))
831 continue;
832
833 /*
834 * No need to call mmu_notifier_invalidate_range() as we are
835 * downgrading page table protection not changing it to point
836 * to a new page.
837 *
838 * See Documentation/vm/mmu_notifier.rst
839 */
840 if (pmdp) {
841#ifdef CONFIG_FS_DAX_PMD
842 pmd_t pmd;
843
844 if (pfn != pmd_pfn(*pmdp))
845 goto unlock_pmd;
846 if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
847 goto unlock_pmd;
848
849 flush_cache_page(vma, address, pfn);
850 pmd = pmdp_invalidate(vma, address, pmdp);
851 pmd = pmd_wrprotect(pmd);
852 pmd = pmd_mkclean(pmd);
853 set_pmd_at(vma->vm_mm, address, pmdp, pmd);
854unlock_pmd:
855#endif
856 spin_unlock(ptl);
857 } else {
858 if (pfn != pte_pfn(*ptep))
859 goto unlock_pte;
860 if (!pte_dirty(*ptep) && !pte_write(*ptep))
861 goto unlock_pte;
862
863 flush_cache_page(vma, address, pfn);
864 pte = ptep_clear_flush(vma, address, ptep);
865 pte = pte_wrprotect(pte);
866 pte = pte_mkclean(pte);
867 set_pte_at(vma->vm_mm, address, ptep, pte);
868unlock_pte:
869 pte_unmap_unlock(ptep, ptl);
870 }
871
872 mmu_notifier_invalidate_range_end(&range);
873 }
874 i_mmap_unlock_read(mapping);
875}
876
877static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
878 struct address_space *mapping, void *entry)
879{
880 unsigned long pfn, index, count;
881 long ret = 0;
882
883 /*
884 * A page got tagged dirty in DAX mapping? Something is seriously
885 * wrong.
886 */
887 if (WARN_ON(!xa_is_value(entry)))
888 return -EIO;
889
890 if (unlikely(dax_is_locked(entry))) {
891 void *old_entry = entry;
892
893 entry = get_unlocked_entry(xas, 0);
894
895 /* Entry got punched out / reallocated? */
896 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
897 goto put_unlocked;
898 /*
899 * Entry got reallocated elsewhere? No need to writeback.
900 * We have to compare pfns as we must not bail out due to
901 * difference in lockbit or entry type.
902 */
903 if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
904 goto put_unlocked;
905 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
906 dax_is_zero_entry(entry))) {
907 ret = -EIO;
908 goto put_unlocked;
909 }
910
911 /* Another fsync thread may have already done this entry */
912 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
913 goto put_unlocked;
914 }
915
916 /* Lock the entry to serialize with page faults */
917 dax_lock_entry(xas, entry);
918
919 /*
920 * We can clear the tag now but we have to be careful so that concurrent
921 * dax_writeback_one() calls for the same index cannot finish before we
922 * actually flush the caches. This is achieved as the calls will look
923 * at the entry only under the i_pages lock and once they do that
924 * they will see the entry locked and wait for it to unlock.
925 */
926 xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
927 xas_unlock_irq(xas);
928
929 /*
930 * If dax_writeback_mapping_range() was given a wbc->range_start
931 * in the middle of a PMD, the 'index' we use needs to be
932 * aligned to the start of the PMD.
933 * This allows us to flush for PMD_SIZE and not have to worry about
934 * partial PMD writebacks.
935 */
936 pfn = dax_to_pfn(entry);
937 count = 1UL << dax_entry_order(entry);
938 index = xas->xa_index & ~(count - 1);
939
940 dax_entry_mkclean(mapping, index, pfn);
941 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
942 /*
943 * After we have flushed the cache, we can clear the dirty tag. There
944 * cannot be new dirty data in the pfn after the flush has completed as
945 * the pfn mappings are writeprotected and fault waits for mapping
946 * entry lock.
947 */
948 xas_reset(xas);
949 xas_lock_irq(xas);
950 xas_store(xas, entry);
951 xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
952 dax_wake_entry(xas, entry, WAKE_NEXT);
953
954 trace_dax_writeback_one(mapping->host, index, count);
955 return ret;
956
957 put_unlocked:
958 put_unlocked_entry(xas, entry, WAKE_NEXT);
959 return ret;
960}
961
962/*
963 * Flush the mapping to the persistent domain within the byte range of [start,
964 * end]. This is required by data integrity operations to ensure file data is
965 * on persistent storage prior to completion of the operation.
966 */
967int dax_writeback_mapping_range(struct address_space *mapping,
968 struct dax_device *dax_dev, struct writeback_control *wbc)
969{
970 XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
971 struct inode *inode = mapping->host;
972 pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
973 void *entry;
974 int ret = 0;
975 unsigned int scanned = 0;
976
977 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
978 return -EIO;
979
980 if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL)
981 return 0;
982
983 trace_dax_writeback_range(inode, xas.xa_index, end_index);
984
985 tag_pages_for_writeback(mapping, xas.xa_index, end_index);
986
987 xas_lock_irq(&xas);
988 xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
989 ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
990 if (ret < 0) {
991 mapping_set_error(mapping, ret);
992 break;
993 }
994 if (++scanned % XA_CHECK_SCHED)
995 continue;
996
997 xas_pause(&xas);
998 xas_unlock_irq(&xas);
999 cond_resched();
1000 xas_lock_irq(&xas);
1001 }
1002 xas_unlock_irq(&xas);
1003 trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
1004 return ret;
1005}
1006EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1007
1008static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
1009{
1010 return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
1011}
1012
1013static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
1014 pfn_t *pfnp)
1015{
1016 const sector_t sector = dax_iomap_sector(iomap, pos);
1017 pgoff_t pgoff;
1018 int id, rc;
1019 long length;
1020
1021 rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
1022 if (rc)
1023 return rc;
1024 id = dax_read_lock();
1025 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1026 NULL, pfnp);
1027 if (length < 0) {
1028 rc = length;
1029 goto out;
1030 }
1031 rc = -EINVAL;
1032 if (PFN_PHYS(length) < size)
1033 goto out;
1034 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1035 goto out;
1036 /* For larger pages we need devmap */
1037 if (length > 1 && !pfn_t_devmap(*pfnp))
1038 goto out;
1039 rc = 0;
1040out:
1041 dax_read_unlock(id);
1042 return rc;
1043}
1044
1045/*
1046 * The user has performed a load from a hole in the file. Allocating a new
1047 * page in the file would cause excessive storage usage for workloads with
1048 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1049 * If this page is ever written to we will re-fault and change the mapping to
1050 * point to real DAX storage instead.
1051 */
1052static vm_fault_t dax_load_hole(struct xa_state *xas,
1053 struct address_space *mapping, void **entry,
1054 struct vm_fault *vmf)
1055{
1056 struct inode *inode = mapping->host;
1057 unsigned long vaddr = vmf->address;
1058 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1059 vm_fault_t ret;
1060
1061 *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1062 DAX_ZERO_PAGE, false);
1063
1064 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1065 trace_dax_load_hole(inode, vmf, ret);
1066 return ret;
1067}
1068
1069s64 dax_iomap_zero(loff_t pos, u64 length, struct iomap *iomap)
1070{
1071 sector_t sector = iomap_sector(iomap, pos & PAGE_MASK);
1072 pgoff_t pgoff;
1073 long rc, id;
1074 void *kaddr;
1075 bool page_aligned = false;
1076 unsigned offset = offset_in_page(pos);
1077 unsigned size = min_t(u64, PAGE_SIZE - offset, length);
1078
1079 if (IS_ALIGNED(sector << SECTOR_SHIFT, PAGE_SIZE) &&
1080 (size == PAGE_SIZE))
1081 page_aligned = true;
1082
1083 rc = bdev_dax_pgoff(iomap->bdev, sector, PAGE_SIZE, &pgoff);
1084 if (rc)
1085 return rc;
1086
1087 id = dax_read_lock();
1088
1089 if (page_aligned)
1090 rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1091 else
1092 rc = dax_direct_access(iomap->dax_dev, pgoff, 1, &kaddr, NULL);
1093 if (rc < 0) {
1094 dax_read_unlock(id);
1095 return rc;
1096 }
1097
1098 if (!page_aligned) {
1099 memset(kaddr + offset, 0, size);
1100 dax_flush(iomap->dax_dev, kaddr + offset, size);
1101 }
1102 dax_read_unlock(id);
1103 return size;
1104}
1105
1106static loff_t
1107dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1108 struct iomap *iomap, struct iomap *srcmap)
1109{
1110 struct block_device *bdev = iomap->bdev;
1111 struct dax_device *dax_dev = iomap->dax_dev;
1112 struct iov_iter *iter = data;
1113 loff_t end = pos + length, done = 0;
1114 ssize_t ret = 0;
1115 size_t xfer;
1116 int id;
1117
1118 if (iov_iter_rw(iter) == READ) {
1119 end = min(end, i_size_read(inode));
1120 if (pos >= end)
1121 return 0;
1122
1123 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1124 return iov_iter_zero(min(length, end - pos), iter);
1125 }
1126
1127 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1128 return -EIO;
1129
1130 /*
1131 * Write can allocate block for an area which has a hole page mapped
1132 * into page tables. We have to tear down these mappings so that data
1133 * written by write(2) is visible in mmap.
1134 */
1135 if (iomap->flags & IOMAP_F_NEW) {
1136 invalidate_inode_pages2_range(inode->i_mapping,
1137 pos >> PAGE_SHIFT,
1138 (end - 1) >> PAGE_SHIFT);
1139 }
1140
1141 id = dax_read_lock();
1142 while (pos < end) {
1143 unsigned offset = pos & (PAGE_SIZE - 1);
1144 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1145 const sector_t sector = dax_iomap_sector(iomap, pos);
1146 ssize_t map_len;
1147 pgoff_t pgoff;
1148 void *kaddr;
1149
1150 if (fatal_signal_pending(current)) {
1151 ret = -EINTR;
1152 break;
1153 }
1154
1155 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1156 if (ret)
1157 break;
1158
1159 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1160 &kaddr, NULL);
1161 if (map_len < 0) {
1162 ret = map_len;
1163 break;
1164 }
1165
1166 map_len = PFN_PHYS(map_len);
1167 kaddr += offset;
1168 map_len -= offset;
1169 if (map_len > end - pos)
1170 map_len = end - pos;
1171
1172 /*
1173 * The userspace address for the memory copy has already been
1174 * validated via access_ok() in either vfs_read() or
1175 * vfs_write(), depending on which operation we are doing.
1176 */
1177 if (iov_iter_rw(iter) == WRITE)
1178 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1179 map_len, iter);
1180 else
1181 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1182 map_len, iter);
1183
1184 pos += xfer;
1185 length -= xfer;
1186 done += xfer;
1187
1188 if (xfer == 0)
1189 ret = -EFAULT;
1190 if (xfer < map_len)
1191 break;
1192 }
1193 dax_read_unlock(id);
1194
1195 return done ? done : ret;
1196}
1197
1198/**
1199 * dax_iomap_rw - Perform I/O to a DAX file
1200 * @iocb: The control block for this I/O
1201 * @iter: The addresses to do I/O from or to
1202 * @ops: iomap ops passed from the file system
1203 *
1204 * This function performs read and write operations to directly mapped
1205 * persistent memory. The callers needs to take care of read/write exclusion
1206 * and evicting any page cache pages in the region under I/O.
1207 */
1208ssize_t
1209dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1210 const struct iomap_ops *ops)
1211{
1212 struct address_space *mapping = iocb->ki_filp->f_mapping;
1213 struct inode *inode = mapping->host;
1214 loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1215 unsigned flags = 0;
1216
1217 if (iov_iter_rw(iter) == WRITE) {
1218 lockdep_assert_held_write(&inode->i_rwsem);
1219 flags |= IOMAP_WRITE;
1220 } else {
1221 lockdep_assert_held(&inode->i_rwsem);
1222 }
1223
1224 if (iocb->ki_flags & IOCB_NOWAIT)
1225 flags |= IOMAP_NOWAIT;
1226
1227 while (iov_iter_count(iter)) {
1228 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1229 iter, dax_iomap_actor);
1230 if (ret <= 0)
1231 break;
1232 pos += ret;
1233 done += ret;
1234 }
1235
1236 iocb->ki_pos += done;
1237 return done ? done : ret;
1238}
1239EXPORT_SYMBOL_GPL(dax_iomap_rw);
1240
1241static vm_fault_t dax_fault_return(int error)
1242{
1243 if (error == 0)
1244 return VM_FAULT_NOPAGE;
1245 return vmf_error(error);
1246}
1247
1248/*
1249 * MAP_SYNC on a dax mapping guarantees dirty metadata is
1250 * flushed on write-faults (non-cow), but not read-faults.
1251 */
1252static bool dax_fault_is_synchronous(unsigned long flags,
1253 struct vm_area_struct *vma, struct iomap *iomap)
1254{
1255 return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1256 && (iomap->flags & IOMAP_F_DIRTY);
1257}
1258
1259static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1260 int *iomap_errp, const struct iomap_ops *ops)
1261{
1262 struct vm_area_struct *vma = vmf->vma;
1263 struct address_space *mapping = vma->vm_file->f_mapping;
1264 XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1265 struct inode *inode = mapping->host;
1266 unsigned long vaddr = vmf->address;
1267 loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1268 struct iomap iomap = { .type = IOMAP_HOLE };
1269 struct iomap srcmap = { .type = IOMAP_HOLE };
1270 unsigned flags = IOMAP_FAULT;
1271 int error, major = 0;
1272 bool write = vmf->flags & FAULT_FLAG_WRITE;
1273 bool sync;
1274 vm_fault_t ret = 0;
1275 void *entry;
1276 pfn_t pfn;
1277
1278 trace_dax_pte_fault(inode, vmf, ret);
1279 /*
1280 * Check whether offset isn't beyond end of file now. Caller is supposed
1281 * to hold locks serializing us with truncate / punch hole so this is
1282 * a reliable test.
1283 */
1284 if (pos >= i_size_read(inode)) {
1285 ret = VM_FAULT_SIGBUS;
1286 goto out;
1287 }
1288
1289 if (write && !vmf->cow_page)
1290 flags |= IOMAP_WRITE;
1291
1292 entry = grab_mapping_entry(&xas, mapping, 0);
1293 if (xa_is_internal(entry)) {
1294 ret = xa_to_internal(entry);
1295 goto out;
1296 }
1297
1298 /*
1299 * It is possible, particularly with mixed reads & writes to private
1300 * mappings, that we have raced with a PMD fault that overlaps with
1301 * the PTE we need to set up. If so just return and the fault will be
1302 * retried.
1303 */
1304 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1305 ret = VM_FAULT_NOPAGE;
1306 goto unlock_entry;
1307 }
1308
1309 /*
1310 * Note that we don't bother to use iomap_apply here: DAX required
1311 * the file system block size to be equal the page size, which means
1312 * that we never have to deal with more than a single extent here.
1313 */
1314 error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap, &srcmap);
1315 if (iomap_errp)
1316 *iomap_errp = error;
1317 if (error) {
1318 ret = dax_fault_return(error);
1319 goto unlock_entry;
1320 }
1321 if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1322 error = -EIO; /* fs corruption? */
1323 goto error_finish_iomap;
1324 }
1325
1326 if (vmf->cow_page) {
1327 sector_t sector = dax_iomap_sector(&iomap, pos);
1328
1329 switch (iomap.type) {
1330 case IOMAP_HOLE:
1331 case IOMAP_UNWRITTEN:
1332 clear_user_highpage(vmf->cow_page, vaddr);
1333 break;
1334 case IOMAP_MAPPED:
1335 error = copy_cow_page_dax(iomap.bdev, iomap.dax_dev,
1336 sector, vmf->cow_page, vaddr);
1337 break;
1338 default:
1339 WARN_ON_ONCE(1);
1340 error = -EIO;
1341 break;
1342 }
1343
1344 if (error)
1345 goto error_finish_iomap;
1346
1347 __SetPageUptodate(vmf->cow_page);
1348 ret = finish_fault(vmf);
1349 if (!ret)
1350 ret = VM_FAULT_DONE_COW;
1351 goto finish_iomap;
1352 }
1353
1354 sync = dax_fault_is_synchronous(flags, vma, &iomap);
1355
1356 switch (iomap.type) {
1357 case IOMAP_MAPPED:
1358 if (iomap.flags & IOMAP_F_NEW) {
1359 count_vm_event(PGMAJFAULT);
1360 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1361 major = VM_FAULT_MAJOR;
1362 }
1363 error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1364 if (error < 0)
1365 goto error_finish_iomap;
1366
1367 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1368 0, write && !sync);
1369
1370 /*
1371 * If we are doing synchronous page fault and inode needs fsync,
1372 * we can insert PTE into page tables only after that happens.
1373 * Skip insertion for now and return the pfn so that caller can
1374 * insert it after fsync is done.
1375 */
1376 if (sync) {
1377 if (WARN_ON_ONCE(!pfnp)) {
1378 error = -EIO;
1379 goto error_finish_iomap;
1380 }
1381 *pfnp = pfn;
1382 ret = VM_FAULT_NEEDDSYNC | major;
1383 goto finish_iomap;
1384 }
1385 trace_dax_insert_mapping(inode, vmf, entry);
1386 if (write)
1387 ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn);
1388 else
1389 ret = vmf_insert_mixed(vma, vaddr, pfn);
1390
1391 goto finish_iomap;
1392 case IOMAP_UNWRITTEN:
1393 case IOMAP_HOLE:
1394 if (!write) {
1395 ret = dax_load_hole(&xas, mapping, &entry, vmf);
1396 goto finish_iomap;
1397 }
1398 fallthrough;
1399 default:
1400 WARN_ON_ONCE(1);
1401 error = -EIO;
1402 break;
1403 }
1404
1405 error_finish_iomap:
1406 ret = dax_fault_return(error);
1407 finish_iomap:
1408 if (ops->iomap_end) {
1409 int copied = PAGE_SIZE;
1410
1411 if (ret & VM_FAULT_ERROR)
1412 copied = 0;
1413 /*
1414 * The fault is done by now and there's no way back (other
1415 * thread may be already happily using PTE we have installed).
1416 * Just ignore error from ->iomap_end since we cannot do much
1417 * with it.
1418 */
1419 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1420 }
1421 unlock_entry:
1422 dax_unlock_entry(&xas, entry);
1423 out:
1424 trace_dax_pte_fault_done(inode, vmf, ret);
1425 return ret | major;
1426}
1427
1428#ifdef CONFIG_FS_DAX_PMD
1429static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1430 struct iomap *iomap, void **entry)
1431{
1432 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1433 unsigned long pmd_addr = vmf->address & PMD_MASK;
1434 struct vm_area_struct *vma = vmf->vma;
1435 struct inode *inode = mapping->host;
1436 pgtable_t pgtable = NULL;
1437 struct page *zero_page;
1438 spinlock_t *ptl;
1439 pmd_t pmd_entry;
1440 pfn_t pfn;
1441
1442 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1443
1444 if (unlikely(!zero_page))
1445 goto fallback;
1446
1447 pfn = page_to_pfn_t(zero_page);
1448 *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1449 DAX_PMD | DAX_ZERO_PAGE, false);
1450
1451 if (arch_needs_pgtable_deposit()) {
1452 pgtable = pte_alloc_one(vma->vm_mm);
1453 if (!pgtable)
1454 return VM_FAULT_OOM;
1455 }
1456
1457 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1458 if (!pmd_none(*(vmf->pmd))) {
1459 spin_unlock(ptl);
1460 goto fallback;
1461 }
1462
1463 if (pgtable) {
1464 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1465 mm_inc_nr_ptes(vma->vm_mm);
1466 }
1467 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1468 pmd_entry = pmd_mkhuge(pmd_entry);
1469 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1470 spin_unlock(ptl);
1471 trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1472 return VM_FAULT_NOPAGE;
1473
1474fallback:
1475 if (pgtable)
1476 pte_free(vma->vm_mm, pgtable);
1477 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1478 return VM_FAULT_FALLBACK;
1479}
1480
1481static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1482 const struct iomap_ops *ops)
1483{
1484 struct vm_area_struct *vma = vmf->vma;
1485 struct address_space *mapping = vma->vm_file->f_mapping;
1486 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1487 unsigned long pmd_addr = vmf->address & PMD_MASK;
1488 bool write = vmf->flags & FAULT_FLAG_WRITE;
1489 bool sync;
1490 unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1491 struct inode *inode = mapping->host;
1492 vm_fault_t result = VM_FAULT_FALLBACK;
1493 struct iomap iomap = { .type = IOMAP_HOLE };
1494 struct iomap srcmap = { .type = IOMAP_HOLE };
1495 pgoff_t max_pgoff;
1496 void *entry;
1497 loff_t pos;
1498 int error;
1499 pfn_t pfn;
1500
1501 /*
1502 * Check whether offset isn't beyond end of file now. Caller is
1503 * supposed to hold locks serializing us with truncate / punch hole so
1504 * this is a reliable test.
1505 */
1506 max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1507
1508 trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1509
1510 /*
1511 * Make sure that the faulting address's PMD offset (color) matches
1512 * the PMD offset from the start of the file. This is necessary so
1513 * that a PMD range in the page table overlaps exactly with a PMD
1514 * range in the page cache.
1515 */
1516 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1517 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1518 goto fallback;
1519
1520 /* Fall back to PTEs if we're going to COW */
1521 if (write && !(vma->vm_flags & VM_SHARED))
1522 goto fallback;
1523
1524 /* If the PMD would extend outside the VMA */
1525 if (pmd_addr < vma->vm_start)
1526 goto fallback;
1527 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1528 goto fallback;
1529
1530 if (xas.xa_index >= max_pgoff) {
1531 result = VM_FAULT_SIGBUS;
1532 goto out;
1533 }
1534
1535 /* If the PMD would extend beyond the file size */
1536 if ((xas.xa_index | PG_PMD_COLOUR) >= max_pgoff)
1537 goto fallback;
1538
1539 /*
1540 * grab_mapping_entry() will make sure we get an empty PMD entry,
1541 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1542 * entry is already in the array, for instance), it will return
1543 * VM_FAULT_FALLBACK.
1544 */
1545 entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1546 if (xa_is_internal(entry)) {
1547 result = xa_to_internal(entry);
1548 goto fallback;
1549 }
1550
1551 /*
1552 * It is possible, particularly with mixed reads & writes to private
1553 * mappings, that we have raced with a PTE fault that overlaps with
1554 * the PMD we need to set up. If so just return and the fault will be
1555 * retried.
1556 */
1557 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1558 !pmd_devmap(*vmf->pmd)) {
1559 result = 0;
1560 goto unlock_entry;
1561 }
1562
1563 /*
1564 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1565 * setting up a mapping, so really we're using iomap_begin() as a way
1566 * to look up our filesystem block.
1567 */
1568 pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1569 error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap,
1570 &srcmap);
1571 if (error)
1572 goto unlock_entry;
1573
1574 if (iomap.offset + iomap.length < pos + PMD_SIZE)
1575 goto finish_iomap;
1576
1577 sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1578
1579 switch (iomap.type) {
1580 case IOMAP_MAPPED:
1581 error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1582 if (error < 0)
1583 goto finish_iomap;
1584
1585 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1586 DAX_PMD, write && !sync);
1587
1588 /*
1589 * If we are doing synchronous page fault and inode needs fsync,
1590 * we can insert PMD into page tables only after that happens.
1591 * Skip insertion for now and return the pfn so that caller can
1592 * insert it after fsync is done.
1593 */
1594 if (sync) {
1595 if (WARN_ON_ONCE(!pfnp))
1596 goto finish_iomap;
1597 *pfnp = pfn;
1598 result = VM_FAULT_NEEDDSYNC;
1599 goto finish_iomap;
1600 }
1601
1602 trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
1603 result = vmf_insert_pfn_pmd(vmf, pfn, write);
1604 break;
1605 case IOMAP_UNWRITTEN:
1606 case IOMAP_HOLE:
1607 if (WARN_ON_ONCE(write))
1608 break;
1609 result = dax_pmd_load_hole(&xas, vmf, &iomap, &entry);
1610 break;
1611 default:
1612 WARN_ON_ONCE(1);
1613 break;
1614 }
1615
1616 finish_iomap:
1617 if (ops->iomap_end) {
1618 int copied = PMD_SIZE;
1619
1620 if (result == VM_FAULT_FALLBACK)
1621 copied = 0;
1622 /*
1623 * The fault is done by now and there's no way back (other
1624 * thread may be already happily using PMD we have installed).
1625 * Just ignore error from ->iomap_end since we cannot do much
1626 * with it.
1627 */
1628 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1629 &iomap);
1630 }
1631 unlock_entry:
1632 dax_unlock_entry(&xas, entry);
1633 fallback:
1634 if (result == VM_FAULT_FALLBACK) {
1635 split_huge_pmd(vma, vmf->pmd, vmf->address);
1636 count_vm_event(THP_FAULT_FALLBACK);
1637 }
1638out:
1639 trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1640 return result;
1641}
1642#else
1643static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1644 const struct iomap_ops *ops)
1645{
1646 return VM_FAULT_FALLBACK;
1647}
1648#endif /* CONFIG_FS_DAX_PMD */
1649
1650/**
1651 * dax_iomap_fault - handle a page fault on a DAX file
1652 * @vmf: The description of the fault
1653 * @pe_size: Size of the page to fault in
1654 * @pfnp: PFN to insert for synchronous faults if fsync is required
1655 * @iomap_errp: Storage for detailed error code in case of error
1656 * @ops: Iomap ops passed from the file system
1657 *
1658 * When a page fault occurs, filesystems may call this helper in
1659 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1660 * has done all the necessary locking for page fault to proceed
1661 * successfully.
1662 */
1663vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1664 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1665{
1666 switch (pe_size) {
1667 case PE_SIZE_PTE:
1668 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1669 case PE_SIZE_PMD:
1670 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1671 default:
1672 return VM_FAULT_FALLBACK;
1673 }
1674}
1675EXPORT_SYMBOL_GPL(dax_iomap_fault);
1676
1677/*
1678 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1679 * @vmf: The description of the fault
1680 * @pfn: PFN to insert
1681 * @order: Order of entry to insert.
1682 *
1683 * This function inserts a writeable PTE or PMD entry into the page tables
1684 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1685 */
1686static vm_fault_t
1687dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1688{
1689 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1690 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1691 void *entry;
1692 vm_fault_t ret;
1693
1694 xas_lock_irq(&xas);
1695 entry = get_unlocked_entry(&xas, order);
1696 /* Did we race with someone splitting entry or so? */
1697 if (!entry || dax_is_conflict(entry) ||
1698 (order == 0 && !dax_is_pte_entry(entry))) {
1699 put_unlocked_entry(&xas, entry, WAKE_NEXT);
1700 xas_unlock_irq(&xas);
1701 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1702 VM_FAULT_NOPAGE);
1703 return VM_FAULT_NOPAGE;
1704 }
1705 xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1706 dax_lock_entry(&xas, entry);
1707 xas_unlock_irq(&xas);
1708 if (order == 0)
1709 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1710#ifdef CONFIG_FS_DAX_PMD
1711 else if (order == PMD_ORDER)
1712 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1713#endif
1714 else
1715 ret = VM_FAULT_FALLBACK;
1716 dax_unlock_entry(&xas, entry);
1717 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1718 return ret;
1719}
1720
1721/**
1722 * dax_finish_sync_fault - finish synchronous page fault
1723 * @vmf: The description of the fault
1724 * @pe_size: Size of entry to be inserted
1725 * @pfn: PFN to insert
1726 *
1727 * This function ensures that the file range touched by the page fault is
1728 * stored persistently on the media and handles inserting of appropriate page
1729 * table entry.
1730 */
1731vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1732 enum page_entry_size pe_size, pfn_t pfn)
1733{
1734 int err;
1735 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1736 unsigned int order = pe_order(pe_size);
1737 size_t len = PAGE_SIZE << order;
1738
1739 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1740 if (err)
1741 return VM_FAULT_SIGBUS;
1742 return dax_insert_pfn_mkwrite(vmf, pfn, order);
1743}
1744EXPORT_SYMBOL_GPL(dax_finish_sync_fault);