~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/dax.c

Version: ~ [ linux-4.19-rc4 ] ~ [ linux-4.18.8 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.70 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.127 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.156 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.122 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.57 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.31.14 ] ~ [ linux-2.6.30.10 ] ~ [ linux-2.6.29.6 ] ~ [ linux-2.6.28.10 ] ~ [ linux-2.6.27.62 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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, bool (*wait_fn)(void))
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                 bool revalidate;
241 
242                 entry = __radix_tree_lookup(&mapping->i_pages, index, NULL,
243                                           &slot);
244                 if (!entry ||
245                     WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)) ||
246                     !slot_locked(mapping, slot)) {
247                         if (slotp)
248                                 *slotp = slot;
249                         return entry;
250                 }
251 
252                 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
253                 prepare_to_wait_exclusive(wq, &ewait.wait,
254                                           TASK_UNINTERRUPTIBLE);
255                 xa_unlock_irq(&mapping->i_pages);
256                 revalidate = wait_fn();
257                 finish_wait(wq, &ewait.wait);
258                 xa_lock_irq(&mapping->i_pages);
259                 if (revalidate)
260                         return ERR_PTR(-EAGAIN);
261         }
262 }
263 
264 static bool entry_wait(void)
265 {
266         schedule();
267         /*
268          * Never return an ERR_PTR() from
269          * __get_unlocked_mapping_entry(), just keep looping.
270          */
271         return false;
272 }
273 
274 static void *get_unlocked_mapping_entry(struct address_space *mapping,
275                 pgoff_t index, void ***slotp)
276 {
277         return __get_unlocked_mapping_entry(mapping, index, slotp, entry_wait);
278 }
279 
280 static void unlock_mapping_entry(struct address_space *mapping, pgoff_t index)
281 {
282         void *entry, **slot;
283 
284         xa_lock_irq(&mapping->i_pages);
285         entry = __radix_tree_lookup(&mapping->i_pages, index, NULL, &slot);
286         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
287                          !slot_locked(mapping, slot))) {
288                 xa_unlock_irq(&mapping->i_pages);
289                 return;
290         }
291         unlock_slot(mapping, slot);
292         xa_unlock_irq(&mapping->i_pages);
293         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
294 }
295 
296 static void put_locked_mapping_entry(struct address_space *mapping,
297                 pgoff_t index)
298 {
299         unlock_mapping_entry(mapping, index);
300 }
301 
302 /*
303  * Called when we are done with radix tree entry we looked up via
304  * get_unlocked_mapping_entry() and which we didn't lock in the end.
305  */
306 static void put_unlocked_mapping_entry(struct address_space *mapping,
307                                        pgoff_t index, void *entry)
308 {
309         if (!entry)
310                 return;
311 
312         /* We have to wake up next waiter for the radix tree entry lock */
313         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
314 }
315 
316 static unsigned long dax_entry_size(void *entry)
317 {
318         if (dax_is_zero_entry(entry))
319                 return 0;
320         else if (dax_is_empty_entry(entry))
321                 return 0;
322         else if (dax_is_pmd_entry(entry))
323                 return PMD_SIZE;
324         else
325                 return PAGE_SIZE;
326 }
327 
328 static unsigned long dax_radix_end_pfn(void *entry)
329 {
330         return dax_radix_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
331 }
332 
333 /*
334  * Iterate through all mapped pfns represented by an entry, i.e. skip
335  * 'empty' and 'zero' entries.
336  */
337 #define for_each_mapped_pfn(entry, pfn) \
338         for (pfn = dax_radix_pfn(entry); \
339                         pfn < dax_radix_end_pfn(entry); pfn++)
340 
341 /*
342  * TODO: for reflink+dax we need a way to associate a single page with
343  * multiple address_space instances at different linear_page_index()
344  * offsets.
345  */
346 static void dax_associate_entry(void *entry, struct address_space *mapping,
347                 struct vm_area_struct *vma, unsigned long address)
348 {
349         unsigned long size = dax_entry_size(entry), pfn, index;
350         int i = 0;
351 
352         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
353                 return;
354 
355         index = linear_page_index(vma, address & ~(size - 1));
356         for_each_mapped_pfn(entry, pfn) {
357                 struct page *page = pfn_to_page(pfn);
358 
359                 WARN_ON_ONCE(page->mapping);
360                 page->mapping = mapping;
361                 page->index = index + i++;
362         }
363 }
364 
365 static void dax_disassociate_entry(void *entry, struct address_space *mapping,
366                 bool trunc)
367 {
368         unsigned long pfn;
369 
370         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
371                 return;
372 
373         for_each_mapped_pfn(entry, pfn) {
374                 struct page *page = pfn_to_page(pfn);
375 
376                 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
377                 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
378                 page->mapping = NULL;
379                 page->index = 0;
380         }
381 }
382 
383 static struct page *dax_busy_page(void *entry)
384 {
385         unsigned long pfn;
386 
387         for_each_mapped_pfn(entry, pfn) {
388                 struct page *page = pfn_to_page(pfn);
389 
390                 if (page_ref_count(page) > 1)
391                         return page;
392         }
393         return NULL;
394 }
395 
396 static bool entry_wait_revalidate(void)
397 {
398         rcu_read_unlock();
399         schedule();
400         rcu_read_lock();
401 
402         /*
403          * Tell __get_unlocked_mapping_entry() to take a break, we need
404          * to revalidate page->mapping after dropping locks
405          */
406         return true;
407 }
408 
409 bool dax_lock_mapping_entry(struct page *page)
410 {
411         pgoff_t index;
412         struct inode *inode;
413         bool did_lock = false;
414         void *entry = NULL, **slot;
415         struct address_space *mapping;
416 
417         rcu_read_lock();
418         for (;;) {
419                 mapping = READ_ONCE(page->mapping);
420 
421                 if (!dax_mapping(mapping))
422                         break;
423 
424                 /*
425                  * In the device-dax case there's no need to lock, a
426                  * struct dev_pagemap pin is sufficient to keep the
427                  * inode alive, and we assume we have dev_pagemap pin
428                  * otherwise we would not have a valid pfn_to_page()
429                  * translation.
430                  */
431                 inode = mapping->host;
432                 if (S_ISCHR(inode->i_mode)) {
433                         did_lock = true;
434                         break;
435                 }
436 
437                 xa_lock_irq(&mapping->i_pages);
438                 if (mapping != page->mapping) {
439                         xa_unlock_irq(&mapping->i_pages);
440                         continue;
441                 }
442                 index = page->index;
443 
444                 entry = __get_unlocked_mapping_entry(mapping, index, &slot,
445                                 entry_wait_revalidate);
446                 if (!entry) {
447                         xa_unlock_irq(&mapping->i_pages);
448                         break;
449                 } else if (IS_ERR(entry)) {
450                         WARN_ON_ONCE(PTR_ERR(entry) != -EAGAIN);
451                         continue;
452                 }
453                 lock_slot(mapping, slot);
454                 did_lock = true;
455                 xa_unlock_irq(&mapping->i_pages);
456                 break;
457         }
458         rcu_read_unlock();
459 
460         return did_lock;
461 }
462 
463 void dax_unlock_mapping_entry(struct page *page)
464 {
465         struct address_space *mapping = page->mapping;
466         struct inode *inode = mapping->host;
467 
468         if (S_ISCHR(inode->i_mode))
469                 return;
470 
471         unlock_mapping_entry(mapping, page->index);
472 }
473 
474 /*
475  * Find radix tree entry at given index. If it points to an exceptional entry,
476  * return it with the radix tree entry locked. If the radix tree doesn't
477  * contain given index, create an empty exceptional entry for the index and
478  * return with it locked.
479  *
480  * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
481  * either return that locked entry or will return an error.  This error will
482  * happen if there are any 4k entries within the 2MiB range that we are
483  * requesting.
484  *
485  * We always favor 4k entries over 2MiB entries. There isn't a flow where we
486  * evict 4k entries in order to 'upgrade' them to a 2MiB entry.  A 2MiB
487  * insertion will fail if it finds any 4k entries already in the tree, and a
488  * 4k insertion will cause an existing 2MiB entry to be unmapped and
489  * downgraded to 4k entries.  This happens for both 2MiB huge zero pages as
490  * well as 2MiB empty entries.
491  *
492  * The exception to this downgrade path is for 2MiB DAX PMD entries that have
493  * real storage backing them.  We will leave these real 2MiB DAX entries in
494  * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
495  *
496  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
497  * persistent memory the benefit is doubtful. We can add that later if we can
498  * show it helps.
499  */
500 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
501                 unsigned long size_flag)
502 {
503         bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
504         void *entry, **slot;
505 
506 restart:
507         xa_lock_irq(&mapping->i_pages);
508         entry = get_unlocked_mapping_entry(mapping, index, &slot);
509 
510         if (WARN_ON_ONCE(entry && !radix_tree_exceptional_entry(entry))) {
511                 entry = ERR_PTR(-EIO);
512                 goto out_unlock;
513         }
514 
515         if (entry) {
516                 if (size_flag & RADIX_DAX_PMD) {
517                         if (dax_is_pte_entry(entry)) {
518                                 put_unlocked_mapping_entry(mapping, index,
519                                                 entry);
520                                 entry = ERR_PTR(-EEXIST);
521                                 goto out_unlock;
522                         }
523                 } else { /* trying to grab a PTE entry */
524                         if (dax_is_pmd_entry(entry) &&
525                             (dax_is_zero_entry(entry) ||
526                              dax_is_empty_entry(entry))) {
527                                 pmd_downgrade = true;
528                         }
529                 }
530         }
531 
532         /* No entry for given index? Make sure radix tree is big enough. */
533         if (!entry || pmd_downgrade) {
534                 int err;
535 
536                 if (pmd_downgrade) {
537                         /*
538                          * Make sure 'entry' remains valid while we drop
539                          * the i_pages lock.
540                          */
541                         entry = lock_slot(mapping, slot);
542                 }
543 
544                 xa_unlock_irq(&mapping->i_pages);
545                 /*
546                  * Besides huge zero pages the only other thing that gets
547                  * downgraded are empty entries which don't need to be
548                  * unmapped.
549                  */
550                 if (pmd_downgrade && dax_is_zero_entry(entry))
551                         unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
552                                                         PG_PMD_NR, false);
553 
554                 err = radix_tree_preload(
555                                 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
556                 if (err) {
557                         if (pmd_downgrade)
558                                 put_locked_mapping_entry(mapping, index);
559                         return ERR_PTR(err);
560                 }
561                 xa_lock_irq(&mapping->i_pages);
562 
563                 if (!entry) {
564                         /*
565                          * We needed to drop the i_pages lock while calling
566                          * radix_tree_preload() and we didn't have an entry to
567                          * lock.  See if another thread inserted an entry at
568                          * our index during this time.
569                          */
570                         entry = __radix_tree_lookup(&mapping->i_pages, index,
571                                         NULL, &slot);
572                         if (entry) {
573                                 radix_tree_preload_end();
574                                 xa_unlock_irq(&mapping->i_pages);
575                                 goto restart;
576                         }
577                 }
578 
579                 if (pmd_downgrade) {
580                         dax_disassociate_entry(entry, mapping, false);
581                         radix_tree_delete(&mapping->i_pages, index);
582                         mapping->nrexceptional--;
583                         dax_wake_mapping_entry_waiter(mapping, index, entry,
584                                         true);
585                 }
586 
587                 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
588 
589                 err = __radix_tree_insert(&mapping->i_pages, index,
590                                 dax_radix_order(entry), entry);
591                 radix_tree_preload_end();
592                 if (err) {
593                         xa_unlock_irq(&mapping->i_pages);
594                         /*
595                          * Our insertion of a DAX entry failed, most likely
596                          * because we were inserting a PMD entry and it
597                          * collided with a PTE sized entry at a different
598                          * index in the PMD range.  We haven't inserted
599                          * anything into the radix tree and have no waiters to
600                          * wake.
601                          */
602                         return ERR_PTR(err);
603                 }
604                 /* Good, we have inserted empty locked entry into the tree. */
605                 mapping->nrexceptional++;
606                 xa_unlock_irq(&mapping->i_pages);
607                 return entry;
608         }
609         entry = lock_slot(mapping, slot);
610  out_unlock:
611         xa_unlock_irq(&mapping->i_pages);
612         return entry;
613 }
614 
615 /**
616  * dax_layout_busy_page - find first pinned page in @mapping
617  * @mapping: address space to scan for a page with ref count > 1
618  *
619  * DAX requires ZONE_DEVICE mapped pages. These pages are never
620  * 'onlined' to the page allocator so they are considered idle when
621  * page->count == 1. A filesystem uses this interface to determine if
622  * any page in the mapping is busy, i.e. for DMA, or other
623  * get_user_pages() usages.
624  *
625  * It is expected that the filesystem is holding locks to block the
626  * establishment of new mappings in this address_space. I.e. it expects
627  * to be able to run unmap_mapping_range() and subsequently not race
628  * mapping_mapped() becoming true.
629  */
630 struct page *dax_layout_busy_page(struct address_space *mapping)
631 {
632         pgoff_t indices[PAGEVEC_SIZE];
633         struct page *page = NULL;
634         struct pagevec pvec;
635         pgoff_t index, end;
636         unsigned i;
637 
638         /*
639          * In the 'limited' case get_user_pages() for dax is disabled.
640          */
641         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
642                 return NULL;
643 
644         if (!dax_mapping(mapping) || !mapping_mapped(mapping))
645                 return NULL;
646 
647         pagevec_init(&pvec);
648         index = 0;
649         end = -1;
650 
651         /*
652          * If we race get_user_pages_fast() here either we'll see the
653          * elevated page count in the pagevec_lookup and wait, or
654          * get_user_pages_fast() will see that the page it took a reference
655          * against is no longer mapped in the page tables and bail to the
656          * get_user_pages() slow path.  The slow path is protected by
657          * pte_lock() and pmd_lock(). New references are not taken without
658          * holding those locks, and unmap_mapping_range() will not zero the
659          * pte or pmd without holding the respective lock, so we are
660          * guaranteed to either see new references or prevent new
661          * references from being established.
662          */
663         unmap_mapping_range(mapping, 0, 0, 1);
664 
665         while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
666                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
667                                 indices)) {
668                 for (i = 0; i < pagevec_count(&pvec); i++) {
669                         struct page *pvec_ent = pvec.pages[i];
670                         void *entry;
671 
672                         index = indices[i];
673                         if (index >= end)
674                                 break;
675 
676                         if (WARN_ON_ONCE(
677                              !radix_tree_exceptional_entry(pvec_ent)))
678                                 continue;
679 
680                         xa_lock_irq(&mapping->i_pages);
681                         entry = get_unlocked_mapping_entry(mapping, index, NULL);
682                         if (entry)
683                                 page = dax_busy_page(entry);
684                         put_unlocked_mapping_entry(mapping, index, entry);
685                         xa_unlock_irq(&mapping->i_pages);
686                         if (page)
687                                 break;
688                 }
689 
690                 /*
691                  * We don't expect normal struct page entries to exist in our
692                  * tree, but we keep these pagevec calls so that this code is
693                  * consistent with the common pattern for handling pagevecs
694                  * throughout the kernel.
695                  */
696                 pagevec_remove_exceptionals(&pvec);
697                 pagevec_release(&pvec);
698                 index++;
699 
700                 if (page)
701                         break;
702         }
703         return page;
704 }
705 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
706 
707 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
708                                           pgoff_t index, bool trunc)
709 {
710         int ret = 0;
711         void *entry;
712         struct radix_tree_root *pages = &mapping->i_pages;
713 
714         xa_lock_irq(pages);
715         entry = get_unlocked_mapping_entry(mapping, index, NULL);
716         if (!entry || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)))
717                 goto out;
718         if (!trunc &&
719             (radix_tree_tag_get(pages, index, PAGECACHE_TAG_DIRTY) ||
720              radix_tree_tag_get(pages, index, PAGECACHE_TAG_TOWRITE)))
721                 goto out;
722         dax_disassociate_entry(entry, mapping, trunc);
723         radix_tree_delete(pages, index);
724         mapping->nrexceptional--;
725         ret = 1;
726 out:
727         put_unlocked_mapping_entry(mapping, index, entry);
728         xa_unlock_irq(pages);
729         return ret;
730 }
731 /*
732  * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
733  * entry to get unlocked before deleting it.
734  */
735 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
736 {
737         int ret = __dax_invalidate_mapping_entry(mapping, index, true);
738 
739         /*
740          * This gets called from truncate / punch_hole path. As such, the caller
741          * must hold locks protecting against concurrent modifications of the
742          * radix tree (usually fs-private i_mmap_sem for writing). Since the
743          * caller has seen exceptional entry for this index, we better find it
744          * at that index as well...
745          */
746         WARN_ON_ONCE(!ret);
747         return ret;
748 }
749 
750 /*
751  * Invalidate exceptional DAX entry if it is clean.
752  */
753 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
754                                       pgoff_t index)
755 {
756         return __dax_invalidate_mapping_entry(mapping, index, false);
757 }
758 
759 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
760                 sector_t sector, size_t size, struct page *to,
761                 unsigned long vaddr)
762 {
763         void *vto, *kaddr;
764         pgoff_t pgoff;
765         long rc;
766         int id;
767 
768         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
769         if (rc)
770                 return rc;
771 
772         id = dax_read_lock();
773         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, NULL);
774         if (rc < 0) {
775                 dax_read_unlock(id);
776                 return rc;
777         }
778         vto = kmap_atomic(to);
779         copy_user_page(vto, (void __force *)kaddr, vaddr, to);
780         kunmap_atomic(vto);
781         dax_read_unlock(id);
782         return 0;
783 }
784 
785 /*
786  * By this point grab_mapping_entry() has ensured that we have a locked entry
787  * of the appropriate size so we don't have to worry about downgrading PMDs to
788  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
789  * already in the tree, we will skip the insertion and just dirty the PMD as
790  * appropriate.
791  */
792 static void *dax_insert_mapping_entry(struct address_space *mapping,
793                                       struct vm_fault *vmf,
794                                       void *entry, pfn_t pfn_t,
795                                       unsigned long flags, bool dirty)
796 {
797         struct radix_tree_root *pages = &mapping->i_pages;
798         unsigned long pfn = pfn_t_to_pfn(pfn_t);
799         pgoff_t index = vmf->pgoff;
800         void *new_entry;
801 
802         if (dirty)
803                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
804 
805         if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_ZERO_PAGE)) {
806                 /* we are replacing a zero page with block mapping */
807                 if (dax_is_pmd_entry(entry))
808                         unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
809                                                         PG_PMD_NR, false);
810                 else /* pte entry */
811                         unmap_mapping_pages(mapping, vmf->pgoff, 1, false);
812         }
813 
814         xa_lock_irq(pages);
815         new_entry = dax_radix_locked_entry(pfn, flags);
816         if (dax_entry_size(entry) != dax_entry_size(new_entry)) {
817                 dax_disassociate_entry(entry, mapping, false);
818                 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address);
819         }
820 
821         if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
822                 /*
823                  * Only swap our new entry into the radix tree if the current
824                  * entry is a zero page or an empty entry.  If a normal PTE or
825                  * PMD entry is already in the tree, we leave it alone.  This
826                  * means that if we are trying to insert a PTE and the
827                  * existing entry is a PMD, we will just leave the PMD in the
828                  * tree and dirty it if necessary.
829                  */
830                 struct radix_tree_node *node;
831                 void **slot;
832                 void *ret;
833 
834                 ret = __radix_tree_lookup(pages, index, &node, &slot);
835                 WARN_ON_ONCE(ret != entry);
836                 __radix_tree_replace(pages, node, slot,
837                                      new_entry, NULL);
838                 entry = new_entry;
839         }
840 
841         if (dirty)
842                 radix_tree_tag_set(pages, index, PAGECACHE_TAG_DIRTY);
843 
844         xa_unlock_irq(pages);
845         return entry;
846 }
847 
848 static inline unsigned long
849 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
850 {
851         unsigned long address;
852 
853         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
854         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
855         return address;
856 }
857 
858 /* Walk all mappings of a given index of a file and writeprotect them */
859 static void dax_mapping_entry_mkclean(struct address_space *mapping,
860                                       pgoff_t index, unsigned long pfn)
861 {
862         struct vm_area_struct *vma;
863         pte_t pte, *ptep = NULL;
864         pmd_t *pmdp = NULL;
865         spinlock_t *ptl;
866 
867         i_mmap_lock_read(mapping);
868         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
869                 unsigned long address, start, end;
870 
871                 cond_resched();
872 
873                 if (!(vma->vm_flags & VM_SHARED))
874                         continue;
875 
876                 address = pgoff_address(index, vma);
877 
878                 /*
879                  * Note because we provide start/end to follow_pte_pmd it will
880                  * call mmu_notifier_invalidate_range_start() on our behalf
881                  * before taking any lock.
882                  */
883                 if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl))
884                         continue;
885 
886                 /*
887                  * No need to call mmu_notifier_invalidate_range() as we are
888                  * downgrading page table protection not changing it to point
889                  * to a new page.
890                  *
891                  * See Documentation/vm/mmu_notifier.rst
892                  */
893                 if (pmdp) {
894 #ifdef CONFIG_FS_DAX_PMD
895                         pmd_t pmd;
896 
897                         if (pfn != pmd_pfn(*pmdp))
898                                 goto unlock_pmd;
899                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
900                                 goto unlock_pmd;
901 
902                         flush_cache_page(vma, address, pfn);
903                         pmd = pmdp_huge_clear_flush(vma, address, pmdp);
904                         pmd = pmd_wrprotect(pmd);
905                         pmd = pmd_mkclean(pmd);
906                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
907 unlock_pmd:
908 #endif
909                         spin_unlock(ptl);
910                 } else {
911                         if (pfn != pte_pfn(*ptep))
912                                 goto unlock_pte;
913                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
914                                 goto unlock_pte;
915 
916                         flush_cache_page(vma, address, pfn);
917                         pte = ptep_clear_flush(vma, address, ptep);
918                         pte = pte_wrprotect(pte);
919                         pte = pte_mkclean(pte);
920                         set_pte_at(vma->vm_mm, address, ptep, pte);
921 unlock_pte:
922                         pte_unmap_unlock(ptep, ptl);
923                 }
924 
925                 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
926         }
927         i_mmap_unlock_read(mapping);
928 }
929 
930 static int dax_writeback_one(struct dax_device *dax_dev,
931                 struct address_space *mapping, pgoff_t index, void *entry)
932 {
933         struct radix_tree_root *pages = &mapping->i_pages;
934         void *entry2, **slot;
935         unsigned long pfn;
936         long ret = 0;
937         size_t size;
938 
939         /*
940          * A page got tagged dirty in DAX mapping? Something is seriously
941          * wrong.
942          */
943         if (WARN_ON(!radix_tree_exceptional_entry(entry)))
944                 return -EIO;
945 
946         xa_lock_irq(pages);
947         entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
948         /* Entry got punched out / reallocated? */
949         if (!entry2 || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry2)))
950                 goto put_unlocked;
951         /*
952          * Entry got reallocated elsewhere? No need to writeback. We have to
953          * compare pfns as we must not bail out due to difference in lockbit
954          * or entry type.
955          */
956         if (dax_radix_pfn(entry2) != dax_radix_pfn(entry))
957                 goto put_unlocked;
958         if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
959                                 dax_is_zero_entry(entry))) {
960                 ret = -EIO;
961                 goto put_unlocked;
962         }
963 
964         /* Another fsync thread may have already written back this entry */
965         if (!radix_tree_tag_get(pages, index, PAGECACHE_TAG_TOWRITE))
966                 goto put_unlocked;
967         /* Lock the entry to serialize with page faults */
968         entry = lock_slot(mapping, slot);
969         /*
970          * We can clear the tag now but we have to be careful so that concurrent
971          * dax_writeback_one() calls for the same index cannot finish before we
972          * actually flush the caches. This is achieved as the calls will look
973          * at the entry only under the i_pages lock and once they do that
974          * they will see the entry locked and wait for it to unlock.
975          */
976         radix_tree_tag_clear(pages, index, PAGECACHE_TAG_TOWRITE);
977         xa_unlock_irq(pages);
978 
979         /*
980          * Even if dax_writeback_mapping_range() was given a wbc->range_start
981          * in the middle of a PMD, the 'index' we are given will be aligned to
982          * the start index of the PMD, as will the pfn we pull from 'entry'.
983          * This allows us to flush for PMD_SIZE and not have to worry about
984          * partial PMD writebacks.
985          */
986         pfn = dax_radix_pfn(entry);
987         size = PAGE_SIZE << dax_radix_order(entry);
988 
989         dax_mapping_entry_mkclean(mapping, index, pfn);
990         dax_flush(dax_dev, page_address(pfn_to_page(pfn)), size);
991         /*
992          * After we have flushed the cache, we can clear the dirty tag. There
993          * cannot be new dirty data in the pfn after the flush has completed as
994          * the pfn mappings are writeprotected and fault waits for mapping
995          * entry lock.
996          */
997         xa_lock_irq(pages);
998         radix_tree_tag_clear(pages, index, PAGECACHE_TAG_DIRTY);
999         xa_unlock_irq(pages);
1000         trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT);
1001         put_locked_mapping_entry(mapping, index);
1002         return ret;
1003 
1004  put_unlocked:
1005         put_unlocked_mapping_entry(mapping, index, entry2);
1006         xa_unlock_irq(pages);
1007         return ret;
1008 }
1009 
1010 /*
1011  * Flush the mapping to the persistent domain within the byte range of [start,
1012  * end]. This is required by data integrity operations to ensure file data is
1013  * on persistent storage prior to completion of the operation.
1014  */
1015 int dax_writeback_mapping_range(struct address_space *mapping,
1016                 struct block_device *bdev, struct writeback_control *wbc)
1017 {
1018         struct inode *inode = mapping->host;
1019         pgoff_t start_index, end_index;
1020         pgoff_t indices[PAGEVEC_SIZE];
1021         struct dax_device *dax_dev;
1022         struct pagevec pvec;
1023         bool done = false;
1024         int i, ret = 0;
1025 
1026         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
1027                 return -EIO;
1028 
1029         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
1030                 return 0;
1031 
1032         dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
1033         if (!dax_dev)
1034                 return -EIO;
1035 
1036         start_index = wbc->range_start >> PAGE_SHIFT;
1037         end_index = wbc->range_end >> PAGE_SHIFT;
1038 
1039         trace_dax_writeback_range(inode, start_index, end_index);
1040 
1041         tag_pages_for_writeback(mapping, start_index, end_index);
1042 
1043         pagevec_init(&pvec);
1044         while (!done) {
1045                 pvec.nr = find_get_entries_tag(mapping, start_index,
1046                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
1047                                 pvec.pages, indices);
1048 
1049                 if (pvec.nr == 0)
1050                         break;
1051 
1052                 for (i = 0; i < pvec.nr; i++) {
1053                         if (indices[i] > end_index) {
1054                                 done = true;
1055                                 break;
1056                         }
1057 
1058                         ret = dax_writeback_one(dax_dev, mapping, indices[i],
1059                                         pvec.pages[i]);
1060                         if (ret < 0) {
1061                                 mapping_set_error(mapping, ret);
1062                                 goto out;
1063                         }
1064                 }
1065                 start_index = indices[pvec.nr - 1] + 1;
1066         }
1067 out:
1068         put_dax(dax_dev);
1069         trace_dax_writeback_range_done(inode, start_index, end_index);
1070         return (ret < 0 ? ret : 0);
1071 }
1072 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1073 
1074 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
1075 {
1076         return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
1077 }
1078 
1079 static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
1080                          pfn_t *pfnp)
1081 {
1082         const sector_t sector = dax_iomap_sector(iomap, pos);
1083         pgoff_t pgoff;
1084         int id, rc;
1085         long length;
1086 
1087         rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
1088         if (rc)
1089                 return rc;
1090         id = dax_read_lock();
1091         length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1092                                    NULL, pfnp);
1093         if (length < 0) {
1094                 rc = length;
1095                 goto out;
1096         }
1097         rc = -EINVAL;
1098         if (PFN_PHYS(length) < size)
1099                 goto out;
1100         if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1101                 goto out;
1102         /* For larger pages we need devmap */
1103         if (length > 1 && !pfn_t_devmap(*pfnp))
1104                 goto out;
1105         rc = 0;
1106 out:
1107         dax_read_unlock(id);
1108         return rc;
1109 }
1110 
1111 /*
1112  * The user has performed a load from a hole in the file.  Allocating a new
1113  * page in the file would cause excessive storage usage for workloads with
1114  * sparse files.  Instead we insert a read-only mapping of the 4k zero page.
1115  * If this page is ever written to we will re-fault and change the mapping to
1116  * point to real DAX storage instead.
1117  */
1118 static vm_fault_t dax_load_hole(struct address_space *mapping, void *entry,
1119                          struct vm_fault *vmf)
1120 {
1121         struct inode *inode = mapping->host;
1122         unsigned long vaddr = vmf->address;
1123         vm_fault_t ret = VM_FAULT_NOPAGE;
1124         struct page *zero_page;
1125         pfn_t pfn;
1126 
1127         zero_page = ZERO_PAGE(0);
1128         if (unlikely(!zero_page)) {
1129                 ret = VM_FAULT_OOM;
1130                 goto out;
1131         }
1132 
1133         pfn = page_to_pfn_t(zero_page);
1134         dax_insert_mapping_entry(mapping, vmf, entry, pfn, RADIX_DAX_ZERO_PAGE,
1135                         false);
1136         ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1137 out:
1138         trace_dax_load_hole(inode, vmf, ret);
1139         return ret;
1140 }
1141 
1142 static bool dax_range_is_aligned(struct block_device *bdev,
1143                                  unsigned int offset, unsigned int length)
1144 {
1145         unsigned short sector_size = bdev_logical_block_size(bdev);
1146 
1147         if (!IS_ALIGNED(offset, sector_size))
1148                 return false;
1149         if (!IS_ALIGNED(length, sector_size))
1150                 return false;
1151 
1152         return true;
1153 }
1154 
1155 int __dax_zero_page_range(struct block_device *bdev,
1156                 struct dax_device *dax_dev, sector_t sector,
1157                 unsigned int offset, unsigned int size)
1158 {
1159         if (dax_range_is_aligned(bdev, offset, size)) {
1160                 sector_t start_sector = sector + (offset >> 9);
1161 
1162                 return blkdev_issue_zeroout(bdev, start_sector,
1163                                 size >> 9, GFP_NOFS, 0);
1164         } else {
1165                 pgoff_t pgoff;
1166                 long rc, id;
1167                 void *kaddr;
1168 
1169                 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
1170                 if (rc)
1171                         return rc;
1172 
1173                 id = dax_read_lock();
1174                 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, NULL);
1175                 if (rc < 0) {
1176                         dax_read_unlock(id);
1177                         return rc;
1178                 }
1179                 memset(kaddr + offset, 0, size);
1180                 dax_flush(dax_dev, kaddr + offset, size);
1181                 dax_read_unlock(id);
1182         }
1183         return 0;
1184 }
1185 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1186 
1187 static loff_t
1188 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1189                 struct iomap *iomap)
1190 {
1191         struct block_device *bdev = iomap->bdev;
1192         struct dax_device *dax_dev = iomap->dax_dev;
1193         struct iov_iter *iter = data;
1194         loff_t end = pos + length, done = 0;
1195         ssize_t ret = 0;
1196         size_t xfer;
1197         int id;
1198 
1199         if (iov_iter_rw(iter) == READ) {
1200                 end = min(end, i_size_read(inode));
1201                 if (pos >= end)
1202                         return 0;
1203 
1204                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1205                         return iov_iter_zero(min(length, end - pos), iter);
1206         }
1207 
1208         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1209                 return -EIO;
1210 
1211         /*
1212          * Write can allocate block for an area which has a hole page mapped
1213          * into page tables. We have to tear down these mappings so that data
1214          * written by write(2) is visible in mmap.
1215          */
1216         if (iomap->flags & IOMAP_F_NEW) {
1217                 invalidate_inode_pages2_range(inode->i_mapping,
1218                                               pos >> PAGE_SHIFT,
1219                                               (end - 1) >> PAGE_SHIFT);
1220         }
1221 
1222         id = dax_read_lock();
1223         while (pos < end) {
1224                 unsigned offset = pos & (PAGE_SIZE - 1);
1225                 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1226                 const sector_t sector = dax_iomap_sector(iomap, pos);
1227                 ssize_t map_len;
1228                 pgoff_t pgoff;
1229                 void *kaddr;
1230 
1231                 if (fatal_signal_pending(current)) {
1232                         ret = -EINTR;
1233                         break;
1234                 }
1235 
1236                 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1237                 if (ret)
1238                         break;
1239 
1240                 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1241                                 &kaddr, NULL);
1242                 if (map_len < 0) {
1243                         ret = map_len;
1244                         break;
1245                 }
1246 
1247                 map_len = PFN_PHYS(map_len);
1248                 kaddr += offset;
1249                 map_len -= offset;
1250                 if (map_len > end - pos)
1251                         map_len = end - pos;
1252 
1253                 /*
1254                  * The userspace address for the memory copy has already been
1255                  * validated via access_ok() in either vfs_read() or
1256                  * vfs_write(), depending on which operation we are doing.
1257                  */
1258                 if (iov_iter_rw(iter) == WRITE)
1259                         xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1260                                         map_len, iter);
1261                 else
1262                         xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1263                                         map_len, iter);
1264 
1265                 pos += xfer;
1266                 length -= xfer;
1267                 done += xfer;
1268 
1269                 if (xfer == 0)
1270                         ret = -EFAULT;
1271                 if (xfer < map_len)
1272                         break;
1273         }
1274         dax_read_unlock(id);
1275 
1276         return done ? done : ret;
1277 }
1278 
1279 /**
1280  * dax_iomap_rw - Perform I/O to a DAX file
1281  * @iocb:       The control block for this I/O
1282  * @iter:       The addresses to do I/O from or to
1283  * @ops:        iomap ops passed from the file system
1284  *
1285  * This function performs read and write operations to directly mapped
1286  * persistent memory.  The callers needs to take care of read/write exclusion
1287  * and evicting any page cache pages in the region under I/O.
1288  */
1289 ssize_t
1290 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1291                 const struct iomap_ops *ops)
1292 {
1293         struct address_space *mapping = iocb->ki_filp->f_mapping;
1294         struct inode *inode = mapping->host;
1295         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1296         unsigned flags = 0;
1297 
1298         if (iov_iter_rw(iter) == WRITE) {
1299                 lockdep_assert_held_exclusive(&inode->i_rwsem);
1300                 flags |= IOMAP_WRITE;
1301         } else {
1302                 lockdep_assert_held(&inode->i_rwsem);
1303         }
1304 
1305         while (iov_iter_count(iter)) {
1306                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1307                                 iter, dax_iomap_actor);
1308                 if (ret <= 0)
1309                         break;
1310                 pos += ret;
1311                 done += ret;
1312         }
1313 
1314         iocb->ki_pos += done;
1315         return done ? done : ret;
1316 }
1317 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1318 
1319 static vm_fault_t dax_fault_return(int error)
1320 {
1321         if (error == 0)
1322                 return VM_FAULT_NOPAGE;
1323         if (error == -ENOMEM)
1324                 return VM_FAULT_OOM;
1325         return VM_FAULT_SIGBUS;
1326 }
1327 
1328 /*
1329  * MAP_SYNC on a dax mapping guarantees dirty metadata is
1330  * flushed on write-faults (non-cow), but not read-faults.
1331  */
1332 static bool dax_fault_is_synchronous(unsigned long flags,
1333                 struct vm_area_struct *vma, struct iomap *iomap)
1334 {
1335         return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1336                 && (iomap->flags & IOMAP_F_DIRTY);
1337 }
1338 
1339 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1340                                int *iomap_errp, const struct iomap_ops *ops)
1341 {
1342         struct vm_area_struct *vma = vmf->vma;
1343         struct address_space *mapping = vma->vm_file->f_mapping;
1344         struct inode *inode = mapping->host;
1345         unsigned long vaddr = vmf->address;
1346         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1347         struct iomap iomap = { 0 };
1348         unsigned flags = IOMAP_FAULT;
1349         int error, major = 0;
1350         bool write = vmf->flags & FAULT_FLAG_WRITE;
1351         bool sync;
1352         vm_fault_t ret = 0;
1353         void *entry;
1354         pfn_t pfn;
1355 
1356         trace_dax_pte_fault(inode, vmf, ret);
1357         /*
1358          * Check whether offset isn't beyond end of file now. Caller is supposed
1359          * to hold locks serializing us with truncate / punch hole so this is
1360          * a reliable test.
1361          */
1362         if (pos >= i_size_read(inode)) {
1363                 ret = VM_FAULT_SIGBUS;
1364                 goto out;
1365         }
1366 
1367         if (write && !vmf->cow_page)
1368                 flags |= IOMAP_WRITE;
1369 
1370         entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1371         if (IS_ERR(entry)) {
1372                 ret = dax_fault_return(PTR_ERR(entry));
1373                 goto out;
1374         }
1375 
1376         /*
1377          * It is possible, particularly with mixed reads & writes to private
1378          * mappings, that we have raced with a PMD fault that overlaps with
1379          * the PTE we need to set up.  If so just return and the fault will be
1380          * retried.
1381          */
1382         if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1383                 ret = VM_FAULT_NOPAGE;
1384                 goto unlock_entry;
1385         }
1386 
1387         /*
1388          * Note that we don't bother to use iomap_apply here: DAX required
1389          * the file system block size to be equal the page size, which means
1390          * that we never have to deal with more than a single extent here.
1391          */
1392         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1393         if (iomap_errp)
1394                 *iomap_errp = error;
1395         if (error) {
1396                 ret = dax_fault_return(error);
1397                 goto unlock_entry;
1398         }
1399         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1400                 error = -EIO;   /* fs corruption? */
1401                 goto error_finish_iomap;
1402         }
1403 
1404         if (vmf->cow_page) {
1405                 sector_t sector = dax_iomap_sector(&iomap, pos);
1406 
1407                 switch (iomap.type) {
1408                 case IOMAP_HOLE:
1409                 case IOMAP_UNWRITTEN:
1410                         clear_user_highpage(vmf->cow_page, vaddr);
1411                         break;
1412                 case IOMAP_MAPPED:
1413                         error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1414                                         sector, PAGE_SIZE, vmf->cow_page, vaddr);
1415                         break;
1416                 default:
1417                         WARN_ON_ONCE(1);
1418                         error = -EIO;
1419                         break;
1420                 }
1421 
1422                 if (error)
1423                         goto error_finish_iomap;
1424 
1425                 __SetPageUptodate(vmf->cow_page);
1426                 ret = finish_fault(vmf);
1427                 if (!ret)
1428                         ret = VM_FAULT_DONE_COW;
1429                 goto finish_iomap;
1430         }
1431 
1432         sync = dax_fault_is_synchronous(flags, vma, &iomap);
1433 
1434         switch (iomap.type) {
1435         case IOMAP_MAPPED:
1436                 if (iomap.flags & IOMAP_F_NEW) {
1437                         count_vm_event(PGMAJFAULT);
1438                         count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1439                         major = VM_FAULT_MAJOR;
1440                 }
1441                 error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1442                 if (error < 0)
1443                         goto error_finish_iomap;
1444 
1445                 entry = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1446                                                  0, write && !sync);
1447 
1448                 /*
1449                  * If we are doing synchronous page fault and inode needs fsync,
1450                  * we can insert PTE into page tables only after that happens.
1451                  * Skip insertion for now and return the pfn so that caller can
1452                  * insert it after fsync is done.
1453                  */
1454                 if (sync) {
1455                         if (WARN_ON_ONCE(!pfnp)) {
1456                                 error = -EIO;
1457                                 goto error_finish_iomap;
1458                         }
1459                         *pfnp = pfn;
1460                         ret = VM_FAULT_NEEDDSYNC | major;
1461                         goto finish_iomap;
1462                 }
1463                 trace_dax_insert_mapping(inode, vmf, entry);
1464                 if (write)
1465                         ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn);
1466                 else
1467                         ret = vmf_insert_mixed(vma, vaddr, pfn);
1468 
1469                 goto finish_iomap;
1470         case IOMAP_UNWRITTEN:
1471         case IOMAP_HOLE:
1472                 if (!write) {
1473                         ret = dax_load_hole(mapping, entry, vmf);
1474                         goto finish_iomap;
1475                 }
1476                 /*FALLTHRU*/
1477         default:
1478                 WARN_ON_ONCE(1);
1479                 error = -EIO;
1480                 break;
1481         }
1482 
1483  error_finish_iomap:
1484         ret = dax_fault_return(error);
1485  finish_iomap:
1486         if (ops->iomap_end) {
1487                 int copied = PAGE_SIZE;
1488 
1489                 if (ret & VM_FAULT_ERROR)
1490                         copied = 0;
1491                 /*
1492                  * The fault is done by now and there's no way back (other
1493                  * thread may be already happily using PTE we have installed).
1494                  * Just ignore error from ->iomap_end since we cannot do much
1495                  * with it.
1496                  */
1497                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1498         }
1499  unlock_entry:
1500         put_locked_mapping_entry(mapping, vmf->pgoff);
1501  out:
1502         trace_dax_pte_fault_done(inode, vmf, ret);
1503         return ret | major;
1504 }
1505 
1506 #ifdef CONFIG_FS_DAX_PMD
1507 static vm_fault_t dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
1508                 void *entry)
1509 {
1510         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1511         unsigned long pmd_addr = vmf->address & PMD_MASK;
1512         struct inode *inode = mapping->host;
1513         struct page *zero_page;
1514         void *ret = NULL;
1515         spinlock_t *ptl;
1516         pmd_t pmd_entry;
1517         pfn_t pfn;
1518 
1519         zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1520 
1521         if (unlikely(!zero_page))
1522                 goto fallback;
1523 
1524         pfn = page_to_pfn_t(zero_page);
1525         ret = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1526                         RADIX_DAX_PMD | RADIX_DAX_ZERO_PAGE, false);
1527 
1528         ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1529         if (!pmd_none(*(vmf->pmd))) {
1530                 spin_unlock(ptl);
1531                 goto fallback;
1532         }
1533 
1534         pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1535         pmd_entry = pmd_mkhuge(pmd_entry);
1536         set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1537         spin_unlock(ptl);
1538         trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
1539         return VM_FAULT_NOPAGE;
1540 
1541 fallback:
1542         trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
1543         return VM_FAULT_FALLBACK;
1544 }
1545 
1546 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1547                                const struct iomap_ops *ops)
1548 {
1549         struct vm_area_struct *vma = vmf->vma;
1550         struct address_space *mapping = vma->vm_file->f_mapping;
1551         unsigned long pmd_addr = vmf->address & PMD_MASK;
1552         bool write = vmf->flags & FAULT_FLAG_WRITE;
1553         bool sync;
1554         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1555         struct inode *inode = mapping->host;
1556         vm_fault_t result = VM_FAULT_FALLBACK;
1557         struct iomap iomap = { 0 };
1558         pgoff_t max_pgoff, pgoff;
1559         void *entry;
1560         loff_t pos;
1561         int error;
1562         pfn_t pfn;
1563 
1564         /*
1565          * Check whether offset isn't beyond end of file now. Caller is
1566          * supposed to hold locks serializing us with truncate / punch hole so
1567          * this is a reliable test.
1568          */
1569         pgoff = linear_page_index(vma, pmd_addr);
1570         max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1571 
1572         trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1573 
1574         /*
1575          * Make sure that the faulting address's PMD offset (color) matches
1576          * the PMD offset from the start of the file.  This is necessary so
1577          * that a PMD range in the page table overlaps exactly with a PMD
1578          * range in the radix tree.
1579          */
1580         if ((vmf->pgoff & PG_PMD_COLOUR) !=
1581             ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1582                 goto fallback;
1583 
1584         /* Fall back to PTEs if we're going to COW */
1585         if (write && !(vma->vm_flags & VM_SHARED))
1586                 goto fallback;
1587 
1588         /* If the PMD would extend outside the VMA */
1589         if (pmd_addr < vma->vm_start)
1590                 goto fallback;
1591         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1592                 goto fallback;
1593 
1594         if (pgoff >= max_pgoff) {
1595                 result = VM_FAULT_SIGBUS;
1596                 goto out;
1597         }
1598 
1599         /* If the PMD would extend beyond the file size */
1600         if ((pgoff | PG_PMD_COLOUR) >= max_pgoff)
1601                 goto fallback;
1602 
1603         /*
1604          * grab_mapping_entry() will make sure we get a 2MiB empty entry, a
1605          * 2MiB zero page entry or a DAX PMD.  If it can't (because a 4k page
1606          * is already in the tree, for instance), it will return -EEXIST and
1607          * we just fall back to 4k entries.
1608          */
1609         entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1610         if (IS_ERR(entry))
1611                 goto fallback;
1612 
1613         /*
1614          * It is possible, particularly with mixed reads & writes to private
1615          * mappings, that we have raced with a PTE fault that overlaps with
1616          * the PMD we need to set up.  If so just return and the fault will be
1617          * retried.
1618          */
1619         if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1620                         !pmd_devmap(*vmf->pmd)) {
1621                 result = 0;
1622                 goto unlock_entry;
1623         }
1624 
1625         /*
1626          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1627          * setting up a mapping, so really we're using iomap_begin() as a way
1628          * to look up our filesystem block.
1629          */
1630         pos = (loff_t)pgoff << PAGE_SHIFT;
1631         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1632         if (error)
1633                 goto unlock_entry;
1634 
1635         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1636                 goto finish_iomap;
1637 
1638         sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1639 
1640         switch (iomap.type) {
1641         case IOMAP_MAPPED:
1642                 error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1643                 if (error < 0)
1644                         goto finish_iomap;
1645 
1646                 entry = dax_insert_mapping_entry(mapping, vmf, entry, pfn,
1647                                                 RADIX_DAX_PMD, write && !sync);
1648 
1649                 /*
1650                  * If we are doing synchronous page fault and inode needs fsync,
1651                  * we can insert PMD into page tables only after that happens.
1652                  * Skip insertion for now and return the pfn so that caller can
1653                  * insert it after fsync is done.
1654                  */
1655                 if (sync) {
1656                         if (WARN_ON_ONCE(!pfnp))
1657                                 goto finish_iomap;
1658                         *pfnp = pfn;
1659                         result = VM_FAULT_NEEDDSYNC;
1660                         goto finish_iomap;
1661                 }
1662 
1663                 trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
1664                 result = vmf_insert_pfn_pmd(vma, vmf->address, vmf->pmd, pfn,
1665                                             write);
1666                 break;
1667         case IOMAP_UNWRITTEN:
1668         case IOMAP_HOLE:
1669                 if (WARN_ON_ONCE(write))
1670                         break;
1671                 result = dax_pmd_load_hole(vmf, &iomap, entry);
1672                 break;
1673         default:
1674                 WARN_ON_ONCE(1);
1675                 break;
1676         }
1677 
1678  finish_iomap:
1679         if (ops->iomap_end) {
1680                 int copied = PMD_SIZE;
1681 
1682                 if (result == VM_FAULT_FALLBACK)
1683                         copied = 0;
1684                 /*
1685                  * The fault is done by now and there's no way back (other
1686                  * thread may be already happily using PMD we have installed).
1687                  * Just ignore error from ->iomap_end since we cannot do much
1688                  * with it.
1689                  */
1690                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1691                                 &iomap);
1692         }
1693  unlock_entry:
1694         put_locked_mapping_entry(mapping, pgoff);
1695  fallback:
1696         if (result == VM_FAULT_FALLBACK) {
1697                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1698                 count_vm_event(THP_FAULT_FALLBACK);
1699         }
1700 out:
1701         trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1702         return result;
1703 }
1704 #else
1705 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1706                                const struct iomap_ops *ops)
1707 {
1708         return VM_FAULT_FALLBACK;
1709 }
1710 #endif /* CONFIG_FS_DAX_PMD */
1711 
1712 /**
1713  * dax_iomap_fault - handle a page fault on a DAX file
1714  * @vmf: The description of the fault
1715  * @pe_size: Size of the page to fault in
1716  * @pfnp: PFN to insert for synchronous faults if fsync is required
1717  * @iomap_errp: Storage for detailed error code in case of error
1718  * @ops: Iomap ops passed from the file system
1719  *
1720  * When a page fault occurs, filesystems may call this helper in
1721  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1722  * has done all the necessary locking for page fault to proceed
1723  * successfully.
1724  */
1725 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1726                     pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1727 {
1728         switch (pe_size) {
1729         case PE_SIZE_PTE:
1730                 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1731         case PE_SIZE_PMD:
1732                 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1733         default:
1734                 return VM_FAULT_FALLBACK;
1735         }
1736 }
1737 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1738 
1739 /**
1740  * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1741  * @vmf: The description of the fault
1742  * @pe_size: Size of entry to be inserted
1743  * @pfn: PFN to insert
1744  *
1745  * This function inserts writeable PTE or PMD entry into page tables for mmaped
1746  * DAX file.  It takes care of marking corresponding radix tree entry as dirty
1747  * as well.
1748  */
1749 static vm_fault_t dax_insert_pfn_mkwrite(struct vm_fault *vmf,
1750                                   enum page_entry_size pe_size,
1751                                   pfn_t pfn)
1752 {
1753         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1754         void *entry, **slot;
1755         pgoff_t index = vmf->pgoff;
1756         vm_fault_t ret;
1757 
1758         xa_lock_irq(&mapping->i_pages);
1759         entry = get_unlocked_mapping_entry(mapping, index, &slot);
1760         /* Did we race with someone splitting entry or so? */
1761         if (!entry ||
1762             (pe_size == PE_SIZE_PTE && !dax_is_pte_entry(entry)) ||
1763             (pe_size == PE_SIZE_PMD && !dax_is_pmd_entry(entry))) {
1764                 put_unlocked_mapping_entry(mapping, index, entry);
1765                 xa_unlock_irq(&mapping->i_pages);
1766                 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1767                                                       VM_FAULT_NOPAGE);
1768                 return VM_FAULT_NOPAGE;
1769         }
1770         radix_tree_tag_set(&mapping->i_pages, index, PAGECACHE_TAG_DIRTY);
1771         entry = lock_slot(mapping, slot);
1772         xa_unlock_irq(&mapping->i_pages);
1773         switch (pe_size) {
1774         case PE_SIZE_PTE:
1775                 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1776                 break;
1777 #ifdef CONFIG_FS_DAX_PMD
1778         case PE_SIZE_PMD:
1779                 ret = vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
1780                         pfn, true);
1781                 break;
1782 #endif
1783         default:
1784                 ret = VM_FAULT_FALLBACK;
1785         }
1786         put_locked_mapping_entry(mapping, index);
1787         trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1788         return ret;
1789 }
1790 
1791 /**
1792  * dax_finish_sync_fault - finish synchronous page fault
1793  * @vmf: The description of the fault
1794  * @pe_size: Size of entry to be inserted
1795  * @pfn: PFN to insert
1796  *
1797  * This function ensures that the file range touched by the page fault is
1798  * stored persistently on the media and handles inserting of appropriate page
1799  * table entry.
1800  */
1801 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1802                 enum page_entry_size pe_size, pfn_t pfn)
1803 {
1804         int err;
1805         loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1806         size_t len = 0;
1807 
1808         if (pe_size == PE_SIZE_PTE)
1809                 len = PAGE_SIZE;
1810         else if (pe_size == PE_SIZE_PMD)
1811                 len = PMD_SIZE;
1812         else
1813                 WARN_ON_ONCE(1);
1814         err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1815         if (err)
1816                 return VM_FAULT_SIGBUS;
1817         return dax_insert_pfn_mkwrite(vmf, pe_size, pfn);
1818 }
1819 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
1820 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp