e8f61ea690f7e4be84f6a7167235e4cd2ba689dd
[muen/linux.git] / fs / dax.c
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/sched.h>
29 #include <linux/sched/signal.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 #define CREATE_TRACE_POINTS
39 #include <trace/events/fs_dax.h>
40
41 /* We choose 4096 entries - same as per-zone page wait tables */
42 #define DAX_WAIT_TABLE_BITS 12
43 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
44
45 /* The 'colour' (ie low bits) within a PMD of a page offset.  */
46 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
47 #define PG_PMD_NR       (PMD_SIZE >> PAGE_SHIFT)
48
49 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
50
51 static int __init init_dax_wait_table(void)
52 {
53         int i;
54
55         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
56                 init_waitqueue_head(wait_table + i);
57         return 0;
58 }
59 fs_initcall(init_dax_wait_table);
60
61 /*
62  * We use lowest available bit in exceptional entry for locking, one bit for
63  * the entry size (PMD) and two more to tell us if the entry is a zero page or
64  * an empty entry that is just used for locking.  In total four special bits.
65  *
66  * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
67  * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
68  * block allocation.
69  */
70 #define RADIX_DAX_SHIFT         (RADIX_TREE_EXCEPTIONAL_SHIFT + 4)
71 #define RADIX_DAX_ENTRY_LOCK    (1 << RADIX_TREE_EXCEPTIONAL_SHIFT)
72 #define RADIX_DAX_PMD           (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1))
73 #define RADIX_DAX_ZERO_PAGE     (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2))
74 #define RADIX_DAX_EMPTY         (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 3))
75
76 static unsigned long dax_radix_pfn(void *entry)
77 {
78         return (unsigned long)entry >> RADIX_DAX_SHIFT;
79 }
80
81 static void *dax_radix_locked_entry(unsigned long pfn, unsigned long flags)
82 {
83         return (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY | flags |
84                         (pfn << RADIX_DAX_SHIFT) | RADIX_DAX_ENTRY_LOCK);
85 }
86
87 static unsigned int dax_radix_order(void *entry)
88 {
89         if ((unsigned long)entry & RADIX_DAX_PMD)
90                 return PMD_SHIFT - PAGE_SHIFT;
91         return 0;
92 }
93
94 static int dax_is_pmd_entry(void *entry)
95 {
96         return (unsigned long)entry & RADIX_DAX_PMD;
97 }
98
99 static int dax_is_pte_entry(void *entry)
100 {
101         return !((unsigned long)entry & RADIX_DAX_PMD);
102 }
103
104 static int dax_is_zero_entry(void *entry)
105 {
106         return (unsigned long)entry & RADIX_DAX_ZERO_PAGE;
107 }
108
109 static int dax_is_empty_entry(void *entry)
110 {
111         return (unsigned long)entry & RADIX_DAX_EMPTY;
112 }
113
114 /*
115  * DAX radix tree locking
116  */
117 struct exceptional_entry_key {
118         struct address_space *mapping;
119         pgoff_t entry_start;
120 };
121
122 struct wait_exceptional_entry_queue {
123         wait_queue_entry_t wait;
124         struct exceptional_entry_key key;
125 };
126
127 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
128                 pgoff_t index, void *entry, struct exceptional_entry_key *key)
129 {
130         unsigned long hash;
131
132         /*
133          * If 'entry' is a PMD, align the 'index' that we use for the wait
134          * queue to the start of that PMD.  This ensures that all offsets in
135          * the range covered by the PMD map to the same bit lock.
136          */
137         if (dax_is_pmd_entry(entry))
138                 index &= ~PG_PMD_COLOUR;
139
140         key->mapping = mapping;
141         key->entry_start = index;
142
143         hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
144         return wait_table + hash;
145 }
146
147 static int wake_exceptional_entry_func(wait_queue_entry_t *wait, unsigned int mode,
148                                        int sync, void *keyp)
149 {
150         struct exceptional_entry_key *key = keyp;
151         struct wait_exceptional_entry_queue *ewait =
152                 container_of(wait, struct wait_exceptional_entry_queue, wait);
153
154         if (key->mapping != ewait->key.mapping ||
155             key->entry_start != ewait->key.entry_start)
156                 return 0;
157         return autoremove_wake_function(wait, mode, sync, NULL);
158 }
159
160 /*
161  * @entry may no longer be the entry at the index in the mapping.
162  * The important information it's conveying is whether the entry at
163  * this index used to be a PMD entry.
164  */
165 static void dax_wake_mapping_entry_waiter(struct address_space *mapping,
166                 pgoff_t index, void *entry, bool wake_all)
167 {
168         struct exceptional_entry_key key;
169         wait_queue_head_t *wq;
170
171         wq = dax_entry_waitqueue(mapping, index, entry, &key);
172
173         /*
174          * Checking for locked entry and prepare_to_wait_exclusive() happens
175          * under the i_pages lock, ditto for entry handling in our callers.
176          * So at this point all tasks that could have seen our entry locked
177          * must be in the waitqueue and the following check will see them.
178          */
179         if (waitqueue_active(wq))
180                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
181 }
182
183 /*
184  * Check whether the given slot is locked.  Must be called with the i_pages
185  * lock held.
186  */
187 static inline int slot_locked(struct address_space *mapping, void **slot)
188 {
189         unsigned long entry = (unsigned long)
190                 radix_tree_deref_slot_protected(slot, &mapping->i_pages.xa_lock);
191         return entry & RADIX_DAX_ENTRY_LOCK;
192 }
193
194 /*
195  * Mark the given slot as locked.  Must be called with the i_pages lock held.
196  */
197 static inline void *lock_slot(struct address_space *mapping, void **slot)
198 {
199         unsigned long entry = (unsigned long)
200                 radix_tree_deref_slot_protected(slot, &mapping->i_pages.xa_lock);
201
202         entry |= RADIX_DAX_ENTRY_LOCK;
203         radix_tree_replace_slot(&mapping->i_pages, slot, (void *)entry);
204         return (void *)entry;
205 }
206
207 /*
208  * Mark the given slot as unlocked.  Must be called with the i_pages lock held.
209  */
210 static inline void *unlock_slot(struct address_space *mapping, void **slot)
211 {
212         unsigned long entry = (unsigned long)
213                 radix_tree_deref_slot_protected(slot, &mapping->i_pages.xa_lock);
214
215         entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
216         radix_tree_replace_slot(&mapping->i_pages, slot, (void *)entry);
217         return (void *)entry;
218 }
219
220 /*
221  * Lookup entry in radix tree, wait for it to become unlocked if it is
222  * exceptional entry and return it. The caller must call
223  * put_unlocked_mapping_entry() when he decided not to lock the entry or
224  * put_locked_mapping_entry() when he locked the entry and now wants to
225  * unlock it.
226  *
227  * Must be called with the i_pages lock held.
228  */
229 static void *get_unlocked_mapping_entry(struct address_space *mapping,
230                                         pgoff_t index, void ***slotp)
231 {
232         void *entry, **slot;
233         struct wait_exceptional_entry_queue ewait;
234         wait_queue_head_t *wq;
235
236         init_wait(&ewait.wait);
237         ewait.wait.func = wake_exceptional_entry_func;
238
239         for (;;) {
240                 entry = __radix_tree_lookup(&mapping->i_pages, index, NULL,
241                                           &slot);
242                 if (!entry ||
243                     WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)) ||
244                     !slot_locked(mapping, slot)) {
245                         if (slotp)
246                                 *slotp = slot;
247                         return entry;
248                 }
249
250                 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
251                 prepare_to_wait_exclusive(wq, &ewait.wait,
252                                           TASK_UNINTERRUPTIBLE);
253                 xa_unlock_irq(&mapping->i_pages);
254                 schedule();
255                 finish_wait(wq, &ewait.wait);
256                 xa_lock_irq(&mapping->i_pages);
257         }
258 }
259
260 static void dax_unlock_mapping_entry(struct address_space *mapping,
261                                      pgoff_t index)
262 {
263         void *entry, **slot;
264
265         xa_lock_irq(&mapping->i_pages);
266         entry = __radix_tree_lookup(&mapping->i_pages, index, NULL, &slot);
267         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
268                          !slot_locked(mapping, slot))) {
269                 xa_unlock_irq(&mapping->i_pages);
270                 return;
271         }
272         unlock_slot(mapping, slot);
273         xa_unlock_irq(&mapping->i_pages);
274         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
275 }
276
277 static void put_locked_mapping_entry(struct address_space *mapping,
278                 pgoff_t index)
279 {
280         dax_unlock_mapping_entry(mapping, index);
281 }
282
283 /*
284  * Called when we are done with radix tree entry we looked up via
285  * get_unlocked_mapping_entry() and which we didn't lock in the end.
286  */
287 static void put_unlocked_mapping_entry(struct address_space *mapping,
288                                        pgoff_t index, void *entry)
289 {
290         if (!entry)
291                 return;
292
293         /* We have to wake up next waiter for the radix tree entry lock */
294         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
295 }
296
297 static unsigned long dax_entry_size(void *entry)
298 {
299         if (dax_is_zero_entry(entry))
300                 return 0;
301         else if (dax_is_empty_entry(entry))
302                 return 0;
303         else if (dax_is_pmd_entry(entry))
304                 return PMD_SIZE;
305         else
306                 return PAGE_SIZE;
307 }
308
309 static unsigned long dax_radix_end_pfn(void *entry)
310 {
311         return dax_radix_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
312 }
313
314 /*
315  * Iterate through all mapped pfns represented by an entry, i.e. skip
316  * 'empty' and 'zero' entries.
317  */
318 #define for_each_mapped_pfn(entry, pfn) \
319         for (pfn = dax_radix_pfn(entry); \
320                         pfn < dax_radix_end_pfn(entry); pfn++)
321
322 static void dax_associate_entry(void *entry, struct address_space *mapping)
323 {
324         unsigned long pfn;
325
326         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
327                 return;
328
329         for_each_mapped_pfn(entry, pfn) {
330                 struct page *page = pfn_to_page(pfn);
331
332                 WARN_ON_ONCE(page->mapping);
333                 page->mapping = mapping;
334         }
335 }
336
337 static void dax_disassociate_entry(void *entry, struct address_space *mapping,
338                 bool trunc)
339 {
340         unsigned long pfn;
341
342         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
343                 return;
344
345         for_each_mapped_pfn(entry, pfn) {
346                 struct page *page = pfn_to_page(pfn);
347
348                 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
349                 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
350                 page->mapping = NULL;
351         }
352 }
353
354 static struct page *dax_busy_page(void *entry)
355 {
356         unsigned long pfn;
357
358         for_each_mapped_pfn(entry, pfn) {
359                 struct page *page = pfn_to_page(pfn);
360
361                 if (page_ref_count(page) > 1)
362                         return page;
363         }
364         return NULL;
365 }
366
367 /*
368  * Find radix tree entry at given index. If it points to an exceptional entry,
369  * return it with the radix tree entry locked. If the radix tree doesn't
370  * contain given index, create an empty exceptional entry for the index and
371  * return with it locked.
372  *
373  * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
374  * either return that locked entry or will return an error.  This error will
375  * happen if there are any 4k entries within the 2MiB range that we are
376  * requesting.
377  *
378  * We always favor 4k entries over 2MiB entries. There isn't a flow where we
379  * evict 4k entries in order to 'upgrade' them to a 2MiB entry.  A 2MiB
380  * insertion will fail if it finds any 4k entries already in the tree, and a
381  * 4k insertion will cause an existing 2MiB entry to be unmapped and
382  * downgraded to 4k entries.  This happens for both 2MiB huge zero pages as
383  * well as 2MiB empty entries.
384  *
385  * The exception to this downgrade path is for 2MiB DAX PMD entries that have
386  * real storage backing them.  We will leave these real 2MiB DAX entries in
387  * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
388  *
389  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
390  * persistent memory the benefit is doubtful. We can add that later if we can
391  * show it helps.
392  */
393 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
394                 unsigned long size_flag)
395 {
396         bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
397         void *entry, **slot;
398
399 restart:
400         xa_lock_irq(&mapping->i_pages);
401         entry = get_unlocked_mapping_entry(mapping, index, &slot);
402
403         if (WARN_ON_ONCE(entry && !radix_tree_exceptional_entry(entry))) {
404                 entry = ERR_PTR(-EIO);
405                 goto out_unlock;
406         }
407
408         if (entry) {
409                 if (size_flag & RADIX_DAX_PMD) {
410                         if (dax_is_pte_entry(entry)) {
411                                 put_unlocked_mapping_entry(mapping, index,
412                                                 entry);
413                                 entry = ERR_PTR(-EEXIST);
414                                 goto out_unlock;
415                         }
416                 } else { /* trying to grab a PTE entry */
417                         if (dax_is_pmd_entry(entry) &&
418                             (dax_is_zero_entry(entry) ||
419                              dax_is_empty_entry(entry))) {
420                                 pmd_downgrade = true;
421                         }
422                 }
423         }
424
425         /* No entry for given index? Make sure radix tree is big enough. */
426         if (!entry || pmd_downgrade) {
427                 int err;
428
429                 if (pmd_downgrade) {
430                         /*
431                          * Make sure 'entry' remains valid while we drop
432                          * the i_pages lock.
433                          */
434                         entry = lock_slot(mapping, slot);
435                 }
436
437                 xa_unlock_irq(&mapping->i_pages);
438                 /*
439                  * Besides huge zero pages the only other thing that gets
440                  * downgraded are empty entries which don't need to be
441                  * unmapped.
442                  */
443                 if (pmd_downgrade && dax_is_zero_entry(entry))
444                         unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
445                                                         PG_PMD_NR, false);
446
447                 err = radix_tree_preload(
448                                 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
449                 if (err) {
450                         if (pmd_downgrade)
451                                 put_locked_mapping_entry(mapping, index);
452                         return ERR_PTR(err);
453                 }
454                 xa_lock_irq(&mapping->i_pages);
455
456                 if (!entry) {
457                         /*
458                          * We needed to drop the i_pages lock while calling
459                          * radix_tree_preload() and we didn't have an entry to
460                          * lock.  See if another thread inserted an entry at
461                          * our index during this time.
462                          */
463                         entry = __radix_tree_lookup(&mapping->i_pages, index,
464                                         NULL, &slot);
465                         if (entry) {
466                                 radix_tree_preload_end();
467                                 xa_unlock_irq(&mapping->i_pages);
468                                 goto restart;
469                         }
470                 }
471
472                 if (pmd_downgrade) {
473                         dax_disassociate_entry(entry, mapping, false);
474                         radix_tree_delete(&mapping->i_pages, index);
475                         mapping->nrexceptional--;
476                         dax_wake_mapping_entry_waiter(mapping, index, entry,
477                                         true);
478                 }
479
480                 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
481
482                 err = __radix_tree_insert(&mapping->i_pages, index,
483                                 dax_radix_order(entry), entry);
484                 radix_tree_preload_end();
485                 if (err) {
486                         xa_unlock_irq(&mapping->i_pages);
487                         /*
488                          * Our insertion of a DAX entry failed, most likely
489                          * because we were inserting a PMD entry and it
490                          * collided with a PTE sized entry at a different
491                          * index in the PMD range.  We haven't inserted
492                          * anything into the radix tree and have no waiters to
493                          * wake.
494                          */
495                         return ERR_PTR(err);
496                 }
497                 /* Good, we have inserted empty locked entry into the tree. */
498                 mapping->nrexceptional++;
499                 xa_unlock_irq(&mapping->i_pages);
500                 return entry;
501         }
502         entry = lock_slot(mapping, slot);
503  out_unlock:
504         xa_unlock_irq(&mapping->i_pages);
505         return entry;
506 }
507
508 /**
509  * dax_layout_busy_page - find first pinned page in @mapping
510  * @mapping: address space to scan for a page with ref count > 1
511  *
512  * DAX requires ZONE_DEVICE mapped pages. These pages are never
513  * 'onlined' to the page allocator so they are considered idle when
514  * page->count == 1. A filesystem uses this interface to determine if
515  * any page in the mapping is busy, i.e. for DMA, or other
516  * get_user_pages() usages.
517  *
518  * It is expected that the filesystem is holding locks to block the
519  * establishment of new mappings in this address_space. I.e. it expects
520  * to be able to run unmap_mapping_range() and subsequently not race
521  * mapping_mapped() becoming true.
522  */
523 struct page *dax_layout_busy_page(struct address_space *mapping)
524 {
525         pgoff_t indices[PAGEVEC_SIZE];
526         struct page *page = NULL;
527         struct pagevec pvec;
528         pgoff_t index, end;
529         unsigned i;
530
531         /*
532          * In the 'limited' case get_user_pages() for dax is disabled.
533          */
534         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
535                 return NULL;
536
537         if (!dax_mapping(mapping) || !mapping_mapped(mapping))
538                 return NULL;
539
540         pagevec_init(&pvec);
541         index = 0;
542         end = -1;
543
544         /*
545          * If we race get_user_pages_fast() here either we'll see the
546          * elevated page count in the pagevec_lookup and wait, or
547          * get_user_pages_fast() will see that the page it took a reference
548          * against is no longer mapped in the page tables and bail to the
549          * get_user_pages() slow path.  The slow path is protected by
550          * pte_lock() and pmd_lock(). New references are not taken without
551          * holding those locks, and unmap_mapping_range() will not zero the
552          * pte or pmd without holding the respective lock, so we are
553          * guaranteed to either see new references or prevent new
554          * references from being established.
555          */
556         unmap_mapping_range(mapping, 0, 0, 1);
557
558         while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
559                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
560                                 indices)) {
561                 for (i = 0; i < pagevec_count(&pvec); i++) {
562                         struct page *pvec_ent = pvec.pages[i];
563                         void *entry;
564
565                         index = indices[i];
566                         if (index >= end)
567                                 break;
568
569                         if (!radix_tree_exceptional_entry(pvec_ent))
570                                 continue;
571
572                         xa_lock_irq(&mapping->i_pages);
573                         entry = get_unlocked_mapping_entry(mapping, index, NULL);
574                         if (entry)
575                                 page = dax_busy_page(entry);
576                         put_unlocked_mapping_entry(mapping, index, entry);
577                         xa_unlock_irq(&mapping->i_pages);
578                         if (page)
579                                 break;
580                 }
581                 pagevec_remove_exceptionals(&pvec);
582                 pagevec_release(&pvec);
583                 index++;
584
585                 if (page)
586                         break;
587         }
588         return page;
589 }
590 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
591
592 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
593                                           pgoff_t index, bool trunc)
594 {
595         int ret = 0;
596         void *entry;
597         struct radix_tree_root *pages = &mapping->i_pages;
598
599         xa_lock_irq(pages);
600         entry = get_unlocked_mapping_entry(mapping, index, NULL);
601         if (!entry || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)))
602                 goto out;
603         if (!trunc &&
604             (radix_tree_tag_get(pages, index, PAGECACHE_TAG_DIRTY) ||
605              radix_tree_tag_get(pages, index, PAGECACHE_TAG_TOWRITE)))
606                 goto out;
607         dax_disassociate_entry(entry, mapping, trunc);
608         radix_tree_delete(pages, index);
609         mapping->nrexceptional--;
610         ret = 1;
611 out:
612         put_unlocked_mapping_entry(mapping, index, entry);
613         xa_unlock_irq(pages);
614         return ret;
615 }
616 /*
617  * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
618  * entry to get unlocked before deleting it.
619  */
620 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
621 {
622         int ret = __dax_invalidate_mapping_entry(mapping, index, true);
623
624         /*
625          * This gets called from truncate / punch_hole path. As such, the caller
626          * must hold locks protecting against concurrent modifications of the
627          * radix tree (usually fs-private i_mmap_sem for writing). Since the
628          * caller has seen exceptional entry for this index, we better find it
629          * at that index as well...
630          */
631         WARN_ON_ONCE(!ret);
632         return ret;
633 }
634
635 /*
636  * Invalidate exceptional DAX entry if it is clean.
637  */
638 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
639                                       pgoff_t index)
640 {
641         return __dax_invalidate_mapping_entry(mapping, index, false);
642 }
643
644 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
645                 sector_t sector, size_t size, struct page *to,
646                 unsigned long vaddr)
647 {
648         void *vto, *kaddr;
649         pgoff_t pgoff;
650         pfn_t pfn;
651         long rc;
652         int id;
653
654         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
655         if (rc)
656                 return rc;
657
658         id = dax_read_lock();
659         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
660         if (rc < 0) {
661                 dax_read_unlock(id);
662                 return rc;
663         }
664         vto = kmap_atomic(to);
665         copy_user_page(vto, (void __force *)kaddr, vaddr, to);
666         kunmap_atomic(vto);
667         dax_read_unlock(id);
668         return 0;
669 }
670
671 /*
672  * By this point grab_mapping_entry() has ensured that we have a locked entry
673  * of the appropriate size so we don't have to worry about downgrading PMDs to
674  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
675  * already in the tree, we will skip the insertion and just dirty the PMD as
676  * appropriate.
677  */
678 static void *dax_insert_mapping_entry(struct address_space *mapping,
679                                       struct vm_fault *vmf,
680                                       void *entry, pfn_t pfn_t,
681                                       unsigned long flags, bool dirty)
682 {
683         struct radix_tree_root *pages = &mapping->i_pages;
684         unsigned long pfn = pfn_t_to_pfn(pfn_t);
685         pgoff_t index = vmf->pgoff;
686         void *new_entry;
687
688         if (dirty)
689                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
690
691         if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_ZERO_PAGE)) {
692                 /* we are replacing a zero page with block mapping */
693                 if (dax_is_pmd_entry(entry))
694                         unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
695                                                         PG_PMD_NR, false);
696                 else /* pte entry */
697                         unmap_mapping_pages(mapping, vmf->pgoff, 1, false);
698         }
699
700         xa_lock_irq(pages);
701         new_entry = dax_radix_locked_entry(pfn, flags);
702         if (dax_entry_size(entry) != dax_entry_size(new_entry)) {
703                 dax_disassociate_entry(entry, mapping, false);
704                 dax_associate_entry(new_entry, mapping);
705         }
706
707         if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
708                 /*
709                  * Only swap our new entry into the radix tree if the current
710                  * entry is a zero page or an empty entry.  If a normal PTE or
711                  * PMD entry is already in the tree, we leave it alone.  This
712                  * means that if we are trying to insert a PTE and the
713                  * existing entry is a PMD, we will just leave the PMD in the
714                  * tree and dirty it if necessary.
715                  */
716                 struct radix_tree_node *node;
717                 void **slot;
718                 void *ret;
719
720                 ret = __radix_tree_lookup(pages, index, &node, &slot);
721                 WARN_ON_ONCE(ret != entry);
722                 __radix_tree_replace(pages, node, slot,
723                                      new_entry, NULL);
724                 entry = new_entry;
725         }
726
727         if (dirty)
728                 radix_tree_tag_set(pages, index, PAGECACHE_TAG_DIRTY);
729
730         xa_unlock_irq(pages);
731         return entry;
732 }
733
734 static inline unsigned long
735 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
736 {
737         unsigned long address;
738
739         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
740         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
741         return address;
742 }
743
744 /* Walk all mappings of a given index of a file and writeprotect them */
745 static void dax_mapping_entry_mkclean(struct address_space *mapping,
746                                       pgoff_t index, unsigned long pfn)
747 {
748         struct vm_area_struct *vma;
749         pte_t pte, *ptep = NULL;
750         pmd_t *pmdp = NULL;
751         spinlock_t *ptl;
752
753         i_mmap_lock_read(mapping);
754         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
755                 unsigned long address, start, end;
756
757                 cond_resched();
758
759                 if (!(vma->vm_flags & VM_SHARED))
760                         continue;
761
762                 address = pgoff_address(index, vma);
763
764                 /*
765                  * Note because we provide start/end to follow_pte_pmd it will
766                  * call mmu_notifier_invalidate_range_start() on our behalf
767                  * before taking any lock.
768                  */
769                 if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl))
770                         continue;
771
772                 /*
773                  * No need to call mmu_notifier_invalidate_range() as we are
774                  * downgrading page table protection not changing it to point
775                  * to a new page.
776                  *
777                  * See Documentation/vm/mmu_notifier.txt
778                  */
779                 if (pmdp) {
780 #ifdef CONFIG_FS_DAX_PMD
781                         pmd_t pmd;
782
783                         if (pfn != pmd_pfn(*pmdp))
784                                 goto unlock_pmd;
785                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
786                                 goto unlock_pmd;
787
788                         flush_cache_page(vma, address, pfn);
789                         pmd = pmdp_huge_clear_flush(vma, address, pmdp);
790                         pmd = pmd_wrprotect(pmd);
791                         pmd = pmd_mkclean(pmd);
792                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
793 unlock_pmd:
794 #endif
795                         spin_unlock(ptl);
796                 } else {
797                         if (pfn != pte_pfn(*ptep))
798                                 goto unlock_pte;
799                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
800                                 goto unlock_pte;
801
802                         flush_cache_page(vma, address, pfn);
803                         pte = ptep_clear_flush(vma, address, ptep);
804                         pte = pte_wrprotect(pte);
805                         pte = pte_mkclean(pte);
806                         set_pte_at(vma->vm_mm, address, ptep, pte);
807 unlock_pte:
808                         pte_unmap_unlock(ptep, ptl);
809                 }
810
811                 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
812         }
813         i_mmap_unlock_read(mapping);
814 }
815
816 static int dax_writeback_one(struct dax_device *dax_dev,
817                 struct address_space *mapping, pgoff_t index, void *entry)
818 {
819         struct radix_tree_root *pages = &mapping->i_pages;
820         void *entry2, **slot;
821         unsigned long pfn;
822         long ret = 0;
823         size_t size;
824
825         /*
826          * A page got tagged dirty in DAX mapping? Something is seriously
827          * wrong.
828          */
829         if (WARN_ON(!radix_tree_exceptional_entry(entry)))
830                 return -EIO;
831
832         xa_lock_irq(pages);
833         entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
834         /* Entry got punched out / reallocated? */
835         if (!entry2 || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry2)))
836                 goto put_unlocked;
837         /*
838          * Entry got reallocated elsewhere? No need to writeback. We have to
839          * compare pfns as we must not bail out due to difference in lockbit
840          * or entry type.
841          */
842         if (dax_radix_pfn(entry2) != dax_radix_pfn(entry))
843                 goto put_unlocked;
844         if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
845                                 dax_is_zero_entry(entry))) {
846                 ret = -EIO;
847                 goto put_unlocked;
848         }
849
850         /* Another fsync thread may have already written back this entry */
851         if (!radix_tree_tag_get(pages, index, PAGECACHE_TAG_TOWRITE))
852                 goto put_unlocked;
853         /* Lock the entry to serialize with page faults */
854         entry = lock_slot(mapping, slot);
855         /*
856          * We can clear the tag now but we have to be careful so that concurrent
857          * dax_writeback_one() calls for the same index cannot finish before we
858          * actually flush the caches. This is achieved as the calls will look
859          * at the entry only under the i_pages lock and once they do that
860          * they will see the entry locked and wait for it to unlock.
861          */
862         radix_tree_tag_clear(pages, index, PAGECACHE_TAG_TOWRITE);
863         xa_unlock_irq(pages);
864
865         /*
866          * Even if dax_writeback_mapping_range() was given a wbc->range_start
867          * in the middle of a PMD, the 'index' we are given will be aligned to
868          * the start index of the PMD, as will the pfn we pull from 'entry'.
869          * This allows us to flush for PMD_SIZE and not have to worry about
870          * partial PMD writebacks.
871          */
872         pfn = dax_radix_pfn(entry);
873         size = PAGE_SIZE << dax_radix_order(entry);
874
875         dax_mapping_entry_mkclean(mapping, index, pfn);
876         dax_flush(dax_dev, page_address(pfn_to_page(pfn)), size);
877         /*
878          * After we have flushed the cache, we can clear the dirty tag. There
879          * cannot be new dirty data in the pfn after the flush has completed as
880          * the pfn mappings are writeprotected and fault waits for mapping
881          * entry lock.
882          */
883         xa_lock_irq(pages);
884         radix_tree_tag_clear(pages, index, PAGECACHE_TAG_DIRTY);
885         xa_unlock_irq(pages);
886         trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT);
887         put_locked_mapping_entry(mapping, index);
888         return ret;
889
890  put_unlocked:
891         put_unlocked_mapping_entry(mapping, index, entry2);
892         xa_unlock_irq(pages);
893         return ret;
894 }
895
896 /*
897  * Flush the mapping to the persistent domain within the byte range of [start,
898  * end]. This is required by data integrity operations to ensure file data is
899  * on persistent storage prior to completion of the operation.
900  */
901 int dax_writeback_mapping_range(struct address_space *mapping,
902                 struct block_device *bdev, struct writeback_control *wbc)
903 {
904         struct inode *inode = mapping->host;
905         pgoff_t start_index, end_index;
906         pgoff_t indices[PAGEVEC_SIZE];
907         struct dax_device *dax_dev;
908         struct pagevec pvec;
909         bool done = false;
910         int i, ret = 0;
911
912         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
913                 return -EIO;
914
915         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
916                 return 0;
917
918         dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
919         if (!dax_dev)
920                 return -EIO;
921
922         start_index = wbc->range_start >> PAGE_SHIFT;
923         end_index = wbc->range_end >> PAGE_SHIFT;
924
925         trace_dax_writeback_range(inode, start_index, end_index);
926
927         tag_pages_for_writeback(mapping, start_index, end_index);
928
929         pagevec_init(&pvec);
930         while (!done) {
931                 pvec.nr = find_get_entries_tag(mapping, start_index,
932                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
933                                 pvec.pages, indices);
934
935                 if (pvec.nr == 0)
936                         break;
937
938                 for (i = 0; i < pvec.nr; i++) {
939                         if (indices[i] > end_index) {
940                                 done = true;
941                                 break;
942                         }
943
944                         ret = dax_writeback_one(dax_dev, mapping, indices[i],
945                                         pvec.pages[i]);
946                         if (ret < 0) {
947                                 mapping_set_error(mapping, ret);
948                                 goto out;
949                         }
950                 }
951                 start_index = indices[pvec.nr - 1] + 1;
952         }
953 out:
954         put_dax(dax_dev);
955         trace_dax_writeback_range_done(inode, start_index, end_index);
956         return (ret < 0 ? ret : 0);
957 }
958 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
959
960 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
961 {
962         return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
963 }
964
965 static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
966                          pfn_t *pfnp)
967 {
968         const sector_t sector = dax_iomap_sector(iomap, pos);
969         pgoff_t pgoff;
970         void *kaddr;
971         int id, rc;
972         long length;
973
974         rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
975         if (rc)
976                 return rc;
977         id = dax_read_lock();
978         length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
979                                    &kaddr, pfnp);
980         if (length < 0) {
981                 rc = length;
982                 goto out;
983         }
984         rc = -EINVAL;
985         if (PFN_PHYS(length) < size)
986                 goto out;
987         if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
988                 goto out;
989         /* For larger pages we need devmap */
990         if (length > 1 && !pfn_t_devmap(*pfnp))
991                 goto out;
992         rc = 0;
993 out:
994         dax_read_unlock(id);
995         return rc;
996 }
997
998 /*
999  * The user has performed a load from a hole in the file.  Allocating a new
1000  * page in the file would cause excessive storage usage for workloads with
1001  * sparse files.  Instead we insert a read-only mapping of the 4k zero page.
1002  * If this page is ever written to we will re-fault and change the mapping to
1003  * point to real DAX storage instead.
1004  */
1005 static int dax_load_hole(struct address_space *mapping, void *entry,
1006                          struct vm_fault *vmf)
1007 {
1008         struct inode *inode = mapping->host;
1009         unsigned long vaddr = vmf->address;
1010         int ret = VM_FAULT_NOPAGE;
1011         struct page *zero_page;
1012         void *entry2;
1013         pfn_t pfn;
1014
1015         zero_page = ZERO_PAGE(0);
1016         if (unlikely(!zero_page)) {
1017                 ret = VM_FAULT_OOM;
1018                 goto out;
1019         }
1020
1021         pfn = page_to_pfn_t(zero_page);
1022         entry2 = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1023                         RADIX_DAX_ZERO_PAGE, false);
1024         if (IS_ERR(entry2)) {
1025                 ret = VM_FAULT_SIGBUS;
1026                 goto out;
1027         }
1028
1029         vm_insert_mixed(vmf->vma, vaddr, pfn);
1030 out:
1031         trace_dax_load_hole(inode, vmf, ret);
1032         return ret;
1033 }
1034
1035 static bool dax_range_is_aligned(struct block_device *bdev,
1036                                  unsigned int offset, unsigned int length)
1037 {
1038         unsigned short sector_size = bdev_logical_block_size(bdev);
1039
1040         if (!IS_ALIGNED(offset, sector_size))
1041                 return false;
1042         if (!IS_ALIGNED(length, sector_size))
1043                 return false;
1044
1045         return true;
1046 }
1047
1048 int __dax_zero_page_range(struct block_device *bdev,
1049                 struct dax_device *dax_dev, sector_t sector,
1050                 unsigned int offset, unsigned int size)
1051 {
1052         if (dax_range_is_aligned(bdev, offset, size)) {
1053                 sector_t start_sector = sector + (offset >> 9);
1054
1055                 return blkdev_issue_zeroout(bdev, start_sector,
1056                                 size >> 9, GFP_NOFS, 0);
1057         } else {
1058                 pgoff_t pgoff;
1059                 long rc, id;
1060                 void *kaddr;
1061                 pfn_t pfn;
1062
1063                 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
1064                 if (rc)
1065                         return rc;
1066
1067                 id = dax_read_lock();
1068                 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr,
1069                                 &pfn);
1070                 if (rc < 0) {
1071                         dax_read_unlock(id);
1072                         return rc;
1073                 }
1074                 memset(kaddr + offset, 0, size);
1075                 dax_flush(dax_dev, kaddr + offset, size);
1076                 dax_read_unlock(id);
1077         }
1078         return 0;
1079 }
1080 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1081
1082 static loff_t
1083 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1084                 struct iomap *iomap)
1085 {
1086         struct block_device *bdev = iomap->bdev;
1087         struct dax_device *dax_dev = iomap->dax_dev;
1088         struct iov_iter *iter = data;
1089         loff_t end = pos + length, done = 0;
1090         ssize_t ret = 0;
1091         int id;
1092
1093         if (iov_iter_rw(iter) == READ) {
1094                 end = min(end, i_size_read(inode));
1095                 if (pos >= end)
1096                         return 0;
1097
1098                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1099                         return iov_iter_zero(min(length, end - pos), iter);
1100         }
1101
1102         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1103                 return -EIO;
1104
1105         /*
1106          * Write can allocate block for an area which has a hole page mapped
1107          * into page tables. We have to tear down these mappings so that data
1108          * written by write(2) is visible in mmap.
1109          */
1110         if (iomap->flags & IOMAP_F_NEW) {
1111                 invalidate_inode_pages2_range(inode->i_mapping,
1112                                               pos >> PAGE_SHIFT,
1113                                               (end - 1) >> PAGE_SHIFT);
1114         }
1115
1116         id = dax_read_lock();
1117         while (pos < end) {
1118                 unsigned offset = pos & (PAGE_SIZE - 1);
1119                 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1120                 const sector_t sector = dax_iomap_sector(iomap, pos);
1121                 ssize_t map_len;
1122                 pgoff_t pgoff;
1123                 void *kaddr;
1124                 pfn_t pfn;
1125
1126                 if (fatal_signal_pending(current)) {
1127                         ret = -EINTR;
1128                         break;
1129                 }
1130
1131                 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1132                 if (ret)
1133                         break;
1134
1135                 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1136                                 &kaddr, &pfn);
1137                 if (map_len < 0) {
1138                         ret = map_len;
1139                         break;
1140                 }
1141
1142                 map_len = PFN_PHYS(map_len);
1143                 kaddr += offset;
1144                 map_len -= offset;
1145                 if (map_len > end - pos)
1146                         map_len = end - pos;
1147
1148                 /*
1149                  * The userspace address for the memory copy has already been
1150                  * validated via access_ok() in either vfs_read() or
1151                  * vfs_write(), depending on which operation we are doing.
1152                  */
1153                 if (iov_iter_rw(iter) == WRITE)
1154                         map_len = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1155                                         map_len, iter);
1156                 else
1157                         map_len = copy_to_iter(kaddr, map_len, iter);
1158                 if (map_len <= 0) {
1159                         ret = map_len ? map_len : -EFAULT;
1160                         break;
1161                 }
1162
1163                 pos += map_len;
1164                 length -= map_len;
1165                 done += map_len;
1166         }
1167         dax_read_unlock(id);
1168
1169         return done ? done : ret;
1170 }
1171
1172 /**
1173  * dax_iomap_rw - Perform I/O to a DAX file
1174  * @iocb:       The control block for this I/O
1175  * @iter:       The addresses to do I/O from or to
1176  * @ops:        iomap ops passed from the file system
1177  *
1178  * This function performs read and write operations to directly mapped
1179  * persistent memory.  The callers needs to take care of read/write exclusion
1180  * and evicting any page cache pages in the region under I/O.
1181  */
1182 ssize_t
1183 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1184                 const struct iomap_ops *ops)
1185 {
1186         struct address_space *mapping = iocb->ki_filp->f_mapping;
1187         struct inode *inode = mapping->host;
1188         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1189         unsigned flags = 0;
1190
1191         if (iov_iter_rw(iter) == WRITE) {
1192                 lockdep_assert_held_exclusive(&inode->i_rwsem);
1193                 flags |= IOMAP_WRITE;
1194         } else {
1195                 lockdep_assert_held(&inode->i_rwsem);
1196         }
1197
1198         while (iov_iter_count(iter)) {
1199                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1200                                 iter, dax_iomap_actor);
1201                 if (ret <= 0)
1202                         break;
1203                 pos += ret;
1204                 done += ret;
1205         }
1206
1207         iocb->ki_pos += done;
1208         return done ? done : ret;
1209 }
1210 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1211
1212 static int dax_fault_return(int error)
1213 {
1214         if (error == 0)
1215                 return VM_FAULT_NOPAGE;
1216         if (error == -ENOMEM)
1217                 return VM_FAULT_OOM;
1218         return VM_FAULT_SIGBUS;
1219 }
1220
1221 /*
1222  * MAP_SYNC on a dax mapping guarantees dirty metadata is
1223  * flushed on write-faults (non-cow), but not read-faults.
1224  */
1225 static bool dax_fault_is_synchronous(unsigned long flags,
1226                 struct vm_area_struct *vma, struct iomap *iomap)
1227 {
1228         return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1229                 && (iomap->flags & IOMAP_F_DIRTY);
1230 }
1231
1232 static int dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1233                                int *iomap_errp, const struct iomap_ops *ops)
1234 {
1235         struct vm_area_struct *vma = vmf->vma;
1236         struct address_space *mapping = vma->vm_file->f_mapping;
1237         struct inode *inode = mapping->host;
1238         unsigned long vaddr = vmf->address;
1239         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1240         struct iomap iomap = { 0 };
1241         unsigned flags = IOMAP_FAULT;
1242         int error, major = 0;
1243         bool write = vmf->flags & FAULT_FLAG_WRITE;
1244         bool sync;
1245         int vmf_ret = 0;
1246         void *entry;
1247         pfn_t pfn;
1248
1249         trace_dax_pte_fault(inode, vmf, vmf_ret);
1250         /*
1251          * Check whether offset isn't beyond end of file now. Caller is supposed
1252          * to hold locks serializing us with truncate / punch hole so this is
1253          * a reliable test.
1254          */
1255         if (pos >= i_size_read(inode)) {
1256                 vmf_ret = VM_FAULT_SIGBUS;
1257                 goto out;
1258         }
1259
1260         if (write && !vmf->cow_page)
1261                 flags |= IOMAP_WRITE;
1262
1263         entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1264         if (IS_ERR(entry)) {
1265                 vmf_ret = dax_fault_return(PTR_ERR(entry));
1266                 goto out;
1267         }
1268
1269         /*
1270          * It is possible, particularly with mixed reads & writes to private
1271          * mappings, that we have raced with a PMD fault that overlaps with
1272          * the PTE we need to set up.  If so just return and the fault will be
1273          * retried.
1274          */
1275         if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1276                 vmf_ret = VM_FAULT_NOPAGE;
1277                 goto unlock_entry;
1278         }
1279
1280         /*
1281          * Note that we don't bother to use iomap_apply here: DAX required
1282          * the file system block size to be equal the page size, which means
1283          * that we never have to deal with more than a single extent here.
1284          */
1285         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1286         if (iomap_errp)
1287                 *iomap_errp = error;
1288         if (error) {
1289                 vmf_ret = dax_fault_return(error);
1290                 goto unlock_entry;
1291         }
1292         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1293                 error = -EIO;   /* fs corruption? */
1294                 goto error_finish_iomap;
1295         }
1296
1297         if (vmf->cow_page) {
1298                 sector_t sector = dax_iomap_sector(&iomap, pos);
1299
1300                 switch (iomap.type) {
1301                 case IOMAP_HOLE:
1302                 case IOMAP_UNWRITTEN:
1303                         clear_user_highpage(vmf->cow_page, vaddr);
1304                         break;
1305                 case IOMAP_MAPPED:
1306                         error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1307                                         sector, PAGE_SIZE, vmf->cow_page, vaddr);
1308                         break;
1309                 default:
1310                         WARN_ON_ONCE(1);
1311                         error = -EIO;
1312                         break;
1313                 }
1314
1315                 if (error)
1316                         goto error_finish_iomap;
1317
1318                 __SetPageUptodate(vmf->cow_page);
1319                 vmf_ret = finish_fault(vmf);
1320                 if (!vmf_ret)
1321                         vmf_ret = VM_FAULT_DONE_COW;
1322                 goto finish_iomap;
1323         }
1324
1325         sync = dax_fault_is_synchronous(flags, vma, &iomap);
1326
1327         switch (iomap.type) {
1328         case IOMAP_MAPPED:
1329                 if (iomap.flags & IOMAP_F_NEW) {
1330                         count_vm_event(PGMAJFAULT);
1331                         count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1332                         major = VM_FAULT_MAJOR;
1333                 }
1334                 error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1335                 if (error < 0)
1336                         goto error_finish_iomap;
1337
1338                 entry = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1339                                                  0, write && !sync);
1340                 if (IS_ERR(entry)) {
1341                         error = PTR_ERR(entry);
1342                         goto error_finish_iomap;
1343                 }
1344
1345                 /*
1346                  * If we are doing synchronous page fault and inode needs fsync,
1347                  * we can insert PTE into page tables only after that happens.
1348                  * Skip insertion for now and return the pfn so that caller can
1349                  * insert it after fsync is done.
1350                  */
1351                 if (sync) {
1352                         if (WARN_ON_ONCE(!pfnp)) {
1353                                 error = -EIO;
1354                                 goto error_finish_iomap;
1355                         }
1356                         *pfnp = pfn;
1357                         vmf_ret = VM_FAULT_NEEDDSYNC | major;
1358                         goto finish_iomap;
1359                 }
1360                 trace_dax_insert_mapping(inode, vmf, entry);
1361                 if (write)
1362                         error = vm_insert_mixed_mkwrite(vma, vaddr, pfn);
1363                 else
1364                         error = vm_insert_mixed(vma, vaddr, pfn);
1365
1366                 /* -EBUSY is fine, somebody else faulted on the same PTE */
1367                 if (error == -EBUSY)
1368                         error = 0;
1369                 break;
1370         case IOMAP_UNWRITTEN:
1371         case IOMAP_HOLE:
1372                 if (!write) {
1373                         vmf_ret = dax_load_hole(mapping, entry, vmf);
1374                         goto finish_iomap;
1375                 }
1376                 /*FALLTHRU*/
1377         default:
1378                 WARN_ON_ONCE(1);
1379                 error = -EIO;
1380                 break;
1381         }
1382
1383  error_finish_iomap:
1384         vmf_ret = dax_fault_return(error) | major;
1385  finish_iomap:
1386         if (ops->iomap_end) {
1387                 int copied = PAGE_SIZE;
1388
1389                 if (vmf_ret & VM_FAULT_ERROR)
1390                         copied = 0;
1391                 /*
1392                  * The fault is done by now and there's no way back (other
1393                  * thread may be already happily using PTE we have installed).
1394                  * Just ignore error from ->iomap_end since we cannot do much
1395                  * with it.
1396                  */
1397                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1398         }
1399  unlock_entry:
1400         put_locked_mapping_entry(mapping, vmf->pgoff);
1401  out:
1402         trace_dax_pte_fault_done(inode, vmf, vmf_ret);
1403         return vmf_ret;
1404 }
1405
1406 #ifdef CONFIG_FS_DAX_PMD
1407 static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
1408                 void *entry)
1409 {
1410         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1411         unsigned long pmd_addr = vmf->address & PMD_MASK;
1412         struct inode *inode = mapping->host;
1413         struct page *zero_page;
1414         void *ret = NULL;
1415         spinlock_t *ptl;
1416         pmd_t pmd_entry;
1417         pfn_t pfn;
1418
1419         zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1420
1421         if (unlikely(!zero_page))
1422                 goto fallback;
1423
1424         pfn = page_to_pfn_t(zero_page);
1425         ret = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1426                         RADIX_DAX_PMD | RADIX_DAX_ZERO_PAGE, false);
1427         if (IS_ERR(ret))
1428                 goto fallback;
1429
1430         ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1431         if (!pmd_none(*(vmf->pmd))) {
1432                 spin_unlock(ptl);
1433                 goto fallback;
1434         }
1435
1436         pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1437         pmd_entry = pmd_mkhuge(pmd_entry);
1438         set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1439         spin_unlock(ptl);
1440         trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
1441         return VM_FAULT_NOPAGE;
1442
1443 fallback:
1444         trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
1445         return VM_FAULT_FALLBACK;
1446 }
1447
1448 static int dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1449                                const struct iomap_ops *ops)
1450 {
1451         struct vm_area_struct *vma = vmf->vma;
1452         struct address_space *mapping = vma->vm_file->f_mapping;
1453         unsigned long pmd_addr = vmf->address & PMD_MASK;
1454         bool write = vmf->flags & FAULT_FLAG_WRITE;
1455         bool sync;
1456         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1457         struct inode *inode = mapping->host;
1458         int result = VM_FAULT_FALLBACK;
1459         struct iomap iomap = { 0 };
1460         pgoff_t max_pgoff, pgoff;
1461         void *entry;
1462         loff_t pos;
1463         int error;
1464         pfn_t pfn;
1465
1466         /*
1467          * Check whether offset isn't beyond end of file now. Caller is
1468          * supposed to hold locks serializing us with truncate / punch hole so
1469          * this is a reliable test.
1470          */
1471         pgoff = linear_page_index(vma, pmd_addr);
1472         max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1473
1474         trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1475
1476         /*
1477          * Make sure that the faulting address's PMD offset (color) matches
1478          * the PMD offset from the start of the file.  This is necessary so
1479          * that a PMD range in the page table overlaps exactly with a PMD
1480          * range in the radix tree.
1481          */
1482         if ((vmf->pgoff & PG_PMD_COLOUR) !=
1483             ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1484                 goto fallback;
1485
1486         /* Fall back to PTEs if we're going to COW */
1487         if (write && !(vma->vm_flags & VM_SHARED))
1488                 goto fallback;
1489
1490         /* If the PMD would extend outside the VMA */
1491         if (pmd_addr < vma->vm_start)
1492                 goto fallback;
1493         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1494                 goto fallback;
1495
1496         if (pgoff >= max_pgoff) {
1497                 result = VM_FAULT_SIGBUS;
1498                 goto out;
1499         }
1500
1501         /* If the PMD would extend beyond the file size */
1502         if ((pgoff | PG_PMD_COLOUR) >= max_pgoff)
1503                 goto fallback;
1504
1505         /*
1506          * grab_mapping_entry() will make sure we get a 2MiB empty entry, a
1507          * 2MiB zero page entry or a DAX PMD.  If it can't (because a 4k page
1508          * is already in the tree, for instance), it will return -EEXIST and
1509          * we just fall back to 4k entries.
1510          */
1511         entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1512         if (IS_ERR(entry))
1513                 goto fallback;
1514
1515         /*
1516          * It is possible, particularly with mixed reads & writes to private
1517          * mappings, that we have raced with a PTE fault that overlaps with
1518          * the PMD we need to set up.  If so just return and the fault will be
1519          * retried.
1520          */
1521         if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1522                         !pmd_devmap(*vmf->pmd)) {
1523                 result = 0;
1524                 goto unlock_entry;
1525         }
1526
1527         /*
1528          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1529          * setting up a mapping, so really we're using iomap_begin() as a way
1530          * to look up our filesystem block.
1531          */
1532         pos = (loff_t)pgoff << PAGE_SHIFT;
1533         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1534         if (error)
1535                 goto unlock_entry;
1536
1537         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1538                 goto finish_iomap;
1539
1540         sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1541
1542         switch (iomap.type) {
1543         case IOMAP_MAPPED:
1544                 error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1545                 if (error < 0)
1546                         goto finish_iomap;
1547
1548                 entry = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1549                                                 RADIX_DAX_PMD, write && !sync);
1550                 if (IS_ERR(entry))
1551                         goto finish_iomap;
1552
1553                 /*
1554                  * If we are doing synchronous page fault and inode needs fsync,
1555                  * we can insert PMD into page tables only after that happens.
1556                  * Skip insertion for now and return the pfn so that caller can
1557                  * insert it after fsync is done.
1558                  */
1559                 if (sync) {
1560                         if (WARN_ON_ONCE(!pfnp))
1561                                 goto finish_iomap;
1562                         *pfnp = pfn;
1563                         result = VM_FAULT_NEEDDSYNC;
1564                         goto finish_iomap;
1565                 }
1566
1567                 trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
1568                 result = vmf_insert_pfn_pmd(vma, vmf->address, vmf->pmd, pfn,
1569                                             write);
1570                 break;
1571         case IOMAP_UNWRITTEN:
1572         case IOMAP_HOLE:
1573                 if (WARN_ON_ONCE(write))
1574                         break;
1575                 result = dax_pmd_load_hole(vmf, &iomap, entry);
1576                 break;
1577         default:
1578                 WARN_ON_ONCE(1);
1579                 break;
1580         }
1581
1582  finish_iomap:
1583         if (ops->iomap_end) {
1584                 int copied = PMD_SIZE;
1585
1586                 if (result == VM_FAULT_FALLBACK)
1587                         copied = 0;
1588                 /*
1589                  * The fault is done by now and there's no way back (other
1590                  * thread may be already happily using PMD we have installed).
1591                  * Just ignore error from ->iomap_end since we cannot do much
1592                  * with it.
1593                  */
1594                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1595                                 &iomap);
1596         }
1597  unlock_entry:
1598         put_locked_mapping_entry(mapping, pgoff);
1599  fallback:
1600         if (result == VM_FAULT_FALLBACK) {
1601                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1602                 count_vm_event(THP_FAULT_FALLBACK);
1603         }
1604 out:
1605         trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1606         return result;
1607 }
1608 #else
1609 static int dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1610                                const struct iomap_ops *ops)
1611 {
1612         return VM_FAULT_FALLBACK;
1613 }
1614 #endif /* CONFIG_FS_DAX_PMD */
1615
1616 /**
1617  * dax_iomap_fault - handle a page fault on a DAX file
1618  * @vmf: The description of the fault
1619  * @pe_size: Size of the page to fault in
1620  * @pfnp: PFN to insert for synchronous faults if fsync is required
1621  * @iomap_errp: Storage for detailed error code in case of error
1622  * @ops: Iomap ops passed from the file system
1623  *
1624  * When a page fault occurs, filesystems may call this helper in
1625  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1626  * has done all the necessary locking for page fault to proceed
1627  * successfully.
1628  */
1629 int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1630                     pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1631 {
1632         switch (pe_size) {
1633         case PE_SIZE_PTE:
1634                 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1635         case PE_SIZE_PMD:
1636                 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1637         default:
1638                 return VM_FAULT_FALLBACK;
1639         }
1640 }
1641 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1642
1643 /**
1644  * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1645  * @vmf: The description of the fault
1646  * @pe_size: Size of entry to be inserted
1647  * @pfn: PFN to insert
1648  *
1649  * This function inserts writeable PTE or PMD entry into page tables for mmaped
1650  * DAX file.  It takes care of marking corresponding radix tree entry as dirty
1651  * as well.
1652  */
1653 static int dax_insert_pfn_mkwrite(struct vm_fault *vmf,
1654                                   enum page_entry_size pe_size,
1655                                   pfn_t pfn)
1656 {
1657         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1658         void *entry, **slot;
1659         pgoff_t index = vmf->pgoff;
1660         int vmf_ret, error;
1661
1662         xa_lock_irq(&mapping->i_pages);
1663         entry = get_unlocked_mapping_entry(mapping, index, &slot);
1664         /* Did we race with someone splitting entry or so? */
1665         if (!entry ||
1666             (pe_size == PE_SIZE_PTE && !dax_is_pte_entry(entry)) ||
1667             (pe_size == PE_SIZE_PMD && !dax_is_pmd_entry(entry))) {
1668                 put_unlocked_mapping_entry(mapping, index, entry);
1669                 xa_unlock_irq(&mapping->i_pages);
1670                 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1671                                                       VM_FAULT_NOPAGE);
1672                 return VM_FAULT_NOPAGE;
1673         }
1674         radix_tree_tag_set(&mapping->i_pages, index, PAGECACHE_TAG_DIRTY);
1675         entry = lock_slot(mapping, slot);
1676         xa_unlock_irq(&mapping->i_pages);
1677         switch (pe_size) {
1678         case PE_SIZE_PTE:
1679                 error = vm_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1680                 vmf_ret = dax_fault_return(error);
1681                 break;
1682 #ifdef CONFIG_FS_DAX_PMD
1683         case PE_SIZE_PMD:
1684                 vmf_ret = vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
1685                         pfn, true);
1686                 break;
1687 #endif
1688         default:
1689                 vmf_ret = VM_FAULT_FALLBACK;
1690         }
1691         put_locked_mapping_entry(mapping, index);
1692         trace_dax_insert_pfn_mkwrite(mapping->host, vmf, vmf_ret);
1693         return vmf_ret;
1694 }
1695
1696 /**
1697  * dax_finish_sync_fault - finish synchronous page fault
1698  * @vmf: The description of the fault
1699  * @pe_size: Size of entry to be inserted
1700  * @pfn: PFN to insert
1701  *
1702  * This function ensures that the file range touched by the page fault is
1703  * stored persistently on the media and handles inserting of appropriate page
1704  * table entry.
1705  */
1706 int dax_finish_sync_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1707                           pfn_t pfn)
1708 {
1709         int err;
1710         loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1711         size_t len = 0;
1712
1713         if (pe_size == PE_SIZE_PTE)
1714                 len = PAGE_SIZE;
1715         else if (pe_size == PE_SIZE_PMD)
1716                 len = PMD_SIZE;
1717         else
1718                 WARN_ON_ONCE(1);
1719         err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1720         if (err)
1721                 return VM_FAULT_SIGBUS;
1722         return dax_insert_pfn_mkwrite(vmf, pe_size, pfn);
1723 }
1724 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);