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/*
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
35static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
36{
37 struct request_queue *q = bdev->bd_queue;
38 long rc = -EIO;
39
40 dax->addr = (void __pmem *) ERR_PTR(-EIO);
41 if (blk_queue_enter(q, true) != 0)
42 return rc;
43
44 rc = bdev_direct_access(bdev, dax);
45 if (rc < 0) {
46 dax->addr = (void __pmem *) ERR_PTR(rc);
47 blk_queue_exit(q);
48 return rc;
49 }
50 return rc;
51}
52
53static void dax_unmap_atomic(struct block_device *bdev,
54 const struct blk_dax_ctl *dax)
55{
56 if (IS_ERR(dax->addr))
57 return;
58 blk_queue_exit(bdev->bd_queue);
59}
60
61struct page *read_dax_sector(struct block_device *bdev, sector_t n)
62{
63 struct page *page = alloc_pages(GFP_KERNEL, 0);
64 struct blk_dax_ctl dax = {
65 .size = PAGE_SIZE,
66 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
67 };
68 long rc;
69
70 if (!page)
71 return ERR_PTR(-ENOMEM);
72
73 rc = dax_map_atomic(bdev, &dax);
74 if (rc < 0)
75 return ERR_PTR(rc);
76 memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
77 dax_unmap_atomic(bdev, &dax);
78 return page;
79}
80
81/*
82 * dax_clear_sectors() is called from within transaction context from XFS,
83 * and hence this means the stack from this point must follow GFP_NOFS
84 * semantics for all operations.
85 */
86int dax_clear_sectors(struct block_device *bdev, sector_t _sector, long _size)
87{
88 struct blk_dax_ctl dax = {
89 .sector = _sector,
90 .size = _size,
91 };
92
93 might_sleep();
94 do {
95 long count, sz;
96
97 count = dax_map_atomic(bdev, &dax);
98 if (count < 0)
99 return count;
100 sz = min_t(long, count, SZ_128K);
101 clear_pmem(dax.addr, sz);
102 dax.size -= sz;
103 dax.sector += sz / 512;
104 dax_unmap_atomic(bdev, &dax);
105 cond_resched();
106 } while (dax.size);
107
108 wmb_pmem();
109 return 0;
110}
111EXPORT_SYMBOL_GPL(dax_clear_sectors);
112
113/* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
114static void dax_new_buf(void __pmem *addr, unsigned size, unsigned first,
115 loff_t pos, loff_t end)
116{
117 loff_t final = end - pos + first; /* The final byte of the buffer */
118
119 if (first > 0)
120 clear_pmem(addr, first);
121 if (final < size)
122 clear_pmem(addr + final, size - final);
123}
124
125static bool buffer_written(struct buffer_head *bh)
126{
127 return buffer_mapped(bh) && !buffer_unwritten(bh);
128}
129
130/*
131 * When ext4 encounters a hole, it returns without modifying the buffer_head
132 * which means that we can't trust b_size. To cope with this, we set b_state
133 * to 0 before calling get_block and, if any bit is set, we know we can trust
134 * b_size. Unfortunate, really, since ext4 knows precisely how long a hole is
135 * and would save us time calling get_block repeatedly.
136 */
137static bool buffer_size_valid(struct buffer_head *bh)
138{
139 return bh->b_state != 0;
140}
141
142
143static sector_t to_sector(const struct buffer_head *bh,
144 const struct inode *inode)
145{
146 sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
147
148 return sector;
149}
150
151static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
152 loff_t start, loff_t end, get_block_t get_block,
153 struct buffer_head *bh)
154{
155 loff_t pos = start, max = start, bh_max = start;
156 bool hole = false, need_wmb = false;
157 struct block_device *bdev = NULL;
158 int rw = iov_iter_rw(iter), rc;
159 long map_len = 0;
160 struct blk_dax_ctl dax = {
161 .addr = (void __pmem *) ERR_PTR(-EIO),
162 };
163
164 if (rw == READ)
165 end = min(end, i_size_read(inode));
166
167 while (pos < end) {
168 size_t len;
169 if (pos == max) {
170 unsigned blkbits = inode->i_blkbits;
171 long page = pos >> PAGE_SHIFT;
172 sector_t block = page << (PAGE_SHIFT - blkbits);
173 unsigned first = pos - (block << blkbits);
174 long size;
175
176 if (pos == bh_max) {
177 bh->b_size = PAGE_ALIGN(end - pos);
178 bh->b_state = 0;
179 rc = get_block(inode, block, bh, rw == WRITE);
180 if (rc)
181 break;
182 if (!buffer_size_valid(bh))
183 bh->b_size = 1 << blkbits;
184 bh_max = pos - first + bh->b_size;
185 bdev = bh->b_bdev;
186 } else {
187 unsigned done = bh->b_size -
188 (bh_max - (pos - first));
189 bh->b_blocknr += done >> blkbits;
190 bh->b_size -= done;
191 }
192
193 hole = rw == READ && !buffer_written(bh);
194 if (hole) {
195 size = bh->b_size - first;
196 } else {
197 dax_unmap_atomic(bdev, &dax);
198 dax.sector = to_sector(bh, inode);
199 dax.size = bh->b_size;
200 map_len = dax_map_atomic(bdev, &dax);
201 if (map_len < 0) {
202 rc = map_len;
203 break;
204 }
205 if (buffer_unwritten(bh) || buffer_new(bh)) {
206 dax_new_buf(dax.addr, map_len, first,
207 pos, end);
208 need_wmb = true;
209 }
210 dax.addr += first;
211 size = map_len - first;
212 }
213 max = min(pos + size, end);
214 }
215
216 if (iov_iter_rw(iter) == WRITE) {
217 len = copy_from_iter_pmem(dax.addr, max - pos, iter);
218 need_wmb = true;
219 } else if (!hole)
220 len = copy_to_iter((void __force *) dax.addr, max - pos,
221 iter);
222 else
223 len = iov_iter_zero(max - pos, iter);
224
225 if (!len) {
226 rc = -EFAULT;
227 break;
228 }
229
230 pos += len;
231 if (!IS_ERR(dax.addr))
232 dax.addr += len;
233 }
234
235 if (need_wmb)
236 wmb_pmem();
237 dax_unmap_atomic(bdev, &dax);
238
239 return (pos == start) ? rc : pos - start;
240}
241
242/**
243 * dax_do_io - Perform I/O to a DAX file
244 * @iocb: The control block for this I/O
245 * @inode: The file which the I/O is directed at
246 * @iter: The addresses to do I/O from or to
247 * @pos: The file offset where the I/O starts
248 * @get_block: The filesystem method used to translate file offsets to blocks
249 * @end_io: A filesystem callback for I/O completion
250 * @flags: See below
251 *
252 * This function uses the same locking scheme as do_blockdev_direct_IO:
253 * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
254 * caller for writes. For reads, we take and release the i_mutex ourselves.
255 * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
256 * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
257 * is in progress.
258 */
259ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
260 struct iov_iter *iter, loff_t pos, get_block_t get_block,
261 dio_iodone_t end_io, int flags)
262{
263 struct buffer_head bh;
264 ssize_t retval = -EINVAL;
265 loff_t end = pos + iov_iter_count(iter);
266
267 memset(&bh, 0, sizeof(bh));
268 bh.b_bdev = inode->i_sb->s_bdev;
269
270 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) {
271 struct address_space *mapping = inode->i_mapping;
272 inode_lock(inode);
273 retval = filemap_write_and_wait_range(mapping, pos, end - 1);
274 if (retval) {
275 inode_unlock(inode);
276 goto out;
277 }
278 }
279
280 /* Protects against truncate */
281 if (!(flags & DIO_SKIP_DIO_COUNT))
282 inode_dio_begin(inode);
283
284 retval = dax_io(inode, iter, pos, end, get_block, &bh);
285
286 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
287 inode_unlock(inode);
288
289 if (end_io) {
290 int err;
291
292 err = end_io(iocb, pos, retval, bh.b_private);
293 if (err)
294 retval = err;
295 }
296
297 if (!(flags & DIO_SKIP_DIO_COUNT))
298 inode_dio_end(inode);
299 out:
300 return retval;
301}
302EXPORT_SYMBOL_GPL(dax_do_io);
303
304/*
305 * The user has performed a load from a hole in the file. Allocating
306 * a new page in the file would cause excessive storage usage for
307 * workloads with sparse files. We allocate a page cache page instead.
308 * We'll kick it out of the page cache if it's ever written to,
309 * otherwise it will simply fall out of the page cache under memory
310 * pressure without ever having been dirtied.
311 */
312static int dax_load_hole(struct address_space *mapping, struct page *page,
313 struct vm_fault *vmf)
314{
315 unsigned long size;
316 struct inode *inode = mapping->host;
317 if (!page)
318 page = find_or_create_page(mapping, vmf->pgoff,
319 GFP_KERNEL | __GFP_ZERO);
320 if (!page)
321 return VM_FAULT_OOM;
322 /* Recheck i_size under page lock to avoid truncate race */
323 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
324 if (vmf->pgoff >= size) {
325 unlock_page(page);
326 put_page(page);
327 return VM_FAULT_SIGBUS;
328 }
329
330 vmf->page = page;
331 return VM_FAULT_LOCKED;
332}
333
334static int copy_user_bh(struct page *to, struct inode *inode,
335 struct buffer_head *bh, unsigned long vaddr)
336{
337 struct blk_dax_ctl dax = {
338 .sector = to_sector(bh, inode),
339 .size = bh->b_size,
340 };
341 struct block_device *bdev = bh->b_bdev;
342 void *vto;
343
344 if (dax_map_atomic(bdev, &dax) < 0)
345 return PTR_ERR(dax.addr);
346 vto = kmap_atomic(to);
347 copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
348 kunmap_atomic(vto);
349 dax_unmap_atomic(bdev, &dax);
350 return 0;
351}
352
353#define NO_SECTOR -1
354#define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
355
356static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
357 sector_t sector, bool pmd_entry, bool dirty)
358{
359 struct radix_tree_root *page_tree = &mapping->page_tree;
360 pgoff_t pmd_index = DAX_PMD_INDEX(index);
361 int type, error = 0;
362 void *entry;
363
364 WARN_ON_ONCE(pmd_entry && !dirty);
365 if (dirty)
366 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
367
368 spin_lock_irq(&mapping->tree_lock);
369
370 entry = radix_tree_lookup(page_tree, pmd_index);
371 if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
372 index = pmd_index;
373 goto dirty;
374 }
375
376 entry = radix_tree_lookup(page_tree, index);
377 if (entry) {
378 type = RADIX_DAX_TYPE(entry);
379 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
380 type != RADIX_DAX_PMD)) {
381 error = -EIO;
382 goto unlock;
383 }
384
385 if (!pmd_entry || type == RADIX_DAX_PMD)
386 goto dirty;
387
388 /*
389 * We only insert dirty PMD entries into the radix tree. This
390 * means we don't need to worry about removing a dirty PTE
391 * entry and inserting a clean PMD entry, thus reducing the
392 * range we would flush with a follow-up fsync/msync call.
393 */
394 radix_tree_delete(&mapping->page_tree, index);
395 mapping->nrexceptional--;
396 }
397
398 if (sector == NO_SECTOR) {
399 /*
400 * This can happen during correct operation if our pfn_mkwrite
401 * fault raced against a hole punch operation. If this
402 * happens the pte that was hole punched will have been
403 * unmapped and the radix tree entry will have been removed by
404 * the time we are called, but the call will still happen. We
405 * will return all the way up to wp_pfn_shared(), where the
406 * pte_same() check will fail, eventually causing page fault
407 * to be retried by the CPU.
408 */
409 goto unlock;
410 }
411
412 error = radix_tree_insert(page_tree, index,
413 RADIX_DAX_ENTRY(sector, pmd_entry));
414 if (error)
415 goto unlock;
416
417 mapping->nrexceptional++;
418 dirty:
419 if (dirty)
420 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
421 unlock:
422 spin_unlock_irq(&mapping->tree_lock);
423 return error;
424}
425
426static int dax_writeback_one(struct block_device *bdev,
427 struct address_space *mapping, pgoff_t index, void *entry)
428{
429 struct radix_tree_root *page_tree = &mapping->page_tree;
430 int type = RADIX_DAX_TYPE(entry);
431 struct radix_tree_node *node;
432 struct blk_dax_ctl dax;
433 void **slot;
434 int ret = 0;
435
436 spin_lock_irq(&mapping->tree_lock);
437 /*
438 * Regular page slots are stabilized by the page lock even
439 * without the tree itself locked. These unlocked entries
440 * need verification under the tree lock.
441 */
442 if (!__radix_tree_lookup(page_tree, index, &node, &slot))
443 goto unlock;
444 if (*slot != entry)
445 goto unlock;
446
447 /* another fsync thread may have already written back this entry */
448 if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
449 goto unlock;
450
451 if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
452 ret = -EIO;
453 goto unlock;
454 }
455
456 dax.sector = RADIX_DAX_SECTOR(entry);
457 dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
458 spin_unlock_irq(&mapping->tree_lock);
459
460 /*
461 * We cannot hold tree_lock while calling dax_map_atomic() because it
462 * eventually calls cond_resched().
463 */
464 ret = dax_map_atomic(bdev, &dax);
465 if (ret < 0)
466 return ret;
467
468 if (WARN_ON_ONCE(ret < dax.size)) {
469 ret = -EIO;
470 goto unmap;
471 }
472
473 wb_cache_pmem(dax.addr, dax.size);
474
475 spin_lock_irq(&mapping->tree_lock);
476 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
477 spin_unlock_irq(&mapping->tree_lock);
478 unmap:
479 dax_unmap_atomic(bdev, &dax);
480 return ret;
481
482 unlock:
483 spin_unlock_irq(&mapping->tree_lock);
484 return ret;
485}
486
487/*
488 * Flush the mapping to the persistent domain within the byte range of [start,
489 * end]. This is required by data integrity operations to ensure file data is
490 * on persistent storage prior to completion of the operation.
491 */
492int dax_writeback_mapping_range(struct address_space *mapping,
493 struct block_device *bdev, struct writeback_control *wbc)
494{
495 struct inode *inode = mapping->host;
496 pgoff_t start_index, end_index, pmd_index;
497 pgoff_t indices[PAGEVEC_SIZE];
498 struct pagevec pvec;
499 bool done = false;
500 int i, ret = 0;
501 void *entry;
502
503 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
504 return -EIO;
505
506 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
507 return 0;
508
509 start_index = wbc->range_start >> PAGE_SHIFT;
510 end_index = wbc->range_end >> PAGE_SHIFT;
511 pmd_index = DAX_PMD_INDEX(start_index);
512
513 rcu_read_lock();
514 entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
515 rcu_read_unlock();
516
517 /* see if the start of our range is covered by a PMD entry */
518 if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
519 start_index = pmd_index;
520
521 tag_pages_for_writeback(mapping, start_index, end_index);
522
523 pagevec_init(&pvec, 0);
524 while (!done) {
525 pvec.nr = find_get_entries_tag(mapping, start_index,
526 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
527 pvec.pages, indices);
528
529 if (pvec.nr == 0)
530 break;
531
532 for (i = 0; i < pvec.nr; i++) {
533 if (indices[i] > end_index) {
534 done = true;
535 break;
536 }
537
538 ret = dax_writeback_one(bdev, mapping, indices[i],
539 pvec.pages[i]);
540 if (ret < 0)
541 return ret;
542 }
543 }
544 wmb_pmem();
545 return 0;
546}
547EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
548
549static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
550 struct vm_area_struct *vma, struct vm_fault *vmf)
551{
552 unsigned long vaddr = (unsigned long)vmf->virtual_address;
553 struct address_space *mapping = inode->i_mapping;
554 struct block_device *bdev = bh->b_bdev;
555 struct blk_dax_ctl dax = {
556 .sector = to_sector(bh, inode),
557 .size = bh->b_size,
558 };
559 pgoff_t size;
560 int error;
561
562 i_mmap_lock_read(mapping);
563
564 /*
565 * Check truncate didn't happen while we were allocating a block.
566 * If it did, this block may or may not be still allocated to the
567 * file. We can't tell the filesystem to free it because we can't
568 * take i_mutex here. In the worst case, the file still has blocks
569 * allocated past the end of the file.
570 */
571 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
572 if (unlikely(vmf->pgoff >= size)) {
573 error = -EIO;
574 goto out;
575 }
576
577 if (dax_map_atomic(bdev, &dax) < 0) {
578 error = PTR_ERR(dax.addr);
579 goto out;
580 }
581
582 if (buffer_unwritten(bh) || buffer_new(bh)) {
583 clear_pmem(dax.addr, PAGE_SIZE);
584 wmb_pmem();
585 }
586 dax_unmap_atomic(bdev, &dax);
587
588 error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
589 vmf->flags & FAULT_FLAG_WRITE);
590 if (error)
591 goto out;
592
593 error = vm_insert_mixed(vma, vaddr, dax.pfn);
594
595 out:
596 i_mmap_unlock_read(mapping);
597
598 return error;
599}
600
601/**
602 * __dax_fault - handle a page fault on a DAX file
603 * @vma: The virtual memory area where the fault occurred
604 * @vmf: The description of the fault
605 * @get_block: The filesystem method used to translate file offsets to blocks
606 * @complete_unwritten: The filesystem method used to convert unwritten blocks
607 * to written so the data written to them is exposed. This is required for
608 * required by write faults for filesystems that will return unwritten
609 * extent mappings from @get_block, but it is optional for reads as
610 * dax_insert_mapping() will always zero unwritten blocks. If the fs does
611 * not support unwritten extents, the it should pass NULL.
612 *
613 * When a page fault occurs, filesystems may call this helper in their
614 * fault handler for DAX files. __dax_fault() assumes the caller has done all
615 * the necessary locking for the page fault to proceed successfully.
616 */
617int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
618 get_block_t get_block, dax_iodone_t complete_unwritten)
619{
620 struct file *file = vma->vm_file;
621 struct address_space *mapping = file->f_mapping;
622 struct inode *inode = mapping->host;
623 struct page *page;
624 struct buffer_head bh;
625 unsigned long vaddr = (unsigned long)vmf->virtual_address;
626 unsigned blkbits = inode->i_blkbits;
627 sector_t block;
628 pgoff_t size;
629 int error;
630 int major = 0;
631
632 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
633 if (vmf->pgoff >= size)
634 return VM_FAULT_SIGBUS;
635
636 memset(&bh, 0, sizeof(bh));
637 block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
638 bh.b_bdev = inode->i_sb->s_bdev;
639 bh.b_size = PAGE_SIZE;
640
641 repeat:
642 page = find_get_page(mapping, vmf->pgoff);
643 if (page) {
644 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
645 put_page(page);
646 return VM_FAULT_RETRY;
647 }
648 if (unlikely(page->mapping != mapping)) {
649 unlock_page(page);
650 put_page(page);
651 goto repeat;
652 }
653 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
654 if (unlikely(vmf->pgoff >= size)) {
655 /*
656 * We have a struct page covering a hole in the file
657 * from a read fault and we've raced with a truncate
658 */
659 error = -EIO;
660 goto unlock_page;
661 }
662 }
663
664 error = get_block(inode, block, &bh, 0);
665 if (!error && (bh.b_size < PAGE_SIZE))
666 error = -EIO; /* fs corruption? */
667 if (error)
668 goto unlock_page;
669
670 if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
671 if (vmf->flags & FAULT_FLAG_WRITE) {
672 error = get_block(inode, block, &bh, 1);
673 count_vm_event(PGMAJFAULT);
674 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
675 major = VM_FAULT_MAJOR;
676 if (!error && (bh.b_size < PAGE_SIZE))
677 error = -EIO;
678 if (error)
679 goto unlock_page;
680 } else {
681 return dax_load_hole(mapping, page, vmf);
682 }
683 }
684
685 if (vmf->cow_page) {
686 struct page *new_page = vmf->cow_page;
687 if (buffer_written(&bh))
688 error = copy_user_bh(new_page, inode, &bh, vaddr);
689 else
690 clear_user_highpage(new_page, vaddr);
691 if (error)
692 goto unlock_page;
693 vmf->page = page;
694 if (!page) {
695 i_mmap_lock_read(mapping);
696 /* Check we didn't race with truncate */
697 size = (i_size_read(inode) + PAGE_SIZE - 1) >>
698 PAGE_SHIFT;
699 if (vmf->pgoff >= size) {
700 i_mmap_unlock_read(mapping);
701 error = -EIO;
702 goto out;
703 }
704 }
705 return VM_FAULT_LOCKED;
706 }
707
708 /* Check we didn't race with a read fault installing a new page */
709 if (!page && major)
710 page = find_lock_page(mapping, vmf->pgoff);
711
712 if (page) {
713 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
714 PAGE_SIZE, 0);
715 delete_from_page_cache(page);
716 unlock_page(page);
717 put_page(page);
718 page = NULL;
719 }
720
721 /*
722 * If we successfully insert the new mapping over an unwritten extent,
723 * we need to ensure we convert the unwritten extent. If there is an
724 * error inserting the mapping, the filesystem needs to leave it as
725 * unwritten to prevent exposure of the stale underlying data to
726 * userspace, but we still need to call the completion function so
727 * the private resources on the mapping buffer can be released. We
728 * indicate what the callback should do via the uptodate variable, same
729 * as for normal BH based IO completions.
730 */
731 error = dax_insert_mapping(inode, &bh, vma, vmf);
732 if (buffer_unwritten(&bh)) {
733 if (complete_unwritten)
734 complete_unwritten(&bh, !error);
735 else
736 WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
737 }
738
739 out:
740 if (error == -ENOMEM)
741 return VM_FAULT_OOM | major;
742 /* -EBUSY is fine, somebody else faulted on the same PTE */
743 if ((error < 0) && (error != -EBUSY))
744 return VM_FAULT_SIGBUS | major;
745 return VM_FAULT_NOPAGE | major;
746
747 unlock_page:
748 if (page) {
749 unlock_page(page);
750 put_page(page);
751 }
752 goto out;
753}
754EXPORT_SYMBOL(__dax_fault);
755
756/**
757 * dax_fault - handle a page fault on a DAX file
758 * @vma: The virtual memory area where the fault occurred
759 * @vmf: The description of the fault
760 * @get_block: The filesystem method used to translate file offsets to blocks
761 *
762 * When a page fault occurs, filesystems may call this helper in their
763 * fault handler for DAX files.
764 */
765int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
766 get_block_t get_block, dax_iodone_t complete_unwritten)
767{
768 int result;
769 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
770
771 if (vmf->flags & FAULT_FLAG_WRITE) {
772 sb_start_pagefault(sb);
773 file_update_time(vma->vm_file);
774 }
775 result = __dax_fault(vma, vmf, get_block, complete_unwritten);
776 if (vmf->flags & FAULT_FLAG_WRITE)
777 sb_end_pagefault(sb);
778
779 return result;
780}
781EXPORT_SYMBOL_GPL(dax_fault);
782
783#ifdef CONFIG_TRANSPARENT_HUGEPAGE
784/*
785 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
786 * more often than one might expect in the below function.
787 */
788#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
789
790static void __dax_dbg(struct buffer_head *bh, unsigned long address,
791 const char *reason, const char *fn)
792{
793 if (bh) {
794 char bname[BDEVNAME_SIZE];
795 bdevname(bh->b_bdev, bname);
796 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
797 "length %zd fallback: %s\n", fn, current->comm,
798 address, bname, bh->b_state, (u64)bh->b_blocknr,
799 bh->b_size, reason);
800 } else {
801 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
802 current->comm, address, reason);
803 }
804}
805
806#define dax_pmd_dbg(bh, address, reason) __dax_dbg(bh, address, reason, "dax_pmd")
807
808int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
809 pmd_t *pmd, unsigned int flags, get_block_t get_block,
810 dax_iodone_t complete_unwritten)
811{
812 struct file *file = vma->vm_file;
813 struct address_space *mapping = file->f_mapping;
814 struct inode *inode = mapping->host;
815 struct buffer_head bh;
816 unsigned blkbits = inode->i_blkbits;
817 unsigned long pmd_addr = address & PMD_MASK;
818 bool write = flags & FAULT_FLAG_WRITE;
819 struct block_device *bdev;
820 pgoff_t size, pgoff;
821 sector_t block;
822 int error, result = 0;
823 bool alloc = false;
824
825 /* dax pmd mappings require pfn_t_devmap() */
826 if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
827 return VM_FAULT_FALLBACK;
828
829 /* Fall back to PTEs if we're going to COW */
830 if (write && !(vma->vm_flags & VM_SHARED)) {
831 split_huge_pmd(vma, pmd, address);
832 dax_pmd_dbg(NULL, address, "cow write");
833 return VM_FAULT_FALLBACK;
834 }
835 /* If the PMD would extend outside the VMA */
836 if (pmd_addr < vma->vm_start) {
837 dax_pmd_dbg(NULL, address, "vma start unaligned");
838 return VM_FAULT_FALLBACK;
839 }
840 if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
841 dax_pmd_dbg(NULL, address, "vma end unaligned");
842 return VM_FAULT_FALLBACK;
843 }
844
845 pgoff = linear_page_index(vma, pmd_addr);
846 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
847 if (pgoff >= size)
848 return VM_FAULT_SIGBUS;
849 /* If the PMD would cover blocks out of the file */
850 if ((pgoff | PG_PMD_COLOUR) >= size) {
851 dax_pmd_dbg(NULL, address,
852 "offset + huge page size > file size");
853 return VM_FAULT_FALLBACK;
854 }
855
856 memset(&bh, 0, sizeof(bh));
857 bh.b_bdev = inode->i_sb->s_bdev;
858 block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
859
860 bh.b_size = PMD_SIZE;
861
862 if (get_block(inode, block, &bh, 0) != 0)
863 return VM_FAULT_SIGBUS;
864
865 if (!buffer_mapped(&bh) && write) {
866 if (get_block(inode, block, &bh, 1) != 0)
867 return VM_FAULT_SIGBUS;
868 alloc = true;
869 }
870
871 bdev = bh.b_bdev;
872
873 /*
874 * If the filesystem isn't willing to tell us the length of a hole,
875 * just fall back to PTEs. Calling get_block 512 times in a loop
876 * would be silly.
877 */
878 if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
879 dax_pmd_dbg(&bh, address, "allocated block too small");
880 return VM_FAULT_FALLBACK;
881 }
882
883 /*
884 * If we allocated new storage, make sure no process has any
885 * zero pages covering this hole
886 */
887 if (alloc) {
888 loff_t lstart = pgoff << PAGE_SHIFT;
889 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
890
891 truncate_pagecache_range(inode, lstart, lend);
892 }
893
894 i_mmap_lock_read(mapping);
895
896 /*
897 * If a truncate happened while we were allocating blocks, we may
898 * leave blocks allocated to the file that are beyond EOF. We can't
899 * take i_mutex here, so just leave them hanging; they'll be freed
900 * when the file is deleted.
901 */
902 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
903 if (pgoff >= size) {
904 result = VM_FAULT_SIGBUS;
905 goto out;
906 }
907 if ((pgoff | PG_PMD_COLOUR) >= size) {
908 dax_pmd_dbg(&bh, address,
909 "offset + huge page size > file size");
910 goto fallback;
911 }
912
913 if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
914 spinlock_t *ptl;
915 pmd_t entry;
916 struct page *zero_page = get_huge_zero_page();
917
918 if (unlikely(!zero_page)) {
919 dax_pmd_dbg(&bh, address, "no zero page");
920 goto fallback;
921 }
922
923 ptl = pmd_lock(vma->vm_mm, pmd);
924 if (!pmd_none(*pmd)) {
925 spin_unlock(ptl);
926 dax_pmd_dbg(&bh, address, "pmd already present");
927 goto fallback;
928 }
929
930 dev_dbg(part_to_dev(bdev->bd_part),
931 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
932 __func__, current->comm, address,
933 (unsigned long long) to_sector(&bh, inode));
934
935 entry = mk_pmd(zero_page, vma->vm_page_prot);
936 entry = pmd_mkhuge(entry);
937 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
938 result = VM_FAULT_NOPAGE;
939 spin_unlock(ptl);
940 } else {
941 struct blk_dax_ctl dax = {
942 .sector = to_sector(&bh, inode),
943 .size = PMD_SIZE,
944 };
945 long length = dax_map_atomic(bdev, &dax);
946
947 if (length < 0) {
948 result = VM_FAULT_SIGBUS;
949 goto out;
950 }
951 if (length < PMD_SIZE) {
952 dax_pmd_dbg(&bh, address, "dax-length too small");
953 dax_unmap_atomic(bdev, &dax);
954 goto fallback;
955 }
956 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
957 dax_pmd_dbg(&bh, address, "pfn unaligned");
958 dax_unmap_atomic(bdev, &dax);
959 goto fallback;
960 }
961
962 if (!pfn_t_devmap(dax.pfn)) {
963 dax_unmap_atomic(bdev, &dax);
964 dax_pmd_dbg(&bh, address, "pfn not in memmap");
965 goto fallback;
966 }
967
968 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
969 clear_pmem(dax.addr, PMD_SIZE);
970 wmb_pmem();
971 count_vm_event(PGMAJFAULT);
972 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
973 result |= VM_FAULT_MAJOR;
974 }
975 dax_unmap_atomic(bdev, &dax);
976
977 /*
978 * For PTE faults we insert a radix tree entry for reads, and
979 * leave it clean. Then on the first write we dirty the radix
980 * tree entry via the dax_pfn_mkwrite() path. This sequence
981 * allows the dax_pfn_mkwrite() call to be simpler and avoid a
982 * call into get_block() to translate the pgoff to a sector in
983 * order to be able to create a new radix tree entry.
984 *
985 * The PMD path doesn't have an equivalent to
986 * dax_pfn_mkwrite(), though, so for a read followed by a
987 * write we traverse all the way through __dax_pmd_fault()
988 * twice. This means we can just skip inserting a radix tree
989 * entry completely on the initial read and just wait until
990 * the write to insert a dirty entry.
991 */
992 if (write) {
993 error = dax_radix_entry(mapping, pgoff, dax.sector,
994 true, true);
995 if (error) {
996 dax_pmd_dbg(&bh, address,
997 "PMD radix insertion failed");
998 goto fallback;
999 }
1000 }
1001
1002 dev_dbg(part_to_dev(bdev->bd_part),
1003 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
1004 __func__, current->comm, address,
1005 pfn_t_to_pfn(dax.pfn),
1006 (unsigned long long) dax.sector);
1007 result |= vmf_insert_pfn_pmd(vma, address, pmd,
1008 dax.pfn, write);
1009 }
1010
1011 out:
1012 i_mmap_unlock_read(mapping);
1013
1014 if (buffer_unwritten(&bh))
1015 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
1016
1017 return result;
1018
1019 fallback:
1020 count_vm_event(THP_FAULT_FALLBACK);
1021 result = VM_FAULT_FALLBACK;
1022 goto out;
1023}
1024EXPORT_SYMBOL_GPL(__dax_pmd_fault);
1025
1026/**
1027 * dax_pmd_fault - handle a PMD fault on a DAX file
1028 * @vma: The virtual memory area where the fault occurred
1029 * @vmf: The description of the fault
1030 * @get_block: The filesystem method used to translate file offsets to blocks
1031 *
1032 * When a page fault occurs, filesystems may call this helper in their
1033 * pmd_fault handler for DAX files.
1034 */
1035int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1036 pmd_t *pmd, unsigned int flags, get_block_t get_block,
1037 dax_iodone_t complete_unwritten)
1038{
1039 int result;
1040 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
1041
1042 if (flags & FAULT_FLAG_WRITE) {
1043 sb_start_pagefault(sb);
1044 file_update_time(vma->vm_file);
1045 }
1046 result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
1047 complete_unwritten);
1048 if (flags & FAULT_FLAG_WRITE)
1049 sb_end_pagefault(sb);
1050
1051 return result;
1052}
1053EXPORT_SYMBOL_GPL(dax_pmd_fault);
1054#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1055
1056/**
1057 * dax_pfn_mkwrite - handle first write to DAX page
1058 * @vma: The virtual memory area where the fault occurred
1059 * @vmf: The description of the fault
1060 */
1061int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1062{
1063 struct file *file = vma->vm_file;
1064 int error;
1065
1066 /*
1067 * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1068 * RADIX_DAX_PTE entry already exists in the radix tree from a
1069 * previous call to __dax_fault(). We just want to look up that PTE
1070 * entry using vmf->pgoff and make sure the dirty tag is set. This
1071 * saves us from having to make a call to get_block() here to look
1072 * up the sector.
1073 */
1074 error = dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false,
1075 true);
1076
1077 if (error == -ENOMEM)
1078 return VM_FAULT_OOM;
1079 if (error)
1080 return VM_FAULT_SIGBUS;
1081 return VM_FAULT_NOPAGE;
1082}
1083EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1084
1085/**
1086 * dax_zero_page_range - zero a range within a page of a DAX file
1087 * @inode: The file being truncated
1088 * @from: The file offset that is being truncated to
1089 * @length: The number of bytes to zero
1090 * @get_block: The filesystem method used to translate file offsets to blocks
1091 *
1092 * This function can be called by a filesystem when it is zeroing part of a
1093 * page in a DAX file. This is intended for hole-punch operations. If
1094 * you are truncating a file, the helper function dax_truncate_page() may be
1095 * more convenient.
1096 *
1097 * We work in terms of PAGE_SIZE here for commonality with
1098 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1099 * took care of disposing of the unnecessary blocks. Even if the filesystem
1100 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1101 * since the file might be mmapped.
1102 */
1103int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1104 get_block_t get_block)
1105{
1106 struct buffer_head bh;
1107 pgoff_t index = from >> PAGE_SHIFT;
1108 unsigned offset = from & (PAGE_SIZE-1);
1109 int err;
1110
1111 /* Block boundary? Nothing to do */
1112 if (!length)
1113 return 0;
1114 BUG_ON((offset + length) > PAGE_SIZE);
1115
1116 memset(&bh, 0, sizeof(bh));
1117 bh.b_bdev = inode->i_sb->s_bdev;
1118 bh.b_size = PAGE_SIZE;
1119 err = get_block(inode, index, &bh, 0);
1120 if (err < 0)
1121 return err;
1122 if (buffer_written(&bh)) {
1123 struct block_device *bdev = bh.b_bdev;
1124 struct blk_dax_ctl dax = {
1125 .sector = to_sector(&bh, inode),
1126 .size = PAGE_SIZE,
1127 };
1128
1129 if (dax_map_atomic(bdev, &dax) < 0)
1130 return PTR_ERR(dax.addr);
1131 clear_pmem(dax.addr + offset, length);
1132 wmb_pmem();
1133 dax_unmap_atomic(bdev, &dax);
1134 }
1135
1136 return 0;
1137}
1138EXPORT_SYMBOL_GPL(dax_zero_page_range);
1139
1140/**
1141 * dax_truncate_page - handle a partial page being truncated in a DAX file
1142 * @inode: The file being truncated
1143 * @from: The file offset that is being truncated to
1144 * @get_block: The filesystem method used to translate file offsets to blocks
1145 *
1146 * Similar to block_truncate_page(), this function can be called by a
1147 * filesystem when it is truncating a DAX file to handle the partial page.
1148 *
1149 * We work in terms of PAGE_SIZE here for commonality with
1150 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1151 * took care of disposing of the unnecessary blocks. Even if the filesystem
1152 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1153 * since the file might be mmapped.
1154 */
1155int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1156{
1157 unsigned length = PAGE_ALIGN(from) - from;
1158 return dax_zero_page_range(inode, from, length, get_block);
1159}
1160EXPORT_SYMBOL_GPL(dax_truncate_page);