HID: hid-input: add mapping for emoji picker key
[linux-2.6-microblaze.git] / fs / dax.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * fs/dax.c - Direct Access filesystem code
4  * Copyright (c) 2013-2014 Intel Corporation
5  * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
6  * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
7  */
8
9 #include <linux/atomic.h>
10 #include <linux/blkdev.h>
11 #include <linux/buffer_head.h>
12 #include <linux/dax.h>
13 #include <linux/fs.h>
14 #include <linux/genhd.h>
15 #include <linux/highmem.h>
16 #include <linux/memcontrol.h>
17 #include <linux/mm.h>
18 #include <linux/mutex.h>
19 #include <linux/pagevec.h>
20 #include <linux/sched.h>
21 #include <linux/sched/signal.h>
22 #include <linux/uio.h>
23 #include <linux/vmstat.h>
24 #include <linux/pfn_t.h>
25 #include <linux/sizes.h>
26 #include <linux/mmu_notifier.h>
27 #include <linux/iomap.h>
28 #include <asm/pgalloc.h>
29
30 #define CREATE_TRACE_POINTS
31 #include <trace/events/fs_dax.h>
32
33 static inline unsigned int pe_order(enum page_entry_size pe_size)
34 {
35         if (pe_size == PE_SIZE_PTE)
36                 return PAGE_SHIFT - PAGE_SHIFT;
37         if (pe_size == PE_SIZE_PMD)
38                 return PMD_SHIFT - PAGE_SHIFT;
39         if (pe_size == PE_SIZE_PUD)
40                 return PUD_SHIFT - PAGE_SHIFT;
41         return ~0;
42 }
43
44 /* We choose 4096 entries - same as per-zone page wait tables */
45 #define DAX_WAIT_TABLE_BITS 12
46 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
47
48 /* The 'colour' (ie low bits) within a PMD of a page offset.  */
49 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
50 #define PG_PMD_NR       (PMD_SIZE >> PAGE_SHIFT)
51
52 /* The order of a PMD entry */
53 #define PMD_ORDER       (PMD_SHIFT - PAGE_SHIFT)
54
55 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
56
57 static int __init init_dax_wait_table(void)
58 {
59         int i;
60
61         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
62                 init_waitqueue_head(wait_table + i);
63         return 0;
64 }
65 fs_initcall(init_dax_wait_table);
66
67 /*
68  * DAX pagecache entries use XArray value entries so they can't be mistaken
69  * for pages.  We use one bit for locking, one bit for the entry size (PMD)
70  * and two more to tell us if the entry is a zero page or an empty entry that
71  * is just used for locking.  In total four special bits.
72  *
73  * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
74  * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
75  * block allocation.
76  */
77 #define DAX_SHIFT       (4)
78 #define DAX_LOCKED      (1UL << 0)
79 #define DAX_PMD         (1UL << 1)
80 #define DAX_ZERO_PAGE   (1UL << 2)
81 #define DAX_EMPTY       (1UL << 3)
82
83 static unsigned long dax_to_pfn(void *entry)
84 {
85         return xa_to_value(entry) >> DAX_SHIFT;
86 }
87
88 static void *dax_make_entry(pfn_t pfn, unsigned long flags)
89 {
90         return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
91 }
92
93 static bool dax_is_locked(void *entry)
94 {
95         return xa_to_value(entry) & DAX_LOCKED;
96 }
97
98 static unsigned int dax_entry_order(void *entry)
99 {
100         if (xa_to_value(entry) & DAX_PMD)
101                 return PMD_ORDER;
102         return 0;
103 }
104
105 static unsigned long dax_is_pmd_entry(void *entry)
106 {
107         return xa_to_value(entry) & DAX_PMD;
108 }
109
110 static bool dax_is_pte_entry(void *entry)
111 {
112         return !(xa_to_value(entry) & DAX_PMD);
113 }
114
115 static int dax_is_zero_entry(void *entry)
116 {
117         return xa_to_value(entry) & DAX_ZERO_PAGE;
118 }
119
120 static int dax_is_empty_entry(void *entry)
121 {
122         return xa_to_value(entry) & DAX_EMPTY;
123 }
124
125 /*
126  * true if the entry that was found is of a smaller order than the entry
127  * we were looking for
128  */
129 static bool dax_is_conflict(void *entry)
130 {
131         return entry == XA_RETRY_ENTRY;
132 }
133
134 /*
135  * DAX page cache entry locking
136  */
137 struct exceptional_entry_key {
138         struct xarray *xa;
139         pgoff_t entry_start;
140 };
141
142 struct wait_exceptional_entry_queue {
143         wait_queue_entry_t wait;
144         struct exceptional_entry_key key;
145 };
146
147 static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
148                 void *entry, struct exceptional_entry_key *key)
149 {
150         unsigned long hash;
151         unsigned long index = xas->xa_index;
152
153         /*
154          * If 'entry' is a PMD, align the 'index' that we use for the wait
155          * queue to the start of that PMD.  This ensures that all offsets in
156          * the range covered by the PMD map to the same bit lock.
157          */
158         if (dax_is_pmd_entry(entry))
159                 index &= ~PG_PMD_COLOUR;
160         key->xa = xas->xa;
161         key->entry_start = index;
162
163         hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
164         return wait_table + hash;
165 }
166
167 static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
168                 unsigned int mode, int sync, void *keyp)
169 {
170         struct exceptional_entry_key *key = keyp;
171         struct wait_exceptional_entry_queue *ewait =
172                 container_of(wait, struct wait_exceptional_entry_queue, wait);
173
174         if (key->xa != ewait->key.xa ||
175             key->entry_start != ewait->key.entry_start)
176                 return 0;
177         return autoremove_wake_function(wait, mode, sync, NULL);
178 }
179
180 /*
181  * @entry may no longer be the entry at the index in the mapping.
182  * The important information it's conveying is whether the entry at
183  * this index used to be a PMD entry.
184  */
185 static void dax_wake_entry(struct xa_state *xas, void *entry, bool wake_all)
186 {
187         struct exceptional_entry_key key;
188         wait_queue_head_t *wq;
189
190         wq = dax_entry_waitqueue(xas, entry, &key);
191
192         /*
193          * Checking for locked entry and prepare_to_wait_exclusive() happens
194          * under the i_pages lock, ditto for entry handling in our callers.
195          * So at this point all tasks that could have seen our entry locked
196          * must be in the waitqueue and the following check will see them.
197          */
198         if (waitqueue_active(wq))
199                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
200 }
201
202 /*
203  * Look up entry in page cache, wait for it to become unlocked if it
204  * is a DAX entry and return it.  The caller must subsequently call
205  * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
206  * if it did.  The entry returned may have a larger order than @order.
207  * If @order is larger than the order of the entry found in i_pages, this
208  * function returns a dax_is_conflict entry.
209  *
210  * Must be called with the i_pages lock held.
211  */
212 static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
213 {
214         void *entry;
215         struct wait_exceptional_entry_queue ewait;
216         wait_queue_head_t *wq;
217
218         init_wait(&ewait.wait);
219         ewait.wait.func = wake_exceptional_entry_func;
220
221         for (;;) {
222                 entry = xas_find_conflict(xas);
223                 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
224                         return entry;
225                 if (dax_entry_order(entry) < order)
226                         return XA_RETRY_ENTRY;
227                 if (!dax_is_locked(entry))
228                         return entry;
229
230                 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
231                 prepare_to_wait_exclusive(wq, &ewait.wait,
232                                           TASK_UNINTERRUPTIBLE);
233                 xas_unlock_irq(xas);
234                 xas_reset(xas);
235                 schedule();
236                 finish_wait(wq, &ewait.wait);
237                 xas_lock_irq(xas);
238         }
239 }
240
241 /*
242  * The only thing keeping the address space around is the i_pages lock
243  * (it's cycled in clear_inode() after removing the entries from i_pages)
244  * After we call xas_unlock_irq(), we cannot touch xas->xa.
245  */
246 static void wait_entry_unlocked(struct xa_state *xas, void *entry)
247 {
248         struct wait_exceptional_entry_queue ewait;
249         wait_queue_head_t *wq;
250
251         init_wait(&ewait.wait);
252         ewait.wait.func = wake_exceptional_entry_func;
253
254         wq = dax_entry_waitqueue(xas, entry, &ewait.key);
255         /*
256          * Unlike get_unlocked_entry() there is no guarantee that this
257          * path ever successfully retrieves an unlocked entry before an
258          * inode dies. Perform a non-exclusive wait in case this path
259          * never successfully performs its own wake up.
260          */
261         prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
262         xas_unlock_irq(xas);
263         schedule();
264         finish_wait(wq, &ewait.wait);
265 }
266
267 static void put_unlocked_entry(struct xa_state *xas, void *entry)
268 {
269         /* If we were the only waiter woken, wake the next one */
270         if (entry && !dax_is_conflict(entry))
271                 dax_wake_entry(xas, entry, false);
272 }
273
274 /*
275  * We used the xa_state to get the entry, but then we locked the entry and
276  * dropped the xa_lock, so we know the xa_state is stale and must be reset
277  * before use.
278  */
279 static void dax_unlock_entry(struct xa_state *xas, void *entry)
280 {
281         void *old;
282
283         BUG_ON(dax_is_locked(entry));
284         xas_reset(xas);
285         xas_lock_irq(xas);
286         old = xas_store(xas, entry);
287         xas_unlock_irq(xas);
288         BUG_ON(!dax_is_locked(old));
289         dax_wake_entry(xas, entry, false);
290 }
291
292 /*
293  * Return: The entry stored at this location before it was locked.
294  */
295 static void *dax_lock_entry(struct xa_state *xas, void *entry)
296 {
297         unsigned long v = xa_to_value(entry);
298         return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
299 }
300
301 static unsigned long dax_entry_size(void *entry)
302 {
303         if (dax_is_zero_entry(entry))
304                 return 0;
305         else if (dax_is_empty_entry(entry))
306                 return 0;
307         else if (dax_is_pmd_entry(entry))
308                 return PMD_SIZE;
309         else
310                 return PAGE_SIZE;
311 }
312
313 static unsigned long dax_end_pfn(void *entry)
314 {
315         return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
316 }
317
318 /*
319  * Iterate through all mapped pfns represented by an entry, i.e. skip
320  * 'empty' and 'zero' entries.
321  */
322 #define for_each_mapped_pfn(entry, pfn) \
323         for (pfn = dax_to_pfn(entry); \
324                         pfn < dax_end_pfn(entry); pfn++)
325
326 /*
327  * TODO: for reflink+dax we need a way to associate a single page with
328  * multiple address_space instances at different linear_page_index()
329  * offsets.
330  */
331 static void dax_associate_entry(void *entry, struct address_space *mapping,
332                 struct vm_area_struct *vma, unsigned long address)
333 {
334         unsigned long size = dax_entry_size(entry), pfn, index;
335         int i = 0;
336
337         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
338                 return;
339
340         index = linear_page_index(vma, address & ~(size - 1));
341         for_each_mapped_pfn(entry, pfn) {
342                 struct page *page = pfn_to_page(pfn);
343
344                 WARN_ON_ONCE(page->mapping);
345                 page->mapping = mapping;
346                 page->index = index + i++;
347         }
348 }
349
350 static void dax_disassociate_entry(void *entry, struct address_space *mapping,
351                 bool trunc)
352 {
353         unsigned long pfn;
354
355         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
356                 return;
357
358         for_each_mapped_pfn(entry, pfn) {
359                 struct page *page = pfn_to_page(pfn);
360
361                 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
362                 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
363                 page->mapping = NULL;
364                 page->index = 0;
365         }
366 }
367
368 static struct page *dax_busy_page(void *entry)
369 {
370         unsigned long pfn;
371
372         for_each_mapped_pfn(entry, pfn) {
373                 struct page *page = pfn_to_page(pfn);
374
375                 if (page_ref_count(page) > 1)
376                         return page;
377         }
378         return NULL;
379 }
380
381 /*
382  * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page
383  * @page: The page whose entry we want to lock
384  *
385  * Context: Process context.
386  * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
387  * not be locked.
388  */
389 dax_entry_t dax_lock_page(struct page *page)
390 {
391         XA_STATE(xas, NULL, 0);
392         void *entry;
393
394         /* Ensure page->mapping isn't freed while we look at it */
395         rcu_read_lock();
396         for (;;) {
397                 struct address_space *mapping = READ_ONCE(page->mapping);
398
399                 entry = NULL;
400                 if (!mapping || !dax_mapping(mapping))
401                         break;
402
403                 /*
404                  * In the device-dax case there's no need to lock, a
405                  * struct dev_pagemap pin is sufficient to keep the
406                  * inode alive, and we assume we have dev_pagemap pin
407                  * otherwise we would not have a valid pfn_to_page()
408                  * translation.
409                  */
410                 entry = (void *)~0UL;
411                 if (S_ISCHR(mapping->host->i_mode))
412                         break;
413
414                 xas.xa = &mapping->i_pages;
415                 xas_lock_irq(&xas);
416                 if (mapping != page->mapping) {
417                         xas_unlock_irq(&xas);
418                         continue;
419                 }
420                 xas_set(&xas, page->index);
421                 entry = xas_load(&xas);
422                 if (dax_is_locked(entry)) {
423                         rcu_read_unlock();
424                         wait_entry_unlocked(&xas, entry);
425                         rcu_read_lock();
426                         continue;
427                 }
428                 dax_lock_entry(&xas, entry);
429                 xas_unlock_irq(&xas);
430                 break;
431         }
432         rcu_read_unlock();
433         return (dax_entry_t)entry;
434 }
435
436 void dax_unlock_page(struct page *page, dax_entry_t cookie)
437 {
438         struct address_space *mapping = page->mapping;
439         XA_STATE(xas, &mapping->i_pages, page->index);
440
441         if (S_ISCHR(mapping->host->i_mode))
442                 return;
443
444         dax_unlock_entry(&xas, (void *)cookie);
445 }
446
447 /*
448  * Find page cache entry at given index. If it is a DAX entry, return it
449  * with the entry locked. If the page cache doesn't contain an entry at
450  * that index, add a locked empty entry.
451  *
452  * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
453  * either return that locked entry or will return VM_FAULT_FALLBACK.
454  * This will happen if there are any PTE entries within the PMD range
455  * that we are requesting.
456  *
457  * We always favor PTE entries over PMD entries. There isn't a flow where we
458  * evict PTE entries in order to 'upgrade' them to a PMD entry.  A PMD
459  * insertion will fail if it finds any PTE entries already in the tree, and a
460  * PTE insertion will cause an existing PMD entry to be unmapped and
461  * downgraded to PTE entries.  This happens for both PMD zero pages as
462  * well as PMD empty entries.
463  *
464  * The exception to this downgrade path is for PMD entries that have
465  * real storage backing them.  We will leave these real PMD entries in
466  * the tree, and PTE writes will simply dirty the entire PMD entry.
467  *
468  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
469  * persistent memory the benefit is doubtful. We can add that later if we can
470  * show it helps.
471  *
472  * On error, this function does not return an ERR_PTR.  Instead it returns
473  * a VM_FAULT code, encoded as an xarray internal entry.  The ERR_PTR values
474  * overlap with xarray value entries.
475  */
476 static void *grab_mapping_entry(struct xa_state *xas,
477                 struct address_space *mapping, unsigned int order)
478 {
479         unsigned long index = xas->xa_index;
480         bool pmd_downgrade = false; /* splitting PMD entry into PTE entries? */
481         void *entry;
482
483 retry:
484         xas_lock_irq(xas);
485         entry = get_unlocked_entry(xas, order);
486
487         if (entry) {
488                 if (dax_is_conflict(entry))
489                         goto fallback;
490                 if (!xa_is_value(entry)) {
491                         xas_set_err(xas, -EIO);
492                         goto out_unlock;
493                 }
494
495                 if (order == 0) {
496                         if (dax_is_pmd_entry(entry) &&
497                             (dax_is_zero_entry(entry) ||
498                              dax_is_empty_entry(entry))) {
499                                 pmd_downgrade = true;
500                         }
501                 }
502         }
503
504         if (pmd_downgrade) {
505                 /*
506                  * Make sure 'entry' remains valid while we drop
507                  * the i_pages lock.
508                  */
509                 dax_lock_entry(xas, entry);
510
511                 /*
512                  * Besides huge zero pages the only other thing that gets
513                  * downgraded are empty entries which don't need to be
514                  * unmapped.
515                  */
516                 if (dax_is_zero_entry(entry)) {
517                         xas_unlock_irq(xas);
518                         unmap_mapping_pages(mapping,
519                                         xas->xa_index & ~PG_PMD_COLOUR,
520                                         PG_PMD_NR, false);
521                         xas_reset(xas);
522                         xas_lock_irq(xas);
523                 }
524
525                 dax_disassociate_entry(entry, mapping, false);
526                 xas_store(xas, NULL);   /* undo the PMD join */
527                 dax_wake_entry(xas, entry, true);
528                 mapping->nrexceptional--;
529                 entry = NULL;
530                 xas_set(xas, index);
531         }
532
533         if (entry) {
534                 dax_lock_entry(xas, entry);
535         } else {
536                 unsigned long flags = DAX_EMPTY;
537
538                 if (order > 0)
539                         flags |= DAX_PMD;
540                 entry = dax_make_entry(pfn_to_pfn_t(0), flags);
541                 dax_lock_entry(xas, entry);
542                 if (xas_error(xas))
543                         goto out_unlock;
544                 mapping->nrexceptional++;
545         }
546
547 out_unlock:
548         xas_unlock_irq(xas);
549         if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
550                 goto retry;
551         if (xas->xa_node == XA_ERROR(-ENOMEM))
552                 return xa_mk_internal(VM_FAULT_OOM);
553         if (xas_error(xas))
554                 return xa_mk_internal(VM_FAULT_SIGBUS);
555         return entry;
556 fallback:
557         xas_unlock_irq(xas);
558         return xa_mk_internal(VM_FAULT_FALLBACK);
559 }
560
561 /**
562  * dax_layout_busy_page_range - find first pinned page in @mapping
563  * @mapping: address space to scan for a page with ref count > 1
564  * @start: Starting offset. Page containing 'start' is included.
565  * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
566  *       pages from 'start' till the end of file are included.
567  *
568  * DAX requires ZONE_DEVICE mapped pages. These pages are never
569  * 'onlined' to the page allocator so they are considered idle when
570  * page->count == 1. A filesystem uses this interface to determine if
571  * any page in the mapping is busy, i.e. for DMA, or other
572  * get_user_pages() usages.
573  *
574  * It is expected that the filesystem is holding locks to block the
575  * establishment of new mappings in this address_space. I.e. it expects
576  * to be able to run unmap_mapping_range() and subsequently not race
577  * mapping_mapped() becoming true.
578  */
579 struct page *dax_layout_busy_page_range(struct address_space *mapping,
580                                         loff_t start, loff_t end)
581 {
582         void *entry;
583         unsigned int scanned = 0;
584         struct page *page = NULL;
585         pgoff_t start_idx = start >> PAGE_SHIFT;
586         pgoff_t end_idx;
587         XA_STATE(xas, &mapping->i_pages, start_idx);
588
589         /*
590          * In the 'limited' case get_user_pages() for dax is disabled.
591          */
592         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
593                 return NULL;
594
595         if (!dax_mapping(mapping) || !mapping_mapped(mapping))
596                 return NULL;
597
598         /* If end == LLONG_MAX, all pages from start to till end of file */
599         if (end == LLONG_MAX)
600                 end_idx = ULONG_MAX;
601         else
602                 end_idx = end >> PAGE_SHIFT;
603         /*
604          * If we race get_user_pages_fast() here either we'll see the
605          * elevated page count in the iteration and wait, or
606          * get_user_pages_fast() will see that the page it took a reference
607          * against is no longer mapped in the page tables and bail to the
608          * get_user_pages() slow path.  The slow path is protected by
609          * pte_lock() and pmd_lock(). New references are not taken without
610          * holding those locks, and unmap_mapping_pages() will not zero the
611          * pte or pmd without holding the respective lock, so we are
612          * guaranteed to either see new references or prevent new
613          * references from being established.
614          */
615         unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0);
616
617         xas_lock_irq(&xas);
618         xas_for_each(&xas, entry, end_idx) {
619                 if (WARN_ON_ONCE(!xa_is_value(entry)))
620                         continue;
621                 if (unlikely(dax_is_locked(entry)))
622                         entry = get_unlocked_entry(&xas, 0);
623                 if (entry)
624                         page = dax_busy_page(entry);
625                 put_unlocked_entry(&xas, entry);
626                 if (page)
627                         break;
628                 if (++scanned % XA_CHECK_SCHED)
629                         continue;
630
631                 xas_pause(&xas);
632                 xas_unlock_irq(&xas);
633                 cond_resched();
634                 xas_lock_irq(&xas);
635         }
636         xas_unlock_irq(&xas);
637         return page;
638 }
639 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
640
641 struct page *dax_layout_busy_page(struct address_space *mapping)
642 {
643         return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
644 }
645 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
646
647 static int __dax_invalidate_entry(struct address_space *mapping,
648                                           pgoff_t index, bool trunc)
649 {
650         XA_STATE(xas, &mapping->i_pages, index);
651         int ret = 0;
652         void *entry;
653
654         xas_lock_irq(&xas);
655         entry = get_unlocked_entry(&xas, 0);
656         if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
657                 goto out;
658         if (!trunc &&
659             (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
660              xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
661                 goto out;
662         dax_disassociate_entry(entry, mapping, trunc);
663         xas_store(&xas, NULL);
664         mapping->nrexceptional--;
665         ret = 1;
666 out:
667         put_unlocked_entry(&xas, entry);
668         xas_unlock_irq(&xas);
669         return ret;
670 }
671
672 /*
673  * Delete DAX entry at @index from @mapping.  Wait for it
674  * to be unlocked before deleting it.
675  */
676 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
677 {
678         int ret = __dax_invalidate_entry(mapping, index, true);
679
680         /*
681          * This gets called from truncate / punch_hole path. As such, the caller
682          * must hold locks protecting against concurrent modifications of the
683          * page cache (usually fs-private i_mmap_sem for writing). Since the
684          * caller has seen a DAX entry for this index, we better find it
685          * at that index as well...
686          */
687         WARN_ON_ONCE(!ret);
688         return ret;
689 }
690
691 /*
692  * Invalidate DAX entry if it is clean.
693  */
694 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
695                                       pgoff_t index)
696 {
697         return __dax_invalidate_entry(mapping, index, false);
698 }
699
700 static int copy_cow_page_dax(struct block_device *bdev, struct dax_device *dax_dev,
701                              sector_t sector, struct page *to, unsigned long vaddr)
702 {
703         void *vto, *kaddr;
704         pgoff_t pgoff;
705         long rc;
706         int id;
707
708         rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
709         if (rc)
710                 return rc;
711
712         id = dax_read_lock();
713         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(PAGE_SIZE), &kaddr, NULL);
714         if (rc < 0) {
715                 dax_read_unlock(id);
716                 return rc;
717         }
718         vto = kmap_atomic(to);
719         copy_user_page(vto, (void __force *)kaddr, vaddr, to);
720         kunmap_atomic(vto);
721         dax_read_unlock(id);
722         return 0;
723 }
724
725 /*
726  * By this point grab_mapping_entry() has ensured that we have a locked entry
727  * of the appropriate size so we don't have to worry about downgrading PMDs to
728  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
729  * already in the tree, we will skip the insertion and just dirty the PMD as
730  * appropriate.
731  */
732 static void *dax_insert_entry(struct xa_state *xas,
733                 struct address_space *mapping, struct vm_fault *vmf,
734                 void *entry, pfn_t pfn, unsigned long flags, bool dirty)
735 {
736         void *new_entry = dax_make_entry(pfn, flags);
737
738         if (dirty)
739                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
740
741         if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) {
742                 unsigned long index = xas->xa_index;
743                 /* we are replacing a zero page with block mapping */
744                 if (dax_is_pmd_entry(entry))
745                         unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
746                                         PG_PMD_NR, false);
747                 else /* pte entry */
748                         unmap_mapping_pages(mapping, index, 1, false);
749         }
750
751         xas_reset(xas);
752         xas_lock_irq(xas);
753         if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
754                 void *old;
755
756                 dax_disassociate_entry(entry, mapping, false);
757                 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address);
758                 /*
759                  * Only swap our new entry into the page cache if the current
760                  * entry is a zero page or an empty entry.  If a normal PTE or
761                  * PMD entry is already in the cache, we leave it alone.  This
762                  * means that if we are trying to insert a PTE and the
763                  * existing entry is a PMD, we will just leave the PMD in the
764                  * tree and dirty it if necessary.
765                  */
766                 old = dax_lock_entry(xas, new_entry);
767                 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
768                                         DAX_LOCKED));
769                 entry = new_entry;
770         } else {
771                 xas_load(xas);  /* Walk the xa_state */
772         }
773
774         if (dirty)
775                 xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
776
777         xas_unlock_irq(xas);
778         return entry;
779 }
780
781 static inline
782 unsigned long pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
783 {
784         unsigned long address;
785
786         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
787         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
788         return address;
789 }
790
791 /* Walk all mappings of a given index of a file and writeprotect them */
792 static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
793                 unsigned long pfn)
794 {
795         struct vm_area_struct *vma;
796         pte_t pte, *ptep = NULL;
797         pmd_t *pmdp = NULL;
798         spinlock_t *ptl;
799
800         i_mmap_lock_read(mapping);
801         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
802                 struct mmu_notifier_range range;
803                 unsigned long address;
804
805                 cond_resched();
806
807                 if (!(vma->vm_flags & VM_SHARED))
808                         continue;
809
810                 address = pgoff_address(index, vma);
811
812                 /*
813                  * follow_invalidate_pte() will use the range to call
814                  * mmu_notifier_invalidate_range_start() on our behalf before
815                  * taking any lock.
816                  */
817                 if (follow_invalidate_pte(vma->vm_mm, address, &range, &ptep,
818                                           &pmdp, &ptl))
819                         continue;
820
821                 /*
822                  * No need to call mmu_notifier_invalidate_range() as we are
823                  * downgrading page table protection not changing it to point
824                  * to a new page.
825                  *
826                  * See Documentation/vm/mmu_notifier.rst
827                  */
828                 if (pmdp) {
829 #ifdef CONFIG_FS_DAX_PMD
830                         pmd_t pmd;
831
832                         if (pfn != pmd_pfn(*pmdp))
833                                 goto unlock_pmd;
834                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
835                                 goto unlock_pmd;
836
837                         flush_cache_page(vma, address, pfn);
838                         pmd = pmdp_invalidate(vma, address, pmdp);
839                         pmd = pmd_wrprotect(pmd);
840                         pmd = pmd_mkclean(pmd);
841                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
842 unlock_pmd:
843 #endif
844                         spin_unlock(ptl);
845                 } else {
846                         if (pfn != pte_pfn(*ptep))
847                                 goto unlock_pte;
848                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
849                                 goto unlock_pte;
850
851                         flush_cache_page(vma, address, pfn);
852                         pte = ptep_clear_flush(vma, address, ptep);
853                         pte = pte_wrprotect(pte);
854                         pte = pte_mkclean(pte);
855                         set_pte_at(vma->vm_mm, address, ptep, pte);
856 unlock_pte:
857                         pte_unmap_unlock(ptep, ptl);
858                 }
859
860                 mmu_notifier_invalidate_range_end(&range);
861         }
862         i_mmap_unlock_read(mapping);
863 }
864
865 static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
866                 struct address_space *mapping, void *entry)
867 {
868         unsigned long pfn, index, count;
869         long ret = 0;
870
871         /*
872          * A page got tagged dirty in DAX mapping? Something is seriously
873          * wrong.
874          */
875         if (WARN_ON(!xa_is_value(entry)))
876                 return -EIO;
877
878         if (unlikely(dax_is_locked(entry))) {
879                 void *old_entry = entry;
880
881                 entry = get_unlocked_entry(xas, 0);
882
883                 /* Entry got punched out / reallocated? */
884                 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
885                         goto put_unlocked;
886                 /*
887                  * Entry got reallocated elsewhere? No need to writeback.
888                  * We have to compare pfns as we must not bail out due to
889                  * difference in lockbit or entry type.
890                  */
891                 if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
892                         goto put_unlocked;
893                 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
894                                         dax_is_zero_entry(entry))) {
895                         ret = -EIO;
896                         goto put_unlocked;
897                 }
898
899                 /* Another fsync thread may have already done this entry */
900                 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
901                         goto put_unlocked;
902         }
903
904         /* Lock the entry to serialize with page faults */
905         dax_lock_entry(xas, entry);
906
907         /*
908          * We can clear the tag now but we have to be careful so that concurrent
909          * dax_writeback_one() calls for the same index cannot finish before we
910          * actually flush the caches. This is achieved as the calls will look
911          * at the entry only under the i_pages lock and once they do that
912          * they will see the entry locked and wait for it to unlock.
913          */
914         xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
915         xas_unlock_irq(xas);
916
917         /*
918          * If dax_writeback_mapping_range() was given a wbc->range_start
919          * in the middle of a PMD, the 'index' we use needs to be
920          * aligned to the start of the PMD.
921          * This allows us to flush for PMD_SIZE and not have to worry about
922          * partial PMD writebacks.
923          */
924         pfn = dax_to_pfn(entry);
925         count = 1UL << dax_entry_order(entry);
926         index = xas->xa_index & ~(count - 1);
927
928         dax_entry_mkclean(mapping, index, pfn);
929         dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
930         /*
931          * After we have flushed the cache, we can clear the dirty tag. There
932          * cannot be new dirty data in the pfn after the flush has completed as
933          * the pfn mappings are writeprotected and fault waits for mapping
934          * entry lock.
935          */
936         xas_reset(xas);
937         xas_lock_irq(xas);
938         xas_store(xas, entry);
939         xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
940         dax_wake_entry(xas, entry, false);
941
942         trace_dax_writeback_one(mapping->host, index, count);
943         return ret;
944
945  put_unlocked:
946         put_unlocked_entry(xas, entry);
947         return ret;
948 }
949
950 /*
951  * Flush the mapping to the persistent domain within the byte range of [start,
952  * end]. This is required by data integrity operations to ensure file data is
953  * on persistent storage prior to completion of the operation.
954  */
955 int dax_writeback_mapping_range(struct address_space *mapping,
956                 struct dax_device *dax_dev, struct writeback_control *wbc)
957 {
958         XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
959         struct inode *inode = mapping->host;
960         pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
961         void *entry;
962         int ret = 0;
963         unsigned int scanned = 0;
964
965         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
966                 return -EIO;
967
968         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
969                 return 0;
970
971         trace_dax_writeback_range(inode, xas.xa_index, end_index);
972
973         tag_pages_for_writeback(mapping, xas.xa_index, end_index);
974
975         xas_lock_irq(&xas);
976         xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
977                 ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
978                 if (ret < 0) {
979                         mapping_set_error(mapping, ret);
980                         break;
981                 }
982                 if (++scanned % XA_CHECK_SCHED)
983                         continue;
984
985                 xas_pause(&xas);
986                 xas_unlock_irq(&xas);
987                 cond_resched();
988                 xas_lock_irq(&xas);
989         }
990         xas_unlock_irq(&xas);
991         trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
992         return ret;
993 }
994 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
995
996 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
997 {
998         return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
999 }
1000
1001 static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
1002                          pfn_t *pfnp)
1003 {
1004         const sector_t sector = dax_iomap_sector(iomap, pos);
1005         pgoff_t pgoff;
1006         int id, rc;
1007         long length;
1008
1009         rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
1010         if (rc)
1011                 return rc;
1012         id = dax_read_lock();
1013         length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1014                                    NULL, pfnp);
1015         if (length < 0) {
1016                 rc = length;
1017                 goto out;
1018         }
1019         rc = -EINVAL;
1020         if (PFN_PHYS(length) < size)
1021                 goto out;
1022         if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1023                 goto out;
1024         /* For larger pages we need devmap */
1025         if (length > 1 && !pfn_t_devmap(*pfnp))
1026                 goto out;
1027         rc = 0;
1028 out:
1029         dax_read_unlock(id);
1030         return rc;
1031 }
1032
1033 /*
1034  * The user has performed a load from a hole in the file.  Allocating a new
1035  * page in the file would cause excessive storage usage for workloads with
1036  * sparse files.  Instead we insert a read-only mapping of the 4k zero page.
1037  * If this page is ever written to we will re-fault and change the mapping to
1038  * point to real DAX storage instead.
1039  */
1040 static vm_fault_t dax_load_hole(struct xa_state *xas,
1041                 struct address_space *mapping, void **entry,
1042                 struct vm_fault *vmf)
1043 {
1044         struct inode *inode = mapping->host;
1045         unsigned long vaddr = vmf->address;
1046         pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1047         vm_fault_t ret;
1048
1049         *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1050                         DAX_ZERO_PAGE, false);
1051
1052         ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1053         trace_dax_load_hole(inode, vmf, ret);
1054         return ret;
1055 }
1056
1057 s64 dax_iomap_zero(loff_t pos, u64 length, struct iomap *iomap)
1058 {
1059         sector_t sector = iomap_sector(iomap, pos & PAGE_MASK);
1060         pgoff_t pgoff;
1061         long rc, id;
1062         void *kaddr;
1063         bool page_aligned = false;
1064         unsigned offset = offset_in_page(pos);
1065         unsigned size = min_t(u64, PAGE_SIZE - offset, length);
1066
1067         if (IS_ALIGNED(sector << SECTOR_SHIFT, PAGE_SIZE) &&
1068             (size == PAGE_SIZE))
1069                 page_aligned = true;
1070
1071         rc = bdev_dax_pgoff(iomap->bdev, sector, PAGE_SIZE, &pgoff);
1072         if (rc)
1073                 return rc;
1074
1075         id = dax_read_lock();
1076
1077         if (page_aligned)
1078                 rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1079         else
1080                 rc = dax_direct_access(iomap->dax_dev, pgoff, 1, &kaddr, NULL);
1081         if (rc < 0) {
1082                 dax_read_unlock(id);
1083                 return rc;
1084         }
1085
1086         if (!page_aligned) {
1087                 memset(kaddr + offset, 0, size);
1088                 dax_flush(iomap->dax_dev, kaddr + offset, size);
1089         }
1090         dax_read_unlock(id);
1091         return size;
1092 }
1093
1094 static loff_t
1095 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1096                 struct iomap *iomap, struct iomap *srcmap)
1097 {
1098         struct block_device *bdev = iomap->bdev;
1099         struct dax_device *dax_dev = iomap->dax_dev;
1100         struct iov_iter *iter = data;
1101         loff_t end = pos + length, done = 0;
1102         ssize_t ret = 0;
1103         size_t xfer;
1104         int id;
1105
1106         if (iov_iter_rw(iter) == READ) {
1107                 end = min(end, i_size_read(inode));
1108                 if (pos >= end)
1109                         return 0;
1110
1111                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1112                         return iov_iter_zero(min(length, end - pos), iter);
1113         }
1114
1115         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1116                 return -EIO;
1117
1118         /*
1119          * Write can allocate block for an area which has a hole page mapped
1120          * into page tables. We have to tear down these mappings so that data
1121          * written by write(2) is visible in mmap.
1122          */
1123         if (iomap->flags & IOMAP_F_NEW) {
1124                 invalidate_inode_pages2_range(inode->i_mapping,
1125                                               pos >> PAGE_SHIFT,
1126                                               (end - 1) >> PAGE_SHIFT);
1127         }
1128
1129         id = dax_read_lock();
1130         while (pos < end) {
1131                 unsigned offset = pos & (PAGE_SIZE - 1);
1132                 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1133                 const sector_t sector = dax_iomap_sector(iomap, pos);
1134                 ssize_t map_len;
1135                 pgoff_t pgoff;
1136                 void *kaddr;
1137
1138                 if (fatal_signal_pending(current)) {
1139                         ret = -EINTR;
1140                         break;
1141                 }
1142
1143                 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1144                 if (ret)
1145                         break;
1146
1147                 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1148                                 &kaddr, NULL);
1149                 if (map_len < 0) {
1150                         ret = map_len;
1151                         break;
1152                 }
1153
1154                 map_len = PFN_PHYS(map_len);
1155                 kaddr += offset;
1156                 map_len -= offset;
1157                 if (map_len > end - pos)
1158                         map_len = end - pos;
1159
1160                 /*
1161                  * The userspace address for the memory copy has already been
1162                  * validated via access_ok() in either vfs_read() or
1163                  * vfs_write(), depending on which operation we are doing.
1164                  */
1165                 if (iov_iter_rw(iter) == WRITE)
1166                         xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1167                                         map_len, iter);
1168                 else
1169                         xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1170                                         map_len, iter);
1171
1172                 pos += xfer;
1173                 length -= xfer;
1174                 done += xfer;
1175
1176                 if (xfer == 0)
1177                         ret = -EFAULT;
1178                 if (xfer < map_len)
1179                         break;
1180         }
1181         dax_read_unlock(id);
1182
1183         return done ? done : ret;
1184 }
1185
1186 /**
1187  * dax_iomap_rw - Perform I/O to a DAX file
1188  * @iocb:       The control block for this I/O
1189  * @iter:       The addresses to do I/O from or to
1190  * @ops:        iomap ops passed from the file system
1191  *
1192  * This function performs read and write operations to directly mapped
1193  * persistent memory.  The callers needs to take care of read/write exclusion
1194  * and evicting any page cache pages in the region under I/O.
1195  */
1196 ssize_t
1197 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1198                 const struct iomap_ops *ops)
1199 {
1200         struct address_space *mapping = iocb->ki_filp->f_mapping;
1201         struct inode *inode = mapping->host;
1202         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1203         unsigned flags = 0;
1204
1205         if (iov_iter_rw(iter) == WRITE) {
1206                 lockdep_assert_held_write(&inode->i_rwsem);
1207                 flags |= IOMAP_WRITE;
1208         } else {
1209                 lockdep_assert_held(&inode->i_rwsem);
1210         }
1211
1212         if (iocb->ki_flags & IOCB_NOWAIT)
1213                 flags |= IOMAP_NOWAIT;
1214
1215         while (iov_iter_count(iter)) {
1216                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1217                                 iter, dax_iomap_actor);
1218                 if (ret <= 0)
1219                         break;
1220                 pos += ret;
1221                 done += ret;
1222         }
1223
1224         iocb->ki_pos += done;
1225         return done ? done : ret;
1226 }
1227 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1228
1229 static vm_fault_t dax_fault_return(int error)
1230 {
1231         if (error == 0)
1232                 return VM_FAULT_NOPAGE;
1233         return vmf_error(error);
1234 }
1235
1236 /*
1237  * MAP_SYNC on a dax mapping guarantees dirty metadata is
1238  * flushed on write-faults (non-cow), but not read-faults.
1239  */
1240 static bool dax_fault_is_synchronous(unsigned long flags,
1241                 struct vm_area_struct *vma, struct iomap *iomap)
1242 {
1243         return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1244                 && (iomap->flags & IOMAP_F_DIRTY);
1245 }
1246
1247 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1248                                int *iomap_errp, const struct iomap_ops *ops)
1249 {
1250         struct vm_area_struct *vma = vmf->vma;
1251         struct address_space *mapping = vma->vm_file->f_mapping;
1252         XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1253         struct inode *inode = mapping->host;
1254         unsigned long vaddr = vmf->address;
1255         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1256         struct iomap iomap = { .type = IOMAP_HOLE };
1257         struct iomap srcmap = { .type = IOMAP_HOLE };
1258         unsigned flags = IOMAP_FAULT;
1259         int error, major = 0;
1260         bool write = vmf->flags & FAULT_FLAG_WRITE;
1261         bool sync;
1262         vm_fault_t ret = 0;
1263         void *entry;
1264         pfn_t pfn;
1265
1266         trace_dax_pte_fault(inode, vmf, ret);
1267         /*
1268          * Check whether offset isn't beyond end of file now. Caller is supposed
1269          * to hold locks serializing us with truncate / punch hole so this is
1270          * a reliable test.
1271          */
1272         if (pos >= i_size_read(inode)) {
1273                 ret = VM_FAULT_SIGBUS;
1274                 goto out;
1275         }
1276
1277         if (write && !vmf->cow_page)
1278                 flags |= IOMAP_WRITE;
1279
1280         entry = grab_mapping_entry(&xas, mapping, 0);
1281         if (xa_is_internal(entry)) {
1282                 ret = xa_to_internal(entry);
1283                 goto out;
1284         }
1285
1286         /*
1287          * It is possible, particularly with mixed reads & writes to private
1288          * mappings, that we have raced with a PMD fault that overlaps with
1289          * the PTE we need to set up.  If so just return and the fault will be
1290          * retried.
1291          */
1292         if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1293                 ret = VM_FAULT_NOPAGE;
1294                 goto unlock_entry;
1295         }
1296
1297         /*
1298          * Note that we don't bother to use iomap_apply here: DAX required
1299          * the file system block size to be equal the page size, which means
1300          * that we never have to deal with more than a single extent here.
1301          */
1302         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap, &srcmap);
1303         if (iomap_errp)
1304                 *iomap_errp = error;
1305         if (error) {
1306                 ret = dax_fault_return(error);
1307                 goto unlock_entry;
1308         }
1309         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1310                 error = -EIO;   /* fs corruption? */
1311                 goto error_finish_iomap;
1312         }
1313
1314         if (vmf->cow_page) {
1315                 sector_t sector = dax_iomap_sector(&iomap, pos);
1316
1317                 switch (iomap.type) {
1318                 case IOMAP_HOLE:
1319                 case IOMAP_UNWRITTEN:
1320                         clear_user_highpage(vmf->cow_page, vaddr);
1321                         break;
1322                 case IOMAP_MAPPED:
1323                         error = copy_cow_page_dax(iomap.bdev, iomap.dax_dev,
1324                                                   sector, vmf->cow_page, vaddr);
1325                         break;
1326                 default:
1327                         WARN_ON_ONCE(1);
1328                         error = -EIO;
1329                         break;
1330                 }
1331
1332                 if (error)
1333                         goto error_finish_iomap;
1334
1335                 __SetPageUptodate(vmf->cow_page);
1336                 ret = finish_fault(vmf);
1337                 if (!ret)
1338                         ret = VM_FAULT_DONE_COW;
1339                 goto finish_iomap;
1340         }
1341
1342         sync = dax_fault_is_synchronous(flags, vma, &iomap);
1343
1344         switch (iomap.type) {
1345         case IOMAP_MAPPED:
1346                 if (iomap.flags & IOMAP_F_NEW) {
1347                         count_vm_event(PGMAJFAULT);
1348                         count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1349                         major = VM_FAULT_MAJOR;
1350                 }
1351                 error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1352                 if (error < 0)
1353                         goto error_finish_iomap;
1354
1355                 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1356                                                  0, write && !sync);
1357
1358                 /*
1359                  * If we are doing synchronous page fault and inode needs fsync,
1360                  * we can insert PTE into page tables only after that happens.
1361                  * Skip insertion for now and return the pfn so that caller can
1362                  * insert it after fsync is done.
1363                  */
1364                 if (sync) {
1365                         if (WARN_ON_ONCE(!pfnp)) {
1366                                 error = -EIO;
1367                                 goto error_finish_iomap;
1368                         }
1369                         *pfnp = pfn;
1370                         ret = VM_FAULT_NEEDDSYNC | major;
1371                         goto finish_iomap;
1372                 }
1373                 trace_dax_insert_mapping(inode, vmf, entry);
1374                 if (write)
1375                         ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn);
1376                 else
1377                         ret = vmf_insert_mixed(vma, vaddr, pfn);
1378
1379                 goto finish_iomap;
1380         case IOMAP_UNWRITTEN:
1381         case IOMAP_HOLE:
1382                 if (!write) {
1383                         ret = dax_load_hole(&xas, mapping, &entry, vmf);
1384                         goto finish_iomap;
1385                 }
1386                 fallthrough;
1387         default:
1388                 WARN_ON_ONCE(1);
1389                 error = -EIO;
1390                 break;
1391         }
1392
1393  error_finish_iomap:
1394         ret = dax_fault_return(error);
1395  finish_iomap:
1396         if (ops->iomap_end) {
1397                 int copied = PAGE_SIZE;
1398
1399                 if (ret & VM_FAULT_ERROR)
1400                         copied = 0;
1401                 /*
1402                  * The fault is done by now and there's no way back (other
1403                  * thread may be already happily using PTE we have installed).
1404                  * Just ignore error from ->iomap_end since we cannot do much
1405                  * with it.
1406                  */
1407                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1408         }
1409  unlock_entry:
1410         dax_unlock_entry(&xas, entry);
1411  out:
1412         trace_dax_pte_fault_done(inode, vmf, ret);
1413         return ret | major;
1414 }
1415
1416 #ifdef CONFIG_FS_DAX_PMD
1417 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1418                 struct iomap *iomap, void **entry)
1419 {
1420         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1421         unsigned long pmd_addr = vmf->address & PMD_MASK;
1422         struct vm_area_struct *vma = vmf->vma;
1423         struct inode *inode = mapping->host;
1424         pgtable_t pgtable = NULL;
1425         struct page *zero_page;
1426         spinlock_t *ptl;
1427         pmd_t pmd_entry;
1428         pfn_t pfn;
1429
1430         zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1431
1432         if (unlikely(!zero_page))
1433                 goto fallback;
1434
1435         pfn = page_to_pfn_t(zero_page);
1436         *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1437                         DAX_PMD | DAX_ZERO_PAGE, false);
1438
1439         if (arch_needs_pgtable_deposit()) {
1440                 pgtable = pte_alloc_one(vma->vm_mm);
1441                 if (!pgtable)
1442                         return VM_FAULT_OOM;
1443         }
1444
1445         ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1446         if (!pmd_none(*(vmf->pmd))) {
1447                 spin_unlock(ptl);
1448                 goto fallback;
1449         }
1450
1451         if (pgtable) {
1452                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1453                 mm_inc_nr_ptes(vma->vm_mm);
1454         }
1455         pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1456         pmd_entry = pmd_mkhuge(pmd_entry);
1457         set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1458         spin_unlock(ptl);
1459         trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1460         return VM_FAULT_NOPAGE;
1461
1462 fallback:
1463         if (pgtable)
1464                 pte_free(vma->vm_mm, pgtable);
1465         trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1466         return VM_FAULT_FALLBACK;
1467 }
1468
1469 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1470                                const struct iomap_ops *ops)
1471 {
1472         struct vm_area_struct *vma = vmf->vma;
1473         struct address_space *mapping = vma->vm_file->f_mapping;
1474         XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1475         unsigned long pmd_addr = vmf->address & PMD_MASK;
1476         bool write = vmf->flags & FAULT_FLAG_WRITE;
1477         bool sync;
1478         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1479         struct inode *inode = mapping->host;
1480         vm_fault_t result = VM_FAULT_FALLBACK;
1481         struct iomap iomap = { .type = IOMAP_HOLE };
1482         struct iomap srcmap = { .type = IOMAP_HOLE };
1483         pgoff_t max_pgoff;
1484         void *entry;
1485         loff_t pos;
1486         int error;
1487         pfn_t pfn;
1488
1489         /*
1490          * Check whether offset isn't beyond end of file now. Caller is
1491          * supposed to hold locks serializing us with truncate / punch hole so
1492          * this is a reliable test.
1493          */
1494         max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1495
1496         trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1497
1498         /*
1499          * Make sure that the faulting address's PMD offset (color) matches
1500          * the PMD offset from the start of the file.  This is necessary so
1501          * that a PMD range in the page table overlaps exactly with a PMD
1502          * range in the page cache.
1503          */
1504         if ((vmf->pgoff & PG_PMD_COLOUR) !=
1505             ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1506                 goto fallback;
1507
1508         /* Fall back to PTEs if we're going to COW */
1509         if (write && !(vma->vm_flags & VM_SHARED))
1510                 goto fallback;
1511
1512         /* If the PMD would extend outside the VMA */
1513         if (pmd_addr < vma->vm_start)
1514                 goto fallback;
1515         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1516                 goto fallback;
1517
1518         if (xas.xa_index >= max_pgoff) {
1519                 result = VM_FAULT_SIGBUS;
1520                 goto out;
1521         }
1522
1523         /* If the PMD would extend beyond the file size */
1524         if ((xas.xa_index | PG_PMD_COLOUR) >= max_pgoff)
1525                 goto fallback;
1526
1527         /*
1528          * grab_mapping_entry() will make sure we get an empty PMD entry,
1529          * a zero PMD entry or a DAX PMD.  If it can't (because a PTE
1530          * entry is already in the array, for instance), it will return
1531          * VM_FAULT_FALLBACK.
1532          */
1533         entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1534         if (xa_is_internal(entry)) {
1535                 result = xa_to_internal(entry);
1536                 goto fallback;
1537         }
1538
1539         /*
1540          * It is possible, particularly with mixed reads & writes to private
1541          * mappings, that we have raced with a PTE fault that overlaps with
1542          * the PMD we need to set up.  If so just return and the fault will be
1543          * retried.
1544          */
1545         if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1546                         !pmd_devmap(*vmf->pmd)) {
1547                 result = 0;
1548                 goto unlock_entry;
1549         }
1550
1551         /*
1552          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1553          * setting up a mapping, so really we're using iomap_begin() as a way
1554          * to look up our filesystem block.
1555          */
1556         pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1557         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap,
1558                         &srcmap);
1559         if (error)
1560                 goto unlock_entry;
1561
1562         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1563                 goto finish_iomap;
1564
1565         sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1566
1567         switch (iomap.type) {
1568         case IOMAP_MAPPED:
1569                 error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1570                 if (error < 0)
1571                         goto finish_iomap;
1572
1573                 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1574                                                 DAX_PMD, write && !sync);
1575
1576                 /*
1577                  * If we are doing synchronous page fault and inode needs fsync,
1578                  * we can insert PMD into page tables only after that happens.
1579                  * Skip insertion for now and return the pfn so that caller can
1580                  * insert it after fsync is done.
1581                  */
1582                 if (sync) {
1583                         if (WARN_ON_ONCE(!pfnp))
1584                                 goto finish_iomap;
1585                         *pfnp = pfn;
1586                         result = VM_FAULT_NEEDDSYNC;
1587                         goto finish_iomap;
1588                 }
1589
1590                 trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
1591                 result = vmf_insert_pfn_pmd(vmf, pfn, write);
1592                 break;
1593         case IOMAP_UNWRITTEN:
1594         case IOMAP_HOLE:
1595                 if (WARN_ON_ONCE(write))
1596                         break;
1597                 result = dax_pmd_load_hole(&xas, vmf, &iomap, &entry);
1598                 break;
1599         default:
1600                 WARN_ON_ONCE(1);
1601                 break;
1602         }
1603
1604  finish_iomap:
1605         if (ops->iomap_end) {
1606                 int copied = PMD_SIZE;
1607
1608                 if (result == VM_FAULT_FALLBACK)
1609                         copied = 0;
1610                 /*
1611                  * The fault is done by now and there's no way back (other
1612                  * thread may be already happily using PMD we have installed).
1613                  * Just ignore error from ->iomap_end since we cannot do much
1614                  * with it.
1615                  */
1616                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1617                                 &iomap);
1618         }
1619  unlock_entry:
1620         dax_unlock_entry(&xas, entry);
1621  fallback:
1622         if (result == VM_FAULT_FALLBACK) {
1623                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1624                 count_vm_event(THP_FAULT_FALLBACK);
1625         }
1626 out:
1627         trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1628         return result;
1629 }
1630 #else
1631 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1632                                const struct iomap_ops *ops)
1633 {
1634         return VM_FAULT_FALLBACK;
1635 }
1636 #endif /* CONFIG_FS_DAX_PMD */
1637
1638 /**
1639  * dax_iomap_fault - handle a page fault on a DAX file
1640  * @vmf: The description of the fault
1641  * @pe_size: Size of the page to fault in
1642  * @pfnp: PFN to insert for synchronous faults if fsync is required
1643  * @iomap_errp: Storage for detailed error code in case of error
1644  * @ops: Iomap ops passed from the file system
1645  *
1646  * When a page fault occurs, filesystems may call this helper in
1647  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1648  * has done all the necessary locking for page fault to proceed
1649  * successfully.
1650  */
1651 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1652                     pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1653 {
1654         switch (pe_size) {
1655         case PE_SIZE_PTE:
1656                 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1657         case PE_SIZE_PMD:
1658                 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1659         default:
1660                 return VM_FAULT_FALLBACK;
1661         }
1662 }
1663 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1664
1665 /*
1666  * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1667  * @vmf: The description of the fault
1668  * @pfn: PFN to insert
1669  * @order: Order of entry to insert.
1670  *
1671  * This function inserts a writeable PTE or PMD entry into the page tables
1672  * for an mmaped DAX file.  It also marks the page cache entry as dirty.
1673  */
1674 static vm_fault_t
1675 dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1676 {
1677         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1678         XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1679         void *entry;
1680         vm_fault_t ret;
1681
1682         xas_lock_irq(&xas);
1683         entry = get_unlocked_entry(&xas, order);
1684         /* Did we race with someone splitting entry or so? */
1685         if (!entry || dax_is_conflict(entry) ||
1686             (order == 0 && !dax_is_pte_entry(entry))) {
1687                 put_unlocked_entry(&xas, entry);
1688                 xas_unlock_irq(&xas);
1689                 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1690                                                       VM_FAULT_NOPAGE);
1691                 return VM_FAULT_NOPAGE;
1692         }
1693         xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1694         dax_lock_entry(&xas, entry);
1695         xas_unlock_irq(&xas);
1696         if (order == 0)
1697                 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1698 #ifdef CONFIG_FS_DAX_PMD
1699         else if (order == PMD_ORDER)
1700                 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1701 #endif
1702         else
1703                 ret = VM_FAULT_FALLBACK;
1704         dax_unlock_entry(&xas, entry);
1705         trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1706         return ret;
1707 }
1708
1709 /**
1710  * dax_finish_sync_fault - finish synchronous page fault
1711  * @vmf: The description of the fault
1712  * @pe_size: Size of entry to be inserted
1713  * @pfn: PFN to insert
1714  *
1715  * This function ensures that the file range touched by the page fault is
1716  * stored persistently on the media and handles inserting of appropriate page
1717  * table entry.
1718  */
1719 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1720                 enum page_entry_size pe_size, pfn_t pfn)
1721 {
1722         int err;
1723         loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1724         unsigned int order = pe_order(pe_size);
1725         size_t len = PAGE_SIZE << order;
1726
1727         err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1728         if (err)
1729                 return VM_FAULT_SIGBUS;
1730         return dax_insert_pfn_mkwrite(vmf, pfn, order);
1731 }
1732 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);