]>
Commit | Line | Data |
---|---|---|
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/sched/signal.h> | |
31 | #include <linux/uio.h> | |
32 | #include <linux/vmstat.h> | |
33 | #include <linux/pfn_t.h> | |
34 | #include <linux/sizes.h> | |
35 | #include <linux/mmu_notifier.h> | |
36 | #include <linux/iomap.h> | |
37 | #include "internal.h" | |
38 | ||
39 | #define CREATE_TRACE_POINTS | |
40 | #include <trace/events/fs_dax.h> | |
41 | ||
42 | /* We choose 4096 entries - same as per-zone page wait tables */ | |
43 | #define DAX_WAIT_TABLE_BITS 12 | |
44 | #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) | |
45 | ||
46 | static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; | |
47 | ||
48 | static int __init init_dax_wait_table(void) | |
49 | { | |
50 | int i; | |
51 | ||
52 | for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) | |
53 | init_waitqueue_head(wait_table + i); | |
54 | return 0; | |
55 | } | |
56 | fs_initcall(init_dax_wait_table); | |
57 | ||
58 | static int dax_is_pmd_entry(void *entry) | |
59 | { | |
60 | return (unsigned long)entry & RADIX_DAX_PMD; | |
61 | } | |
62 | ||
63 | static int dax_is_pte_entry(void *entry) | |
64 | { | |
65 | return !((unsigned long)entry & RADIX_DAX_PMD); | |
66 | } | |
67 | ||
68 | static int dax_is_zero_entry(void *entry) | |
69 | { | |
70 | return (unsigned long)entry & RADIX_DAX_HZP; | |
71 | } | |
72 | ||
73 | static int dax_is_empty_entry(void *entry) | |
74 | { | |
75 | return (unsigned long)entry & RADIX_DAX_EMPTY; | |
76 | } | |
77 | ||
78 | /* | |
79 | * DAX radix tree locking | |
80 | */ | |
81 | struct exceptional_entry_key { | |
82 | struct address_space *mapping; | |
83 | pgoff_t entry_start; | |
84 | }; | |
85 | ||
86 | struct wait_exceptional_entry_queue { | |
87 | wait_queue_t wait; | |
88 | struct exceptional_entry_key key; | |
89 | }; | |
90 | ||
91 | static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping, | |
92 | pgoff_t index, void *entry, struct exceptional_entry_key *key) | |
93 | { | |
94 | unsigned long hash; | |
95 | ||
96 | /* | |
97 | * If 'entry' is a PMD, align the 'index' that we use for the wait | |
98 | * queue to the start of that PMD. This ensures that all offsets in | |
99 | * the range covered by the PMD map to the same bit lock. | |
100 | */ | |
101 | if (dax_is_pmd_entry(entry)) | |
102 | index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1); | |
103 | ||
104 | key->mapping = mapping; | |
105 | key->entry_start = index; | |
106 | ||
107 | hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS); | |
108 | return wait_table + hash; | |
109 | } | |
110 | ||
111 | static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode, | |
112 | int sync, void *keyp) | |
113 | { | |
114 | struct exceptional_entry_key *key = keyp; | |
115 | struct wait_exceptional_entry_queue *ewait = | |
116 | container_of(wait, struct wait_exceptional_entry_queue, wait); | |
117 | ||
118 | if (key->mapping != ewait->key.mapping || | |
119 | key->entry_start != ewait->key.entry_start) | |
120 | return 0; | |
121 | return autoremove_wake_function(wait, mode, sync, NULL); | |
122 | } | |
123 | ||
124 | /* | |
125 | * Check whether the given slot is locked. The function must be called with | |
126 | * mapping->tree_lock held | |
127 | */ | |
128 | static inline int slot_locked(struct address_space *mapping, void **slot) | |
129 | { | |
130 | unsigned long entry = (unsigned long) | |
131 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
132 | return entry & RADIX_DAX_ENTRY_LOCK; | |
133 | } | |
134 | ||
135 | /* | |
136 | * Mark the given slot is locked. The function must be called with | |
137 | * mapping->tree_lock held | |
138 | */ | |
139 | static inline void *lock_slot(struct address_space *mapping, void **slot) | |
140 | { | |
141 | unsigned long entry = (unsigned long) | |
142 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
143 | ||
144 | entry |= RADIX_DAX_ENTRY_LOCK; | |
145 | radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); | |
146 | return (void *)entry; | |
147 | } | |
148 | ||
149 | /* | |
150 | * Mark the given slot is unlocked. The function must be called with | |
151 | * mapping->tree_lock held | |
152 | */ | |
153 | static inline void *unlock_slot(struct address_space *mapping, void **slot) | |
154 | { | |
155 | unsigned long entry = (unsigned long) | |
156 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
157 | ||
158 | entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK; | |
159 | radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); | |
160 | return (void *)entry; | |
161 | } | |
162 | ||
163 | /* | |
164 | * Lookup entry in radix tree, wait for it to become unlocked if it is | |
165 | * exceptional entry and return it. The caller must call | |
166 | * put_unlocked_mapping_entry() when he decided not to lock the entry or | |
167 | * put_locked_mapping_entry() when he locked the entry and now wants to | |
168 | * unlock it. | |
169 | * | |
170 | * The function must be called with mapping->tree_lock held. | |
171 | */ | |
172 | static void *get_unlocked_mapping_entry(struct address_space *mapping, | |
173 | pgoff_t index, void ***slotp) | |
174 | { | |
175 | void *entry, **slot; | |
176 | struct wait_exceptional_entry_queue ewait; | |
177 | wait_queue_head_t *wq; | |
178 | ||
179 | init_wait(&ewait.wait); | |
180 | ewait.wait.func = wake_exceptional_entry_func; | |
181 | ||
182 | for (;;) { | |
183 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, | |
184 | &slot); | |
185 | if (!entry || !radix_tree_exceptional_entry(entry) || | |
186 | !slot_locked(mapping, slot)) { | |
187 | if (slotp) | |
188 | *slotp = slot; | |
189 | return entry; | |
190 | } | |
191 | ||
192 | wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key); | |
193 | prepare_to_wait_exclusive(wq, &ewait.wait, | |
194 | TASK_UNINTERRUPTIBLE); | |
195 | spin_unlock_irq(&mapping->tree_lock); | |
196 | schedule(); | |
197 | finish_wait(wq, &ewait.wait); | |
198 | spin_lock_irq(&mapping->tree_lock); | |
199 | } | |
200 | } | |
201 | ||
202 | static void dax_unlock_mapping_entry(struct address_space *mapping, | |
203 | pgoff_t index) | |
204 | { | |
205 | void *entry, **slot; | |
206 | ||
207 | spin_lock_irq(&mapping->tree_lock); | |
208 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot); | |
209 | if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) || | |
210 | !slot_locked(mapping, slot))) { | |
211 | spin_unlock_irq(&mapping->tree_lock); | |
212 | return; | |
213 | } | |
214 | unlock_slot(mapping, slot); | |
215 | spin_unlock_irq(&mapping->tree_lock); | |
216 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); | |
217 | } | |
218 | ||
219 | static void put_locked_mapping_entry(struct address_space *mapping, | |
220 | pgoff_t index, void *entry) | |
221 | { | |
222 | if (!radix_tree_exceptional_entry(entry)) { | |
223 | unlock_page(entry); | |
224 | put_page(entry); | |
225 | } else { | |
226 | dax_unlock_mapping_entry(mapping, index); | |
227 | } | |
228 | } | |
229 | ||
230 | /* | |
231 | * Called when we are done with radix tree entry we looked up via | |
232 | * get_unlocked_mapping_entry() and which we didn't lock in the end. | |
233 | */ | |
234 | static void put_unlocked_mapping_entry(struct address_space *mapping, | |
235 | pgoff_t index, void *entry) | |
236 | { | |
237 | if (!radix_tree_exceptional_entry(entry)) | |
238 | return; | |
239 | ||
240 | /* We have to wake up next waiter for the radix tree entry lock */ | |
241 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); | |
242 | } | |
243 | ||
244 | /* | |
245 | * Find radix tree entry at given index. If it points to a page, return with | |
246 | * the page locked. If it points to the exceptional entry, return with the | |
247 | * radix tree entry locked. If the radix tree doesn't contain given index, | |
248 | * create empty exceptional entry for the index and return with it locked. | |
249 | * | |
250 | * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will | |
251 | * either return that locked entry or will return an error. This error will | |
252 | * happen if there are any 4k entries (either zero pages or DAX entries) | |
253 | * within the 2MiB range that we are requesting. | |
254 | * | |
255 | * We always favor 4k entries over 2MiB entries. There isn't a flow where we | |
256 | * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB | |
257 | * insertion will fail if it finds any 4k entries already in the tree, and a | |
258 | * 4k insertion will cause an existing 2MiB entry to be unmapped and | |
259 | * downgraded to 4k entries. This happens for both 2MiB huge zero pages as | |
260 | * well as 2MiB empty entries. | |
261 | * | |
262 | * The exception to this downgrade path is for 2MiB DAX PMD entries that have | |
263 | * real storage backing them. We will leave these real 2MiB DAX entries in | |
264 | * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry. | |
265 | * | |
266 | * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For | |
267 | * persistent memory the benefit is doubtful. We can add that later if we can | |
268 | * show it helps. | |
269 | */ | |
270 | static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index, | |
271 | unsigned long size_flag) | |
272 | { | |
273 | bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */ | |
274 | void *entry, **slot; | |
275 | ||
276 | restart: | |
277 | spin_lock_irq(&mapping->tree_lock); | |
278 | entry = get_unlocked_mapping_entry(mapping, index, &slot); | |
279 | ||
280 | if (entry) { | |
281 | if (size_flag & RADIX_DAX_PMD) { | |
282 | if (!radix_tree_exceptional_entry(entry) || | |
283 | dax_is_pte_entry(entry)) { | |
284 | put_unlocked_mapping_entry(mapping, index, | |
285 | entry); | |
286 | entry = ERR_PTR(-EEXIST); | |
287 | goto out_unlock; | |
288 | } | |
289 | } else { /* trying to grab a PTE entry */ | |
290 | if (radix_tree_exceptional_entry(entry) && | |
291 | dax_is_pmd_entry(entry) && | |
292 | (dax_is_zero_entry(entry) || | |
293 | dax_is_empty_entry(entry))) { | |
294 | pmd_downgrade = true; | |
295 | } | |
296 | } | |
297 | } | |
298 | ||
299 | /* No entry for given index? Make sure radix tree is big enough. */ | |
300 | if (!entry || pmd_downgrade) { | |
301 | int err; | |
302 | ||
303 | if (pmd_downgrade) { | |
304 | /* | |
305 | * Make sure 'entry' remains valid while we drop | |
306 | * mapping->tree_lock. | |
307 | */ | |
308 | entry = lock_slot(mapping, slot); | |
309 | } | |
310 | ||
311 | spin_unlock_irq(&mapping->tree_lock); | |
312 | /* | |
313 | * Besides huge zero pages the only other thing that gets | |
314 | * downgraded are empty entries which don't need to be | |
315 | * unmapped. | |
316 | */ | |
317 | if (pmd_downgrade && dax_is_zero_entry(entry)) | |
318 | unmap_mapping_range(mapping, | |
319 | (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0); | |
320 | ||
321 | err = radix_tree_preload( | |
322 | mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM); | |
323 | if (err) { | |
324 | if (pmd_downgrade) | |
325 | put_locked_mapping_entry(mapping, index, entry); | |
326 | return ERR_PTR(err); | |
327 | } | |
328 | spin_lock_irq(&mapping->tree_lock); | |
329 | ||
330 | if (!entry) { | |
331 | /* | |
332 | * We needed to drop the page_tree lock while calling | |
333 | * radix_tree_preload() and we didn't have an entry to | |
334 | * lock. See if another thread inserted an entry at | |
335 | * our index during this time. | |
336 | */ | |
337 | entry = __radix_tree_lookup(&mapping->page_tree, index, | |
338 | NULL, &slot); | |
339 | if (entry) { | |
340 | radix_tree_preload_end(); | |
341 | spin_unlock_irq(&mapping->tree_lock); | |
342 | goto restart; | |
343 | } | |
344 | } | |
345 | ||
346 | if (pmd_downgrade) { | |
347 | radix_tree_delete(&mapping->page_tree, index); | |
348 | mapping->nrexceptional--; | |
349 | dax_wake_mapping_entry_waiter(mapping, index, entry, | |
350 | true); | |
351 | } | |
352 | ||
353 | entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY); | |
354 | ||
355 | err = __radix_tree_insert(&mapping->page_tree, index, | |
356 | dax_radix_order(entry), entry); | |
357 | radix_tree_preload_end(); | |
358 | if (err) { | |
359 | spin_unlock_irq(&mapping->tree_lock); | |
360 | /* | |
361 | * Our insertion of a DAX entry failed, most likely | |
362 | * because we were inserting a PMD entry and it | |
363 | * collided with a PTE sized entry at a different | |
364 | * index in the PMD range. We haven't inserted | |
365 | * anything into the radix tree and have no waiters to | |
366 | * wake. | |
367 | */ | |
368 | return ERR_PTR(err); | |
369 | } | |
370 | /* Good, we have inserted empty locked entry into the tree. */ | |
371 | mapping->nrexceptional++; | |
372 | spin_unlock_irq(&mapping->tree_lock); | |
373 | return entry; | |
374 | } | |
375 | /* Normal page in radix tree? */ | |
376 | if (!radix_tree_exceptional_entry(entry)) { | |
377 | struct page *page = entry; | |
378 | ||
379 | get_page(page); | |
380 | spin_unlock_irq(&mapping->tree_lock); | |
381 | lock_page(page); | |
382 | /* Page got truncated? Retry... */ | |
383 | if (unlikely(page->mapping != mapping)) { | |
384 | unlock_page(page); | |
385 | put_page(page); | |
386 | goto restart; | |
387 | } | |
388 | return page; | |
389 | } | |
390 | entry = lock_slot(mapping, slot); | |
391 | out_unlock: | |
392 | spin_unlock_irq(&mapping->tree_lock); | |
393 | return entry; | |
394 | } | |
395 | ||
396 | /* | |
397 | * We do not necessarily hold the mapping->tree_lock when we call this | |
398 | * function so it is possible that 'entry' is no longer a valid item in the | |
399 | * radix tree. This is okay because all we really need to do is to find the | |
400 | * correct waitqueue where tasks might be waiting for that old 'entry' and | |
401 | * wake them. | |
402 | */ | |
403 | void dax_wake_mapping_entry_waiter(struct address_space *mapping, | |
404 | pgoff_t index, void *entry, bool wake_all) | |
405 | { | |
406 | struct exceptional_entry_key key; | |
407 | wait_queue_head_t *wq; | |
408 | ||
409 | wq = dax_entry_waitqueue(mapping, index, entry, &key); | |
410 | ||
411 | /* | |
412 | * Checking for locked entry and prepare_to_wait_exclusive() happens | |
413 | * under mapping->tree_lock, ditto for entry handling in our callers. | |
414 | * So at this point all tasks that could have seen our entry locked | |
415 | * must be in the waitqueue and the following check will see them. | |
416 | */ | |
417 | if (waitqueue_active(wq)) | |
418 | __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); | |
419 | } | |
420 | ||
421 | static int __dax_invalidate_mapping_entry(struct address_space *mapping, | |
422 | pgoff_t index, bool trunc) | |
423 | { | |
424 | int ret = 0; | |
425 | void *entry; | |
426 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
427 | ||
428 | spin_lock_irq(&mapping->tree_lock); | |
429 | entry = get_unlocked_mapping_entry(mapping, index, NULL); | |
430 | if (!entry || !radix_tree_exceptional_entry(entry)) | |
431 | goto out; | |
432 | if (!trunc && | |
433 | (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) || | |
434 | radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))) | |
435 | goto out; | |
436 | radix_tree_delete(page_tree, index); | |
437 | mapping->nrexceptional--; | |
438 | ret = 1; | |
439 | out: | |
440 | put_unlocked_mapping_entry(mapping, index, entry); | |
441 | spin_unlock_irq(&mapping->tree_lock); | |
442 | return ret; | |
443 | } | |
444 | /* | |
445 | * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree | |
446 | * entry to get unlocked before deleting it. | |
447 | */ | |
448 | int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) | |
449 | { | |
450 | int ret = __dax_invalidate_mapping_entry(mapping, index, true); | |
451 | ||
452 | /* | |
453 | * This gets called from truncate / punch_hole path. As such, the caller | |
454 | * must hold locks protecting against concurrent modifications of the | |
455 | * radix tree (usually fs-private i_mmap_sem for writing). Since the | |
456 | * caller has seen exceptional entry for this index, we better find it | |
457 | * at that index as well... | |
458 | */ | |
459 | WARN_ON_ONCE(!ret); | |
460 | return ret; | |
461 | } | |
462 | ||
463 | /* | |
464 | * Invalidate exceptional DAX entry if easily possible. This handles DAX | |
465 | * entries for invalidate_inode_pages() so we evict the entry only if we can | |
466 | * do so without blocking. | |
467 | */ | |
468 | int dax_invalidate_mapping_entry(struct address_space *mapping, pgoff_t index) | |
469 | { | |
470 | int ret = 0; | |
471 | void *entry, **slot; | |
472 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
473 | ||
474 | spin_lock_irq(&mapping->tree_lock); | |
475 | entry = __radix_tree_lookup(page_tree, index, NULL, &slot); | |
476 | if (!entry || !radix_tree_exceptional_entry(entry) || | |
477 | slot_locked(mapping, slot)) | |
478 | goto out; | |
479 | if (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) || | |
480 | radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)) | |
481 | goto out; | |
482 | radix_tree_delete(page_tree, index); | |
483 | mapping->nrexceptional--; | |
484 | ret = 1; | |
485 | out: | |
486 | spin_unlock_irq(&mapping->tree_lock); | |
487 | if (ret) | |
488 | dax_wake_mapping_entry_waiter(mapping, index, entry, true); | |
489 | return ret; | |
490 | } | |
491 | ||
492 | /* | |
493 | * Invalidate exceptional DAX entry if it is clean. | |
494 | */ | |
495 | int dax_invalidate_mapping_entry_sync(struct address_space *mapping, | |
496 | pgoff_t index) | |
497 | { | |
498 | return __dax_invalidate_mapping_entry(mapping, index, false); | |
499 | } | |
500 | ||
501 | /* | |
502 | * The user has performed a load from a hole in the file. Allocating | |
503 | * a new page in the file would cause excessive storage usage for | |
504 | * workloads with sparse files. We allocate a page cache page instead. | |
505 | * We'll kick it out of the page cache if it's ever written to, | |
506 | * otherwise it will simply fall out of the page cache under memory | |
507 | * pressure without ever having been dirtied. | |
508 | */ | |
509 | static int dax_load_hole(struct address_space *mapping, void **entry, | |
510 | struct vm_fault *vmf) | |
511 | { | |
512 | struct inode *inode = mapping->host; | |
513 | struct page *page; | |
514 | int ret; | |
515 | ||
516 | /* Hole page already exists? Return it... */ | |
517 | if (!radix_tree_exceptional_entry(*entry)) { | |
518 | page = *entry; | |
519 | goto finish_fault; | |
520 | } | |
521 | ||
522 | /* This will replace locked radix tree entry with a hole page */ | |
523 | page = find_or_create_page(mapping, vmf->pgoff, | |
524 | vmf->gfp_mask | __GFP_ZERO); | |
525 | if (!page) { | |
526 | ret = VM_FAULT_OOM; | |
527 | goto out; | |
528 | } | |
529 | ||
530 | finish_fault: | |
531 | vmf->page = page; | |
532 | ret = finish_fault(vmf); | |
533 | vmf->page = NULL; | |
534 | *entry = page; | |
535 | if (!ret) { | |
536 | /* Grab reference for PTE that is now referencing the page */ | |
537 | get_page(page); | |
538 | ret = VM_FAULT_NOPAGE; | |
539 | } | |
540 | out: | |
541 | trace_dax_load_hole(inode, vmf, ret); | |
542 | return ret; | |
543 | } | |
544 | ||
545 | static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev, | |
546 | sector_t sector, size_t size, struct page *to, | |
547 | unsigned long vaddr) | |
548 | { | |
549 | void *vto, *kaddr; | |
550 | pgoff_t pgoff; | |
551 | pfn_t pfn; | |
552 | long rc; | |
553 | int id; | |
554 | ||
555 | rc = bdev_dax_pgoff(bdev, sector, size, &pgoff); | |
556 | if (rc) | |
557 | return rc; | |
558 | ||
559 | id = dax_read_lock(); | |
560 | rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn); | |
561 | if (rc < 0) { | |
562 | dax_read_unlock(id); | |
563 | return rc; | |
564 | } | |
565 | vto = kmap_atomic(to); | |
566 | copy_user_page(vto, (void __force *)kaddr, vaddr, to); | |
567 | kunmap_atomic(vto); | |
568 | dax_read_unlock(id); | |
569 | return 0; | |
570 | } | |
571 | ||
572 | /* | |
573 | * By this point grab_mapping_entry() has ensured that we have a locked entry | |
574 | * of the appropriate size so we don't have to worry about downgrading PMDs to | |
575 | * PTEs. If we happen to be trying to insert a PTE and there is a PMD | |
576 | * already in the tree, we will skip the insertion and just dirty the PMD as | |
577 | * appropriate. | |
578 | */ | |
579 | static void *dax_insert_mapping_entry(struct address_space *mapping, | |
580 | struct vm_fault *vmf, | |
581 | void *entry, sector_t sector, | |
582 | unsigned long flags) | |
583 | { | |
584 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
585 | int error = 0; | |
586 | bool hole_fill = false; | |
587 | void *new_entry; | |
588 | pgoff_t index = vmf->pgoff; | |
589 | ||
590 | if (vmf->flags & FAULT_FLAG_WRITE) | |
591 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
592 | ||
593 | /* Replacing hole page with block mapping? */ | |
594 | if (!radix_tree_exceptional_entry(entry)) { | |
595 | hole_fill = true; | |
596 | /* | |
597 | * Unmap the page now before we remove it from page cache below. | |
598 | * The page is locked so it cannot be faulted in again. | |
599 | */ | |
600 | unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT, | |
601 | PAGE_SIZE, 0); | |
602 | error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM); | |
603 | if (error) | |
604 | return ERR_PTR(error); | |
605 | } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) { | |
606 | /* replacing huge zero page with PMD block mapping */ | |
607 | unmap_mapping_range(mapping, | |
608 | (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0); | |
609 | } | |
610 | ||
611 | spin_lock_irq(&mapping->tree_lock); | |
612 | new_entry = dax_radix_locked_entry(sector, flags); | |
613 | ||
614 | if (hole_fill) { | |
615 | __delete_from_page_cache(entry, NULL); | |
616 | /* Drop pagecache reference */ | |
617 | put_page(entry); | |
618 | error = __radix_tree_insert(page_tree, index, | |
619 | dax_radix_order(new_entry), new_entry); | |
620 | if (error) { | |
621 | new_entry = ERR_PTR(error); | |
622 | goto unlock; | |
623 | } | |
624 | mapping->nrexceptional++; | |
625 | } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { | |
626 | /* | |
627 | * Only swap our new entry into the radix tree if the current | |
628 | * entry is a zero page or an empty entry. If a normal PTE or | |
629 | * PMD entry is already in the tree, we leave it alone. This | |
630 | * means that if we are trying to insert a PTE and the | |
631 | * existing entry is a PMD, we will just leave the PMD in the | |
632 | * tree and dirty it if necessary. | |
633 | */ | |
634 | struct radix_tree_node *node; | |
635 | void **slot; | |
636 | void *ret; | |
637 | ||
638 | ret = __radix_tree_lookup(page_tree, index, &node, &slot); | |
639 | WARN_ON_ONCE(ret != entry); | |
640 | __radix_tree_replace(page_tree, node, slot, | |
641 | new_entry, NULL, NULL); | |
642 | } | |
643 | if (vmf->flags & FAULT_FLAG_WRITE) | |
644 | radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY); | |
645 | unlock: | |
646 | spin_unlock_irq(&mapping->tree_lock); | |
647 | if (hole_fill) { | |
648 | radix_tree_preload_end(); | |
649 | /* | |
650 | * We don't need hole page anymore, it has been replaced with | |
651 | * locked radix tree entry now. | |
652 | */ | |
653 | if (mapping->a_ops->freepage) | |
654 | mapping->a_ops->freepage(entry); | |
655 | unlock_page(entry); | |
656 | put_page(entry); | |
657 | } | |
658 | return new_entry; | |
659 | } | |
660 | ||
661 | static inline unsigned long | |
662 | pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma) | |
663 | { | |
664 | unsigned long address; | |
665 | ||
666 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
667 | VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); | |
668 | return address; | |
669 | } | |
670 | ||
671 | /* Walk all mappings of a given index of a file and writeprotect them */ | |
672 | static void dax_mapping_entry_mkclean(struct address_space *mapping, | |
673 | pgoff_t index, unsigned long pfn) | |
674 | { | |
675 | struct vm_area_struct *vma; | |
676 | pte_t pte, *ptep = NULL; | |
677 | pmd_t *pmdp = NULL; | |
678 | spinlock_t *ptl; | |
679 | bool changed; | |
680 | ||
681 | i_mmap_lock_read(mapping); | |
682 | vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) { | |
683 | unsigned long address; | |
684 | ||
685 | cond_resched(); | |
686 | ||
687 | if (!(vma->vm_flags & VM_SHARED)) | |
688 | continue; | |
689 | ||
690 | address = pgoff_address(index, vma); | |
691 | changed = false; | |
692 | if (follow_pte_pmd(vma->vm_mm, address, &ptep, &pmdp, &ptl)) | |
693 | continue; | |
694 | ||
695 | if (pmdp) { | |
696 | #ifdef CONFIG_FS_DAX_PMD | |
697 | pmd_t pmd; | |
698 | ||
699 | if (pfn != pmd_pfn(*pmdp)) | |
700 | goto unlock_pmd; | |
701 | if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp)) | |
702 | goto unlock_pmd; | |
703 | ||
704 | flush_cache_page(vma, address, pfn); | |
705 | pmd = pmdp_huge_clear_flush(vma, address, pmdp); | |
706 | pmd = pmd_wrprotect(pmd); | |
707 | pmd = pmd_mkclean(pmd); | |
708 | set_pmd_at(vma->vm_mm, address, pmdp, pmd); | |
709 | changed = true; | |
710 | unlock_pmd: | |
711 | spin_unlock(ptl); | |
712 | #endif | |
713 | } else { | |
714 | if (pfn != pte_pfn(*ptep)) | |
715 | goto unlock_pte; | |
716 | if (!pte_dirty(*ptep) && !pte_write(*ptep)) | |
717 | goto unlock_pte; | |
718 | ||
719 | flush_cache_page(vma, address, pfn); | |
720 | pte = ptep_clear_flush(vma, address, ptep); | |
721 | pte = pte_wrprotect(pte); | |
722 | pte = pte_mkclean(pte); | |
723 | set_pte_at(vma->vm_mm, address, ptep, pte); | |
724 | changed = true; | |
725 | unlock_pte: | |
726 | pte_unmap_unlock(ptep, ptl); | |
727 | } | |
728 | ||
729 | if (changed) | |
730 | mmu_notifier_invalidate_page(vma->vm_mm, address); | |
731 | } | |
732 | i_mmap_unlock_read(mapping); | |
733 | } | |
734 | ||
735 | static int dax_writeback_one(struct block_device *bdev, | |
736 | struct dax_device *dax_dev, struct address_space *mapping, | |
737 | pgoff_t index, void *entry) | |
738 | { | |
739 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
740 | void *entry2, **slot, *kaddr; | |
741 | long ret = 0, id; | |
742 | sector_t sector; | |
743 | pgoff_t pgoff; | |
744 | size_t size; | |
745 | pfn_t pfn; | |
746 | ||
747 | /* | |
748 | * A page got tagged dirty in DAX mapping? Something is seriously | |
749 | * wrong. | |
750 | */ | |
751 | if (WARN_ON(!radix_tree_exceptional_entry(entry))) | |
752 | return -EIO; | |
753 | ||
754 | spin_lock_irq(&mapping->tree_lock); | |
755 | entry2 = get_unlocked_mapping_entry(mapping, index, &slot); | |
756 | /* Entry got punched out / reallocated? */ | |
757 | if (!entry2 || !radix_tree_exceptional_entry(entry2)) | |
758 | goto put_unlocked; | |
759 | /* | |
760 | * Entry got reallocated elsewhere? No need to writeback. We have to | |
761 | * compare sectors as we must not bail out due to difference in lockbit | |
762 | * or entry type. | |
763 | */ | |
764 | if (dax_radix_sector(entry2) != dax_radix_sector(entry)) | |
765 | goto put_unlocked; | |
766 | if (WARN_ON_ONCE(dax_is_empty_entry(entry) || | |
767 | dax_is_zero_entry(entry))) { | |
768 | ret = -EIO; | |
769 | goto put_unlocked; | |
770 | } | |
771 | ||
772 | /* Another fsync thread may have already written back this entry */ | |
773 | if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)) | |
774 | goto put_unlocked; | |
775 | /* Lock the entry to serialize with page faults */ | |
776 | entry = lock_slot(mapping, slot); | |
777 | /* | |
778 | * We can clear the tag now but we have to be careful so that concurrent | |
779 | * dax_writeback_one() calls for the same index cannot finish before we | |
780 | * actually flush the caches. This is achieved as the calls will look | |
781 | * at the entry only under tree_lock and once they do that they will | |
782 | * see the entry locked and wait for it to unlock. | |
783 | */ | |
784 | radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE); | |
785 | spin_unlock_irq(&mapping->tree_lock); | |
786 | ||
787 | /* | |
788 | * Even if dax_writeback_mapping_range() was given a wbc->range_start | |
789 | * in the middle of a PMD, the 'index' we are given will be aligned to | |
790 | * the start index of the PMD, as will the sector we pull from | |
791 | * 'entry'. This allows us to flush for PMD_SIZE and not have to | |
792 | * worry about partial PMD writebacks. | |
793 | */ | |
794 | sector = dax_radix_sector(entry); | |
795 | size = PAGE_SIZE << dax_radix_order(entry); | |
796 | ||
797 | id = dax_read_lock(); | |
798 | ret = bdev_dax_pgoff(bdev, sector, size, &pgoff); | |
799 | if (ret) | |
800 | goto dax_unlock; | |
801 | ||
802 | /* | |
803 | * dax_direct_access() may sleep, so cannot hold tree_lock over | |
804 | * its invocation. | |
805 | */ | |
806 | ret = dax_direct_access(dax_dev, pgoff, size / PAGE_SIZE, &kaddr, &pfn); | |
807 | if (ret < 0) | |
808 | goto dax_unlock; | |
809 | ||
810 | if (WARN_ON_ONCE(ret < size / PAGE_SIZE)) { | |
811 | ret = -EIO; | |
812 | goto dax_unlock; | |
813 | } | |
814 | ||
815 | dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(pfn)); | |
816 | wb_cache_pmem(kaddr, size); | |
817 | /* | |
818 | * After we have flushed the cache, we can clear the dirty tag. There | |
819 | * cannot be new dirty data in the pfn after the flush has completed as | |
820 | * the pfn mappings are writeprotected and fault waits for mapping | |
821 | * entry lock. | |
822 | */ | |
823 | spin_lock_irq(&mapping->tree_lock); | |
824 | radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY); | |
825 | spin_unlock_irq(&mapping->tree_lock); | |
826 | trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT); | |
827 | dax_unlock: | |
828 | dax_read_unlock(id); | |
829 | put_locked_mapping_entry(mapping, index, entry); | |
830 | return ret; | |
831 | ||
832 | put_unlocked: | |
833 | put_unlocked_mapping_entry(mapping, index, entry2); | |
834 | spin_unlock_irq(&mapping->tree_lock); | |
835 | return ret; | |
836 | } | |
837 | ||
838 | /* | |
839 | * Flush the mapping to the persistent domain within the byte range of [start, | |
840 | * end]. This is required by data integrity operations to ensure file data is | |
841 | * on persistent storage prior to completion of the operation. | |
842 | */ | |
843 | int dax_writeback_mapping_range(struct address_space *mapping, | |
844 | struct block_device *bdev, struct writeback_control *wbc) | |
845 | { | |
846 | struct inode *inode = mapping->host; | |
847 | pgoff_t start_index, end_index; | |
848 | pgoff_t indices[PAGEVEC_SIZE]; | |
849 | struct dax_device *dax_dev; | |
850 | struct pagevec pvec; | |
851 | bool done = false; | |
852 | int i, ret = 0; | |
853 | ||
854 | if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) | |
855 | return -EIO; | |
856 | ||
857 | if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) | |
858 | return 0; | |
859 | ||
860 | dax_dev = dax_get_by_host(bdev->bd_disk->disk_name); | |
861 | if (!dax_dev) | |
862 | return -EIO; | |
863 | ||
864 | start_index = wbc->range_start >> PAGE_SHIFT; | |
865 | end_index = wbc->range_end >> PAGE_SHIFT; | |
866 | ||
867 | trace_dax_writeback_range(inode, start_index, end_index); | |
868 | ||
869 | tag_pages_for_writeback(mapping, start_index, end_index); | |
870 | ||
871 | pagevec_init(&pvec, 0); | |
872 | while (!done) { | |
873 | pvec.nr = find_get_entries_tag(mapping, start_index, | |
874 | PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE, | |
875 | pvec.pages, indices); | |
876 | ||
877 | if (pvec.nr == 0) | |
878 | break; | |
879 | ||
880 | for (i = 0; i < pvec.nr; i++) { | |
881 | if (indices[i] > end_index) { | |
882 | done = true; | |
883 | break; | |
884 | } | |
885 | ||
886 | ret = dax_writeback_one(bdev, dax_dev, mapping, | |
887 | indices[i], pvec.pages[i]); | |
888 | if (ret < 0) | |
889 | goto out; | |
890 | } | |
891 | } | |
892 | out: | |
893 | put_dax(dax_dev); | |
894 | trace_dax_writeback_range_done(inode, start_index, end_index); | |
895 | return (ret < 0 ? ret : 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, struct dax_device *dax_dev, | |
901 | sector_t sector, size_t size, void **entryp, | |
902 | struct vm_area_struct *vma, struct vm_fault *vmf) | |
903 | { | |
904 | unsigned long vaddr = vmf->address; | |
905 | void *entry = *entryp; | |
906 | void *ret, *kaddr; | |
907 | pgoff_t pgoff; | |
908 | int id, rc; | |
909 | pfn_t pfn; | |
910 | ||
911 | rc = bdev_dax_pgoff(bdev, sector, size, &pgoff); | |
912 | if (rc) | |
913 | return rc; | |
914 | ||
915 | id = dax_read_lock(); | |
916 | rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn); | |
917 | if (rc < 0) { | |
918 | dax_read_unlock(id); | |
919 | return rc; | |
920 | } | |
921 | dax_read_unlock(id); | |
922 | ||
923 | ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, 0); | |
924 | if (IS_ERR(ret)) | |
925 | return PTR_ERR(ret); | |
926 | *entryp = ret; | |
927 | ||
928 | trace_dax_insert_mapping(mapping->host, vmf, ret); | |
929 | return vm_insert_mixed(vma, vaddr, pfn); | |
930 | } | |
931 | ||
932 | /** | |
933 | * dax_pfn_mkwrite - handle first write to DAX page | |
934 | * @vmf: The description of the fault | |
935 | */ | |
936 | int dax_pfn_mkwrite(struct vm_fault *vmf) | |
937 | { | |
938 | struct file *file = vmf->vma->vm_file; | |
939 | struct address_space *mapping = file->f_mapping; | |
940 | struct inode *inode = mapping->host; | |
941 | void *entry, **slot; | |
942 | pgoff_t index = vmf->pgoff; | |
943 | ||
944 | spin_lock_irq(&mapping->tree_lock); | |
945 | entry = get_unlocked_mapping_entry(mapping, index, &slot); | |
946 | if (!entry || !radix_tree_exceptional_entry(entry)) { | |
947 | if (entry) | |
948 | put_unlocked_mapping_entry(mapping, index, entry); | |
949 | spin_unlock_irq(&mapping->tree_lock); | |
950 | trace_dax_pfn_mkwrite_no_entry(inode, vmf, VM_FAULT_NOPAGE); | |
951 | return VM_FAULT_NOPAGE; | |
952 | } | |
953 | radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY); | |
954 | entry = lock_slot(mapping, slot); | |
955 | spin_unlock_irq(&mapping->tree_lock); | |
956 | /* | |
957 | * If we race with somebody updating the PTE and finish_mkwrite_fault() | |
958 | * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry | |
959 | * the fault in either case. | |
960 | */ | |
961 | finish_mkwrite_fault(vmf); | |
962 | put_locked_mapping_entry(mapping, index, entry); | |
963 | trace_dax_pfn_mkwrite(inode, vmf, VM_FAULT_NOPAGE); | |
964 | return VM_FAULT_NOPAGE; | |
965 | } | |
966 | EXPORT_SYMBOL_GPL(dax_pfn_mkwrite); | |
967 | ||
968 | static bool dax_range_is_aligned(struct block_device *bdev, | |
969 | unsigned int offset, unsigned int length) | |
970 | { | |
971 | unsigned short sector_size = bdev_logical_block_size(bdev); | |
972 | ||
973 | if (!IS_ALIGNED(offset, sector_size)) | |
974 | return false; | |
975 | if (!IS_ALIGNED(length, sector_size)) | |
976 | return false; | |
977 | ||
978 | return true; | |
979 | } | |
980 | ||
981 | int __dax_zero_page_range(struct block_device *bdev, | |
982 | struct dax_device *dax_dev, sector_t sector, | |
983 | unsigned int offset, unsigned int size) | |
984 | { | |
985 | if (dax_range_is_aligned(bdev, offset, size)) { | |
986 | sector_t start_sector = sector + (offset >> 9); | |
987 | ||
988 | return blkdev_issue_zeroout(bdev, start_sector, | |
989 | size >> 9, GFP_NOFS, 0); | |
990 | } else { | |
991 | pgoff_t pgoff; | |
992 | long rc, id; | |
993 | void *kaddr; | |
994 | pfn_t pfn; | |
995 | ||
996 | rc = bdev_dax_pgoff(bdev, sector, size, &pgoff); | |
997 | if (rc) | |
998 | return rc; | |
999 | ||
1000 | id = dax_read_lock(); | |
1001 | rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, | |
1002 | &pfn); | |
1003 | if (rc < 0) { | |
1004 | dax_read_unlock(id); | |
1005 | return rc; | |
1006 | } | |
1007 | clear_pmem(kaddr + offset, size); | |
1008 | dax_read_unlock(id); | |
1009 | } | |
1010 | return 0; | |
1011 | } | |
1012 | EXPORT_SYMBOL_GPL(__dax_zero_page_range); | |
1013 | ||
1014 | static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) | |
1015 | { | |
1016 | return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9); | |
1017 | } | |
1018 | ||
1019 | static loff_t | |
1020 | dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, | |
1021 | struct iomap *iomap) | |
1022 | { | |
1023 | struct block_device *bdev = iomap->bdev; | |
1024 | struct dax_device *dax_dev = iomap->dax_dev; | |
1025 | struct iov_iter *iter = data; | |
1026 | loff_t end = pos + length, done = 0; | |
1027 | ssize_t ret = 0; | |
1028 | int id; | |
1029 | ||
1030 | if (iov_iter_rw(iter) == READ) { | |
1031 | end = min(end, i_size_read(inode)); | |
1032 | if (pos >= end) | |
1033 | return 0; | |
1034 | ||
1035 | if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) | |
1036 | return iov_iter_zero(min(length, end - pos), iter); | |
1037 | } | |
1038 | ||
1039 | if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) | |
1040 | return -EIO; | |
1041 | ||
1042 | /* | |
1043 | * Write can allocate block for an area which has a hole page mapped | |
1044 | * into page tables. We have to tear down these mappings so that data | |
1045 | * written by write(2) is visible in mmap. | |
1046 | */ | |
1047 | if ((iomap->flags & IOMAP_F_NEW) && inode->i_mapping->nrpages) { | |
1048 | invalidate_inode_pages2_range(inode->i_mapping, | |
1049 | pos >> PAGE_SHIFT, | |
1050 | (end - 1) >> PAGE_SHIFT); | |
1051 | } | |
1052 | ||
1053 | id = dax_read_lock(); | |
1054 | while (pos < end) { | |
1055 | unsigned offset = pos & (PAGE_SIZE - 1); | |
1056 | const size_t size = ALIGN(length + offset, PAGE_SIZE); | |
1057 | const sector_t sector = dax_iomap_sector(iomap, pos); | |
1058 | ssize_t map_len; | |
1059 | pgoff_t pgoff; | |
1060 | void *kaddr; | |
1061 | pfn_t pfn; | |
1062 | ||
1063 | if (fatal_signal_pending(current)) { | |
1064 | ret = -EINTR; | |
1065 | break; | |
1066 | } | |
1067 | ||
1068 | ret = bdev_dax_pgoff(bdev, sector, size, &pgoff); | |
1069 | if (ret) | |
1070 | break; | |
1071 | ||
1072 | map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), | |
1073 | &kaddr, &pfn); | |
1074 | if (map_len < 0) { | |
1075 | ret = map_len; | |
1076 | break; | |
1077 | } | |
1078 | ||
1079 | map_len = PFN_PHYS(map_len); | |
1080 | kaddr += offset; | |
1081 | map_len -= offset; | |
1082 | if (map_len > end - pos) | |
1083 | map_len = end - pos; | |
1084 | ||
1085 | if (iov_iter_rw(iter) == WRITE) | |
1086 | map_len = copy_from_iter_pmem(kaddr, map_len, iter); | |
1087 | else | |
1088 | map_len = copy_to_iter(kaddr, map_len, iter); | |
1089 | if (map_len <= 0) { | |
1090 | ret = map_len ? map_len : -EFAULT; | |
1091 | break; | |
1092 | } | |
1093 | ||
1094 | pos += map_len; | |
1095 | length -= map_len; | |
1096 | done += map_len; | |
1097 | } | |
1098 | dax_read_unlock(id); | |
1099 | ||
1100 | return done ? done : ret; | |
1101 | } | |
1102 | ||
1103 | /** | |
1104 | * dax_iomap_rw - Perform I/O to a DAX file | |
1105 | * @iocb: The control block for this I/O | |
1106 | * @iter: The addresses to do I/O from or to | |
1107 | * @ops: iomap ops passed from the file system | |
1108 | * | |
1109 | * This function performs read and write operations to directly mapped | |
1110 | * persistent memory. The callers needs to take care of read/write exclusion | |
1111 | * and evicting any page cache pages in the region under I/O. | |
1112 | */ | |
1113 | ssize_t | |
1114 | dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, | |
1115 | const struct iomap_ops *ops) | |
1116 | { | |
1117 | struct address_space *mapping = iocb->ki_filp->f_mapping; | |
1118 | struct inode *inode = mapping->host; | |
1119 | loff_t pos = iocb->ki_pos, ret = 0, done = 0; | |
1120 | unsigned flags = 0; | |
1121 | ||
1122 | if (iov_iter_rw(iter) == WRITE) { | |
1123 | lockdep_assert_held_exclusive(&inode->i_rwsem); | |
1124 | flags |= IOMAP_WRITE; | |
1125 | } else { | |
1126 | lockdep_assert_held(&inode->i_rwsem); | |
1127 | } | |
1128 | ||
1129 | while (iov_iter_count(iter)) { | |
1130 | ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, | |
1131 | iter, dax_iomap_actor); | |
1132 | if (ret <= 0) | |
1133 | break; | |
1134 | pos += ret; | |
1135 | done += ret; | |
1136 | } | |
1137 | ||
1138 | iocb->ki_pos += done; | |
1139 | return done ? done : ret; | |
1140 | } | |
1141 | EXPORT_SYMBOL_GPL(dax_iomap_rw); | |
1142 | ||
1143 | static int dax_fault_return(int error) | |
1144 | { | |
1145 | if (error == 0) | |
1146 | return VM_FAULT_NOPAGE; | |
1147 | if (error == -ENOMEM) | |
1148 | return VM_FAULT_OOM; | |
1149 | return VM_FAULT_SIGBUS; | |
1150 | } | |
1151 | ||
1152 | static int dax_iomap_pte_fault(struct vm_fault *vmf, | |
1153 | const struct iomap_ops *ops) | |
1154 | { | |
1155 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; | |
1156 | struct inode *inode = mapping->host; | |
1157 | unsigned long vaddr = vmf->address; | |
1158 | loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; | |
1159 | sector_t sector; | |
1160 | struct iomap iomap = { 0 }; | |
1161 | unsigned flags = IOMAP_FAULT; | |
1162 | int error, major = 0; | |
1163 | int vmf_ret = 0; | |
1164 | void *entry; | |
1165 | ||
1166 | trace_dax_pte_fault(inode, vmf, vmf_ret); | |
1167 | /* | |
1168 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
1169 | * to hold locks serializing us with truncate / punch hole so this is | |
1170 | * a reliable test. | |
1171 | */ | |
1172 | if (pos >= i_size_read(inode)) { | |
1173 | vmf_ret = VM_FAULT_SIGBUS; | |
1174 | goto out; | |
1175 | } | |
1176 | ||
1177 | if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) | |
1178 | flags |= IOMAP_WRITE; | |
1179 | ||
1180 | /* | |
1181 | * Note that we don't bother to use iomap_apply here: DAX required | |
1182 | * the file system block size to be equal the page size, which means | |
1183 | * that we never have to deal with more than a single extent here. | |
1184 | */ | |
1185 | error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); | |
1186 | if (error) { | |
1187 | vmf_ret = dax_fault_return(error); | |
1188 | goto out; | |
1189 | } | |
1190 | if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { | |
1191 | vmf_ret = dax_fault_return(-EIO); /* fs corruption? */ | |
1192 | goto finish_iomap; | |
1193 | } | |
1194 | ||
1195 | entry = grab_mapping_entry(mapping, vmf->pgoff, 0); | |
1196 | if (IS_ERR(entry)) { | |
1197 | vmf_ret = dax_fault_return(PTR_ERR(entry)); | |
1198 | goto finish_iomap; | |
1199 | } | |
1200 | ||
1201 | sector = dax_iomap_sector(&iomap, pos); | |
1202 | ||
1203 | if (vmf->cow_page) { | |
1204 | switch (iomap.type) { | |
1205 | case IOMAP_HOLE: | |
1206 | case IOMAP_UNWRITTEN: | |
1207 | clear_user_highpage(vmf->cow_page, vaddr); | |
1208 | break; | |
1209 | case IOMAP_MAPPED: | |
1210 | error = copy_user_dax(iomap.bdev, iomap.dax_dev, | |
1211 | sector, PAGE_SIZE, vmf->cow_page, vaddr); | |
1212 | break; | |
1213 | default: | |
1214 | WARN_ON_ONCE(1); | |
1215 | error = -EIO; | |
1216 | break; | |
1217 | } | |
1218 | ||
1219 | if (error) | |
1220 | goto error_unlock_entry; | |
1221 | ||
1222 | __SetPageUptodate(vmf->cow_page); | |
1223 | vmf_ret = finish_fault(vmf); | |
1224 | if (!vmf_ret) | |
1225 | vmf_ret = VM_FAULT_DONE_COW; | |
1226 | goto unlock_entry; | |
1227 | } | |
1228 | ||
1229 | switch (iomap.type) { | |
1230 | case IOMAP_MAPPED: | |
1231 | if (iomap.flags & IOMAP_F_NEW) { | |
1232 | count_vm_event(PGMAJFAULT); | |
1233 | mem_cgroup_count_vm_event(vmf->vma->vm_mm, PGMAJFAULT); | |
1234 | major = VM_FAULT_MAJOR; | |
1235 | } | |
1236 | error = dax_insert_mapping(mapping, iomap.bdev, iomap.dax_dev, | |
1237 | sector, PAGE_SIZE, &entry, vmf->vma, vmf); | |
1238 | /* -EBUSY is fine, somebody else faulted on the same PTE */ | |
1239 | if (error == -EBUSY) | |
1240 | error = 0; | |
1241 | break; | |
1242 | case IOMAP_UNWRITTEN: | |
1243 | case IOMAP_HOLE: | |
1244 | if (!(vmf->flags & FAULT_FLAG_WRITE)) { | |
1245 | vmf_ret = dax_load_hole(mapping, &entry, vmf); | |
1246 | goto unlock_entry; | |
1247 | } | |
1248 | /*FALLTHRU*/ | |
1249 | default: | |
1250 | WARN_ON_ONCE(1); | |
1251 | error = -EIO; | |
1252 | break; | |
1253 | } | |
1254 | ||
1255 | error_unlock_entry: | |
1256 | vmf_ret = dax_fault_return(error) | major; | |
1257 | unlock_entry: | |
1258 | put_locked_mapping_entry(mapping, vmf->pgoff, entry); | |
1259 | finish_iomap: | |
1260 | if (ops->iomap_end) { | |
1261 | int copied = PAGE_SIZE; | |
1262 | ||
1263 | if (vmf_ret & VM_FAULT_ERROR) | |
1264 | copied = 0; | |
1265 | /* | |
1266 | * The fault is done by now and there's no way back (other | |
1267 | * thread may be already happily using PTE we have installed). | |
1268 | * Just ignore error from ->iomap_end since we cannot do much | |
1269 | * with it. | |
1270 | */ | |
1271 | ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap); | |
1272 | } | |
1273 | out: | |
1274 | trace_dax_pte_fault_done(inode, vmf, vmf_ret); | |
1275 | return vmf_ret; | |
1276 | } | |
1277 | ||
1278 | #ifdef CONFIG_FS_DAX_PMD | |
1279 | /* | |
1280 | * The 'colour' (ie low bits) within a PMD of a page offset. This comes up | |
1281 | * more often than one might expect in the below functions. | |
1282 | */ | |
1283 | #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) | |
1284 | ||
1285 | static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap, | |
1286 | loff_t pos, void **entryp) | |
1287 | { | |
1288 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; | |
1289 | const sector_t sector = dax_iomap_sector(iomap, pos); | |
1290 | struct dax_device *dax_dev = iomap->dax_dev; | |
1291 | struct block_device *bdev = iomap->bdev; | |
1292 | struct inode *inode = mapping->host; | |
1293 | const size_t size = PMD_SIZE; | |
1294 | void *ret = NULL, *kaddr; | |
1295 | long length = 0; | |
1296 | pgoff_t pgoff; | |
1297 | pfn_t pfn; | |
1298 | int id; | |
1299 | ||
1300 | if (bdev_dax_pgoff(bdev, sector, size, &pgoff) != 0) | |
1301 | goto fallback; | |
1302 | ||
1303 | id = dax_read_lock(); | |
1304 | length = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn); | |
1305 | if (length < 0) | |
1306 | goto unlock_fallback; | |
1307 | length = PFN_PHYS(length); | |
1308 | ||
1309 | if (length < size) | |
1310 | goto unlock_fallback; | |
1311 | if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR) | |
1312 | goto unlock_fallback; | |
1313 | if (!pfn_t_devmap(pfn)) | |
1314 | goto unlock_fallback; | |
1315 | dax_read_unlock(id); | |
1316 | ||
1317 | ret = dax_insert_mapping_entry(mapping, vmf, *entryp, sector, | |
1318 | RADIX_DAX_PMD); | |
1319 | if (IS_ERR(ret)) | |
1320 | goto fallback; | |
1321 | *entryp = ret; | |
1322 | ||
1323 | trace_dax_pmd_insert_mapping(inode, vmf, length, pfn, ret); | |
1324 | return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd, | |
1325 | pfn, vmf->flags & FAULT_FLAG_WRITE); | |
1326 | ||
1327 | unlock_fallback: | |
1328 | dax_read_unlock(id); | |
1329 | fallback: | |
1330 | trace_dax_pmd_insert_mapping_fallback(inode, vmf, length, pfn, ret); | |
1331 | return VM_FAULT_FALLBACK; | |
1332 | } | |
1333 | ||
1334 | static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap, | |
1335 | void **entryp) | |
1336 | { | |
1337 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; | |
1338 | unsigned long pmd_addr = vmf->address & PMD_MASK; | |
1339 | struct inode *inode = mapping->host; | |
1340 | struct page *zero_page; | |
1341 | void *ret = NULL; | |
1342 | spinlock_t *ptl; | |
1343 | pmd_t pmd_entry; | |
1344 | ||
1345 | zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); | |
1346 | ||
1347 | if (unlikely(!zero_page)) | |
1348 | goto fallback; | |
1349 | ||
1350 | ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0, | |
1351 | RADIX_DAX_PMD | RADIX_DAX_HZP); | |
1352 | if (IS_ERR(ret)) | |
1353 | goto fallback; | |
1354 | *entryp = ret; | |
1355 | ||
1356 | ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); | |
1357 | if (!pmd_none(*(vmf->pmd))) { | |
1358 | spin_unlock(ptl); | |
1359 | goto fallback; | |
1360 | } | |
1361 | ||
1362 | pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); | |
1363 | pmd_entry = pmd_mkhuge(pmd_entry); | |
1364 | set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); | |
1365 | spin_unlock(ptl); | |
1366 | trace_dax_pmd_load_hole(inode, vmf, zero_page, ret); | |
1367 | return VM_FAULT_NOPAGE; | |
1368 | ||
1369 | fallback: | |
1370 | trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret); | |
1371 | return VM_FAULT_FALLBACK; | |
1372 | } | |
1373 | ||
1374 | static int dax_iomap_pmd_fault(struct vm_fault *vmf, | |
1375 | const struct iomap_ops *ops) | |
1376 | { | |
1377 | struct vm_area_struct *vma = vmf->vma; | |
1378 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1379 | unsigned long pmd_addr = vmf->address & PMD_MASK; | |
1380 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
1381 | unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT; | |
1382 | struct inode *inode = mapping->host; | |
1383 | int result = VM_FAULT_FALLBACK; | |
1384 | struct iomap iomap = { 0 }; | |
1385 | pgoff_t max_pgoff, pgoff; | |
1386 | void *entry; | |
1387 | loff_t pos; | |
1388 | int error; | |
1389 | ||
1390 | /* | |
1391 | * Check whether offset isn't beyond end of file now. Caller is | |
1392 | * supposed to hold locks serializing us with truncate / punch hole so | |
1393 | * this is a reliable test. | |
1394 | */ | |
1395 | pgoff = linear_page_index(vma, pmd_addr); | |
1396 | max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT; | |
1397 | ||
1398 | trace_dax_pmd_fault(inode, vmf, max_pgoff, 0); | |
1399 | ||
1400 | /* Fall back to PTEs if we're going to COW */ | |
1401 | if (write && !(vma->vm_flags & VM_SHARED)) | |
1402 | goto fallback; | |
1403 | ||
1404 | /* If the PMD would extend outside the VMA */ | |
1405 | if (pmd_addr < vma->vm_start) | |
1406 | goto fallback; | |
1407 | if ((pmd_addr + PMD_SIZE) > vma->vm_end) | |
1408 | goto fallback; | |
1409 | ||
1410 | if (pgoff > max_pgoff) { | |
1411 | result = VM_FAULT_SIGBUS; | |
1412 | goto out; | |
1413 | } | |
1414 | ||
1415 | /* If the PMD would extend beyond the file size */ | |
1416 | if ((pgoff | PG_PMD_COLOUR) > max_pgoff) | |
1417 | goto fallback; | |
1418 | ||
1419 | /* | |
1420 | * Note that we don't use iomap_apply here. We aren't doing I/O, only | |
1421 | * setting up a mapping, so really we're using iomap_begin() as a way | |
1422 | * to look up our filesystem block. | |
1423 | */ | |
1424 | pos = (loff_t)pgoff << PAGE_SHIFT; | |
1425 | error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap); | |
1426 | if (error) | |
1427 | goto fallback; | |
1428 | ||
1429 | if (iomap.offset + iomap.length < pos + PMD_SIZE) | |
1430 | goto finish_iomap; | |
1431 | ||
1432 | /* | |
1433 | * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX | |
1434 | * PMD or a HZP entry. If it can't (because a 4k page is already in | |
1435 | * the tree, for instance), it will return -EEXIST and we just fall | |
1436 | * back to 4k entries. | |
1437 | */ | |
1438 | entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD); | |
1439 | if (IS_ERR(entry)) | |
1440 | goto finish_iomap; | |
1441 | ||
1442 | switch (iomap.type) { | |
1443 | case IOMAP_MAPPED: | |
1444 | result = dax_pmd_insert_mapping(vmf, &iomap, pos, &entry); | |
1445 | break; | |
1446 | case IOMAP_UNWRITTEN: | |
1447 | case IOMAP_HOLE: | |
1448 | if (WARN_ON_ONCE(write)) | |
1449 | goto unlock_entry; | |
1450 | result = dax_pmd_load_hole(vmf, &iomap, &entry); | |
1451 | break; | |
1452 | default: | |
1453 | WARN_ON_ONCE(1); | |
1454 | break; | |
1455 | } | |
1456 | ||
1457 | unlock_entry: | |
1458 | put_locked_mapping_entry(mapping, pgoff, entry); | |
1459 | finish_iomap: | |
1460 | if (ops->iomap_end) { | |
1461 | int copied = PMD_SIZE; | |
1462 | ||
1463 | if (result == VM_FAULT_FALLBACK) | |
1464 | copied = 0; | |
1465 | /* | |
1466 | * The fault is done by now and there's no way back (other | |
1467 | * thread may be already happily using PMD we have installed). | |
1468 | * Just ignore error from ->iomap_end since we cannot do much | |
1469 | * with it. | |
1470 | */ | |
1471 | ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags, | |
1472 | &iomap); | |
1473 | } | |
1474 | fallback: | |
1475 | if (result == VM_FAULT_FALLBACK) { | |
1476 | split_huge_pmd(vma, vmf->pmd, vmf->address); | |
1477 | count_vm_event(THP_FAULT_FALLBACK); | |
1478 | } | |
1479 | out: | |
1480 | trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result); | |
1481 | return result; | |
1482 | } | |
1483 | #else | |
1484 | static int dax_iomap_pmd_fault(struct vm_fault *vmf, | |
1485 | const struct iomap_ops *ops) | |
1486 | { | |
1487 | return VM_FAULT_FALLBACK; | |
1488 | } | |
1489 | #endif /* CONFIG_FS_DAX_PMD */ | |
1490 | ||
1491 | /** | |
1492 | * dax_iomap_fault - handle a page fault on a DAX file | |
1493 | * @vmf: The description of the fault | |
1494 | * @ops: iomap ops passed from the file system | |
1495 | * | |
1496 | * When a page fault occurs, filesystems may call this helper in | |
1497 | * their fault handler for DAX files. dax_iomap_fault() assumes the caller | |
1498 | * has done all the necessary locking for page fault to proceed | |
1499 | * successfully. | |
1500 | */ | |
1501 | int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, | |
1502 | const struct iomap_ops *ops) | |
1503 | { | |
1504 | switch (pe_size) { | |
1505 | case PE_SIZE_PTE: | |
1506 | return dax_iomap_pte_fault(vmf, ops); | |
1507 | case PE_SIZE_PMD: | |
1508 | return dax_iomap_pmd_fault(vmf, ops); | |
1509 | default: | |
1510 | return VM_FAULT_FALLBACK; | |
1511 | } | |
1512 | } | |
1513 | EXPORT_SYMBOL_GPL(dax_iomap_fault); |