]>
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> |
289c6aed | 28 | #include <linux/sched.h> |
f361bf4a | 29 | #include <linux/sched/signal.h> |
d475c634 | 30 | #include <linux/uio.h> |
f7ca90b1 | 31 | #include <linux/vmstat.h> |
34c0fd54 | 32 | #include <linux/pfn_t.h> |
0e749e54 | 33 | #include <linux/sizes.h> |
4b4bb46d | 34 | #include <linux/mmu_notifier.h> |
a254e568 CH |
35 | #include <linux/iomap.h> |
36 | #include "internal.h" | |
d475c634 | 37 | |
282a8e03 RZ |
38 | #define CREATE_TRACE_POINTS |
39 | #include <trace/events/fs_dax.h> | |
40 | ||
ac401cc7 JK |
41 | /* We choose 4096 entries - same as per-zone page wait tables */ |
42 | #define DAX_WAIT_TABLE_BITS 12 | |
43 | #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) | |
44 | ||
ce95ab0f | 45 | static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; |
ac401cc7 JK |
46 | |
47 | static int __init init_dax_wait_table(void) | |
48 | { | |
49 | int i; | |
50 | ||
51 | for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) | |
52 | init_waitqueue_head(wait_table + i); | |
53 | return 0; | |
54 | } | |
55 | fs_initcall(init_dax_wait_table); | |
56 | ||
527b19d0 RZ |
57 | /* |
58 | * We use lowest available bit in exceptional entry for locking, one bit for | |
59 | * the entry size (PMD) and two more to tell us if the entry is a zero page or | |
60 | * an empty entry that is just used for locking. In total four special bits. | |
61 | * | |
62 | * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE | |
63 | * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem | |
64 | * block allocation. | |
65 | */ | |
66 | #define RADIX_DAX_SHIFT (RADIX_TREE_EXCEPTIONAL_SHIFT + 4) | |
67 | #define RADIX_DAX_ENTRY_LOCK (1 << RADIX_TREE_EXCEPTIONAL_SHIFT) | |
68 | #define RADIX_DAX_PMD (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1)) | |
69 | #define RADIX_DAX_ZERO_PAGE (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2)) | |
70 | #define RADIX_DAX_EMPTY (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 3)) | |
71 | ||
72 | static unsigned long dax_radix_sector(void *entry) | |
73 | { | |
74 | return (unsigned long)entry >> RADIX_DAX_SHIFT; | |
75 | } | |
76 | ||
77 | static void *dax_radix_locked_entry(sector_t sector, unsigned long flags) | |
78 | { | |
79 | return (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY | flags | | |
80 | ((unsigned long)sector << RADIX_DAX_SHIFT) | | |
81 | RADIX_DAX_ENTRY_LOCK); | |
82 | } | |
83 | ||
84 | static unsigned int dax_radix_order(void *entry) | |
85 | { | |
86 | if ((unsigned long)entry & RADIX_DAX_PMD) | |
87 | return PMD_SHIFT - PAGE_SHIFT; | |
88 | return 0; | |
89 | } | |
90 | ||
642261ac | 91 | static int dax_is_pmd_entry(void *entry) |
d1a5f2b4 | 92 | { |
642261ac | 93 | return (unsigned long)entry & RADIX_DAX_PMD; |
d1a5f2b4 DW |
94 | } |
95 | ||
642261ac | 96 | static int dax_is_pte_entry(void *entry) |
d475c634 | 97 | { |
642261ac | 98 | return !((unsigned long)entry & RADIX_DAX_PMD); |
d475c634 MW |
99 | } |
100 | ||
642261ac | 101 | static int dax_is_zero_entry(void *entry) |
d475c634 | 102 | { |
91d25ba8 | 103 | return (unsigned long)entry & RADIX_DAX_ZERO_PAGE; |
d475c634 MW |
104 | } |
105 | ||
642261ac | 106 | static int dax_is_empty_entry(void *entry) |
b2e0d162 | 107 | { |
642261ac | 108 | return (unsigned long)entry & RADIX_DAX_EMPTY; |
b2e0d162 DW |
109 | } |
110 | ||
ac401cc7 JK |
111 | /* |
112 | * DAX radix tree locking | |
113 | */ | |
114 | struct exceptional_entry_key { | |
115 | struct address_space *mapping; | |
63e95b5c | 116 | pgoff_t entry_start; |
ac401cc7 JK |
117 | }; |
118 | ||
119 | struct wait_exceptional_entry_queue { | |
ac6424b9 | 120 | wait_queue_entry_t wait; |
ac401cc7 JK |
121 | struct exceptional_entry_key key; |
122 | }; | |
123 | ||
63e95b5c RZ |
124 | static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping, |
125 | pgoff_t index, void *entry, struct exceptional_entry_key *key) | |
126 | { | |
127 | unsigned long hash; | |
128 | ||
129 | /* | |
130 | * If 'entry' is a PMD, align the 'index' that we use for the wait | |
131 | * queue to the start of that PMD. This ensures that all offsets in | |
132 | * the range covered by the PMD map to the same bit lock. | |
133 | */ | |
642261ac | 134 | if (dax_is_pmd_entry(entry)) |
63e95b5c RZ |
135 | index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1); |
136 | ||
137 | key->mapping = mapping; | |
138 | key->entry_start = index; | |
139 | ||
140 | hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS); | |
141 | return wait_table + hash; | |
142 | } | |
143 | ||
ac6424b9 | 144 | static int wake_exceptional_entry_func(wait_queue_entry_t *wait, unsigned int mode, |
ac401cc7 JK |
145 | int sync, void *keyp) |
146 | { | |
147 | struct exceptional_entry_key *key = keyp; | |
148 | struct wait_exceptional_entry_queue *ewait = | |
149 | container_of(wait, struct wait_exceptional_entry_queue, wait); | |
150 | ||
151 | if (key->mapping != ewait->key.mapping || | |
63e95b5c | 152 | key->entry_start != ewait->key.entry_start) |
ac401cc7 JK |
153 | return 0; |
154 | return autoremove_wake_function(wait, mode, sync, NULL); | |
155 | } | |
156 | ||
e30331ff RZ |
157 | /* |
158 | * We do not necessarily hold the mapping->tree_lock when we call this | |
159 | * function so it is possible that 'entry' is no longer a valid item in the | |
160 | * radix tree. This is okay because all we really need to do is to find the | |
161 | * correct waitqueue where tasks might be waiting for that old 'entry' and | |
162 | * wake them. | |
163 | */ | |
d01ad197 | 164 | static void dax_wake_mapping_entry_waiter(struct address_space *mapping, |
e30331ff RZ |
165 | pgoff_t index, void *entry, bool wake_all) |
166 | { | |
167 | struct exceptional_entry_key key; | |
168 | wait_queue_head_t *wq; | |
169 | ||
170 | wq = dax_entry_waitqueue(mapping, index, entry, &key); | |
171 | ||
172 | /* | |
173 | * Checking for locked entry and prepare_to_wait_exclusive() happens | |
174 | * under mapping->tree_lock, ditto for entry handling in our callers. | |
175 | * So at this point all tasks that could have seen our entry locked | |
176 | * must be in the waitqueue and the following check will see them. | |
177 | */ | |
178 | if (waitqueue_active(wq)) | |
179 | __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); | |
180 | } | |
181 | ||
ac401cc7 JK |
182 | /* |
183 | * Check whether the given slot is locked. The function must be called with | |
184 | * mapping->tree_lock held | |
185 | */ | |
186 | static inline int slot_locked(struct address_space *mapping, void **slot) | |
187 | { | |
188 | unsigned long entry = (unsigned long) | |
189 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
190 | return entry & RADIX_DAX_ENTRY_LOCK; | |
191 | } | |
192 | ||
193 | /* | |
194 | * Mark the given slot is locked. The function must be called with | |
195 | * mapping->tree_lock held | |
196 | */ | |
197 | static inline void *lock_slot(struct address_space *mapping, void **slot) | |
198 | { | |
199 | unsigned long entry = (unsigned long) | |
200 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
201 | ||
202 | entry |= RADIX_DAX_ENTRY_LOCK; | |
6d75f366 | 203 | radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); |
ac401cc7 JK |
204 | return (void *)entry; |
205 | } | |
206 | ||
207 | /* | |
208 | * Mark the given slot is unlocked. The function must be called with | |
209 | * mapping->tree_lock held | |
210 | */ | |
211 | static inline void *unlock_slot(struct address_space *mapping, void **slot) | |
212 | { | |
213 | unsigned long entry = (unsigned long) | |
214 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
215 | ||
216 | entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK; | |
6d75f366 | 217 | radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); |
ac401cc7 JK |
218 | return (void *)entry; |
219 | } | |
220 | ||
221 | /* | |
222 | * Lookup entry in radix tree, wait for it to become unlocked if it is | |
223 | * exceptional entry and return it. The caller must call | |
224 | * put_unlocked_mapping_entry() when he decided not to lock the entry or | |
225 | * put_locked_mapping_entry() when he locked the entry and now wants to | |
226 | * unlock it. | |
227 | * | |
228 | * The function must be called with mapping->tree_lock held. | |
229 | */ | |
230 | static void *get_unlocked_mapping_entry(struct address_space *mapping, | |
231 | pgoff_t index, void ***slotp) | |
232 | { | |
e3ad61c6 | 233 | void *entry, **slot; |
ac401cc7 | 234 | struct wait_exceptional_entry_queue ewait; |
63e95b5c | 235 | wait_queue_head_t *wq; |
ac401cc7 JK |
236 | |
237 | init_wait(&ewait.wait); | |
238 | ewait.wait.func = wake_exceptional_entry_func; | |
ac401cc7 JK |
239 | |
240 | for (;;) { | |
e3ad61c6 | 241 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, |
ac401cc7 | 242 | &slot); |
91d25ba8 RZ |
243 | if (!entry || |
244 | WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)) || | |
ac401cc7 JK |
245 | !slot_locked(mapping, slot)) { |
246 | if (slotp) | |
247 | *slotp = slot; | |
e3ad61c6 | 248 | return entry; |
ac401cc7 | 249 | } |
63e95b5c RZ |
250 | |
251 | wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key); | |
ac401cc7 JK |
252 | prepare_to_wait_exclusive(wq, &ewait.wait, |
253 | TASK_UNINTERRUPTIBLE); | |
254 | spin_unlock_irq(&mapping->tree_lock); | |
255 | schedule(); | |
256 | finish_wait(wq, &ewait.wait); | |
257 | spin_lock_irq(&mapping->tree_lock); | |
258 | } | |
259 | } | |
260 | ||
b1aa812b JK |
261 | static void dax_unlock_mapping_entry(struct address_space *mapping, |
262 | pgoff_t index) | |
263 | { | |
264 | void *entry, **slot; | |
265 | ||
266 | spin_lock_irq(&mapping->tree_lock); | |
267 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot); | |
268 | if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) || | |
269 | !slot_locked(mapping, slot))) { | |
270 | spin_unlock_irq(&mapping->tree_lock); | |
271 | return; | |
272 | } | |
273 | unlock_slot(mapping, slot); | |
274 | spin_unlock_irq(&mapping->tree_lock); | |
275 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); | |
276 | } | |
277 | ||
422476c4 | 278 | static void put_locked_mapping_entry(struct address_space *mapping, |
91d25ba8 | 279 | pgoff_t index) |
422476c4 | 280 | { |
91d25ba8 | 281 | dax_unlock_mapping_entry(mapping, index); |
422476c4 RZ |
282 | } |
283 | ||
284 | /* | |
285 | * Called when we are done with radix tree entry we looked up via | |
286 | * get_unlocked_mapping_entry() and which we didn't lock in the end. | |
287 | */ | |
288 | static void put_unlocked_mapping_entry(struct address_space *mapping, | |
289 | pgoff_t index, void *entry) | |
290 | { | |
91d25ba8 | 291 | if (!entry) |
422476c4 RZ |
292 | return; |
293 | ||
294 | /* We have to wake up next waiter for the radix tree entry lock */ | |
295 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); | |
296 | } | |
297 | ||
ac401cc7 | 298 | /* |
91d25ba8 RZ |
299 | * Find radix tree entry at given index. If it points to an exceptional entry, |
300 | * return it with the radix tree entry locked. If the radix tree doesn't | |
301 | * contain given index, create an empty exceptional entry for the index and | |
302 | * return with it locked. | |
ac401cc7 | 303 | * |
642261ac RZ |
304 | * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will |
305 | * either return that locked entry or will return an error. This error will | |
91d25ba8 RZ |
306 | * happen if there are any 4k entries within the 2MiB range that we are |
307 | * requesting. | |
642261ac RZ |
308 | * |
309 | * We always favor 4k entries over 2MiB entries. There isn't a flow where we | |
310 | * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB | |
311 | * insertion will fail if it finds any 4k entries already in the tree, and a | |
312 | * 4k insertion will cause an existing 2MiB entry to be unmapped and | |
313 | * downgraded to 4k entries. This happens for both 2MiB huge zero pages as | |
314 | * well as 2MiB empty entries. | |
315 | * | |
316 | * The exception to this downgrade path is for 2MiB DAX PMD entries that have | |
317 | * real storage backing them. We will leave these real 2MiB DAX entries in | |
318 | * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry. | |
319 | * | |
ac401cc7 JK |
320 | * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For |
321 | * persistent memory the benefit is doubtful. We can add that later if we can | |
322 | * show it helps. | |
323 | */ | |
642261ac RZ |
324 | static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index, |
325 | unsigned long size_flag) | |
ac401cc7 | 326 | { |
642261ac | 327 | bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */ |
e3ad61c6 | 328 | void *entry, **slot; |
ac401cc7 JK |
329 | |
330 | restart: | |
331 | spin_lock_irq(&mapping->tree_lock); | |
e3ad61c6 | 332 | entry = get_unlocked_mapping_entry(mapping, index, &slot); |
642261ac | 333 | |
91d25ba8 RZ |
334 | if (WARN_ON_ONCE(entry && !radix_tree_exceptional_entry(entry))) { |
335 | entry = ERR_PTR(-EIO); | |
336 | goto out_unlock; | |
337 | } | |
338 | ||
642261ac RZ |
339 | if (entry) { |
340 | if (size_flag & RADIX_DAX_PMD) { | |
91d25ba8 | 341 | if (dax_is_pte_entry(entry)) { |
642261ac RZ |
342 | put_unlocked_mapping_entry(mapping, index, |
343 | entry); | |
344 | entry = ERR_PTR(-EEXIST); | |
345 | goto out_unlock; | |
346 | } | |
347 | } else { /* trying to grab a PTE entry */ | |
91d25ba8 | 348 | if (dax_is_pmd_entry(entry) && |
642261ac RZ |
349 | (dax_is_zero_entry(entry) || |
350 | dax_is_empty_entry(entry))) { | |
351 | pmd_downgrade = true; | |
352 | } | |
353 | } | |
354 | } | |
355 | ||
ac401cc7 | 356 | /* No entry for given index? Make sure radix tree is big enough. */ |
642261ac | 357 | if (!entry || pmd_downgrade) { |
ac401cc7 JK |
358 | int err; |
359 | ||
642261ac RZ |
360 | if (pmd_downgrade) { |
361 | /* | |
362 | * Make sure 'entry' remains valid while we drop | |
363 | * mapping->tree_lock. | |
364 | */ | |
365 | entry = lock_slot(mapping, slot); | |
366 | } | |
367 | ||
ac401cc7 | 368 | spin_unlock_irq(&mapping->tree_lock); |
642261ac RZ |
369 | /* |
370 | * Besides huge zero pages the only other thing that gets | |
371 | * downgraded are empty entries which don't need to be | |
372 | * unmapped. | |
373 | */ | |
374 | if (pmd_downgrade && dax_is_zero_entry(entry)) | |
375 | unmap_mapping_range(mapping, | |
376 | (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0); | |
377 | ||
ac401cc7 JK |
378 | err = radix_tree_preload( |
379 | mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM); | |
0cb80b48 JK |
380 | if (err) { |
381 | if (pmd_downgrade) | |
91d25ba8 | 382 | put_locked_mapping_entry(mapping, index); |
ac401cc7 | 383 | return ERR_PTR(err); |
0cb80b48 | 384 | } |
ac401cc7 | 385 | spin_lock_irq(&mapping->tree_lock); |
642261ac | 386 | |
e11f8b7b RZ |
387 | if (!entry) { |
388 | /* | |
389 | * We needed to drop the page_tree lock while calling | |
390 | * radix_tree_preload() and we didn't have an entry to | |
391 | * lock. See if another thread inserted an entry at | |
392 | * our index during this time. | |
393 | */ | |
394 | entry = __radix_tree_lookup(&mapping->page_tree, index, | |
395 | NULL, &slot); | |
396 | if (entry) { | |
397 | radix_tree_preload_end(); | |
398 | spin_unlock_irq(&mapping->tree_lock); | |
399 | goto restart; | |
400 | } | |
401 | } | |
402 | ||
642261ac RZ |
403 | if (pmd_downgrade) { |
404 | radix_tree_delete(&mapping->page_tree, index); | |
405 | mapping->nrexceptional--; | |
406 | dax_wake_mapping_entry_waiter(mapping, index, entry, | |
407 | true); | |
408 | } | |
409 | ||
410 | entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY); | |
411 | ||
412 | err = __radix_tree_insert(&mapping->page_tree, index, | |
413 | dax_radix_order(entry), entry); | |
ac401cc7 JK |
414 | radix_tree_preload_end(); |
415 | if (err) { | |
416 | spin_unlock_irq(&mapping->tree_lock); | |
642261ac | 417 | /* |
e11f8b7b RZ |
418 | * Our insertion of a DAX entry failed, most likely |
419 | * because we were inserting a PMD entry and it | |
420 | * collided with a PTE sized entry at a different | |
421 | * index in the PMD range. We haven't inserted | |
422 | * anything into the radix tree and have no waiters to | |
423 | * wake. | |
642261ac | 424 | */ |
ac401cc7 JK |
425 | return ERR_PTR(err); |
426 | } | |
427 | /* Good, we have inserted empty locked entry into the tree. */ | |
428 | mapping->nrexceptional++; | |
429 | spin_unlock_irq(&mapping->tree_lock); | |
e3ad61c6 | 430 | return entry; |
ac401cc7 | 431 | } |
e3ad61c6 | 432 | entry = lock_slot(mapping, slot); |
642261ac | 433 | out_unlock: |
ac401cc7 | 434 | spin_unlock_irq(&mapping->tree_lock); |
e3ad61c6 | 435 | return entry; |
ac401cc7 JK |
436 | } |
437 | ||
c6dcf52c JK |
438 | static int __dax_invalidate_mapping_entry(struct address_space *mapping, |
439 | pgoff_t index, bool trunc) | |
440 | { | |
441 | int ret = 0; | |
442 | void *entry; | |
443 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
444 | ||
445 | spin_lock_irq(&mapping->tree_lock); | |
446 | entry = get_unlocked_mapping_entry(mapping, index, NULL); | |
91d25ba8 | 447 | if (!entry || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry))) |
c6dcf52c JK |
448 | goto out; |
449 | if (!trunc && | |
450 | (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) || | |
451 | radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))) | |
452 | goto out; | |
453 | radix_tree_delete(page_tree, index); | |
454 | mapping->nrexceptional--; | |
455 | ret = 1; | |
456 | out: | |
457 | put_unlocked_mapping_entry(mapping, index, entry); | |
458 | spin_unlock_irq(&mapping->tree_lock); | |
459 | return ret; | |
460 | } | |
ac401cc7 JK |
461 | /* |
462 | * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree | |
463 | * entry to get unlocked before deleting it. | |
464 | */ | |
465 | int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) | |
466 | { | |
c6dcf52c | 467 | int ret = __dax_invalidate_mapping_entry(mapping, index, true); |
ac401cc7 | 468 | |
ac401cc7 JK |
469 | /* |
470 | * This gets called from truncate / punch_hole path. As such, the caller | |
471 | * must hold locks protecting against concurrent modifications of the | |
472 | * radix tree (usually fs-private i_mmap_sem for writing). Since the | |
473 | * caller has seen exceptional entry for this index, we better find it | |
474 | * at that index as well... | |
475 | */ | |
c6dcf52c JK |
476 | WARN_ON_ONCE(!ret); |
477 | return ret; | |
478 | } | |
479 | ||
c6dcf52c JK |
480 | /* |
481 | * Invalidate exceptional DAX entry if it is clean. | |
482 | */ | |
483 | int dax_invalidate_mapping_entry_sync(struct address_space *mapping, | |
484 | pgoff_t index) | |
485 | { | |
486 | return __dax_invalidate_mapping_entry(mapping, index, false); | |
ac401cc7 JK |
487 | } |
488 | ||
cccbce67 DW |
489 | static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev, |
490 | sector_t sector, size_t size, struct page *to, | |
491 | unsigned long vaddr) | |
f7ca90b1 | 492 | { |
cccbce67 DW |
493 | void *vto, *kaddr; |
494 | pgoff_t pgoff; | |
495 | pfn_t pfn; | |
496 | long rc; | |
497 | int id; | |
498 | ||
499 | rc = bdev_dax_pgoff(bdev, sector, size, &pgoff); | |
500 | if (rc) | |
501 | return rc; | |
502 | ||
503 | id = dax_read_lock(); | |
504 | rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn); | |
505 | if (rc < 0) { | |
506 | dax_read_unlock(id); | |
507 | return rc; | |
508 | } | |
f7ca90b1 | 509 | vto = kmap_atomic(to); |
cccbce67 | 510 | copy_user_page(vto, (void __force *)kaddr, vaddr, to); |
f7ca90b1 | 511 | kunmap_atomic(vto); |
cccbce67 | 512 | dax_read_unlock(id); |
f7ca90b1 MW |
513 | return 0; |
514 | } | |
515 | ||
642261ac RZ |
516 | /* |
517 | * By this point grab_mapping_entry() has ensured that we have a locked entry | |
518 | * of the appropriate size so we don't have to worry about downgrading PMDs to | |
519 | * PTEs. If we happen to be trying to insert a PTE and there is a PMD | |
520 | * already in the tree, we will skip the insertion and just dirty the PMD as | |
521 | * appropriate. | |
522 | */ | |
ac401cc7 JK |
523 | static void *dax_insert_mapping_entry(struct address_space *mapping, |
524 | struct vm_fault *vmf, | |
642261ac RZ |
525 | void *entry, sector_t sector, |
526 | unsigned long flags) | |
9973c98e RZ |
527 | { |
528 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
ac401cc7 JK |
529 | void *new_entry; |
530 | pgoff_t index = vmf->pgoff; | |
9973c98e | 531 | |
ac401cc7 | 532 | if (vmf->flags & FAULT_FLAG_WRITE) |
d2b2a28e | 533 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); |
9973c98e | 534 | |
91d25ba8 RZ |
535 | if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_ZERO_PAGE)) { |
536 | /* we are replacing a zero page with block mapping */ | |
537 | if (dax_is_pmd_entry(entry)) | |
538 | unmap_mapping_range(mapping, | |
539 | (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, | |
540 | PMD_SIZE, 0); | |
541 | else /* pte entry */ | |
542 | unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT, | |
543 | PAGE_SIZE, 0); | |
9973c98e RZ |
544 | } |
545 | ||
ac401cc7 | 546 | spin_lock_irq(&mapping->tree_lock); |
642261ac RZ |
547 | new_entry = dax_radix_locked_entry(sector, flags); |
548 | ||
91d25ba8 | 549 | if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { |
642261ac RZ |
550 | /* |
551 | * Only swap our new entry into the radix tree if the current | |
552 | * entry is a zero page or an empty entry. If a normal PTE or | |
553 | * PMD entry is already in the tree, we leave it alone. This | |
554 | * means that if we are trying to insert a PTE and the | |
555 | * existing entry is a PMD, we will just leave the PMD in the | |
556 | * tree and dirty it if necessary. | |
557 | */ | |
f7942430 | 558 | struct radix_tree_node *node; |
ac401cc7 JK |
559 | void **slot; |
560 | void *ret; | |
9973c98e | 561 | |
f7942430 | 562 | ret = __radix_tree_lookup(page_tree, index, &node, &slot); |
ac401cc7 | 563 | WARN_ON_ONCE(ret != entry); |
4d693d08 JW |
564 | __radix_tree_replace(page_tree, node, slot, |
565 | new_entry, NULL, NULL); | |
91d25ba8 | 566 | entry = new_entry; |
9973c98e | 567 | } |
91d25ba8 | 568 | |
ac401cc7 | 569 | if (vmf->flags & FAULT_FLAG_WRITE) |
9973c98e | 570 | radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY); |
91d25ba8 | 571 | |
9973c98e | 572 | spin_unlock_irq(&mapping->tree_lock); |
91d25ba8 | 573 | return entry; |
9973c98e RZ |
574 | } |
575 | ||
4b4bb46d JK |
576 | static inline unsigned long |
577 | pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma) | |
578 | { | |
579 | unsigned long address; | |
580 | ||
581 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
582 | VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); | |
583 | return address; | |
584 | } | |
585 | ||
586 | /* Walk all mappings of a given index of a file and writeprotect them */ | |
587 | static void dax_mapping_entry_mkclean(struct address_space *mapping, | |
588 | pgoff_t index, unsigned long pfn) | |
589 | { | |
590 | struct vm_area_struct *vma; | |
f729c8c9 RZ |
591 | pte_t pte, *ptep = NULL; |
592 | pmd_t *pmdp = NULL; | |
4b4bb46d | 593 | spinlock_t *ptl; |
4b4bb46d JK |
594 | |
595 | i_mmap_lock_read(mapping); | |
596 | vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) { | |
a4d1a885 | 597 | unsigned long address, start, end; |
4b4bb46d JK |
598 | |
599 | cond_resched(); | |
600 | ||
601 | if (!(vma->vm_flags & VM_SHARED)) | |
602 | continue; | |
603 | ||
604 | address = pgoff_address(index, vma); | |
a4d1a885 JG |
605 | |
606 | /* | |
607 | * Note because we provide start/end to follow_pte_pmd it will | |
608 | * call mmu_notifier_invalidate_range_start() on our behalf | |
609 | * before taking any lock. | |
610 | */ | |
611 | if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl)) | |
4b4bb46d | 612 | continue; |
4b4bb46d | 613 | |
f729c8c9 RZ |
614 | if (pmdp) { |
615 | #ifdef CONFIG_FS_DAX_PMD | |
616 | pmd_t pmd; | |
617 | ||
618 | if (pfn != pmd_pfn(*pmdp)) | |
619 | goto unlock_pmd; | |
620 | if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp)) | |
621 | goto unlock_pmd; | |
622 | ||
623 | flush_cache_page(vma, address, pfn); | |
624 | pmd = pmdp_huge_clear_flush(vma, address, pmdp); | |
625 | pmd = pmd_wrprotect(pmd); | |
626 | pmd = pmd_mkclean(pmd); | |
627 | set_pmd_at(vma->vm_mm, address, pmdp, pmd); | |
a4d1a885 | 628 | mmu_notifier_invalidate_range(vma->vm_mm, start, end); |
f729c8c9 RZ |
629 | unlock_pmd: |
630 | spin_unlock(ptl); | |
631 | #endif | |
632 | } else { | |
633 | if (pfn != pte_pfn(*ptep)) | |
634 | goto unlock_pte; | |
635 | if (!pte_dirty(*ptep) && !pte_write(*ptep)) | |
636 | goto unlock_pte; | |
637 | ||
638 | flush_cache_page(vma, address, pfn); | |
639 | pte = ptep_clear_flush(vma, address, ptep); | |
640 | pte = pte_wrprotect(pte); | |
641 | pte = pte_mkclean(pte); | |
642 | set_pte_at(vma->vm_mm, address, ptep, pte); | |
a4d1a885 | 643 | mmu_notifier_invalidate_range(vma->vm_mm, start, end); |
f729c8c9 RZ |
644 | unlock_pte: |
645 | pte_unmap_unlock(ptep, ptl); | |
646 | } | |
4b4bb46d | 647 | |
a4d1a885 | 648 | mmu_notifier_invalidate_range_end(vma->vm_mm, start, end); |
4b4bb46d JK |
649 | } |
650 | i_mmap_unlock_read(mapping); | |
651 | } | |
652 | ||
9973c98e | 653 | static int dax_writeback_one(struct block_device *bdev, |
cccbce67 DW |
654 | struct dax_device *dax_dev, struct address_space *mapping, |
655 | pgoff_t index, void *entry) | |
9973c98e RZ |
656 | { |
657 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
cccbce67 DW |
658 | void *entry2, **slot, *kaddr; |
659 | long ret = 0, id; | |
660 | sector_t sector; | |
661 | pgoff_t pgoff; | |
662 | size_t size; | |
663 | pfn_t pfn; | |
9973c98e | 664 | |
9973c98e | 665 | /* |
a6abc2c0 JK |
666 | * A page got tagged dirty in DAX mapping? Something is seriously |
667 | * wrong. | |
9973c98e | 668 | */ |
a6abc2c0 JK |
669 | if (WARN_ON(!radix_tree_exceptional_entry(entry))) |
670 | return -EIO; | |
9973c98e | 671 | |
a6abc2c0 JK |
672 | spin_lock_irq(&mapping->tree_lock); |
673 | entry2 = get_unlocked_mapping_entry(mapping, index, &slot); | |
674 | /* Entry got punched out / reallocated? */ | |
91d25ba8 | 675 | if (!entry2 || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry2))) |
a6abc2c0 JK |
676 | goto put_unlocked; |
677 | /* | |
678 | * Entry got reallocated elsewhere? No need to writeback. We have to | |
679 | * compare sectors as we must not bail out due to difference in lockbit | |
680 | * or entry type. | |
681 | */ | |
682 | if (dax_radix_sector(entry2) != dax_radix_sector(entry)) | |
683 | goto put_unlocked; | |
642261ac RZ |
684 | if (WARN_ON_ONCE(dax_is_empty_entry(entry) || |
685 | dax_is_zero_entry(entry))) { | |
9973c98e | 686 | ret = -EIO; |
a6abc2c0 | 687 | goto put_unlocked; |
9973c98e RZ |
688 | } |
689 | ||
a6abc2c0 JK |
690 | /* Another fsync thread may have already written back this entry */ |
691 | if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)) | |
692 | goto put_unlocked; | |
693 | /* Lock the entry to serialize with page faults */ | |
694 | entry = lock_slot(mapping, slot); | |
695 | /* | |
696 | * We can clear the tag now but we have to be careful so that concurrent | |
697 | * dax_writeback_one() calls for the same index cannot finish before we | |
698 | * actually flush the caches. This is achieved as the calls will look | |
699 | * at the entry only under tree_lock and once they do that they will | |
700 | * see the entry locked and wait for it to unlock. | |
701 | */ | |
702 | radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE); | |
703 | spin_unlock_irq(&mapping->tree_lock); | |
704 | ||
642261ac RZ |
705 | /* |
706 | * Even if dax_writeback_mapping_range() was given a wbc->range_start | |
707 | * in the middle of a PMD, the 'index' we are given will be aligned to | |
708 | * the start index of the PMD, as will the sector we pull from | |
709 | * 'entry'. This allows us to flush for PMD_SIZE and not have to | |
710 | * worry about partial PMD writebacks. | |
711 | */ | |
cccbce67 DW |
712 | sector = dax_radix_sector(entry); |
713 | size = PAGE_SIZE << dax_radix_order(entry); | |
714 | ||
715 | id = dax_read_lock(); | |
716 | ret = bdev_dax_pgoff(bdev, sector, size, &pgoff); | |
717 | if (ret) | |
718 | goto dax_unlock; | |
9973c98e RZ |
719 | |
720 | /* | |
cccbce67 DW |
721 | * dax_direct_access() may sleep, so cannot hold tree_lock over |
722 | * its invocation. | |
9973c98e | 723 | */ |
cccbce67 DW |
724 | ret = dax_direct_access(dax_dev, pgoff, size / PAGE_SIZE, &kaddr, &pfn); |
725 | if (ret < 0) | |
726 | goto dax_unlock; | |
9973c98e | 727 | |
cccbce67 | 728 | if (WARN_ON_ONCE(ret < size / PAGE_SIZE)) { |
9973c98e | 729 | ret = -EIO; |
cccbce67 | 730 | goto dax_unlock; |
9973c98e RZ |
731 | } |
732 | ||
cccbce67 | 733 | dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(pfn)); |
6318770a | 734 | dax_flush(dax_dev, pgoff, kaddr, size); |
4b4bb46d JK |
735 | /* |
736 | * After we have flushed the cache, we can clear the dirty tag. There | |
737 | * cannot be new dirty data in the pfn after the flush has completed as | |
738 | * the pfn mappings are writeprotected and fault waits for mapping | |
739 | * entry lock. | |
740 | */ | |
741 | spin_lock_irq(&mapping->tree_lock); | |
742 | radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY); | |
743 | spin_unlock_irq(&mapping->tree_lock); | |
f9bc3a07 | 744 | trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT); |
cccbce67 DW |
745 | dax_unlock: |
746 | dax_read_unlock(id); | |
91d25ba8 | 747 | put_locked_mapping_entry(mapping, index); |
9973c98e RZ |
748 | return ret; |
749 | ||
a6abc2c0 JK |
750 | put_unlocked: |
751 | put_unlocked_mapping_entry(mapping, index, entry2); | |
9973c98e RZ |
752 | spin_unlock_irq(&mapping->tree_lock); |
753 | return ret; | |
754 | } | |
755 | ||
756 | /* | |
757 | * Flush the mapping to the persistent domain within the byte range of [start, | |
758 | * end]. This is required by data integrity operations to ensure file data is | |
759 | * on persistent storage prior to completion of the operation. | |
760 | */ | |
7f6d5b52 RZ |
761 | int dax_writeback_mapping_range(struct address_space *mapping, |
762 | struct block_device *bdev, struct writeback_control *wbc) | |
9973c98e RZ |
763 | { |
764 | struct inode *inode = mapping->host; | |
642261ac | 765 | pgoff_t start_index, end_index; |
9973c98e | 766 | pgoff_t indices[PAGEVEC_SIZE]; |
cccbce67 | 767 | struct dax_device *dax_dev; |
9973c98e RZ |
768 | struct pagevec pvec; |
769 | bool done = false; | |
770 | int i, ret = 0; | |
9973c98e RZ |
771 | |
772 | if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) | |
773 | return -EIO; | |
774 | ||
7f6d5b52 RZ |
775 | if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) |
776 | return 0; | |
777 | ||
cccbce67 DW |
778 | dax_dev = dax_get_by_host(bdev->bd_disk->disk_name); |
779 | if (!dax_dev) | |
780 | return -EIO; | |
781 | ||
09cbfeaf KS |
782 | start_index = wbc->range_start >> PAGE_SHIFT; |
783 | end_index = wbc->range_end >> PAGE_SHIFT; | |
9973c98e | 784 | |
d14a3f48 RZ |
785 | trace_dax_writeback_range(inode, start_index, end_index); |
786 | ||
9973c98e RZ |
787 | tag_pages_for_writeback(mapping, start_index, end_index); |
788 | ||
789 | pagevec_init(&pvec, 0); | |
790 | while (!done) { | |
791 | pvec.nr = find_get_entries_tag(mapping, start_index, | |
792 | PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE, | |
793 | pvec.pages, indices); | |
794 | ||
795 | if (pvec.nr == 0) | |
796 | break; | |
797 | ||
798 | for (i = 0; i < pvec.nr; i++) { | |
799 | if (indices[i] > end_index) { | |
800 | done = true; | |
801 | break; | |
802 | } | |
803 | ||
cccbce67 DW |
804 | ret = dax_writeback_one(bdev, dax_dev, mapping, |
805 | indices[i], pvec.pages[i]); | |
819ec6b9 JL |
806 | if (ret < 0) { |
807 | mapping_set_error(mapping, ret); | |
d14a3f48 | 808 | goto out; |
819ec6b9 | 809 | } |
9973c98e | 810 | } |
1eb643d0 | 811 | start_index = indices[pvec.nr - 1] + 1; |
9973c98e | 812 | } |
d14a3f48 | 813 | out: |
cccbce67 | 814 | put_dax(dax_dev); |
d14a3f48 RZ |
815 | trace_dax_writeback_range_done(inode, start_index, end_index); |
816 | return (ret < 0 ? ret : 0); | |
9973c98e RZ |
817 | } |
818 | EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); | |
819 | ||
ac401cc7 | 820 | static int dax_insert_mapping(struct address_space *mapping, |
cccbce67 | 821 | struct block_device *bdev, struct dax_device *dax_dev, |
91d25ba8 | 822 | sector_t sector, size_t size, void *entry, |
cccbce67 | 823 | struct vm_area_struct *vma, struct vm_fault *vmf) |
f7ca90b1 | 824 | { |
1a29d85e | 825 | unsigned long vaddr = vmf->address; |
cccbce67 DW |
826 | void *ret, *kaddr; |
827 | pgoff_t pgoff; | |
828 | int id, rc; | |
829 | pfn_t pfn; | |
f7ca90b1 | 830 | |
cccbce67 DW |
831 | rc = bdev_dax_pgoff(bdev, sector, size, &pgoff); |
832 | if (rc) | |
833 | return rc; | |
f7ca90b1 | 834 | |
cccbce67 DW |
835 | id = dax_read_lock(); |
836 | rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn); | |
837 | if (rc < 0) { | |
838 | dax_read_unlock(id); | |
839 | return rc; | |
840 | } | |
841 | dax_read_unlock(id); | |
842 | ||
843 | ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, 0); | |
4d9a2c87 JK |
844 | if (IS_ERR(ret)) |
845 | return PTR_ERR(ret); | |
9973c98e | 846 | |
b4440734 | 847 | trace_dax_insert_mapping(mapping->host, vmf, ret); |
91d25ba8 RZ |
848 | if (vmf->flags & FAULT_FLAG_WRITE) |
849 | return vm_insert_mixed_mkwrite(vma, vaddr, pfn); | |
850 | else | |
851 | return vm_insert_mixed(vma, vaddr, pfn); | |
0e3b210c | 852 | } |
0e3b210c | 853 | |
e30331ff | 854 | /* |
91d25ba8 RZ |
855 | * The user has performed a load from a hole in the file. Allocating a new |
856 | * page in the file would cause excessive storage usage for workloads with | |
857 | * sparse files. Instead we insert a read-only mapping of the 4k zero page. | |
858 | * If this page is ever written to we will re-fault and change the mapping to | |
859 | * point to real DAX storage instead. | |
e30331ff | 860 | */ |
91d25ba8 | 861 | static int dax_load_hole(struct address_space *mapping, void *entry, |
e30331ff RZ |
862 | struct vm_fault *vmf) |
863 | { | |
864 | struct inode *inode = mapping->host; | |
91d25ba8 RZ |
865 | unsigned long vaddr = vmf->address; |
866 | int ret = VM_FAULT_NOPAGE; | |
867 | struct page *zero_page; | |
868 | void *entry2; | |
e30331ff | 869 | |
91d25ba8 RZ |
870 | zero_page = ZERO_PAGE(0); |
871 | if (unlikely(!zero_page)) { | |
e30331ff RZ |
872 | ret = VM_FAULT_OOM; |
873 | goto out; | |
874 | } | |
875 | ||
91d25ba8 RZ |
876 | entry2 = dax_insert_mapping_entry(mapping, vmf, entry, 0, |
877 | RADIX_DAX_ZERO_PAGE); | |
878 | if (IS_ERR(entry2)) { | |
879 | ret = VM_FAULT_SIGBUS; | |
880 | goto out; | |
e30331ff | 881 | } |
91d25ba8 RZ |
882 | |
883 | vm_insert_mixed(vmf->vma, vaddr, page_to_pfn_t(zero_page)); | |
e30331ff RZ |
884 | out: |
885 | trace_dax_load_hole(inode, vmf, ret); | |
886 | return ret; | |
887 | } | |
888 | ||
4b0228fa VV |
889 | static bool dax_range_is_aligned(struct block_device *bdev, |
890 | unsigned int offset, unsigned int length) | |
891 | { | |
892 | unsigned short sector_size = bdev_logical_block_size(bdev); | |
893 | ||
894 | if (!IS_ALIGNED(offset, sector_size)) | |
895 | return false; | |
896 | if (!IS_ALIGNED(length, sector_size)) | |
897 | return false; | |
898 | ||
899 | return true; | |
900 | } | |
901 | ||
cccbce67 DW |
902 | int __dax_zero_page_range(struct block_device *bdev, |
903 | struct dax_device *dax_dev, sector_t sector, | |
904 | unsigned int offset, unsigned int size) | |
679c8bd3 | 905 | { |
cccbce67 DW |
906 | if (dax_range_is_aligned(bdev, offset, size)) { |
907 | sector_t start_sector = sector + (offset >> 9); | |
4b0228fa VV |
908 | |
909 | return blkdev_issue_zeroout(bdev, start_sector, | |
53ef7d0e | 910 | size >> 9, GFP_NOFS, 0); |
4b0228fa | 911 | } else { |
cccbce67 DW |
912 | pgoff_t pgoff; |
913 | long rc, id; | |
914 | void *kaddr; | |
915 | pfn_t pfn; | |
916 | ||
e84b83b9 | 917 | rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff); |
cccbce67 DW |
918 | if (rc) |
919 | return rc; | |
920 | ||
921 | id = dax_read_lock(); | |
e84b83b9 | 922 | rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, |
cccbce67 DW |
923 | &pfn); |
924 | if (rc < 0) { | |
925 | dax_read_unlock(id); | |
926 | return rc; | |
927 | } | |
81f55870 DW |
928 | memset(kaddr + offset, 0, size); |
929 | dax_flush(dax_dev, pgoff, kaddr + offset, size); | |
cccbce67 | 930 | dax_read_unlock(id); |
4b0228fa | 931 | } |
679c8bd3 CH |
932 | return 0; |
933 | } | |
934 | EXPORT_SYMBOL_GPL(__dax_zero_page_range); | |
935 | ||
333ccc97 | 936 | static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) |
25726bc1 | 937 | { |
333ccc97 | 938 | return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9); |
25726bc1 | 939 | } |
a254e568 | 940 | |
a254e568 | 941 | static loff_t |
11c59c92 | 942 | dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, |
a254e568 CH |
943 | struct iomap *iomap) |
944 | { | |
cccbce67 DW |
945 | struct block_device *bdev = iomap->bdev; |
946 | struct dax_device *dax_dev = iomap->dax_dev; | |
a254e568 CH |
947 | struct iov_iter *iter = data; |
948 | loff_t end = pos + length, done = 0; | |
949 | ssize_t ret = 0; | |
cccbce67 | 950 | int id; |
a254e568 CH |
951 | |
952 | if (iov_iter_rw(iter) == READ) { | |
953 | end = min(end, i_size_read(inode)); | |
954 | if (pos >= end) | |
955 | return 0; | |
956 | ||
957 | if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) | |
958 | return iov_iter_zero(min(length, end - pos), iter); | |
959 | } | |
960 | ||
961 | if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) | |
962 | return -EIO; | |
963 | ||
e3fce68c JK |
964 | /* |
965 | * Write can allocate block for an area which has a hole page mapped | |
966 | * into page tables. We have to tear down these mappings so that data | |
967 | * written by write(2) is visible in mmap. | |
968 | */ | |
cd656375 | 969 | if (iomap->flags & IOMAP_F_NEW) { |
e3fce68c JK |
970 | invalidate_inode_pages2_range(inode->i_mapping, |
971 | pos >> PAGE_SHIFT, | |
972 | (end - 1) >> PAGE_SHIFT); | |
973 | } | |
974 | ||
cccbce67 | 975 | id = dax_read_lock(); |
a254e568 CH |
976 | while (pos < end) { |
977 | unsigned offset = pos & (PAGE_SIZE - 1); | |
cccbce67 DW |
978 | const size_t size = ALIGN(length + offset, PAGE_SIZE); |
979 | const sector_t sector = dax_iomap_sector(iomap, pos); | |
a254e568 | 980 | ssize_t map_len; |
cccbce67 DW |
981 | pgoff_t pgoff; |
982 | void *kaddr; | |
983 | pfn_t pfn; | |
a254e568 | 984 | |
d1908f52 MH |
985 | if (fatal_signal_pending(current)) { |
986 | ret = -EINTR; | |
987 | break; | |
988 | } | |
989 | ||
cccbce67 DW |
990 | ret = bdev_dax_pgoff(bdev, sector, size, &pgoff); |
991 | if (ret) | |
992 | break; | |
993 | ||
994 | map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), | |
995 | &kaddr, &pfn); | |
a254e568 CH |
996 | if (map_len < 0) { |
997 | ret = map_len; | |
998 | break; | |
999 | } | |
1000 | ||
cccbce67 DW |
1001 | map_len = PFN_PHYS(map_len); |
1002 | kaddr += offset; | |
a254e568 CH |
1003 | map_len -= offset; |
1004 | if (map_len > end - pos) | |
1005 | map_len = end - pos; | |
1006 | ||
1007 | if (iov_iter_rw(iter) == WRITE) | |
fec53774 DW |
1008 | map_len = dax_copy_from_iter(dax_dev, pgoff, kaddr, |
1009 | map_len, iter); | |
a254e568 | 1010 | else |
cccbce67 | 1011 | map_len = copy_to_iter(kaddr, map_len, iter); |
a254e568 CH |
1012 | if (map_len <= 0) { |
1013 | ret = map_len ? map_len : -EFAULT; | |
1014 | break; | |
1015 | } | |
1016 | ||
1017 | pos += map_len; | |
1018 | length -= map_len; | |
1019 | done += map_len; | |
1020 | } | |
cccbce67 | 1021 | dax_read_unlock(id); |
a254e568 CH |
1022 | |
1023 | return done ? done : ret; | |
1024 | } | |
1025 | ||
1026 | /** | |
11c59c92 | 1027 | * dax_iomap_rw - Perform I/O to a DAX file |
a254e568 CH |
1028 | * @iocb: The control block for this I/O |
1029 | * @iter: The addresses to do I/O from or to | |
1030 | * @ops: iomap ops passed from the file system | |
1031 | * | |
1032 | * This function performs read and write operations to directly mapped | |
1033 | * persistent memory. The callers needs to take care of read/write exclusion | |
1034 | * and evicting any page cache pages in the region under I/O. | |
1035 | */ | |
1036 | ssize_t | |
11c59c92 | 1037 | dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, |
8ff6daa1 | 1038 | const struct iomap_ops *ops) |
a254e568 CH |
1039 | { |
1040 | struct address_space *mapping = iocb->ki_filp->f_mapping; | |
1041 | struct inode *inode = mapping->host; | |
1042 | loff_t pos = iocb->ki_pos, ret = 0, done = 0; | |
1043 | unsigned flags = 0; | |
1044 | ||
168316db CH |
1045 | if (iov_iter_rw(iter) == WRITE) { |
1046 | lockdep_assert_held_exclusive(&inode->i_rwsem); | |
a254e568 | 1047 | flags |= IOMAP_WRITE; |
168316db CH |
1048 | } else { |
1049 | lockdep_assert_held(&inode->i_rwsem); | |
1050 | } | |
a254e568 | 1051 | |
a254e568 CH |
1052 | while (iov_iter_count(iter)) { |
1053 | ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, | |
11c59c92 | 1054 | iter, dax_iomap_actor); |
a254e568 CH |
1055 | if (ret <= 0) |
1056 | break; | |
1057 | pos += ret; | |
1058 | done += ret; | |
1059 | } | |
1060 | ||
1061 | iocb->ki_pos += done; | |
1062 | return done ? done : ret; | |
1063 | } | |
11c59c92 | 1064 | EXPORT_SYMBOL_GPL(dax_iomap_rw); |
a7d73fe6 | 1065 | |
9f141d6e JK |
1066 | static int dax_fault_return(int error) |
1067 | { | |
1068 | if (error == 0) | |
1069 | return VM_FAULT_NOPAGE; | |
1070 | if (error == -ENOMEM) | |
1071 | return VM_FAULT_OOM; | |
1072 | return VM_FAULT_SIGBUS; | |
1073 | } | |
1074 | ||
a2d58167 DJ |
1075 | static int dax_iomap_pte_fault(struct vm_fault *vmf, |
1076 | const struct iomap_ops *ops) | |
a7d73fe6 | 1077 | { |
11bac800 | 1078 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
a7d73fe6 | 1079 | struct inode *inode = mapping->host; |
1a29d85e | 1080 | unsigned long vaddr = vmf->address; |
a7d73fe6 CH |
1081 | loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; |
1082 | sector_t sector; | |
1083 | struct iomap iomap = { 0 }; | |
9484ab1b | 1084 | unsigned flags = IOMAP_FAULT; |
a7d73fe6 | 1085 | int error, major = 0; |
b1aa812b | 1086 | int vmf_ret = 0; |
a7d73fe6 CH |
1087 | void *entry; |
1088 | ||
a9c42b33 | 1089 | trace_dax_pte_fault(inode, vmf, vmf_ret); |
a7d73fe6 CH |
1090 | /* |
1091 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
1092 | * to hold locks serializing us with truncate / punch hole so this is | |
1093 | * a reliable test. | |
1094 | */ | |
a9c42b33 RZ |
1095 | if (pos >= i_size_read(inode)) { |
1096 | vmf_ret = VM_FAULT_SIGBUS; | |
1097 | goto out; | |
1098 | } | |
a7d73fe6 | 1099 | |
a7d73fe6 CH |
1100 | if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) |
1101 | flags |= IOMAP_WRITE; | |
1102 | ||
13e451fd JK |
1103 | entry = grab_mapping_entry(mapping, vmf->pgoff, 0); |
1104 | if (IS_ERR(entry)) { | |
1105 | vmf_ret = dax_fault_return(PTR_ERR(entry)); | |
1106 | goto out; | |
1107 | } | |
1108 | ||
e2093926 RZ |
1109 | /* |
1110 | * It is possible, particularly with mixed reads & writes to private | |
1111 | * mappings, that we have raced with a PMD fault that overlaps with | |
1112 | * the PTE we need to set up. If so just return and the fault will be | |
1113 | * retried. | |
1114 | */ | |
1115 | if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) { | |
1116 | vmf_ret = VM_FAULT_NOPAGE; | |
1117 | goto unlock_entry; | |
1118 | } | |
1119 | ||
a7d73fe6 CH |
1120 | /* |
1121 | * Note that we don't bother to use iomap_apply here: DAX required | |
1122 | * the file system block size to be equal the page size, which means | |
1123 | * that we never have to deal with more than a single extent here. | |
1124 | */ | |
1125 | error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); | |
a9c42b33 RZ |
1126 | if (error) { |
1127 | vmf_ret = dax_fault_return(error); | |
13e451fd | 1128 | goto unlock_entry; |
a9c42b33 | 1129 | } |
a7d73fe6 | 1130 | if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { |
13e451fd JK |
1131 | error = -EIO; /* fs corruption? */ |
1132 | goto error_finish_iomap; | |
a7d73fe6 CH |
1133 | } |
1134 | ||
333ccc97 | 1135 | sector = dax_iomap_sector(&iomap, pos); |
a7d73fe6 CH |
1136 | |
1137 | if (vmf->cow_page) { | |
1138 | switch (iomap.type) { | |
1139 | case IOMAP_HOLE: | |
1140 | case IOMAP_UNWRITTEN: | |
1141 | clear_user_highpage(vmf->cow_page, vaddr); | |
1142 | break; | |
1143 | case IOMAP_MAPPED: | |
cccbce67 DW |
1144 | error = copy_user_dax(iomap.bdev, iomap.dax_dev, |
1145 | sector, PAGE_SIZE, vmf->cow_page, vaddr); | |
a7d73fe6 CH |
1146 | break; |
1147 | default: | |
1148 | WARN_ON_ONCE(1); | |
1149 | error = -EIO; | |
1150 | break; | |
1151 | } | |
1152 | ||
1153 | if (error) | |
13e451fd | 1154 | goto error_finish_iomap; |
b1aa812b JK |
1155 | |
1156 | __SetPageUptodate(vmf->cow_page); | |
1157 | vmf_ret = finish_fault(vmf); | |
1158 | if (!vmf_ret) | |
1159 | vmf_ret = VM_FAULT_DONE_COW; | |
13e451fd | 1160 | goto finish_iomap; |
a7d73fe6 CH |
1161 | } |
1162 | ||
1163 | switch (iomap.type) { | |
1164 | case IOMAP_MAPPED: | |
1165 | if (iomap.flags & IOMAP_F_NEW) { | |
1166 | count_vm_event(PGMAJFAULT); | |
2262185c | 1167 | count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT); |
a7d73fe6 CH |
1168 | major = VM_FAULT_MAJOR; |
1169 | } | |
cccbce67 | 1170 | error = dax_insert_mapping(mapping, iomap.bdev, iomap.dax_dev, |
91d25ba8 | 1171 | sector, PAGE_SIZE, entry, vmf->vma, vmf); |
9f141d6e JK |
1172 | /* -EBUSY is fine, somebody else faulted on the same PTE */ |
1173 | if (error == -EBUSY) | |
1174 | error = 0; | |
a7d73fe6 CH |
1175 | break; |
1176 | case IOMAP_UNWRITTEN: | |
1177 | case IOMAP_HOLE: | |
1550290b | 1178 | if (!(vmf->flags & FAULT_FLAG_WRITE)) { |
91d25ba8 | 1179 | vmf_ret = dax_load_hole(mapping, entry, vmf); |
13e451fd | 1180 | goto finish_iomap; |
1550290b | 1181 | } |
a7d73fe6 CH |
1182 | /*FALLTHRU*/ |
1183 | default: | |
1184 | WARN_ON_ONCE(1); | |
1185 | error = -EIO; | |
1186 | break; | |
1187 | } | |
1188 | ||
13e451fd | 1189 | error_finish_iomap: |
9f141d6e | 1190 | vmf_ret = dax_fault_return(error) | major; |
9f141d6e JK |
1191 | finish_iomap: |
1192 | if (ops->iomap_end) { | |
1193 | int copied = PAGE_SIZE; | |
1194 | ||
1195 | if (vmf_ret & VM_FAULT_ERROR) | |
1196 | copied = 0; | |
1197 | /* | |
1198 | * The fault is done by now and there's no way back (other | |
1199 | * thread may be already happily using PTE we have installed). | |
1200 | * Just ignore error from ->iomap_end since we cannot do much | |
1201 | * with it. | |
1202 | */ | |
1203 | ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap); | |
1550290b | 1204 | } |
13e451fd | 1205 | unlock_entry: |
91d25ba8 | 1206 | put_locked_mapping_entry(mapping, vmf->pgoff); |
13e451fd | 1207 | out: |
a9c42b33 | 1208 | trace_dax_pte_fault_done(inode, vmf, vmf_ret); |
9f141d6e | 1209 | return vmf_ret; |
a7d73fe6 | 1210 | } |
642261ac RZ |
1211 | |
1212 | #ifdef CONFIG_FS_DAX_PMD | |
1213 | /* | |
1214 | * The 'colour' (ie low bits) within a PMD of a page offset. This comes up | |
1215 | * more often than one might expect in the below functions. | |
1216 | */ | |
1217 | #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) | |
1218 | ||
f4200391 | 1219 | static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap, |
91d25ba8 | 1220 | loff_t pos, void *entry) |
642261ac | 1221 | { |
f4200391 | 1222 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
cccbce67 DW |
1223 | const sector_t sector = dax_iomap_sector(iomap, pos); |
1224 | struct dax_device *dax_dev = iomap->dax_dev; | |
642261ac | 1225 | struct block_device *bdev = iomap->bdev; |
27a7ffac | 1226 | struct inode *inode = mapping->host; |
cccbce67 DW |
1227 | const size_t size = PMD_SIZE; |
1228 | void *ret = NULL, *kaddr; | |
1229 | long length = 0; | |
1230 | pgoff_t pgoff; | |
1231 | pfn_t pfn; | |
1232 | int id; | |
1233 | ||
1234 | if (bdev_dax_pgoff(bdev, sector, size, &pgoff) != 0) | |
27a7ffac | 1235 | goto fallback; |
642261ac | 1236 | |
cccbce67 DW |
1237 | id = dax_read_lock(); |
1238 | length = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn); | |
1239 | if (length < 0) | |
1240 | goto unlock_fallback; | |
1241 | length = PFN_PHYS(length); | |
1242 | ||
1243 | if (length < size) | |
1244 | goto unlock_fallback; | |
1245 | if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR) | |
1246 | goto unlock_fallback; | |
1247 | if (!pfn_t_devmap(pfn)) | |
1248 | goto unlock_fallback; | |
1249 | dax_read_unlock(id); | |
1250 | ||
91d25ba8 | 1251 | ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, |
642261ac RZ |
1252 | RADIX_DAX_PMD); |
1253 | if (IS_ERR(ret)) | |
27a7ffac | 1254 | goto fallback; |
642261ac | 1255 | |
cccbce67 | 1256 | trace_dax_pmd_insert_mapping(inode, vmf, length, pfn, ret); |
f4200391 | 1257 | return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd, |
cccbce67 | 1258 | pfn, vmf->flags & FAULT_FLAG_WRITE); |
642261ac | 1259 | |
cccbce67 DW |
1260 | unlock_fallback: |
1261 | dax_read_unlock(id); | |
27a7ffac | 1262 | fallback: |
cccbce67 | 1263 | trace_dax_pmd_insert_mapping_fallback(inode, vmf, length, pfn, ret); |
642261ac RZ |
1264 | return VM_FAULT_FALLBACK; |
1265 | } | |
1266 | ||
f4200391 | 1267 | static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap, |
91d25ba8 | 1268 | void *entry) |
642261ac | 1269 | { |
f4200391 DJ |
1270 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
1271 | unsigned long pmd_addr = vmf->address & PMD_MASK; | |
653b2ea3 | 1272 | struct inode *inode = mapping->host; |
642261ac | 1273 | struct page *zero_page; |
653b2ea3 | 1274 | void *ret = NULL; |
642261ac RZ |
1275 | spinlock_t *ptl; |
1276 | pmd_t pmd_entry; | |
642261ac | 1277 | |
f4200391 | 1278 | zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); |
642261ac RZ |
1279 | |
1280 | if (unlikely(!zero_page)) | |
653b2ea3 | 1281 | goto fallback; |
642261ac | 1282 | |
91d25ba8 RZ |
1283 | ret = dax_insert_mapping_entry(mapping, vmf, entry, 0, |
1284 | RADIX_DAX_PMD | RADIX_DAX_ZERO_PAGE); | |
642261ac | 1285 | if (IS_ERR(ret)) |
653b2ea3 | 1286 | goto fallback; |
642261ac | 1287 | |
f4200391 DJ |
1288 | ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); |
1289 | if (!pmd_none(*(vmf->pmd))) { | |
642261ac | 1290 | spin_unlock(ptl); |
653b2ea3 | 1291 | goto fallback; |
642261ac RZ |
1292 | } |
1293 | ||
f4200391 | 1294 | pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); |
642261ac | 1295 | pmd_entry = pmd_mkhuge(pmd_entry); |
f4200391 | 1296 | set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); |
642261ac | 1297 | spin_unlock(ptl); |
f4200391 | 1298 | trace_dax_pmd_load_hole(inode, vmf, zero_page, ret); |
642261ac | 1299 | return VM_FAULT_NOPAGE; |
653b2ea3 RZ |
1300 | |
1301 | fallback: | |
f4200391 | 1302 | trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret); |
653b2ea3 | 1303 | return VM_FAULT_FALLBACK; |
642261ac RZ |
1304 | } |
1305 | ||
a2d58167 DJ |
1306 | static int dax_iomap_pmd_fault(struct vm_fault *vmf, |
1307 | const struct iomap_ops *ops) | |
642261ac | 1308 | { |
f4200391 | 1309 | struct vm_area_struct *vma = vmf->vma; |
642261ac | 1310 | struct address_space *mapping = vma->vm_file->f_mapping; |
d8a849e1 DJ |
1311 | unsigned long pmd_addr = vmf->address & PMD_MASK; |
1312 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
9484ab1b | 1313 | unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT; |
642261ac RZ |
1314 | struct inode *inode = mapping->host; |
1315 | int result = VM_FAULT_FALLBACK; | |
1316 | struct iomap iomap = { 0 }; | |
1317 | pgoff_t max_pgoff, pgoff; | |
642261ac RZ |
1318 | void *entry; |
1319 | loff_t pos; | |
1320 | int error; | |
1321 | ||
282a8e03 RZ |
1322 | /* |
1323 | * Check whether offset isn't beyond end of file now. Caller is | |
1324 | * supposed to hold locks serializing us with truncate / punch hole so | |
1325 | * this is a reliable test. | |
1326 | */ | |
1327 | pgoff = linear_page_index(vma, pmd_addr); | |
1328 | max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT; | |
1329 | ||
f4200391 | 1330 | trace_dax_pmd_fault(inode, vmf, max_pgoff, 0); |
282a8e03 | 1331 | |
fffa281b RZ |
1332 | /* |
1333 | * Make sure that the faulting address's PMD offset (color) matches | |
1334 | * the PMD offset from the start of the file. This is necessary so | |
1335 | * that a PMD range in the page table overlaps exactly with a PMD | |
1336 | * range in the radix tree. | |
1337 | */ | |
1338 | if ((vmf->pgoff & PG_PMD_COLOUR) != | |
1339 | ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR)) | |
1340 | goto fallback; | |
1341 | ||
642261ac RZ |
1342 | /* Fall back to PTEs if we're going to COW */ |
1343 | if (write && !(vma->vm_flags & VM_SHARED)) | |
1344 | goto fallback; | |
1345 | ||
1346 | /* If the PMD would extend outside the VMA */ | |
1347 | if (pmd_addr < vma->vm_start) | |
1348 | goto fallback; | |
1349 | if ((pmd_addr + PMD_SIZE) > vma->vm_end) | |
1350 | goto fallback; | |
1351 | ||
282a8e03 RZ |
1352 | if (pgoff > max_pgoff) { |
1353 | result = VM_FAULT_SIGBUS; | |
1354 | goto out; | |
1355 | } | |
642261ac RZ |
1356 | |
1357 | /* If the PMD would extend beyond the file size */ | |
1358 | if ((pgoff | PG_PMD_COLOUR) > max_pgoff) | |
1359 | goto fallback; | |
1360 | ||
876f2946 | 1361 | /* |
91d25ba8 RZ |
1362 | * grab_mapping_entry() will make sure we get a 2MiB empty entry, a |
1363 | * 2MiB zero page entry or a DAX PMD. If it can't (because a 4k page | |
1364 | * is already in the tree, for instance), it will return -EEXIST and | |
1365 | * we just fall back to 4k entries. | |
876f2946 RZ |
1366 | */ |
1367 | entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD); | |
1368 | if (IS_ERR(entry)) | |
1369 | goto fallback; | |
1370 | ||
e2093926 RZ |
1371 | /* |
1372 | * It is possible, particularly with mixed reads & writes to private | |
1373 | * mappings, that we have raced with a PTE fault that overlaps with | |
1374 | * the PMD we need to set up. If so just return and the fault will be | |
1375 | * retried. | |
1376 | */ | |
1377 | if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) && | |
1378 | !pmd_devmap(*vmf->pmd)) { | |
1379 | result = 0; | |
1380 | goto unlock_entry; | |
1381 | } | |
1382 | ||
642261ac RZ |
1383 | /* |
1384 | * Note that we don't use iomap_apply here. We aren't doing I/O, only | |
1385 | * setting up a mapping, so really we're using iomap_begin() as a way | |
1386 | * to look up our filesystem block. | |
1387 | */ | |
1388 | pos = (loff_t)pgoff << PAGE_SHIFT; | |
1389 | error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap); | |
1390 | if (error) | |
876f2946 | 1391 | goto unlock_entry; |
9f141d6e | 1392 | |
642261ac RZ |
1393 | if (iomap.offset + iomap.length < pos + PMD_SIZE) |
1394 | goto finish_iomap; | |
1395 | ||
642261ac RZ |
1396 | switch (iomap.type) { |
1397 | case IOMAP_MAPPED: | |
91d25ba8 | 1398 | result = dax_pmd_insert_mapping(vmf, &iomap, pos, entry); |
642261ac RZ |
1399 | break; |
1400 | case IOMAP_UNWRITTEN: | |
1401 | case IOMAP_HOLE: | |
1402 | if (WARN_ON_ONCE(write)) | |
876f2946 | 1403 | break; |
91d25ba8 | 1404 | result = dax_pmd_load_hole(vmf, &iomap, entry); |
642261ac RZ |
1405 | break; |
1406 | default: | |
1407 | WARN_ON_ONCE(1); | |
1408 | break; | |
1409 | } | |
1410 | ||
1411 | finish_iomap: | |
1412 | if (ops->iomap_end) { | |
9f141d6e JK |
1413 | int copied = PMD_SIZE; |
1414 | ||
1415 | if (result == VM_FAULT_FALLBACK) | |
1416 | copied = 0; | |
1417 | /* | |
1418 | * The fault is done by now and there's no way back (other | |
1419 | * thread may be already happily using PMD we have installed). | |
1420 | * Just ignore error from ->iomap_end since we cannot do much | |
1421 | * with it. | |
1422 | */ | |
1423 | ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags, | |
1424 | &iomap); | |
642261ac | 1425 | } |
876f2946 | 1426 | unlock_entry: |
91d25ba8 | 1427 | put_locked_mapping_entry(mapping, pgoff); |
642261ac RZ |
1428 | fallback: |
1429 | if (result == VM_FAULT_FALLBACK) { | |
d8a849e1 | 1430 | split_huge_pmd(vma, vmf->pmd, vmf->address); |
642261ac RZ |
1431 | count_vm_event(THP_FAULT_FALLBACK); |
1432 | } | |
282a8e03 | 1433 | out: |
f4200391 | 1434 | trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result); |
642261ac RZ |
1435 | return result; |
1436 | } | |
a2d58167 | 1437 | #else |
01cddfe9 AB |
1438 | static int dax_iomap_pmd_fault(struct vm_fault *vmf, |
1439 | const struct iomap_ops *ops) | |
a2d58167 DJ |
1440 | { |
1441 | return VM_FAULT_FALLBACK; | |
1442 | } | |
642261ac | 1443 | #endif /* CONFIG_FS_DAX_PMD */ |
a2d58167 DJ |
1444 | |
1445 | /** | |
1446 | * dax_iomap_fault - handle a page fault on a DAX file | |
1447 | * @vmf: The description of the fault | |
1448 | * @ops: iomap ops passed from the file system | |
1449 | * | |
1450 | * When a page fault occurs, filesystems may call this helper in | |
1451 | * their fault handler for DAX files. dax_iomap_fault() assumes the caller | |
1452 | * has done all the necessary locking for page fault to proceed | |
1453 | * successfully. | |
1454 | */ | |
c791ace1 DJ |
1455 | int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, |
1456 | const struct iomap_ops *ops) | |
a2d58167 | 1457 | { |
c791ace1 DJ |
1458 | switch (pe_size) { |
1459 | case PE_SIZE_PTE: | |
a2d58167 | 1460 | return dax_iomap_pte_fault(vmf, ops); |
c791ace1 | 1461 | case PE_SIZE_PMD: |
a2d58167 DJ |
1462 | return dax_iomap_pmd_fault(vmf, ops); |
1463 | default: | |
1464 | return VM_FAULT_FALLBACK; | |
1465 | } | |
1466 | } | |
1467 | EXPORT_SYMBOL_GPL(dax_iomap_fault); |