]>
Commit | Line | Data |
---|---|---|
d475c634 MW |
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> | |
d77e92e2 | 20 | #include <linux/dax.h> |
d475c634 MW |
21 | #include <linux/fs.h> |
22 | #include <linux/genhd.h> | |
f7ca90b1 MW |
23 | #include <linux/highmem.h> |
24 | #include <linux/memcontrol.h> | |
25 | #include <linux/mm.h> | |
d475c634 | 26 | #include <linux/mutex.h> |
9973c98e | 27 | #include <linux/pagevec.h> |
2765cfbb | 28 | #include <linux/pmem.h> |
289c6aed | 29 | #include <linux/sched.h> |
f361bf4a | 30 | #include <linux/sched/signal.h> |
d475c634 | 31 | #include <linux/uio.h> |
f7ca90b1 | 32 | #include <linux/vmstat.h> |
34c0fd54 | 33 | #include <linux/pfn_t.h> |
0e749e54 | 34 | #include <linux/sizes.h> |
4b4bb46d | 35 | #include <linux/mmu_notifier.h> |
a254e568 CH |
36 | #include <linux/iomap.h> |
37 | #include "internal.h" | |
d475c634 | 38 | |
282a8e03 RZ |
39 | #define CREATE_TRACE_POINTS |
40 | #include <trace/events/fs_dax.h> | |
41 | ||
ac401cc7 JK |
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 | ||
ce95ab0f | 46 | static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; |
ac401cc7 JK |
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 | ||
642261ac | 58 | static int dax_is_pmd_entry(void *entry) |
d1a5f2b4 | 59 | { |
642261ac | 60 | return (unsigned long)entry & RADIX_DAX_PMD; |
d1a5f2b4 DW |
61 | } |
62 | ||
642261ac | 63 | static int dax_is_pte_entry(void *entry) |
d475c634 | 64 | { |
642261ac | 65 | return !((unsigned long)entry & RADIX_DAX_PMD); |
d475c634 MW |
66 | } |
67 | ||
642261ac | 68 | static int dax_is_zero_entry(void *entry) |
d475c634 | 69 | { |
642261ac | 70 | return (unsigned long)entry & RADIX_DAX_HZP; |
d475c634 MW |
71 | } |
72 | ||
642261ac | 73 | static int dax_is_empty_entry(void *entry) |
b2e0d162 | 74 | { |
642261ac | 75 | return (unsigned long)entry & RADIX_DAX_EMPTY; |
b2e0d162 DW |
76 | } |
77 | ||
ac401cc7 JK |
78 | /* |
79 | * DAX radix tree locking | |
80 | */ | |
81 | struct exceptional_entry_key { | |
82 | struct address_space *mapping; | |
63e95b5c | 83 | pgoff_t entry_start; |
ac401cc7 JK |
84 | }; |
85 | ||
86 | struct wait_exceptional_entry_queue { | |
87 | wait_queue_t wait; | |
88 | struct exceptional_entry_key key; | |
89 | }; | |
90 | ||
63e95b5c RZ |
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 | */ | |
642261ac | 101 | if (dax_is_pmd_entry(entry)) |
63e95b5c RZ |
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 | ||
ac401cc7 JK |
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 || | |
63e95b5c | 119 | key->entry_start != ewait->key.entry_start) |
ac401cc7 JK |
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; | |
6d75f366 | 145 | radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); |
ac401cc7 JK |
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; | |
6d75f366 | 159 | radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); |
ac401cc7 JK |
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 | { | |
e3ad61c6 | 175 | void *entry, **slot; |
ac401cc7 | 176 | struct wait_exceptional_entry_queue ewait; |
63e95b5c | 177 | wait_queue_head_t *wq; |
ac401cc7 JK |
178 | |
179 | init_wait(&ewait.wait); | |
180 | ewait.wait.func = wake_exceptional_entry_func; | |
ac401cc7 JK |
181 | |
182 | for (;;) { | |
e3ad61c6 | 183 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, |
ac401cc7 | 184 | &slot); |
e3ad61c6 | 185 | if (!entry || !radix_tree_exceptional_entry(entry) || |
ac401cc7 JK |
186 | !slot_locked(mapping, slot)) { |
187 | if (slotp) | |
188 | *slotp = slot; | |
e3ad61c6 | 189 | return entry; |
ac401cc7 | 190 | } |
63e95b5c RZ |
191 | |
192 | wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key); | |
ac401cc7 JK |
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 | ||
b1aa812b JK |
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 | ||
422476c4 RZ |
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 | ||
ac401cc7 JK |
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 | * | |
642261ac RZ |
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 | * | |
ac401cc7 JK |
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 | */ | |
642261ac RZ |
270 | static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index, |
271 | unsigned long size_flag) | |
ac401cc7 | 272 | { |
642261ac | 273 | bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */ |
e3ad61c6 | 274 | void *entry, **slot; |
ac401cc7 JK |
275 | |
276 | restart: | |
277 | spin_lock_irq(&mapping->tree_lock); | |
e3ad61c6 | 278 | entry = get_unlocked_mapping_entry(mapping, index, &slot); |
642261ac RZ |
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 | ||
ac401cc7 | 299 | /* No entry for given index? Make sure radix tree is big enough. */ |
642261ac | 300 | if (!entry || pmd_downgrade) { |
ac401cc7 JK |
301 | int err; |
302 | ||
642261ac RZ |
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 | ||
ac401cc7 | 311 | spin_unlock_irq(&mapping->tree_lock); |
642261ac RZ |
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 | ||
ac401cc7 JK |
321 | err = radix_tree_preload( |
322 | mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM); | |
0cb80b48 JK |
323 | if (err) { |
324 | if (pmd_downgrade) | |
325 | put_locked_mapping_entry(mapping, index, entry); | |
ac401cc7 | 326 | return ERR_PTR(err); |
0cb80b48 | 327 | } |
ac401cc7 | 328 | spin_lock_irq(&mapping->tree_lock); |
642261ac | 329 | |
e11f8b7b RZ |
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 | ||
642261ac RZ |
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); | |
ac401cc7 JK |
357 | radix_tree_preload_end(); |
358 | if (err) { | |
359 | spin_unlock_irq(&mapping->tree_lock); | |
642261ac | 360 | /* |
e11f8b7b RZ |
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. | |
642261ac | 367 | */ |
ac401cc7 JK |
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); | |
e3ad61c6 | 373 | return entry; |
ac401cc7 JK |
374 | } |
375 | /* Normal page in radix tree? */ | |
e3ad61c6 RZ |
376 | if (!radix_tree_exceptional_entry(entry)) { |
377 | struct page *page = entry; | |
ac401cc7 JK |
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 | } | |
e3ad61c6 | 390 | entry = lock_slot(mapping, slot); |
642261ac | 391 | out_unlock: |
ac401cc7 | 392 | spin_unlock_irq(&mapping->tree_lock); |
e3ad61c6 | 393 | return entry; |
ac401cc7 JK |
394 | } |
395 | ||
63e95b5c RZ |
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 | |
642261ac RZ |
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. | |
63e95b5c | 402 | */ |
ac401cc7 | 403 | void dax_wake_mapping_entry_waiter(struct address_space *mapping, |
63e95b5c | 404 | pgoff_t index, void *entry, bool wake_all) |
ac401cc7 | 405 | { |
63e95b5c RZ |
406 | struct exceptional_entry_key key; |
407 | wait_queue_head_t *wq; | |
408 | ||
409 | wq = dax_entry_waitqueue(mapping, index, entry, &key); | |
ac401cc7 JK |
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 | */ | |
63e95b5c | 417 | if (waitqueue_active(wq)) |
ac401cc7 | 418 | __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); |
ac401cc7 JK |
419 | } |
420 | ||
c6dcf52c JK |
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 | } | |
ac401cc7 JK |
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 | { | |
c6dcf52c | 450 | int ret = __dax_invalidate_mapping_entry(mapping, index, true); |
ac401cc7 | 451 | |
ac401cc7 JK |
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 | */ | |
c6dcf52c JK |
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); | |
ac401cc7 | 483 | mapping->nrexceptional--; |
c6dcf52c JK |
484 | ret = 1; |
485 | out: | |
ac401cc7 | 486 | spin_unlock_irq(&mapping->tree_lock); |
c6dcf52c JK |
487 | if (ret) |
488 | dax_wake_mapping_entry_waiter(mapping, index, entry, true); | |
489 | return ret; | |
490 | } | |
ac401cc7 | 491 | |
c6dcf52c JK |
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); | |
ac401cc7 JK |
499 | } |
500 | ||
f7ca90b1 MW |
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 | */ | |
f449b936 | 509 | static int dax_load_hole(struct address_space *mapping, void **entry, |
ac401cc7 | 510 | struct vm_fault *vmf) |
f7ca90b1 | 511 | { |
678c9fd0 | 512 | struct inode *inode = mapping->host; |
ac401cc7 | 513 | struct page *page; |
f449b936 | 514 | int ret; |
f7ca90b1 | 515 | |
ac401cc7 | 516 | /* Hole page already exists? Return it... */ |
f449b936 JK |
517 | if (!radix_tree_exceptional_entry(*entry)) { |
518 | page = *entry; | |
678c9fd0 | 519 | goto finish_fault; |
ac401cc7 | 520 | } |
f7ca90b1 | 521 | |
ac401cc7 JK |
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); | |
678c9fd0 RZ |
525 | if (!page) { |
526 | ret = VM_FAULT_OOM; | |
527 | goto out; | |
528 | } | |
529 | ||
530 | finish_fault: | |
f7ca90b1 | 531 | vmf->page = page; |
f449b936 JK |
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); | |
678c9fd0 | 538 | ret = VM_FAULT_NOPAGE; |
f449b936 | 539 | } |
678c9fd0 RZ |
540 | out: |
541 | trace_dax_load_hole(inode, vmf, ret); | |
f449b936 | 542 | return ret; |
f7ca90b1 MW |
543 | } |
544 | ||
cccbce67 DW |
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) | |
f7ca90b1 | 548 | { |
cccbce67 DW |
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 | } | |
f7ca90b1 | 565 | vto = kmap_atomic(to); |
cccbce67 | 566 | copy_user_page(vto, (void __force *)kaddr, vaddr, to); |
f7ca90b1 | 567 | kunmap_atomic(vto); |
cccbce67 | 568 | dax_read_unlock(id); |
f7ca90b1 MW |
569 | return 0; |
570 | } | |
571 | ||
642261ac RZ |
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 | */ | |
ac401cc7 JK |
579 | static void *dax_insert_mapping_entry(struct address_space *mapping, |
580 | struct vm_fault *vmf, | |
642261ac RZ |
581 | void *entry, sector_t sector, |
582 | unsigned long flags) | |
9973c98e RZ |
583 | { |
584 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
ac401cc7 JK |
585 | int error = 0; |
586 | bool hole_fill = false; | |
587 | void *new_entry; | |
588 | pgoff_t index = vmf->pgoff; | |
9973c98e | 589 | |
ac401cc7 | 590 | if (vmf->flags & FAULT_FLAG_WRITE) |
d2b2a28e | 591 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); |
9973c98e | 592 | |
ac401cc7 JK |
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); | |
642261ac RZ |
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); | |
9973c98e RZ |
609 | } |
610 | ||
ac401cc7 | 611 | spin_lock_irq(&mapping->tree_lock); |
642261ac RZ |
612 | new_entry = dax_radix_locked_entry(sector, flags); |
613 | ||
ac401cc7 JK |
614 | if (hole_fill) { |
615 | __delete_from_page_cache(entry, NULL); | |
616 | /* Drop pagecache reference */ | |
617 | put_page(entry); | |
642261ac RZ |
618 | error = __radix_tree_insert(page_tree, index, |
619 | dax_radix_order(new_entry), new_entry); | |
ac401cc7 JK |
620 | if (error) { |
621 | new_entry = ERR_PTR(error); | |
9973c98e RZ |
622 | goto unlock; |
623 | } | |
ac401cc7 | 624 | mapping->nrexceptional++; |
642261ac RZ |
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 | */ | |
f7942430 | 634 | struct radix_tree_node *node; |
ac401cc7 JK |
635 | void **slot; |
636 | void *ret; | |
9973c98e | 637 | |
f7942430 | 638 | ret = __radix_tree_lookup(page_tree, index, &node, &slot); |
ac401cc7 | 639 | WARN_ON_ONCE(ret != entry); |
4d693d08 JW |
640 | __radix_tree_replace(page_tree, node, slot, |
641 | new_entry, NULL, NULL); | |
9973c98e | 642 | } |
ac401cc7 | 643 | if (vmf->flags & FAULT_FLAG_WRITE) |
9973c98e RZ |
644 | radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY); |
645 | unlock: | |
646 | spin_unlock_irq(&mapping->tree_lock); | |
ac401cc7 JK |
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; | |
9973c98e RZ |
659 | } |
660 | ||
4b4bb46d JK |
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; | |
f729c8c9 RZ |
676 | pte_t pte, *ptep = NULL; |
677 | pmd_t *pmdp = NULL; | |
4b4bb46d JK |
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; | |
f729c8c9 | 692 | if (follow_pte_pmd(vma->vm_mm, address, &ptep, &pmdp, &ptl)) |
4b4bb46d | 693 | continue; |
4b4bb46d | 694 | |
f729c8c9 RZ |
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 | } | |
4b4bb46d JK |
728 | |
729 | if (changed) | |
730 | mmu_notifier_invalidate_page(vma->vm_mm, address); | |
731 | } | |
732 | i_mmap_unlock_read(mapping); | |
733 | } | |
734 | ||
9973c98e | 735 | static int dax_writeback_one(struct block_device *bdev, |
cccbce67 DW |
736 | struct dax_device *dax_dev, struct address_space *mapping, |
737 | pgoff_t index, void *entry) | |
9973c98e RZ |
738 | { |
739 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
cccbce67 DW |
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; | |
9973c98e | 746 | |
9973c98e | 747 | /* |
a6abc2c0 JK |
748 | * A page got tagged dirty in DAX mapping? Something is seriously |
749 | * wrong. | |
9973c98e | 750 | */ |
a6abc2c0 JK |
751 | if (WARN_ON(!radix_tree_exceptional_entry(entry))) |
752 | return -EIO; | |
9973c98e | 753 | |
a6abc2c0 JK |
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; | |
642261ac RZ |
766 | if (WARN_ON_ONCE(dax_is_empty_entry(entry) || |
767 | dax_is_zero_entry(entry))) { | |
9973c98e | 768 | ret = -EIO; |
a6abc2c0 | 769 | goto put_unlocked; |
9973c98e RZ |
770 | } |
771 | ||
a6abc2c0 JK |
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 | ||
642261ac RZ |
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 | */ | |
cccbce67 DW |
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; | |
9973c98e RZ |
801 | |
802 | /* | |
cccbce67 DW |
803 | * dax_direct_access() may sleep, so cannot hold tree_lock over |
804 | * its invocation. | |
9973c98e | 805 | */ |
cccbce67 DW |
806 | ret = dax_direct_access(dax_dev, pgoff, size / PAGE_SIZE, &kaddr, &pfn); |
807 | if (ret < 0) | |
808 | goto dax_unlock; | |
9973c98e | 809 | |
cccbce67 | 810 | if (WARN_ON_ONCE(ret < size / PAGE_SIZE)) { |
9973c98e | 811 | ret = -EIO; |
cccbce67 | 812 | goto dax_unlock; |
9973c98e RZ |
813 | } |
814 | ||
cccbce67 DW |
815 | dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(pfn)); |
816 | wb_cache_pmem(kaddr, size); | |
4b4bb46d JK |
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); | |
cccbce67 DW |
826 | dax_unlock: |
827 | dax_read_unlock(id); | |
a6abc2c0 | 828 | put_locked_mapping_entry(mapping, index, entry); |
9973c98e RZ |
829 | return ret; |
830 | ||
a6abc2c0 JK |
831 | put_unlocked: |
832 | put_unlocked_mapping_entry(mapping, index, entry2); | |
9973c98e RZ |
833 | spin_unlock_irq(&mapping->tree_lock); |
834 | return ret; | |
835 | } | |
836 | ||
837 | /* | |
838 | * Flush the mapping to the persistent domain within the byte range of [start, | |
839 | * end]. This is required by data integrity operations to ensure file data is | |
840 | * on persistent storage prior to completion of the operation. | |
841 | */ | |
7f6d5b52 RZ |
842 | int dax_writeback_mapping_range(struct address_space *mapping, |
843 | struct block_device *bdev, struct writeback_control *wbc) | |
9973c98e RZ |
844 | { |
845 | struct inode *inode = mapping->host; | |
642261ac | 846 | pgoff_t start_index, end_index; |
9973c98e | 847 | pgoff_t indices[PAGEVEC_SIZE]; |
cccbce67 | 848 | struct dax_device *dax_dev; |
9973c98e RZ |
849 | struct pagevec pvec; |
850 | bool done = false; | |
851 | int i, ret = 0; | |
9973c98e RZ |
852 | |
853 | if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) | |
854 | return -EIO; | |
855 | ||
7f6d5b52 RZ |
856 | if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) |
857 | return 0; | |
858 | ||
cccbce67 DW |
859 | dax_dev = dax_get_by_host(bdev->bd_disk->disk_name); |
860 | if (!dax_dev) | |
861 | return -EIO; | |
862 | ||
09cbfeaf KS |
863 | start_index = wbc->range_start >> PAGE_SHIFT; |
864 | end_index = wbc->range_end >> PAGE_SHIFT; | |
9973c98e RZ |
865 | |
866 | tag_pages_for_writeback(mapping, start_index, end_index); | |
867 | ||
868 | pagevec_init(&pvec, 0); | |
869 | while (!done) { | |
870 | pvec.nr = find_get_entries_tag(mapping, start_index, | |
871 | PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE, | |
872 | pvec.pages, indices); | |
873 | ||
874 | if (pvec.nr == 0) | |
875 | break; | |
876 | ||
877 | for (i = 0; i < pvec.nr; i++) { | |
878 | if (indices[i] > end_index) { | |
879 | done = true; | |
880 | break; | |
881 | } | |
882 | ||
cccbce67 DW |
883 | ret = dax_writeback_one(bdev, dax_dev, mapping, |
884 | indices[i], pvec.pages[i]); | |
885 | if (ret < 0) { | |
886 | put_dax(dax_dev); | |
9973c98e | 887 | return ret; |
cccbce67 | 888 | } |
9973c98e RZ |
889 | } |
890 | } | |
cccbce67 | 891 | put_dax(dax_dev); |
9973c98e RZ |
892 | return 0; |
893 | } | |
894 | EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); | |
895 | ||
ac401cc7 | 896 | static int dax_insert_mapping(struct address_space *mapping, |
cccbce67 DW |
897 | struct block_device *bdev, struct dax_device *dax_dev, |
898 | sector_t sector, size_t size, void **entryp, | |
899 | struct vm_area_struct *vma, struct vm_fault *vmf) | |
f7ca90b1 | 900 | { |
1a29d85e | 901 | unsigned long vaddr = vmf->address; |
ac401cc7 | 902 | void *entry = *entryp; |
cccbce67 DW |
903 | void *ret, *kaddr; |
904 | pgoff_t pgoff; | |
905 | int id, rc; | |
906 | pfn_t pfn; | |
f7ca90b1 | 907 | |
cccbce67 DW |
908 | rc = bdev_dax_pgoff(bdev, sector, size, &pgoff); |
909 | if (rc) | |
910 | return rc; | |
f7ca90b1 | 911 | |
cccbce67 DW |
912 | id = dax_read_lock(); |
913 | rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn); | |
914 | if (rc < 0) { | |
915 | dax_read_unlock(id); | |
916 | return rc; | |
917 | } | |
918 | dax_read_unlock(id); | |
919 | ||
920 | ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, 0); | |
4d9a2c87 JK |
921 | if (IS_ERR(ret)) |
922 | return PTR_ERR(ret); | |
ac401cc7 | 923 | *entryp = ret; |
9973c98e | 924 | |
cccbce67 | 925 | return vm_insert_mixed(vma, vaddr, pfn); |
f7ca90b1 MW |
926 | } |
927 | ||
0e3b210c BH |
928 | /** |
929 | * dax_pfn_mkwrite - handle first write to DAX page | |
0e3b210c | 930 | * @vmf: The description of the fault |
0e3b210c | 931 | */ |
11bac800 | 932 | int dax_pfn_mkwrite(struct vm_fault *vmf) |
0e3b210c | 933 | { |
11bac800 | 934 | struct file *file = vmf->vma->vm_file; |
ac401cc7 | 935 | struct address_space *mapping = file->f_mapping; |
c3ff68d7 | 936 | struct inode *inode = mapping->host; |
2f89dc12 | 937 | void *entry, **slot; |
ac401cc7 | 938 | pgoff_t index = vmf->pgoff; |
30f471fd | 939 | |
ac401cc7 | 940 | spin_lock_irq(&mapping->tree_lock); |
2f89dc12 JK |
941 | entry = get_unlocked_mapping_entry(mapping, index, &slot); |
942 | if (!entry || !radix_tree_exceptional_entry(entry)) { | |
943 | if (entry) | |
944 | put_unlocked_mapping_entry(mapping, index, entry); | |
945 | spin_unlock_irq(&mapping->tree_lock); | |
c3ff68d7 | 946 | trace_dax_pfn_mkwrite_no_entry(inode, vmf, VM_FAULT_NOPAGE); |
2f89dc12 JK |
947 | return VM_FAULT_NOPAGE; |
948 | } | |
ac401cc7 | 949 | radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY); |
2f89dc12 | 950 | entry = lock_slot(mapping, slot); |
ac401cc7 | 951 | spin_unlock_irq(&mapping->tree_lock); |
2f89dc12 JK |
952 | /* |
953 | * If we race with somebody updating the PTE and finish_mkwrite_fault() | |
954 | * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry | |
955 | * the fault in either case. | |
956 | */ | |
957 | finish_mkwrite_fault(vmf); | |
958 | put_locked_mapping_entry(mapping, index, entry); | |
c3ff68d7 | 959 | trace_dax_pfn_mkwrite(inode, vmf, VM_FAULT_NOPAGE); |
0e3b210c BH |
960 | return VM_FAULT_NOPAGE; |
961 | } | |
962 | EXPORT_SYMBOL_GPL(dax_pfn_mkwrite); | |
963 | ||
4b0228fa VV |
964 | static bool dax_range_is_aligned(struct block_device *bdev, |
965 | unsigned int offset, unsigned int length) | |
966 | { | |
967 | unsigned short sector_size = bdev_logical_block_size(bdev); | |
968 | ||
969 | if (!IS_ALIGNED(offset, sector_size)) | |
970 | return false; | |
971 | if (!IS_ALIGNED(length, sector_size)) | |
972 | return false; | |
973 | ||
974 | return true; | |
975 | } | |
976 | ||
cccbce67 DW |
977 | int __dax_zero_page_range(struct block_device *bdev, |
978 | struct dax_device *dax_dev, sector_t sector, | |
979 | unsigned int offset, unsigned int size) | |
679c8bd3 | 980 | { |
cccbce67 DW |
981 | if (dax_range_is_aligned(bdev, offset, size)) { |
982 | sector_t start_sector = sector + (offset >> 9); | |
4b0228fa VV |
983 | |
984 | return blkdev_issue_zeroout(bdev, start_sector, | |
53ef7d0e | 985 | size >> 9, GFP_NOFS, 0); |
4b0228fa | 986 | } else { |
cccbce67 DW |
987 | pgoff_t pgoff; |
988 | long rc, id; | |
989 | void *kaddr; | |
990 | pfn_t pfn; | |
991 | ||
992 | rc = bdev_dax_pgoff(bdev, sector, size, &pgoff); | |
993 | if (rc) | |
994 | return rc; | |
995 | ||
996 | id = dax_read_lock(); | |
997 | rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, | |
998 | &pfn); | |
999 | if (rc < 0) { | |
1000 | dax_read_unlock(id); | |
1001 | return rc; | |
1002 | } | |
1003 | clear_pmem(kaddr + offset, size); | |
1004 | dax_read_unlock(id); | |
4b0228fa | 1005 | } |
679c8bd3 CH |
1006 | return 0; |
1007 | } | |
1008 | EXPORT_SYMBOL_GPL(__dax_zero_page_range); | |
1009 | ||
333ccc97 | 1010 | static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) |
25726bc1 | 1011 | { |
333ccc97 | 1012 | return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9); |
25726bc1 | 1013 | } |
a254e568 | 1014 | |
a254e568 | 1015 | static loff_t |
11c59c92 | 1016 | dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, |
a254e568 CH |
1017 | struct iomap *iomap) |
1018 | { | |
cccbce67 DW |
1019 | struct block_device *bdev = iomap->bdev; |
1020 | struct dax_device *dax_dev = iomap->dax_dev; | |
a254e568 CH |
1021 | struct iov_iter *iter = data; |
1022 | loff_t end = pos + length, done = 0; | |
1023 | ssize_t ret = 0; | |
cccbce67 | 1024 | int id; |
a254e568 CH |
1025 | |
1026 | if (iov_iter_rw(iter) == READ) { | |
1027 | end = min(end, i_size_read(inode)); | |
1028 | if (pos >= end) | |
1029 | return 0; | |
1030 | ||
1031 | if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) | |
1032 | return iov_iter_zero(min(length, end - pos), iter); | |
1033 | } | |
1034 | ||
1035 | if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) | |
1036 | return -EIO; | |
1037 | ||
e3fce68c JK |
1038 | /* |
1039 | * Write can allocate block for an area which has a hole page mapped | |
1040 | * into page tables. We have to tear down these mappings so that data | |
1041 | * written by write(2) is visible in mmap. | |
1042 | */ | |
1043 | if ((iomap->flags & IOMAP_F_NEW) && inode->i_mapping->nrpages) { | |
1044 | invalidate_inode_pages2_range(inode->i_mapping, | |
1045 | pos >> PAGE_SHIFT, | |
1046 | (end - 1) >> PAGE_SHIFT); | |
1047 | } | |
1048 | ||
cccbce67 | 1049 | id = dax_read_lock(); |
a254e568 CH |
1050 | while (pos < end) { |
1051 | unsigned offset = pos & (PAGE_SIZE - 1); | |
cccbce67 DW |
1052 | const size_t size = ALIGN(length + offset, PAGE_SIZE); |
1053 | const sector_t sector = dax_iomap_sector(iomap, pos); | |
a254e568 | 1054 | ssize_t map_len; |
cccbce67 DW |
1055 | pgoff_t pgoff; |
1056 | void *kaddr; | |
1057 | pfn_t pfn; | |
a254e568 | 1058 | |
d1908f52 MH |
1059 | if (fatal_signal_pending(current)) { |
1060 | ret = -EINTR; | |
1061 | break; | |
1062 | } | |
1063 | ||
cccbce67 DW |
1064 | ret = bdev_dax_pgoff(bdev, sector, size, &pgoff); |
1065 | if (ret) | |
1066 | break; | |
1067 | ||
1068 | map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), | |
1069 | &kaddr, &pfn); | |
a254e568 CH |
1070 | if (map_len < 0) { |
1071 | ret = map_len; | |
1072 | break; | |
1073 | } | |
1074 | ||
cccbce67 DW |
1075 | map_len = PFN_PHYS(map_len); |
1076 | kaddr += offset; | |
a254e568 CH |
1077 | map_len -= offset; |
1078 | if (map_len > end - pos) | |
1079 | map_len = end - pos; | |
1080 | ||
1081 | if (iov_iter_rw(iter) == WRITE) | |
cccbce67 | 1082 | map_len = copy_from_iter_pmem(kaddr, map_len, iter); |
a254e568 | 1083 | else |
cccbce67 | 1084 | map_len = copy_to_iter(kaddr, map_len, iter); |
a254e568 CH |
1085 | if (map_len <= 0) { |
1086 | ret = map_len ? map_len : -EFAULT; | |
1087 | break; | |
1088 | } | |
1089 | ||
1090 | pos += map_len; | |
1091 | length -= map_len; | |
1092 | done += map_len; | |
1093 | } | |
cccbce67 | 1094 | dax_read_unlock(id); |
a254e568 CH |
1095 | |
1096 | return done ? done : ret; | |
1097 | } | |
1098 | ||
1099 | /** | |
11c59c92 | 1100 | * dax_iomap_rw - Perform I/O to a DAX file |
a254e568 CH |
1101 | * @iocb: The control block for this I/O |
1102 | * @iter: The addresses to do I/O from or to | |
1103 | * @ops: iomap ops passed from the file system | |
1104 | * | |
1105 | * This function performs read and write operations to directly mapped | |
1106 | * persistent memory. The callers needs to take care of read/write exclusion | |
1107 | * and evicting any page cache pages in the region under I/O. | |
1108 | */ | |
1109 | ssize_t | |
11c59c92 | 1110 | dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, |
8ff6daa1 | 1111 | const struct iomap_ops *ops) |
a254e568 CH |
1112 | { |
1113 | struct address_space *mapping = iocb->ki_filp->f_mapping; | |
1114 | struct inode *inode = mapping->host; | |
1115 | loff_t pos = iocb->ki_pos, ret = 0, done = 0; | |
1116 | unsigned flags = 0; | |
1117 | ||
168316db CH |
1118 | if (iov_iter_rw(iter) == WRITE) { |
1119 | lockdep_assert_held_exclusive(&inode->i_rwsem); | |
a254e568 | 1120 | flags |= IOMAP_WRITE; |
168316db CH |
1121 | } else { |
1122 | lockdep_assert_held(&inode->i_rwsem); | |
1123 | } | |
a254e568 | 1124 | |
a254e568 CH |
1125 | while (iov_iter_count(iter)) { |
1126 | ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, | |
11c59c92 | 1127 | iter, dax_iomap_actor); |
a254e568 CH |
1128 | if (ret <= 0) |
1129 | break; | |
1130 | pos += ret; | |
1131 | done += ret; | |
1132 | } | |
1133 | ||
1134 | iocb->ki_pos += done; | |
1135 | return done ? done : ret; | |
1136 | } | |
11c59c92 | 1137 | EXPORT_SYMBOL_GPL(dax_iomap_rw); |
a7d73fe6 | 1138 | |
9f141d6e JK |
1139 | static int dax_fault_return(int error) |
1140 | { | |
1141 | if (error == 0) | |
1142 | return VM_FAULT_NOPAGE; | |
1143 | if (error == -ENOMEM) | |
1144 | return VM_FAULT_OOM; | |
1145 | return VM_FAULT_SIGBUS; | |
1146 | } | |
1147 | ||
a2d58167 DJ |
1148 | static int dax_iomap_pte_fault(struct vm_fault *vmf, |
1149 | const struct iomap_ops *ops) | |
a7d73fe6 | 1150 | { |
11bac800 | 1151 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
a7d73fe6 | 1152 | struct inode *inode = mapping->host; |
1a29d85e | 1153 | unsigned long vaddr = vmf->address; |
a7d73fe6 CH |
1154 | loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; |
1155 | sector_t sector; | |
1156 | struct iomap iomap = { 0 }; | |
9484ab1b | 1157 | unsigned flags = IOMAP_FAULT; |
a7d73fe6 | 1158 | int error, major = 0; |
b1aa812b | 1159 | int vmf_ret = 0; |
a7d73fe6 CH |
1160 | void *entry; |
1161 | ||
a9c42b33 | 1162 | trace_dax_pte_fault(inode, vmf, vmf_ret); |
a7d73fe6 CH |
1163 | /* |
1164 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
1165 | * to hold locks serializing us with truncate / punch hole so this is | |
1166 | * a reliable test. | |
1167 | */ | |
a9c42b33 RZ |
1168 | if (pos >= i_size_read(inode)) { |
1169 | vmf_ret = VM_FAULT_SIGBUS; | |
1170 | goto out; | |
1171 | } | |
a7d73fe6 | 1172 | |
a7d73fe6 CH |
1173 | if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) |
1174 | flags |= IOMAP_WRITE; | |
1175 | ||
1176 | /* | |
1177 | * Note that we don't bother to use iomap_apply here: DAX required | |
1178 | * the file system block size to be equal the page size, which means | |
1179 | * that we never have to deal with more than a single extent here. | |
1180 | */ | |
1181 | error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); | |
a9c42b33 RZ |
1182 | if (error) { |
1183 | vmf_ret = dax_fault_return(error); | |
1184 | goto out; | |
1185 | } | |
a7d73fe6 | 1186 | if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { |
9f141d6e JK |
1187 | vmf_ret = dax_fault_return(-EIO); /* fs corruption? */ |
1188 | goto finish_iomap; | |
1189 | } | |
1190 | ||
1191 | entry = grab_mapping_entry(mapping, vmf->pgoff, 0); | |
1192 | if (IS_ERR(entry)) { | |
1193 | vmf_ret = dax_fault_return(PTR_ERR(entry)); | |
1550290b | 1194 | goto finish_iomap; |
a7d73fe6 CH |
1195 | } |
1196 | ||
333ccc97 | 1197 | sector = dax_iomap_sector(&iomap, pos); |
a7d73fe6 CH |
1198 | |
1199 | if (vmf->cow_page) { | |
1200 | switch (iomap.type) { | |
1201 | case IOMAP_HOLE: | |
1202 | case IOMAP_UNWRITTEN: | |
1203 | clear_user_highpage(vmf->cow_page, vaddr); | |
1204 | break; | |
1205 | case IOMAP_MAPPED: | |
cccbce67 DW |
1206 | error = copy_user_dax(iomap.bdev, iomap.dax_dev, |
1207 | sector, PAGE_SIZE, vmf->cow_page, vaddr); | |
a7d73fe6 CH |
1208 | break; |
1209 | default: | |
1210 | WARN_ON_ONCE(1); | |
1211 | error = -EIO; | |
1212 | break; | |
1213 | } | |
1214 | ||
1215 | if (error) | |
9f141d6e | 1216 | goto error_unlock_entry; |
b1aa812b JK |
1217 | |
1218 | __SetPageUptodate(vmf->cow_page); | |
1219 | vmf_ret = finish_fault(vmf); | |
1220 | if (!vmf_ret) | |
1221 | vmf_ret = VM_FAULT_DONE_COW; | |
9f141d6e | 1222 | goto unlock_entry; |
a7d73fe6 CH |
1223 | } |
1224 | ||
1225 | switch (iomap.type) { | |
1226 | case IOMAP_MAPPED: | |
1227 | if (iomap.flags & IOMAP_F_NEW) { | |
1228 | count_vm_event(PGMAJFAULT); | |
11bac800 | 1229 | mem_cgroup_count_vm_event(vmf->vma->vm_mm, PGMAJFAULT); |
a7d73fe6 CH |
1230 | major = VM_FAULT_MAJOR; |
1231 | } | |
cccbce67 DW |
1232 | error = dax_insert_mapping(mapping, iomap.bdev, iomap.dax_dev, |
1233 | sector, PAGE_SIZE, &entry, vmf->vma, vmf); | |
9f141d6e JK |
1234 | /* -EBUSY is fine, somebody else faulted on the same PTE */ |
1235 | if (error == -EBUSY) | |
1236 | error = 0; | |
a7d73fe6 CH |
1237 | break; |
1238 | case IOMAP_UNWRITTEN: | |
1239 | case IOMAP_HOLE: | |
1550290b | 1240 | if (!(vmf->flags & FAULT_FLAG_WRITE)) { |
f449b936 | 1241 | vmf_ret = dax_load_hole(mapping, &entry, vmf); |
9f141d6e | 1242 | goto unlock_entry; |
1550290b | 1243 | } |
a7d73fe6 CH |
1244 | /*FALLTHRU*/ |
1245 | default: | |
1246 | WARN_ON_ONCE(1); | |
1247 | error = -EIO; | |
1248 | break; | |
1249 | } | |
1250 | ||
9f141d6e JK |
1251 | error_unlock_entry: |
1252 | vmf_ret = dax_fault_return(error) | major; | |
a7d73fe6 | 1253 | unlock_entry: |
f449b936 | 1254 | put_locked_mapping_entry(mapping, vmf->pgoff, entry); |
9f141d6e JK |
1255 | finish_iomap: |
1256 | if (ops->iomap_end) { | |
1257 | int copied = PAGE_SIZE; | |
1258 | ||
1259 | if (vmf_ret & VM_FAULT_ERROR) | |
1260 | copied = 0; | |
1261 | /* | |
1262 | * The fault is done by now and there's no way back (other | |
1263 | * thread may be already happily using PTE we have installed). | |
1264 | * Just ignore error from ->iomap_end since we cannot do much | |
1265 | * with it. | |
1266 | */ | |
1267 | ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap); | |
1550290b | 1268 | } |
a9c42b33 RZ |
1269 | out: |
1270 | trace_dax_pte_fault_done(inode, vmf, vmf_ret); | |
9f141d6e | 1271 | return vmf_ret; |
a7d73fe6 | 1272 | } |
642261ac RZ |
1273 | |
1274 | #ifdef CONFIG_FS_DAX_PMD | |
1275 | /* | |
1276 | * The 'colour' (ie low bits) within a PMD of a page offset. This comes up | |
1277 | * more often than one might expect in the below functions. | |
1278 | */ | |
1279 | #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) | |
1280 | ||
f4200391 DJ |
1281 | static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap, |
1282 | loff_t pos, void **entryp) | |
642261ac | 1283 | { |
f4200391 | 1284 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
cccbce67 DW |
1285 | const sector_t sector = dax_iomap_sector(iomap, pos); |
1286 | struct dax_device *dax_dev = iomap->dax_dev; | |
642261ac | 1287 | struct block_device *bdev = iomap->bdev; |
27a7ffac | 1288 | struct inode *inode = mapping->host; |
cccbce67 DW |
1289 | const size_t size = PMD_SIZE; |
1290 | void *ret = NULL, *kaddr; | |
1291 | long length = 0; | |
1292 | pgoff_t pgoff; | |
1293 | pfn_t pfn; | |
1294 | int id; | |
1295 | ||
1296 | if (bdev_dax_pgoff(bdev, sector, size, &pgoff) != 0) | |
27a7ffac | 1297 | goto fallback; |
642261ac | 1298 | |
cccbce67 DW |
1299 | id = dax_read_lock(); |
1300 | length = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn); | |
1301 | if (length < 0) | |
1302 | goto unlock_fallback; | |
1303 | length = PFN_PHYS(length); | |
1304 | ||
1305 | if (length < size) | |
1306 | goto unlock_fallback; | |
1307 | if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR) | |
1308 | goto unlock_fallback; | |
1309 | if (!pfn_t_devmap(pfn)) | |
1310 | goto unlock_fallback; | |
1311 | dax_read_unlock(id); | |
1312 | ||
1313 | ret = dax_insert_mapping_entry(mapping, vmf, *entryp, sector, | |
642261ac RZ |
1314 | RADIX_DAX_PMD); |
1315 | if (IS_ERR(ret)) | |
27a7ffac | 1316 | goto fallback; |
642261ac RZ |
1317 | *entryp = ret; |
1318 | ||
cccbce67 | 1319 | trace_dax_pmd_insert_mapping(inode, vmf, length, pfn, ret); |
f4200391 | 1320 | return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd, |
cccbce67 | 1321 | pfn, vmf->flags & FAULT_FLAG_WRITE); |
642261ac | 1322 | |
cccbce67 DW |
1323 | unlock_fallback: |
1324 | dax_read_unlock(id); | |
27a7ffac | 1325 | fallback: |
cccbce67 | 1326 | trace_dax_pmd_insert_mapping_fallback(inode, vmf, length, pfn, ret); |
642261ac RZ |
1327 | return VM_FAULT_FALLBACK; |
1328 | } | |
1329 | ||
f4200391 DJ |
1330 | static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap, |
1331 | void **entryp) | |
642261ac | 1332 | { |
f4200391 DJ |
1333 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
1334 | unsigned long pmd_addr = vmf->address & PMD_MASK; | |
653b2ea3 | 1335 | struct inode *inode = mapping->host; |
642261ac | 1336 | struct page *zero_page; |
653b2ea3 | 1337 | void *ret = NULL; |
642261ac RZ |
1338 | spinlock_t *ptl; |
1339 | pmd_t pmd_entry; | |
642261ac | 1340 | |
f4200391 | 1341 | zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); |
642261ac RZ |
1342 | |
1343 | if (unlikely(!zero_page)) | |
653b2ea3 | 1344 | goto fallback; |
642261ac RZ |
1345 | |
1346 | ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0, | |
1347 | RADIX_DAX_PMD | RADIX_DAX_HZP); | |
1348 | if (IS_ERR(ret)) | |
653b2ea3 | 1349 | goto fallback; |
642261ac RZ |
1350 | *entryp = ret; |
1351 | ||
f4200391 DJ |
1352 | ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); |
1353 | if (!pmd_none(*(vmf->pmd))) { | |
642261ac | 1354 | spin_unlock(ptl); |
653b2ea3 | 1355 | goto fallback; |
642261ac RZ |
1356 | } |
1357 | ||
f4200391 | 1358 | pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); |
642261ac | 1359 | pmd_entry = pmd_mkhuge(pmd_entry); |
f4200391 | 1360 | set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); |
642261ac | 1361 | spin_unlock(ptl); |
f4200391 | 1362 | trace_dax_pmd_load_hole(inode, vmf, zero_page, ret); |
642261ac | 1363 | return VM_FAULT_NOPAGE; |
653b2ea3 RZ |
1364 | |
1365 | fallback: | |
f4200391 | 1366 | trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret); |
653b2ea3 | 1367 | return VM_FAULT_FALLBACK; |
642261ac RZ |
1368 | } |
1369 | ||
a2d58167 DJ |
1370 | static int dax_iomap_pmd_fault(struct vm_fault *vmf, |
1371 | const struct iomap_ops *ops) | |
642261ac | 1372 | { |
f4200391 | 1373 | struct vm_area_struct *vma = vmf->vma; |
642261ac | 1374 | struct address_space *mapping = vma->vm_file->f_mapping; |
d8a849e1 DJ |
1375 | unsigned long pmd_addr = vmf->address & PMD_MASK; |
1376 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
9484ab1b | 1377 | unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT; |
642261ac RZ |
1378 | struct inode *inode = mapping->host; |
1379 | int result = VM_FAULT_FALLBACK; | |
1380 | struct iomap iomap = { 0 }; | |
1381 | pgoff_t max_pgoff, pgoff; | |
642261ac RZ |
1382 | void *entry; |
1383 | loff_t pos; | |
1384 | int error; | |
1385 | ||
282a8e03 RZ |
1386 | /* |
1387 | * Check whether offset isn't beyond end of file now. Caller is | |
1388 | * supposed to hold locks serializing us with truncate / punch hole so | |
1389 | * this is a reliable test. | |
1390 | */ | |
1391 | pgoff = linear_page_index(vma, pmd_addr); | |
1392 | max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT; | |
1393 | ||
f4200391 | 1394 | trace_dax_pmd_fault(inode, vmf, max_pgoff, 0); |
282a8e03 | 1395 | |
642261ac RZ |
1396 | /* Fall back to PTEs if we're going to COW */ |
1397 | if (write && !(vma->vm_flags & VM_SHARED)) | |
1398 | goto fallback; | |
1399 | ||
1400 | /* If the PMD would extend outside the VMA */ | |
1401 | if (pmd_addr < vma->vm_start) | |
1402 | goto fallback; | |
1403 | if ((pmd_addr + PMD_SIZE) > vma->vm_end) | |
1404 | goto fallback; | |
1405 | ||
282a8e03 RZ |
1406 | if (pgoff > max_pgoff) { |
1407 | result = VM_FAULT_SIGBUS; | |
1408 | goto out; | |
1409 | } | |
642261ac RZ |
1410 | |
1411 | /* If the PMD would extend beyond the file size */ | |
1412 | if ((pgoff | PG_PMD_COLOUR) > max_pgoff) | |
1413 | goto fallback; | |
1414 | ||
642261ac RZ |
1415 | /* |
1416 | * Note that we don't use iomap_apply here. We aren't doing I/O, only | |
1417 | * setting up a mapping, so really we're using iomap_begin() as a way | |
1418 | * to look up our filesystem block. | |
1419 | */ | |
1420 | pos = (loff_t)pgoff << PAGE_SHIFT; | |
1421 | error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap); | |
1422 | if (error) | |
9f141d6e JK |
1423 | goto fallback; |
1424 | ||
642261ac RZ |
1425 | if (iomap.offset + iomap.length < pos + PMD_SIZE) |
1426 | goto finish_iomap; | |
1427 | ||
9f141d6e JK |
1428 | /* |
1429 | * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX | |
1430 | * PMD or a HZP entry. If it can't (because a 4k page is already in | |
1431 | * the tree, for instance), it will return -EEXIST and we just fall | |
1432 | * back to 4k entries. | |
1433 | */ | |
1434 | entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD); | |
1435 | if (IS_ERR(entry)) | |
1436 | goto finish_iomap; | |
1437 | ||
642261ac RZ |
1438 | switch (iomap.type) { |
1439 | case IOMAP_MAPPED: | |
f4200391 | 1440 | result = dax_pmd_insert_mapping(vmf, &iomap, pos, &entry); |
642261ac RZ |
1441 | break; |
1442 | case IOMAP_UNWRITTEN: | |
1443 | case IOMAP_HOLE: | |
1444 | if (WARN_ON_ONCE(write)) | |
9f141d6e | 1445 | goto unlock_entry; |
f4200391 | 1446 | result = dax_pmd_load_hole(vmf, &iomap, &entry); |
642261ac RZ |
1447 | break; |
1448 | default: | |
1449 | WARN_ON_ONCE(1); | |
1450 | break; | |
1451 | } | |
1452 | ||
9f141d6e JK |
1453 | unlock_entry: |
1454 | put_locked_mapping_entry(mapping, pgoff, entry); | |
642261ac RZ |
1455 | finish_iomap: |
1456 | if (ops->iomap_end) { | |
9f141d6e JK |
1457 | int copied = PMD_SIZE; |
1458 | ||
1459 | if (result == VM_FAULT_FALLBACK) | |
1460 | copied = 0; | |
1461 | /* | |
1462 | * The fault is done by now and there's no way back (other | |
1463 | * thread may be already happily using PMD we have installed). | |
1464 | * Just ignore error from ->iomap_end since we cannot do much | |
1465 | * with it. | |
1466 | */ | |
1467 | ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags, | |
1468 | &iomap); | |
642261ac | 1469 | } |
642261ac RZ |
1470 | fallback: |
1471 | if (result == VM_FAULT_FALLBACK) { | |
d8a849e1 | 1472 | split_huge_pmd(vma, vmf->pmd, vmf->address); |
642261ac RZ |
1473 | count_vm_event(THP_FAULT_FALLBACK); |
1474 | } | |
282a8e03 | 1475 | out: |
f4200391 | 1476 | trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result); |
642261ac RZ |
1477 | return result; |
1478 | } | |
a2d58167 | 1479 | #else |
01cddfe9 AB |
1480 | static int dax_iomap_pmd_fault(struct vm_fault *vmf, |
1481 | const struct iomap_ops *ops) | |
a2d58167 DJ |
1482 | { |
1483 | return VM_FAULT_FALLBACK; | |
1484 | } | |
642261ac | 1485 | #endif /* CONFIG_FS_DAX_PMD */ |
a2d58167 DJ |
1486 | |
1487 | /** | |
1488 | * dax_iomap_fault - handle a page fault on a DAX file | |
1489 | * @vmf: The description of the fault | |
1490 | * @ops: iomap ops passed from the file system | |
1491 | * | |
1492 | * When a page fault occurs, filesystems may call this helper in | |
1493 | * their fault handler for DAX files. dax_iomap_fault() assumes the caller | |
1494 | * has done all the necessary locking for page fault to proceed | |
1495 | * successfully. | |
1496 | */ | |
c791ace1 DJ |
1497 | int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, |
1498 | const struct iomap_ops *ops) | |
a2d58167 | 1499 | { |
c791ace1 DJ |
1500 | switch (pe_size) { |
1501 | case PE_SIZE_PTE: | |
a2d58167 | 1502 | return dax_iomap_pte_fault(vmf, ops); |
c791ace1 | 1503 | case PE_SIZE_PMD: |
a2d58167 DJ |
1504 | return dax_iomap_pmd_fault(vmf, ops); |
1505 | default: | |
1506 | return VM_FAULT_FALLBACK; | |
1507 | } | |
1508 | } | |
1509 | EXPORT_SYMBOL_GPL(dax_iomap_fault); |