Version:  2.0.40 2.2.26 2.4.37 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10

Linux/fs/dax.c

  1 /*
  2  * fs/dax.c - Direct Access filesystem code
  3  * Copyright (c) 2013-2014 Intel Corporation
  4  * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
  5  * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
  6  *
  7  * This program is free software; you can redistribute it and/or modify it
  8  * under the terms and conditions of the GNU General Public License,
  9  * version 2, as published by the Free Software Foundation.
 10  *
 11  * This program is distributed in the hope it will be useful, but WITHOUT
 12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
 14  * more details.
 15  */
 16 
 17 #include <linux/atomic.h>
 18 #include <linux/blkdev.h>
 19 #include <linux/buffer_head.h>
 20 #include <linux/dax.h>
 21 #include <linux/fs.h>
 22 #include <linux/genhd.h>
 23 #include <linux/highmem.h>
 24 #include <linux/memcontrol.h>
 25 #include <linux/mm.h>
 26 #include <linux/mutex.h>
 27 #include <linux/pagevec.h>
 28 #include <linux/pmem.h>
 29 #include <linux/sched.h>
 30 #include <linux/uio.h>
 31 #include <linux/vmstat.h>
 32 #include <linux/pfn_t.h>
 33 #include <linux/sizes.h>
 34 #include <linux/mmu_notifier.h>
 35 #include <linux/iomap.h>
 36 #include "internal.h"
 37 
 38 /* We choose 4096 entries - same as per-zone page wait tables */
 39 #define DAX_WAIT_TABLE_BITS 12
 40 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
 41 
 42 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
 43 
 44 static int __init init_dax_wait_table(void)
 45 {
 46         int i;
 47 
 48         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
 49                 init_waitqueue_head(wait_table + i);
 50         return 0;
 51 }
 52 fs_initcall(init_dax_wait_table);
 53 
 54 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
 55 {
 56         struct request_queue *q = bdev->bd_queue;
 57         long rc = -EIO;
 58 
 59         dax->addr = ERR_PTR(-EIO);
 60         if (blk_queue_enter(q, true) != 0)
 61                 return rc;
 62 
 63         rc = bdev_direct_access(bdev, dax);
 64         if (rc < 0) {
 65                 dax->addr = ERR_PTR(rc);
 66                 blk_queue_exit(q);
 67                 return rc;
 68         }
 69         return rc;
 70 }
 71 
 72 static void dax_unmap_atomic(struct block_device *bdev,
 73                 const struct blk_dax_ctl *dax)
 74 {
 75         if (IS_ERR(dax->addr))
 76                 return;
 77         blk_queue_exit(bdev->bd_queue);
 78 }
 79 
 80 static int dax_is_pmd_entry(void *entry)
 81 {
 82         return (unsigned long)entry & RADIX_DAX_PMD;
 83 }
 84 
 85 static int dax_is_pte_entry(void *entry)
 86 {
 87         return !((unsigned long)entry & RADIX_DAX_PMD);
 88 }
 89 
 90 static int dax_is_zero_entry(void *entry)
 91 {
 92         return (unsigned long)entry & RADIX_DAX_HZP;
 93 }
 94 
 95 static int dax_is_empty_entry(void *entry)
 96 {
 97         return (unsigned long)entry & RADIX_DAX_EMPTY;
 98 }
 99 
100 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
101 {
102         struct page *page = alloc_pages(GFP_KERNEL, 0);
103         struct blk_dax_ctl dax = {
104                 .size = PAGE_SIZE,
105                 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
106         };
107         long rc;
108 
109         if (!page)
110                 return ERR_PTR(-ENOMEM);
111 
112         rc = dax_map_atomic(bdev, &dax);
113         if (rc < 0)
114                 return ERR_PTR(rc);
115         memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
116         dax_unmap_atomic(bdev, &dax);
117         return page;
118 }
119 
120 /*
121  * DAX radix tree locking
122  */
123 struct exceptional_entry_key {
124         struct address_space *mapping;
125         pgoff_t entry_start;
126 };
127 
128 struct wait_exceptional_entry_queue {
129         wait_queue_t wait;
130         struct exceptional_entry_key key;
131 };
132 
133 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
134                 pgoff_t index, void *entry, struct exceptional_entry_key *key)
135 {
136         unsigned long hash;
137 
138         /*
139          * If 'entry' is a PMD, align the 'index' that we use for the wait
140          * queue to the start of that PMD.  This ensures that all offsets in
141          * the range covered by the PMD map to the same bit lock.
142          */
143         if (dax_is_pmd_entry(entry))
144                 index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
145 
146         key->mapping = mapping;
147         key->entry_start = index;
148 
149         hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
150         return wait_table + hash;
151 }
152 
153 static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
154                                        int sync, void *keyp)
155 {
156         struct exceptional_entry_key *key = keyp;
157         struct wait_exceptional_entry_queue *ewait =
158                 container_of(wait, struct wait_exceptional_entry_queue, wait);
159 
160         if (key->mapping != ewait->key.mapping ||
161             key->entry_start != ewait->key.entry_start)
162                 return 0;
163         return autoremove_wake_function(wait, mode, sync, NULL);
164 }
165 
166 /*
167  * Check whether the given slot is locked. The function must be called with
168  * mapping->tree_lock held
169  */
170 static inline int slot_locked(struct address_space *mapping, void **slot)
171 {
172         unsigned long entry = (unsigned long)
173                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
174         return entry & RADIX_DAX_ENTRY_LOCK;
175 }
176 
177 /*
178  * Mark the given slot is locked. The function must be called with
179  * mapping->tree_lock held
180  */
181 static inline void *lock_slot(struct address_space *mapping, void **slot)
182 {
183         unsigned long entry = (unsigned long)
184                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
185 
186         entry |= RADIX_DAX_ENTRY_LOCK;
187         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
188         return (void *)entry;
189 }
190 
191 /*
192  * Mark the given slot is unlocked. The function must be called with
193  * mapping->tree_lock held
194  */
195 static inline void *unlock_slot(struct address_space *mapping, void **slot)
196 {
197         unsigned long entry = (unsigned long)
198                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
199 
200         entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
201         radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
202         return (void *)entry;
203 }
204 
205 /*
206  * Lookup entry in radix tree, wait for it to become unlocked if it is
207  * exceptional entry and return it. The caller must call
208  * put_unlocked_mapping_entry() when he decided not to lock the entry or
209  * put_locked_mapping_entry() when he locked the entry and now wants to
210  * unlock it.
211  *
212  * The function must be called with mapping->tree_lock held.
213  */
214 static void *get_unlocked_mapping_entry(struct address_space *mapping,
215                                         pgoff_t index, void ***slotp)
216 {
217         void *entry, **slot;
218         struct wait_exceptional_entry_queue ewait;
219         wait_queue_head_t *wq;
220 
221         init_wait(&ewait.wait);
222         ewait.wait.func = wake_exceptional_entry_func;
223 
224         for (;;) {
225                 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
226                                           &slot);
227                 if (!entry || !radix_tree_exceptional_entry(entry) ||
228                     !slot_locked(mapping, slot)) {
229                         if (slotp)
230                                 *slotp = slot;
231                         return entry;
232                 }
233 
234                 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
235                 prepare_to_wait_exclusive(wq, &ewait.wait,
236                                           TASK_UNINTERRUPTIBLE);
237                 spin_unlock_irq(&mapping->tree_lock);
238                 schedule();
239                 finish_wait(wq, &ewait.wait);
240                 spin_lock_irq(&mapping->tree_lock);
241         }
242 }
243 
244 static void dax_unlock_mapping_entry(struct address_space *mapping,
245                                      pgoff_t index)
246 {
247         void *entry, **slot;
248 
249         spin_lock_irq(&mapping->tree_lock);
250         entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
251         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
252                          !slot_locked(mapping, slot))) {
253                 spin_unlock_irq(&mapping->tree_lock);
254                 return;
255         }
256         unlock_slot(mapping, slot);
257         spin_unlock_irq(&mapping->tree_lock);
258         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
259 }
260 
261 static void put_locked_mapping_entry(struct address_space *mapping,
262                                      pgoff_t index, void *entry)
263 {
264         if (!radix_tree_exceptional_entry(entry)) {
265                 unlock_page(entry);
266                 put_page(entry);
267         } else {
268                 dax_unlock_mapping_entry(mapping, index);
269         }
270 }
271 
272 /*
273  * Called when we are done with radix tree entry we looked up via
274  * get_unlocked_mapping_entry() and which we didn't lock in the end.
275  */
276 static void put_unlocked_mapping_entry(struct address_space *mapping,
277                                        pgoff_t index, void *entry)
278 {
279         if (!radix_tree_exceptional_entry(entry))
280                 return;
281 
282         /* We have to wake up next waiter for the radix tree entry lock */
283         dax_wake_mapping_entry_waiter(mapping, index, entry, false);
284 }
285 
286 /*
287  * Find radix tree entry at given index. If it points to a page, return with
288  * the page locked. If it points to the exceptional entry, return with the
289  * radix tree entry locked. If the radix tree doesn't contain given index,
290  * create empty exceptional entry for the index and return with it locked.
291  *
292  * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
293  * either return that locked entry or will return an error.  This error will
294  * happen if there are any 4k entries (either zero pages or DAX entries)
295  * within the 2MiB range that we are requesting.
296  *
297  * We always favor 4k entries over 2MiB entries. There isn't a flow where we
298  * evict 4k entries in order to 'upgrade' them to a 2MiB entry.  A 2MiB
299  * insertion will fail if it finds any 4k entries already in the tree, and a
300  * 4k insertion will cause an existing 2MiB entry to be unmapped and
301  * downgraded to 4k entries.  This happens for both 2MiB huge zero pages as
302  * well as 2MiB empty entries.
303  *
304  * The exception to this downgrade path is for 2MiB DAX PMD entries that have
305  * real storage backing them.  We will leave these real 2MiB DAX entries in
306  * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
307  *
308  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
309  * persistent memory the benefit is doubtful. We can add that later if we can
310  * show it helps.
311  */
312 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
313                 unsigned long size_flag)
314 {
315         bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
316         void *entry, **slot;
317 
318 restart:
319         spin_lock_irq(&mapping->tree_lock);
320         entry = get_unlocked_mapping_entry(mapping, index, &slot);
321 
322         if (entry) {
323                 if (size_flag & RADIX_DAX_PMD) {
324                         if (!radix_tree_exceptional_entry(entry) ||
325                             dax_is_pte_entry(entry)) {
326                                 put_unlocked_mapping_entry(mapping, index,
327                                                 entry);
328                                 entry = ERR_PTR(-EEXIST);
329                                 goto out_unlock;
330                         }
331                 } else { /* trying to grab a PTE entry */
332                         if (radix_tree_exceptional_entry(entry) &&
333                             dax_is_pmd_entry(entry) &&
334                             (dax_is_zero_entry(entry) ||
335                              dax_is_empty_entry(entry))) {
336                                 pmd_downgrade = true;
337                         }
338                 }
339         }
340 
341         /* No entry for given index? Make sure radix tree is big enough. */
342         if (!entry || pmd_downgrade) {
343                 int err;
344 
345                 if (pmd_downgrade) {
346                         /*
347                          * Make sure 'entry' remains valid while we drop
348                          * mapping->tree_lock.
349                          */
350                         entry = lock_slot(mapping, slot);
351                 }
352 
353                 spin_unlock_irq(&mapping->tree_lock);
354                 /*
355                  * Besides huge zero pages the only other thing that gets
356                  * downgraded are empty entries which don't need to be
357                  * unmapped.
358                  */
359                 if (pmd_downgrade && dax_is_zero_entry(entry))
360                         unmap_mapping_range(mapping,
361                                 (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
362 
363                 err = radix_tree_preload(
364                                 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
365                 if (err) {
366                         if (pmd_downgrade)
367                                 put_locked_mapping_entry(mapping, index, entry);
368                         return ERR_PTR(err);
369                 }
370                 spin_lock_irq(&mapping->tree_lock);
371 
372                 if (pmd_downgrade) {
373                         radix_tree_delete(&mapping->page_tree, index);
374                         mapping->nrexceptional--;
375                         dax_wake_mapping_entry_waiter(mapping, index, entry,
376                                         true);
377                 }
378 
379                 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
380 
381                 err = __radix_tree_insert(&mapping->page_tree, index,
382                                 dax_radix_order(entry), entry);
383                 radix_tree_preload_end();
384                 if (err) {
385                         spin_unlock_irq(&mapping->tree_lock);
386                         /*
387                          * Someone already created the entry?  This is a
388                          * normal failure when inserting PMDs in a range
389                          * that already contains PTEs.  In that case we want
390                          * to return -EEXIST immediately.
391                          */
392                         if (err == -EEXIST && !(size_flag & RADIX_DAX_PMD))
393                                 goto restart;
394                         /*
395                          * Our insertion of a DAX PMD entry failed, most
396                          * likely because it collided with a PTE sized entry
397                          * at a different index in the PMD range.  We haven't
398                          * inserted anything into the radix tree and have no
399                          * waiters to wake.
400                          */
401                         return ERR_PTR(err);
402                 }
403                 /* Good, we have inserted empty locked entry into the tree. */
404                 mapping->nrexceptional++;
405                 spin_unlock_irq(&mapping->tree_lock);
406                 return entry;
407         }
408         /* Normal page in radix tree? */
409         if (!radix_tree_exceptional_entry(entry)) {
410                 struct page *page = entry;
411 
412                 get_page(page);
413                 spin_unlock_irq(&mapping->tree_lock);
414                 lock_page(page);
415                 /* Page got truncated? Retry... */
416                 if (unlikely(page->mapping != mapping)) {
417                         unlock_page(page);
418                         put_page(page);
419                         goto restart;
420                 }
421                 return page;
422         }
423         entry = lock_slot(mapping, slot);
424  out_unlock:
425         spin_unlock_irq(&mapping->tree_lock);
426         return entry;
427 }
428 
429 /*
430  * We do not necessarily hold the mapping->tree_lock when we call this
431  * function so it is possible that 'entry' is no longer a valid item in the
432  * radix tree.  This is okay because all we really need to do is to find the
433  * correct waitqueue where tasks might be waiting for that old 'entry' and
434  * wake them.
435  */
436 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
437                 pgoff_t index, void *entry, bool wake_all)
438 {
439         struct exceptional_entry_key key;
440         wait_queue_head_t *wq;
441 
442         wq = dax_entry_waitqueue(mapping, index, entry, &key);
443 
444         /*
445          * Checking for locked entry and prepare_to_wait_exclusive() happens
446          * under mapping->tree_lock, ditto for entry handling in our callers.
447          * So at this point all tasks that could have seen our entry locked
448          * must be in the waitqueue and the following check will see them.
449          */
450         if (waitqueue_active(wq))
451                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
452 }
453 
454 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
455                                           pgoff_t index, bool trunc)
456 {
457         int ret = 0;
458         void *entry;
459         struct radix_tree_root *page_tree = &mapping->page_tree;
460 
461         spin_lock_irq(&mapping->tree_lock);
462         entry = get_unlocked_mapping_entry(mapping, index, NULL);
463         if (!entry || !radix_tree_exceptional_entry(entry))
464                 goto out;
465         if (!trunc &&
466             (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
467              radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)))
468                 goto out;
469         radix_tree_delete(page_tree, index);
470         mapping->nrexceptional--;
471         ret = 1;
472 out:
473         put_unlocked_mapping_entry(mapping, index, entry);
474         spin_unlock_irq(&mapping->tree_lock);
475         return ret;
476 }
477 /*
478  * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
479  * entry to get unlocked before deleting it.
480  */
481 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
482 {
483         int ret = __dax_invalidate_mapping_entry(mapping, index, true);
484 
485         /*
486          * This gets called from truncate / punch_hole path. As such, the caller
487          * must hold locks protecting against concurrent modifications of the
488          * radix tree (usually fs-private i_mmap_sem for writing). Since the
489          * caller has seen exceptional entry for this index, we better find it
490          * at that index as well...
491          */
492         WARN_ON_ONCE(!ret);
493         return ret;
494 }
495 
496 /*
497  * Invalidate exceptional DAX entry if easily possible. This handles DAX
498  * entries for invalidate_inode_pages() so we evict the entry only if we can
499  * do so without blocking.
500  */
501 int dax_invalidate_mapping_entry(struct address_space *mapping, pgoff_t index)
502 {
503         int ret = 0;
504         void *entry, **slot;
505         struct radix_tree_root *page_tree = &mapping->page_tree;
506 
507         spin_lock_irq(&mapping->tree_lock);
508         entry = __radix_tree_lookup(page_tree, index, NULL, &slot);
509         if (!entry || !radix_tree_exceptional_entry(entry) ||
510             slot_locked(mapping, slot))
511                 goto out;
512         if (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
513             radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
514                 goto out;
515         radix_tree_delete(page_tree, index);
516         mapping->nrexceptional--;
517         ret = 1;
518 out:
519         spin_unlock_irq(&mapping->tree_lock);
520         if (ret)
521                 dax_wake_mapping_entry_waiter(mapping, index, entry, true);
522         return ret;
523 }
524 
525 /*
526  * Invalidate exceptional DAX entry if it is clean.
527  */
528 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
529                                       pgoff_t index)
530 {
531         return __dax_invalidate_mapping_entry(mapping, index, false);
532 }
533 
534 /*
535  * The user has performed a load from a hole in the file.  Allocating
536  * a new page in the file would cause excessive storage usage for
537  * workloads with sparse files.  We allocate a page cache page instead.
538  * We'll kick it out of the page cache if it's ever written to,
539  * otherwise it will simply fall out of the page cache under memory
540  * pressure without ever having been dirtied.
541  */
542 static int dax_load_hole(struct address_space *mapping, void **entry,
543                          struct vm_fault *vmf)
544 {
545         struct page *page;
546         int ret;
547 
548         /* Hole page already exists? Return it...  */
549         if (!radix_tree_exceptional_entry(*entry)) {
550                 page = *entry;
551                 goto out;
552         }
553 
554         /* This will replace locked radix tree entry with a hole page */
555         page = find_or_create_page(mapping, vmf->pgoff,
556                                    vmf->gfp_mask | __GFP_ZERO);
557         if (!page)
558                 return VM_FAULT_OOM;
559  out:
560         vmf->page = page;
561         ret = finish_fault(vmf);
562         vmf->page = NULL;
563         *entry = page;
564         if (!ret) {
565                 /* Grab reference for PTE that is now referencing the page */
566                 get_page(page);
567                 return VM_FAULT_NOPAGE;
568         }
569         return ret;
570 }
571 
572 static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size,
573                 struct page *to, unsigned long vaddr)
574 {
575         struct blk_dax_ctl dax = {
576                 .sector = sector,
577                 .size = size,
578         };
579         void *vto;
580 
581         if (dax_map_atomic(bdev, &dax) < 0)
582                 return PTR_ERR(dax.addr);
583         vto = kmap_atomic(to);
584         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
585         kunmap_atomic(vto);
586         dax_unmap_atomic(bdev, &dax);
587         return 0;
588 }
589 
590 /*
591  * By this point grab_mapping_entry() has ensured that we have a locked entry
592  * of the appropriate size so we don't have to worry about downgrading PMDs to
593  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
594  * already in the tree, we will skip the insertion and just dirty the PMD as
595  * appropriate.
596  */
597 static void *dax_insert_mapping_entry(struct address_space *mapping,
598                                       struct vm_fault *vmf,
599                                       void *entry, sector_t sector,
600                                       unsigned long flags)
601 {
602         struct radix_tree_root *page_tree = &mapping->page_tree;
603         int error = 0;
604         bool hole_fill = false;
605         void *new_entry;
606         pgoff_t index = vmf->pgoff;
607 
608         if (vmf->flags & FAULT_FLAG_WRITE)
609                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
610 
611         /* Replacing hole page with block mapping? */
612         if (!radix_tree_exceptional_entry(entry)) {
613                 hole_fill = true;
614                 /*
615                  * Unmap the page now before we remove it from page cache below.
616                  * The page is locked so it cannot be faulted in again.
617                  */
618                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
619                                     PAGE_SIZE, 0);
620                 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
621                 if (error)
622                         return ERR_PTR(error);
623         } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) {
624                 /* replacing huge zero page with PMD block mapping */
625                 unmap_mapping_range(mapping,
626                         (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
627         }
628 
629         spin_lock_irq(&mapping->tree_lock);
630         new_entry = dax_radix_locked_entry(sector, flags);
631 
632         if (hole_fill) {
633                 __delete_from_page_cache(entry, NULL);
634                 /* Drop pagecache reference */
635                 put_page(entry);
636                 error = __radix_tree_insert(page_tree, index,
637                                 dax_radix_order(new_entry), new_entry);
638                 if (error) {
639                         new_entry = ERR_PTR(error);
640                         goto unlock;
641                 }
642                 mapping->nrexceptional++;
643         } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
644                 /*
645                  * Only swap our new entry into the radix tree if the current
646                  * entry is a zero page or an empty entry.  If a normal PTE or
647                  * PMD entry is already in the tree, we leave it alone.  This
648                  * means that if we are trying to insert a PTE and the
649                  * existing entry is a PMD, we will just leave the PMD in the
650                  * tree and dirty it if necessary.
651                  */
652                 struct radix_tree_node *node;
653                 void **slot;
654                 void *ret;
655 
656                 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
657                 WARN_ON_ONCE(ret != entry);
658                 __radix_tree_replace(page_tree, node, slot,
659                                      new_entry, NULL, NULL);
660         }
661         if (vmf->flags & FAULT_FLAG_WRITE)
662                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
663  unlock:
664         spin_unlock_irq(&mapping->tree_lock);
665         if (hole_fill) {
666                 radix_tree_preload_end();
667                 /*
668                  * We don't need hole page anymore, it has been replaced with
669                  * locked radix tree entry now.
670                  */
671                 if (mapping->a_ops->freepage)
672                         mapping->a_ops->freepage(entry);
673                 unlock_page(entry);
674                 put_page(entry);
675         }
676         return new_entry;
677 }
678 
679 static inline unsigned long
680 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
681 {
682         unsigned long address;
683 
684         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
685         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
686         return address;
687 }
688 
689 /* Walk all mappings of a given index of a file and writeprotect them */
690 static void dax_mapping_entry_mkclean(struct address_space *mapping,
691                                       pgoff_t index, unsigned long pfn)
692 {
693         struct vm_area_struct *vma;
694         pte_t pte, *ptep = NULL;
695         pmd_t *pmdp = NULL;
696         spinlock_t *ptl;
697         bool changed;
698 
699         i_mmap_lock_read(mapping);
700         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
701                 unsigned long address;
702 
703                 cond_resched();
704 
705                 if (!(vma->vm_flags & VM_SHARED))
706                         continue;
707 
708                 address = pgoff_address(index, vma);
709                 changed = false;
710                 if (follow_pte_pmd(vma->vm_mm, address, &ptep, &pmdp, &ptl))
711                         continue;
712 
713                 if (pmdp) {
714 #ifdef CONFIG_FS_DAX_PMD
715                         pmd_t pmd;
716 
717                         if (pfn != pmd_pfn(*pmdp))
718                                 goto unlock_pmd;
719                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
720                                 goto unlock_pmd;
721 
722                         flush_cache_page(vma, address, pfn);
723                         pmd = pmdp_huge_clear_flush(vma, address, pmdp);
724                         pmd = pmd_wrprotect(pmd);
725                         pmd = pmd_mkclean(pmd);
726                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
727                         changed = true;
728 unlock_pmd:
729                         spin_unlock(ptl);
730 #endif
731                 } else {
732                         if (pfn != pte_pfn(*ptep))
733                                 goto unlock_pte;
734                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
735                                 goto unlock_pte;
736 
737                         flush_cache_page(vma, address, pfn);
738                         pte = ptep_clear_flush(vma, address, ptep);
739                         pte = pte_wrprotect(pte);
740                         pte = pte_mkclean(pte);
741                         set_pte_at(vma->vm_mm, address, ptep, pte);
742                         changed = true;
743 unlock_pte:
744                         pte_unmap_unlock(ptep, ptl);
745                 }
746 
747                 if (changed)
748                         mmu_notifier_invalidate_page(vma->vm_mm, address);
749         }
750         i_mmap_unlock_read(mapping);
751 }
752 
753 static int dax_writeback_one(struct block_device *bdev,
754                 struct address_space *mapping, pgoff_t index, void *entry)
755 {
756         struct radix_tree_root *page_tree = &mapping->page_tree;
757         struct blk_dax_ctl dax;
758         void *entry2, **slot;
759         int ret = 0;
760 
761         /*
762          * A page got tagged dirty in DAX mapping? Something is seriously
763          * wrong.
764          */
765         if (WARN_ON(!radix_tree_exceptional_entry(entry)))
766                 return -EIO;
767 
768         spin_lock_irq(&mapping->tree_lock);
769         entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
770         /* Entry got punched out / reallocated? */
771         if (!entry2 || !radix_tree_exceptional_entry(entry2))
772                 goto put_unlocked;
773         /*
774          * Entry got reallocated elsewhere? No need to writeback. We have to
775          * compare sectors as we must not bail out due to difference in lockbit
776          * or entry type.
777          */
778         if (dax_radix_sector(entry2) != dax_radix_sector(entry))
779                 goto put_unlocked;
780         if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
781                                 dax_is_zero_entry(entry))) {
782                 ret = -EIO;
783                 goto put_unlocked;
784         }
785 
786         /* Another fsync thread may have already written back this entry */
787         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
788                 goto put_unlocked;
789         /* Lock the entry to serialize with page faults */
790         entry = lock_slot(mapping, slot);
791         /*
792          * We can clear the tag now but we have to be careful so that concurrent
793          * dax_writeback_one() calls for the same index cannot finish before we
794          * actually flush the caches. This is achieved as the calls will look
795          * at the entry only under tree_lock and once they do that they will
796          * see the entry locked and wait for it to unlock.
797          */
798         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
799         spin_unlock_irq(&mapping->tree_lock);
800 
801         /*
802          * Even if dax_writeback_mapping_range() was given a wbc->range_start
803          * in the middle of a PMD, the 'index' we are given will be aligned to
804          * the start index of the PMD, as will the sector we pull from
805          * 'entry'.  This allows us to flush for PMD_SIZE and not have to
806          * worry about partial PMD writebacks.
807          */
808         dax.sector = dax_radix_sector(entry);
809         dax.size = PAGE_SIZE << dax_radix_order(entry);
810 
811         /*
812          * We cannot hold tree_lock while calling dax_map_atomic() because it
813          * eventually calls cond_resched().
814          */
815         ret = dax_map_atomic(bdev, &dax);
816         if (ret < 0) {
817                 put_locked_mapping_entry(mapping, index, entry);
818                 return ret;
819         }
820 
821         if (WARN_ON_ONCE(ret < dax.size)) {
822                 ret = -EIO;
823                 goto unmap;
824         }
825 
826         dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(dax.pfn));
827         wb_cache_pmem(dax.addr, dax.size);
828         /*
829          * After we have flushed the cache, we can clear the dirty tag. There
830          * cannot be new dirty data in the pfn after the flush has completed as
831          * the pfn mappings are writeprotected and fault waits for mapping
832          * entry lock.
833          */
834         spin_lock_irq(&mapping->tree_lock);
835         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
836         spin_unlock_irq(&mapping->tree_lock);
837  unmap:
838         dax_unmap_atomic(bdev, &dax);
839         put_locked_mapping_entry(mapping, index, entry);
840         return ret;
841 
842  put_unlocked:
843         put_unlocked_mapping_entry(mapping, index, entry2);
844         spin_unlock_irq(&mapping->tree_lock);
845         return ret;
846 }
847 
848 /*
849  * Flush the mapping to the persistent domain within the byte range of [start,
850  * end]. This is required by data integrity operations to ensure file data is
851  * on persistent storage prior to completion of the operation.
852  */
853 int dax_writeback_mapping_range(struct address_space *mapping,
854                 struct block_device *bdev, struct writeback_control *wbc)
855 {
856         struct inode *inode = mapping->host;
857         pgoff_t start_index, end_index;
858         pgoff_t indices[PAGEVEC_SIZE];
859         struct pagevec pvec;
860         bool done = false;
861         int i, ret = 0;
862 
863         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
864                 return -EIO;
865 
866         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
867                 return 0;
868 
869         start_index = wbc->range_start >> PAGE_SHIFT;
870         end_index = wbc->range_end >> PAGE_SHIFT;
871 
872         tag_pages_for_writeback(mapping, start_index, end_index);
873 
874         pagevec_init(&pvec, 0);
875         while (!done) {
876                 pvec.nr = find_get_entries_tag(mapping, start_index,
877                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
878                                 pvec.pages, indices);
879 
880                 if (pvec.nr == 0)
881                         break;
882 
883                 for (i = 0; i < pvec.nr; i++) {
884                         if (indices[i] > end_index) {
885                                 done = true;
886                                 break;
887                         }
888 
889                         ret = dax_writeback_one(bdev, mapping, indices[i],
890                                         pvec.pages[i]);
891                         if (ret < 0)
892                                 return ret;
893                 }
894         }
895         return 0;
896 }
897 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
898 
899 static int dax_insert_mapping(struct address_space *mapping,
900                 struct block_device *bdev, sector_t sector, size_t size,
901                 void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf)
902 {
903         unsigned long vaddr = vmf->address;
904         struct blk_dax_ctl dax = {
905                 .sector = sector,
906                 .size = size,
907         };
908         void *ret;
909         void *entry = *entryp;
910 
911         if (dax_map_atomic(bdev, &dax) < 0)
912                 return PTR_ERR(dax.addr);
913         dax_unmap_atomic(bdev, &dax);
914 
915         ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector, 0);
916         if (IS_ERR(ret))
917                 return PTR_ERR(ret);
918         *entryp = ret;
919 
920         return vm_insert_mixed(vma, vaddr, dax.pfn);
921 }
922 
923 /**
924  * dax_pfn_mkwrite - handle first write to DAX page
925  * @vma: The virtual memory area where the fault occurred
926  * @vmf: The description of the fault
927  */
928 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
929 {
930         struct file *file = vma->vm_file;
931         struct address_space *mapping = file->f_mapping;
932         void *entry, **slot;
933         pgoff_t index = vmf->pgoff;
934 
935         spin_lock_irq(&mapping->tree_lock);
936         entry = get_unlocked_mapping_entry(mapping, index, &slot);
937         if (!entry || !radix_tree_exceptional_entry(entry)) {
938                 if (entry)
939                         put_unlocked_mapping_entry(mapping, index, entry);
940                 spin_unlock_irq(&mapping->tree_lock);
941                 return VM_FAULT_NOPAGE;
942         }
943         radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
944         entry = lock_slot(mapping, slot);
945         spin_unlock_irq(&mapping->tree_lock);
946         /*
947          * If we race with somebody updating the PTE and finish_mkwrite_fault()
948          * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry
949          * the fault in either case.
950          */
951         finish_mkwrite_fault(vmf);
952         put_locked_mapping_entry(mapping, index, entry);
953         return VM_FAULT_NOPAGE;
954 }
955 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
956 
957 static bool dax_range_is_aligned(struct block_device *bdev,
958                                  unsigned int offset, unsigned int length)
959 {
960         unsigned short sector_size = bdev_logical_block_size(bdev);
961 
962         if (!IS_ALIGNED(offset, sector_size))
963                 return false;
964         if (!IS_ALIGNED(length, sector_size))
965                 return false;
966 
967         return true;
968 }
969 
970 int __dax_zero_page_range(struct block_device *bdev, sector_t sector,
971                 unsigned int offset, unsigned int length)
972 {
973         struct blk_dax_ctl dax = {
974                 .sector         = sector,
975                 .size           = PAGE_SIZE,
976         };
977 
978         if (dax_range_is_aligned(bdev, offset, length)) {
979                 sector_t start_sector = dax.sector + (offset >> 9);
980 
981                 return blkdev_issue_zeroout(bdev, start_sector,
982                                 length >> 9, GFP_NOFS, true);
983         } else {
984                 if (dax_map_atomic(bdev, &dax) < 0)
985                         return PTR_ERR(dax.addr);
986                 clear_pmem(dax.addr + offset, length);
987                 dax_unmap_atomic(bdev, &dax);
988         }
989         return 0;
990 }
991 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
992 
993 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
994 {
995         return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
996 }
997 
998 static loff_t
999 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1000                 struct iomap *iomap)
1001 {
1002         struct iov_iter *iter = data;
1003         loff_t end = pos + length, done = 0;
1004         ssize_t ret = 0;
1005 
1006         if (iov_iter_rw(iter) == READ) {
1007                 end = min(end, i_size_read(inode));
1008                 if (pos >= end)
1009                         return 0;
1010 
1011                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1012                         return iov_iter_zero(min(length, end - pos), iter);
1013         }
1014 
1015         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1016                 return -EIO;
1017 
1018         /*
1019          * Write can allocate block for an area which has a hole page mapped
1020          * into page tables. We have to tear down these mappings so that data
1021          * written by write(2) is visible in mmap.
1022          */
1023         if ((iomap->flags & IOMAP_F_NEW) && inode->i_mapping->nrpages) {
1024                 invalidate_inode_pages2_range(inode->i_mapping,
1025                                               pos >> PAGE_SHIFT,
1026                                               (end - 1) >> PAGE_SHIFT);
1027         }
1028 
1029         while (pos < end) {
1030                 unsigned offset = pos & (PAGE_SIZE - 1);
1031                 struct blk_dax_ctl dax = { 0 };
1032                 ssize_t map_len;
1033 
1034                 if (fatal_signal_pending(current)) {
1035                         ret = -EINTR;
1036                         break;
1037                 }
1038 
1039                 dax.sector = dax_iomap_sector(iomap, pos);
1040                 dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK;
1041                 map_len = dax_map_atomic(iomap->bdev, &dax);
1042                 if (map_len < 0) {
1043                         ret = map_len;
1044                         break;
1045                 }
1046 
1047                 dax.addr += offset;
1048                 map_len -= offset;
1049                 if (map_len > end - pos)
1050                         map_len = end - pos;
1051 
1052                 if (iov_iter_rw(iter) == WRITE)
1053                         map_len = copy_from_iter_pmem(dax.addr, map_len, iter);
1054                 else
1055                         map_len = copy_to_iter(dax.addr, map_len, iter);
1056                 dax_unmap_atomic(iomap->bdev, &dax);
1057                 if (map_len <= 0) {
1058                         ret = map_len ? map_len : -EFAULT;
1059                         break;
1060                 }
1061 
1062                 pos += map_len;
1063                 length -= map_len;
1064                 done += map_len;
1065         }
1066 
1067         return done ? done : ret;
1068 }
1069 
1070 /**
1071  * dax_iomap_rw - Perform I/O to a DAX file
1072  * @iocb:       The control block for this I/O
1073  * @iter:       The addresses to do I/O from or to
1074  * @ops:        iomap ops passed from the file system
1075  *
1076  * This function performs read and write operations to directly mapped
1077  * persistent memory.  The callers needs to take care of read/write exclusion
1078  * and evicting any page cache pages in the region under I/O.
1079  */
1080 ssize_t
1081 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1082                 struct iomap_ops *ops)
1083 {
1084         struct address_space *mapping = iocb->ki_filp->f_mapping;
1085         struct inode *inode = mapping->host;
1086         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1087         unsigned flags = 0;
1088 
1089         if (iov_iter_rw(iter) == WRITE)
1090                 flags |= IOMAP_WRITE;
1091 
1092         while (iov_iter_count(iter)) {
1093                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1094                                 iter, dax_iomap_actor);
1095                 if (ret <= 0)
1096                         break;
1097                 pos += ret;
1098                 done += ret;
1099         }
1100 
1101         iocb->ki_pos += done;
1102         return done ? done : ret;
1103 }
1104 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1105 
1106 static int dax_fault_return(int error)
1107 {
1108         if (error == 0)
1109                 return VM_FAULT_NOPAGE;
1110         if (error == -ENOMEM)
1111                 return VM_FAULT_OOM;
1112         return VM_FAULT_SIGBUS;
1113 }
1114 
1115 /**
1116  * dax_iomap_fault - handle a page fault on a DAX file
1117  * @vma: The virtual memory area where the fault occurred
1118  * @vmf: The description of the fault
1119  * @ops: iomap ops passed from the file system
1120  *
1121  * When a page fault occurs, filesystems may call this helper in their fault
1122  * or mkwrite handler for DAX files. Assumes the caller has done all the
1123  * necessary locking for the page fault to proceed successfully.
1124  */
1125 int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
1126                         struct iomap_ops *ops)
1127 {
1128         struct address_space *mapping = vma->vm_file->f_mapping;
1129         struct inode *inode = mapping->host;
1130         unsigned long vaddr = vmf->address;
1131         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1132         sector_t sector;
1133         struct iomap iomap = { 0 };
1134         unsigned flags = IOMAP_FAULT;
1135         int error, major = 0;
1136         int vmf_ret = 0;
1137         void *entry;
1138 
1139         /*
1140          * Check whether offset isn't beyond end of file now. Caller is supposed
1141          * to hold locks serializing us with truncate / punch hole so this is
1142          * a reliable test.
1143          */
1144         if (pos >= i_size_read(inode))
1145                 return VM_FAULT_SIGBUS;
1146 
1147         if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1148                 flags |= IOMAP_WRITE;
1149 
1150         /*
1151          * Note that we don't bother to use iomap_apply here: DAX required
1152          * the file system block size to be equal the page size, which means
1153          * that we never have to deal with more than a single extent here.
1154          */
1155         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1156         if (error)
1157                 return dax_fault_return(error);
1158         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1159                 vmf_ret = dax_fault_return(-EIO);       /* fs corruption? */
1160                 goto finish_iomap;
1161         }
1162 
1163         entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1164         if (IS_ERR(entry)) {
1165                 vmf_ret = dax_fault_return(PTR_ERR(entry));
1166                 goto finish_iomap;
1167         }
1168 
1169         sector = dax_iomap_sector(&iomap, pos);
1170 
1171         if (vmf->cow_page) {
1172                 switch (iomap.type) {
1173                 case IOMAP_HOLE:
1174                 case IOMAP_UNWRITTEN:
1175                         clear_user_highpage(vmf->cow_page, vaddr);
1176                         break;
1177                 case IOMAP_MAPPED:
1178                         error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE,
1179                                         vmf->cow_page, vaddr);
1180                         break;
1181                 default:
1182                         WARN_ON_ONCE(1);
1183                         error = -EIO;
1184                         break;
1185                 }
1186 
1187                 if (error)
1188                         goto error_unlock_entry;
1189 
1190                 __SetPageUptodate(vmf->cow_page);
1191                 vmf_ret = finish_fault(vmf);
1192                 if (!vmf_ret)
1193                         vmf_ret = VM_FAULT_DONE_COW;
1194                 goto unlock_entry;
1195         }
1196 
1197         switch (iomap.type) {
1198         case IOMAP_MAPPED:
1199                 if (iomap.flags & IOMAP_F_NEW) {
1200                         count_vm_event(PGMAJFAULT);
1201                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1202                         major = VM_FAULT_MAJOR;
1203                 }
1204                 error = dax_insert_mapping(mapping, iomap.bdev, sector,
1205                                 PAGE_SIZE, &entry, vma, vmf);
1206                 /* -EBUSY is fine, somebody else faulted on the same PTE */
1207                 if (error == -EBUSY)
1208                         error = 0;
1209                 break;
1210         case IOMAP_UNWRITTEN:
1211         case IOMAP_HOLE:
1212                 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
1213                         vmf_ret = dax_load_hole(mapping, &entry, vmf);
1214                         goto unlock_entry;
1215                 }
1216                 /*FALLTHRU*/
1217         default:
1218                 WARN_ON_ONCE(1);
1219                 error = -EIO;
1220                 break;
1221         }
1222 
1223  error_unlock_entry:
1224         vmf_ret = dax_fault_return(error) | major;
1225  unlock_entry:
1226         put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1227  finish_iomap:
1228         if (ops->iomap_end) {
1229                 int copied = PAGE_SIZE;
1230 
1231                 if (vmf_ret & VM_FAULT_ERROR)
1232                         copied = 0;
1233                 /*
1234                  * The fault is done by now and there's no way back (other
1235                  * thread may be already happily using PTE we have installed).
1236                  * Just ignore error from ->iomap_end since we cannot do much
1237                  * with it.
1238                  */
1239                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1240         }
1241         return vmf_ret;
1242 }
1243 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1244 
1245 #ifdef CONFIG_FS_DAX_PMD
1246 /*
1247  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
1248  * more often than one might expect in the below functions.
1249  */
1250 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
1251 
1252 static int dax_pmd_insert_mapping(struct vm_area_struct *vma, pmd_t *pmd,
1253                 struct vm_fault *vmf, unsigned long address,
1254                 struct iomap *iomap, loff_t pos, bool write, void **entryp)
1255 {
1256         struct address_space *mapping = vma->vm_file->f_mapping;
1257         struct block_device *bdev = iomap->bdev;
1258         struct blk_dax_ctl dax = {
1259                 .sector = dax_iomap_sector(iomap, pos),
1260                 .size = PMD_SIZE,
1261         };
1262         long length = dax_map_atomic(bdev, &dax);
1263         void *ret;
1264 
1265         if (length < 0) /* dax_map_atomic() failed */
1266                 return VM_FAULT_FALLBACK;
1267         if (length < PMD_SIZE)
1268                 goto unmap_fallback;
1269         if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR)
1270                 goto unmap_fallback;
1271         if (!pfn_t_devmap(dax.pfn))
1272                 goto unmap_fallback;
1273 
1274         dax_unmap_atomic(bdev, &dax);
1275 
1276         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, dax.sector,
1277                         RADIX_DAX_PMD);
1278         if (IS_ERR(ret))
1279                 return VM_FAULT_FALLBACK;
1280         *entryp = ret;
1281 
1282         return vmf_insert_pfn_pmd(vma, address, pmd, dax.pfn, write);
1283 
1284  unmap_fallback:
1285         dax_unmap_atomic(bdev, &dax);
1286         return VM_FAULT_FALLBACK;
1287 }
1288 
1289 static int dax_pmd_load_hole(struct vm_area_struct *vma, pmd_t *pmd,
1290                 struct vm_fault *vmf, unsigned long address,
1291                 struct iomap *iomap, void **entryp)
1292 {
1293         struct address_space *mapping = vma->vm_file->f_mapping;
1294         unsigned long pmd_addr = address & PMD_MASK;
1295         struct page *zero_page;
1296         spinlock_t *ptl;
1297         pmd_t pmd_entry;
1298         void *ret;
1299 
1300         zero_page = mm_get_huge_zero_page(vma->vm_mm);
1301 
1302         if (unlikely(!zero_page))
1303                 return VM_FAULT_FALLBACK;
1304 
1305         ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0,
1306                         RADIX_DAX_PMD | RADIX_DAX_HZP);
1307         if (IS_ERR(ret))
1308                 return VM_FAULT_FALLBACK;
1309         *entryp = ret;
1310 
1311         ptl = pmd_lock(vma->vm_mm, pmd);
1312         if (!pmd_none(*pmd)) {
1313                 spin_unlock(ptl);
1314                 return VM_FAULT_FALLBACK;
1315         }
1316 
1317         pmd_entry = mk_pmd(zero_page, vma->vm_page_prot);
1318         pmd_entry = pmd_mkhuge(pmd_entry);
1319         set_pmd_at(vma->vm_mm, pmd_addr, pmd, pmd_entry);
1320         spin_unlock(ptl);
1321         return VM_FAULT_NOPAGE;
1322 }
1323 
1324 int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1325                 pmd_t *pmd, unsigned int flags, struct iomap_ops *ops)
1326 {
1327         struct address_space *mapping = vma->vm_file->f_mapping;
1328         unsigned long pmd_addr = address & PMD_MASK;
1329         bool write = flags & FAULT_FLAG_WRITE;
1330         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1331         struct inode *inode = mapping->host;
1332         int result = VM_FAULT_FALLBACK;
1333         struct iomap iomap = { 0 };
1334         pgoff_t max_pgoff, pgoff;
1335         struct vm_fault vmf;
1336         void *entry;
1337         loff_t pos;
1338         int error;
1339 
1340         /* Fall back to PTEs if we're going to COW */
1341         if (write && !(vma->vm_flags & VM_SHARED))
1342                 goto fallback;
1343 
1344         /* If the PMD would extend outside the VMA */
1345         if (pmd_addr < vma->vm_start)
1346                 goto fallback;
1347         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1348                 goto fallback;
1349 
1350         /*
1351          * Check whether offset isn't beyond end of file now. Caller is
1352          * supposed to hold locks serializing us with truncate / punch hole so
1353          * this is a reliable test.
1354          */
1355         pgoff = linear_page_index(vma, pmd_addr);
1356         max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1357 
1358         if (pgoff > max_pgoff)
1359                 return VM_FAULT_SIGBUS;
1360 
1361         /* If the PMD would extend beyond the file size */
1362         if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1363                 goto fallback;
1364 
1365         /*
1366          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1367          * setting up a mapping, so really we're using iomap_begin() as a way
1368          * to look up our filesystem block.
1369          */
1370         pos = (loff_t)pgoff << PAGE_SHIFT;
1371         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1372         if (error)
1373                 goto fallback;
1374 
1375         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1376                 goto finish_iomap;
1377 
1378         /*
1379          * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1380          * PMD or a HZP entry.  If it can't (because a 4k page is already in
1381          * the tree, for instance), it will return -EEXIST and we just fall
1382          * back to 4k entries.
1383          */
1384         entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1385         if (IS_ERR(entry))
1386                 goto finish_iomap;
1387 
1388         vmf.pgoff = pgoff;
1389         vmf.flags = flags;
1390         vmf.gfp_mask = mapping_gfp_mask(mapping) | __GFP_IO;
1391 
1392         switch (iomap.type) {
1393         case IOMAP_MAPPED:
1394                 result = dax_pmd_insert_mapping(vma, pmd, &vmf, address,
1395                                 &iomap, pos, write, &entry);
1396                 break;
1397         case IOMAP_UNWRITTEN:
1398         case IOMAP_HOLE:
1399                 if (WARN_ON_ONCE(write))
1400                         goto unlock_entry;
1401                 result = dax_pmd_load_hole(vma, pmd, &vmf, address, &iomap,
1402                                 &entry);
1403                 break;
1404         default:
1405                 WARN_ON_ONCE(1);
1406                 break;
1407         }
1408 
1409  unlock_entry:
1410         put_locked_mapping_entry(mapping, pgoff, entry);
1411  finish_iomap:
1412         if (ops->iomap_end) {
1413                 int copied = PMD_SIZE;
1414 
1415                 if (result == VM_FAULT_FALLBACK)
1416                         copied = 0;
1417                 /*
1418                  * The fault is done by now and there's no way back (other
1419                  * thread may be already happily using PMD we have installed).
1420                  * Just ignore error from ->iomap_end since we cannot do much
1421                  * with it.
1422                  */
1423                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1424                                 &iomap);
1425         }
1426  fallback:
1427         if (result == VM_FAULT_FALLBACK) {
1428                 split_huge_pmd(vma, pmd, address);
1429                 count_vm_event(THP_FAULT_FALLBACK);
1430         }
1431         return result;
1432 }
1433 EXPORT_SYMBOL_GPL(dax_iomap_pmd_fault);
1434 #endif /* CONFIG_FS_DAX_PMD */
1435 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us