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 | ||
917f3452 RZ |
45 | /* The 'colour' (ie low bits) within a PMD of a page offset. */ |
46 | #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) | |
977fbdcd | 47 | #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT) |
917f3452 | 48 | |
ce95ab0f | 49 | static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; |
ac401cc7 JK |
50 | |
51 | static int __init init_dax_wait_table(void) | |
52 | { | |
53 | int i; | |
54 | ||
55 | for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) | |
56 | init_waitqueue_head(wait_table + i); | |
57 | return 0; | |
58 | } | |
59 | fs_initcall(init_dax_wait_table); | |
60 | ||
527b19d0 RZ |
61 | /* |
62 | * We use lowest available bit in exceptional entry for locking, one bit for | |
63 | * the entry size (PMD) and two more to tell us if the entry is a zero page or | |
64 | * an empty entry that is just used for locking. In total four special bits. | |
65 | * | |
66 | * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE | |
67 | * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem | |
68 | * block allocation. | |
69 | */ | |
70 | #define RADIX_DAX_SHIFT (RADIX_TREE_EXCEPTIONAL_SHIFT + 4) | |
71 | #define RADIX_DAX_ENTRY_LOCK (1 << RADIX_TREE_EXCEPTIONAL_SHIFT) | |
72 | #define RADIX_DAX_PMD (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1)) | |
73 | #define RADIX_DAX_ZERO_PAGE (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2)) | |
74 | #define RADIX_DAX_EMPTY (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 3)) | |
75 | ||
3fe0791c | 76 | static unsigned long dax_radix_pfn(void *entry) |
527b19d0 RZ |
77 | { |
78 | return (unsigned long)entry >> RADIX_DAX_SHIFT; | |
79 | } | |
80 | ||
3fe0791c | 81 | static void *dax_radix_locked_entry(unsigned long pfn, unsigned long flags) |
527b19d0 RZ |
82 | { |
83 | return (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY | flags | | |
3fe0791c | 84 | (pfn << RADIX_DAX_SHIFT) | RADIX_DAX_ENTRY_LOCK); |
527b19d0 RZ |
85 | } |
86 | ||
87 | static unsigned int dax_radix_order(void *entry) | |
88 | { | |
89 | if ((unsigned long)entry & RADIX_DAX_PMD) | |
90 | return PMD_SHIFT - PAGE_SHIFT; | |
91 | return 0; | |
92 | } | |
93 | ||
642261ac | 94 | static int dax_is_pmd_entry(void *entry) |
d1a5f2b4 | 95 | { |
642261ac | 96 | return (unsigned long)entry & RADIX_DAX_PMD; |
d1a5f2b4 DW |
97 | } |
98 | ||
642261ac | 99 | static int dax_is_pte_entry(void *entry) |
d475c634 | 100 | { |
642261ac | 101 | return !((unsigned long)entry & RADIX_DAX_PMD); |
d475c634 MW |
102 | } |
103 | ||
642261ac | 104 | static int dax_is_zero_entry(void *entry) |
d475c634 | 105 | { |
91d25ba8 | 106 | return (unsigned long)entry & RADIX_DAX_ZERO_PAGE; |
d475c634 MW |
107 | } |
108 | ||
642261ac | 109 | static int dax_is_empty_entry(void *entry) |
b2e0d162 | 110 | { |
642261ac | 111 | return (unsigned long)entry & RADIX_DAX_EMPTY; |
b2e0d162 DW |
112 | } |
113 | ||
ac401cc7 JK |
114 | /* |
115 | * DAX radix tree locking | |
116 | */ | |
117 | struct exceptional_entry_key { | |
118 | struct address_space *mapping; | |
63e95b5c | 119 | pgoff_t entry_start; |
ac401cc7 JK |
120 | }; |
121 | ||
122 | struct wait_exceptional_entry_queue { | |
ac6424b9 | 123 | wait_queue_entry_t wait; |
ac401cc7 JK |
124 | struct exceptional_entry_key key; |
125 | }; | |
126 | ||
63e95b5c RZ |
127 | static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping, |
128 | pgoff_t index, void *entry, struct exceptional_entry_key *key) | |
129 | { | |
130 | unsigned long hash; | |
131 | ||
132 | /* | |
133 | * If 'entry' is a PMD, align the 'index' that we use for the wait | |
134 | * queue to the start of that PMD. This ensures that all offsets in | |
135 | * the range covered by the PMD map to the same bit lock. | |
136 | */ | |
642261ac | 137 | if (dax_is_pmd_entry(entry)) |
917f3452 | 138 | index &= ~PG_PMD_COLOUR; |
63e95b5c RZ |
139 | |
140 | key->mapping = mapping; | |
141 | key->entry_start = index; | |
142 | ||
143 | hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS); | |
144 | return wait_table + hash; | |
145 | } | |
146 | ||
ac6424b9 | 147 | static int wake_exceptional_entry_func(wait_queue_entry_t *wait, unsigned int mode, |
ac401cc7 JK |
148 | int sync, void *keyp) |
149 | { | |
150 | struct exceptional_entry_key *key = keyp; | |
151 | struct wait_exceptional_entry_queue *ewait = | |
152 | container_of(wait, struct wait_exceptional_entry_queue, wait); | |
153 | ||
154 | if (key->mapping != ewait->key.mapping || | |
63e95b5c | 155 | key->entry_start != ewait->key.entry_start) |
ac401cc7 JK |
156 | return 0; |
157 | return autoremove_wake_function(wait, mode, sync, NULL); | |
158 | } | |
159 | ||
e30331ff RZ |
160 | /* |
161 | * We do not necessarily hold the mapping->tree_lock when we call this | |
162 | * function so it is possible that 'entry' is no longer a valid item in the | |
163 | * radix tree. This is okay because all we really need to do is to find the | |
164 | * correct waitqueue where tasks might be waiting for that old 'entry' and | |
165 | * wake them. | |
166 | */ | |
d01ad197 | 167 | static void dax_wake_mapping_entry_waiter(struct address_space *mapping, |
e30331ff RZ |
168 | pgoff_t index, void *entry, bool wake_all) |
169 | { | |
170 | struct exceptional_entry_key key; | |
171 | wait_queue_head_t *wq; | |
172 | ||
173 | wq = dax_entry_waitqueue(mapping, index, entry, &key); | |
174 | ||
175 | /* | |
176 | * Checking for locked entry and prepare_to_wait_exclusive() happens | |
177 | * under mapping->tree_lock, ditto for entry handling in our callers. | |
178 | * So at this point all tasks that could have seen our entry locked | |
179 | * must be in the waitqueue and the following check will see them. | |
180 | */ | |
181 | if (waitqueue_active(wq)) | |
182 | __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); | |
183 | } | |
184 | ||
ac401cc7 JK |
185 | /* |
186 | * Check whether the given slot is locked. The function must be called with | |
187 | * mapping->tree_lock held | |
188 | */ | |
189 | static inline int slot_locked(struct address_space *mapping, void **slot) | |
190 | { | |
191 | unsigned long entry = (unsigned long) | |
192 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
193 | return entry & RADIX_DAX_ENTRY_LOCK; | |
194 | } | |
195 | ||
196 | /* | |
197 | * Mark the given slot is locked. The function must be called with | |
198 | * mapping->tree_lock held | |
199 | */ | |
200 | static inline void *lock_slot(struct address_space *mapping, void **slot) | |
201 | { | |
202 | unsigned long entry = (unsigned long) | |
203 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
204 | ||
205 | entry |= RADIX_DAX_ENTRY_LOCK; | |
6d75f366 | 206 | radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); |
ac401cc7 JK |
207 | return (void *)entry; |
208 | } | |
209 | ||
210 | /* | |
211 | * Mark the given slot is unlocked. The function must be called with | |
212 | * mapping->tree_lock held | |
213 | */ | |
214 | static inline void *unlock_slot(struct address_space *mapping, void **slot) | |
215 | { | |
216 | unsigned long entry = (unsigned long) | |
217 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
218 | ||
219 | entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK; | |
6d75f366 | 220 | radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); |
ac401cc7 JK |
221 | return (void *)entry; |
222 | } | |
223 | ||
224 | /* | |
225 | * Lookup entry in radix tree, wait for it to become unlocked if it is | |
226 | * exceptional entry and return it. The caller must call | |
227 | * put_unlocked_mapping_entry() when he decided not to lock the entry or | |
228 | * put_locked_mapping_entry() when he locked the entry and now wants to | |
229 | * unlock it. | |
230 | * | |
231 | * The function must be called with mapping->tree_lock held. | |
232 | */ | |
233 | static void *get_unlocked_mapping_entry(struct address_space *mapping, | |
234 | pgoff_t index, void ***slotp) | |
235 | { | |
e3ad61c6 | 236 | void *entry, **slot; |
ac401cc7 | 237 | struct wait_exceptional_entry_queue ewait; |
63e95b5c | 238 | wait_queue_head_t *wq; |
ac401cc7 JK |
239 | |
240 | init_wait(&ewait.wait); | |
241 | ewait.wait.func = wake_exceptional_entry_func; | |
ac401cc7 JK |
242 | |
243 | for (;;) { | |
e3ad61c6 | 244 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, |
ac401cc7 | 245 | &slot); |
91d25ba8 RZ |
246 | if (!entry || |
247 | WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)) || | |
ac401cc7 JK |
248 | !slot_locked(mapping, slot)) { |
249 | if (slotp) | |
250 | *slotp = slot; | |
e3ad61c6 | 251 | return entry; |
ac401cc7 | 252 | } |
63e95b5c RZ |
253 | |
254 | wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key); | |
ac401cc7 JK |
255 | prepare_to_wait_exclusive(wq, &ewait.wait, |
256 | TASK_UNINTERRUPTIBLE); | |
257 | spin_unlock_irq(&mapping->tree_lock); | |
258 | schedule(); | |
259 | finish_wait(wq, &ewait.wait); | |
260 | spin_lock_irq(&mapping->tree_lock); | |
261 | } | |
262 | } | |
263 | ||
b1aa812b JK |
264 | static void dax_unlock_mapping_entry(struct address_space *mapping, |
265 | pgoff_t index) | |
266 | { | |
267 | void *entry, **slot; | |
268 | ||
269 | spin_lock_irq(&mapping->tree_lock); | |
270 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot); | |
271 | if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) || | |
272 | !slot_locked(mapping, slot))) { | |
273 | spin_unlock_irq(&mapping->tree_lock); | |
274 | return; | |
275 | } | |
276 | unlock_slot(mapping, slot); | |
277 | spin_unlock_irq(&mapping->tree_lock); | |
278 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); | |
279 | } | |
280 | ||
422476c4 | 281 | static void put_locked_mapping_entry(struct address_space *mapping, |
91d25ba8 | 282 | pgoff_t index) |
422476c4 | 283 | { |
91d25ba8 | 284 | dax_unlock_mapping_entry(mapping, index); |
422476c4 RZ |
285 | } |
286 | ||
287 | /* | |
288 | * Called when we are done with radix tree entry we looked up via | |
289 | * get_unlocked_mapping_entry() and which we didn't lock in the end. | |
290 | */ | |
291 | static void put_unlocked_mapping_entry(struct address_space *mapping, | |
292 | pgoff_t index, void *entry) | |
293 | { | |
91d25ba8 | 294 | if (!entry) |
422476c4 RZ |
295 | return; |
296 | ||
297 | /* We have to wake up next waiter for the radix tree entry lock */ | |
298 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); | |
299 | } | |
300 | ||
d2c997c0 DW |
301 | static unsigned long dax_entry_size(void *entry) |
302 | { | |
303 | if (dax_is_zero_entry(entry)) | |
304 | return 0; | |
305 | else if (dax_is_empty_entry(entry)) | |
306 | return 0; | |
307 | else if (dax_is_pmd_entry(entry)) | |
308 | return PMD_SIZE; | |
309 | else | |
310 | return PAGE_SIZE; | |
311 | } | |
312 | ||
313 | static unsigned long dax_radix_end_pfn(void *entry) | |
314 | { | |
315 | return dax_radix_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE; | |
316 | } | |
317 | ||
318 | /* | |
319 | * Iterate through all mapped pfns represented by an entry, i.e. skip | |
320 | * 'empty' and 'zero' entries. | |
321 | */ | |
322 | #define for_each_mapped_pfn(entry, pfn) \ | |
323 | for (pfn = dax_radix_pfn(entry); \ | |
324 | pfn < dax_radix_end_pfn(entry); pfn++) | |
325 | ||
326 | static void dax_associate_entry(void *entry, struct address_space *mapping) | |
327 | { | |
328 | unsigned long pfn; | |
329 | ||
330 | if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) | |
331 | return; | |
332 | ||
333 | for_each_mapped_pfn(entry, pfn) { | |
334 | struct page *page = pfn_to_page(pfn); | |
335 | ||
336 | WARN_ON_ONCE(page->mapping); | |
337 | page->mapping = mapping; | |
338 | } | |
339 | } | |
340 | ||
341 | static void dax_disassociate_entry(void *entry, struct address_space *mapping, | |
342 | bool trunc) | |
343 | { | |
344 | unsigned long pfn; | |
345 | ||
346 | if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) | |
347 | return; | |
348 | ||
349 | for_each_mapped_pfn(entry, pfn) { | |
350 | struct page *page = pfn_to_page(pfn); | |
351 | ||
352 | WARN_ON_ONCE(trunc && page_ref_count(page) > 1); | |
353 | WARN_ON_ONCE(page->mapping && page->mapping != mapping); | |
354 | page->mapping = NULL; | |
355 | } | |
356 | } | |
357 | ||
ac401cc7 | 358 | /* |
91d25ba8 RZ |
359 | * Find radix tree entry at given index. If it points to an exceptional entry, |
360 | * return it with the radix tree entry locked. If the radix tree doesn't | |
361 | * contain given index, create an empty exceptional entry for the index and | |
362 | * return with it locked. | |
ac401cc7 | 363 | * |
642261ac RZ |
364 | * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will |
365 | * either return that locked entry or will return an error. This error will | |
91d25ba8 RZ |
366 | * happen if there are any 4k entries within the 2MiB range that we are |
367 | * requesting. | |
642261ac RZ |
368 | * |
369 | * We always favor 4k entries over 2MiB entries. There isn't a flow where we | |
370 | * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB | |
371 | * insertion will fail if it finds any 4k entries already in the tree, and a | |
372 | * 4k insertion will cause an existing 2MiB entry to be unmapped and | |
373 | * downgraded to 4k entries. This happens for both 2MiB huge zero pages as | |
374 | * well as 2MiB empty entries. | |
375 | * | |
376 | * The exception to this downgrade path is for 2MiB DAX PMD entries that have | |
377 | * real storage backing them. We will leave these real 2MiB DAX entries in | |
378 | * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry. | |
379 | * | |
ac401cc7 JK |
380 | * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For |
381 | * persistent memory the benefit is doubtful. We can add that later if we can | |
382 | * show it helps. | |
383 | */ | |
642261ac RZ |
384 | static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index, |
385 | unsigned long size_flag) | |
ac401cc7 | 386 | { |
642261ac | 387 | bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */ |
e3ad61c6 | 388 | void *entry, **slot; |
ac401cc7 JK |
389 | |
390 | restart: | |
391 | spin_lock_irq(&mapping->tree_lock); | |
e3ad61c6 | 392 | entry = get_unlocked_mapping_entry(mapping, index, &slot); |
642261ac | 393 | |
91d25ba8 RZ |
394 | if (WARN_ON_ONCE(entry && !radix_tree_exceptional_entry(entry))) { |
395 | entry = ERR_PTR(-EIO); | |
396 | goto out_unlock; | |
397 | } | |
398 | ||
642261ac RZ |
399 | if (entry) { |
400 | if (size_flag & RADIX_DAX_PMD) { | |
91d25ba8 | 401 | if (dax_is_pte_entry(entry)) { |
642261ac RZ |
402 | put_unlocked_mapping_entry(mapping, index, |
403 | entry); | |
404 | entry = ERR_PTR(-EEXIST); | |
405 | goto out_unlock; | |
406 | } | |
407 | } else { /* trying to grab a PTE entry */ | |
91d25ba8 | 408 | if (dax_is_pmd_entry(entry) && |
642261ac RZ |
409 | (dax_is_zero_entry(entry) || |
410 | dax_is_empty_entry(entry))) { | |
411 | pmd_downgrade = true; | |
412 | } | |
413 | } | |
414 | } | |
415 | ||
ac401cc7 | 416 | /* No entry for given index? Make sure radix tree is big enough. */ |
642261ac | 417 | if (!entry || pmd_downgrade) { |
ac401cc7 JK |
418 | int err; |
419 | ||
642261ac RZ |
420 | if (pmd_downgrade) { |
421 | /* | |
422 | * Make sure 'entry' remains valid while we drop | |
423 | * mapping->tree_lock. | |
424 | */ | |
425 | entry = lock_slot(mapping, slot); | |
426 | } | |
427 | ||
ac401cc7 | 428 | spin_unlock_irq(&mapping->tree_lock); |
642261ac RZ |
429 | /* |
430 | * Besides huge zero pages the only other thing that gets | |
431 | * downgraded are empty entries which don't need to be | |
432 | * unmapped. | |
433 | */ | |
434 | if (pmd_downgrade && dax_is_zero_entry(entry)) | |
977fbdcd MW |
435 | unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR, |
436 | PG_PMD_NR, false); | |
642261ac | 437 | |
ac401cc7 JK |
438 | err = radix_tree_preload( |
439 | mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM); | |
0cb80b48 JK |
440 | if (err) { |
441 | if (pmd_downgrade) | |
91d25ba8 | 442 | put_locked_mapping_entry(mapping, index); |
ac401cc7 | 443 | return ERR_PTR(err); |
0cb80b48 | 444 | } |
ac401cc7 | 445 | spin_lock_irq(&mapping->tree_lock); |
642261ac | 446 | |
e11f8b7b RZ |
447 | if (!entry) { |
448 | /* | |
449 | * We needed to drop the page_tree lock while calling | |
450 | * radix_tree_preload() and we didn't have an entry to | |
451 | * lock. See if another thread inserted an entry at | |
452 | * our index during this time. | |
453 | */ | |
454 | entry = __radix_tree_lookup(&mapping->page_tree, index, | |
455 | NULL, &slot); | |
456 | if (entry) { | |
457 | radix_tree_preload_end(); | |
458 | spin_unlock_irq(&mapping->tree_lock); | |
459 | goto restart; | |
460 | } | |
461 | } | |
462 | ||
642261ac | 463 | if (pmd_downgrade) { |
d2c997c0 | 464 | dax_disassociate_entry(entry, mapping, false); |
642261ac RZ |
465 | radix_tree_delete(&mapping->page_tree, index); |
466 | mapping->nrexceptional--; | |
467 | dax_wake_mapping_entry_waiter(mapping, index, entry, | |
468 | true); | |
469 | } | |
470 | ||
471 | entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY); | |
472 | ||
473 | err = __radix_tree_insert(&mapping->page_tree, index, | |
474 | dax_radix_order(entry), entry); | |
ac401cc7 JK |
475 | radix_tree_preload_end(); |
476 | if (err) { | |
477 | spin_unlock_irq(&mapping->tree_lock); | |
642261ac | 478 | /* |
e11f8b7b RZ |
479 | * Our insertion of a DAX entry failed, most likely |
480 | * because we were inserting a PMD entry and it | |
481 | * collided with a PTE sized entry at a different | |
482 | * index in the PMD range. We haven't inserted | |
483 | * anything into the radix tree and have no waiters to | |
484 | * wake. | |
642261ac | 485 | */ |
ac401cc7 JK |
486 | return ERR_PTR(err); |
487 | } | |
488 | /* Good, we have inserted empty locked entry into the tree. */ | |
489 | mapping->nrexceptional++; | |
490 | spin_unlock_irq(&mapping->tree_lock); | |
e3ad61c6 | 491 | return entry; |
ac401cc7 | 492 | } |
e3ad61c6 | 493 | entry = lock_slot(mapping, slot); |
642261ac | 494 | out_unlock: |
ac401cc7 | 495 | spin_unlock_irq(&mapping->tree_lock); |
e3ad61c6 | 496 | return entry; |
ac401cc7 JK |
497 | } |
498 | ||
c6dcf52c JK |
499 | static int __dax_invalidate_mapping_entry(struct address_space *mapping, |
500 | pgoff_t index, bool trunc) | |
501 | { | |
502 | int ret = 0; | |
503 | void *entry; | |
504 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
505 | ||
506 | spin_lock_irq(&mapping->tree_lock); | |
507 | entry = get_unlocked_mapping_entry(mapping, index, NULL); | |
91d25ba8 | 508 | if (!entry || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry))) |
c6dcf52c JK |
509 | goto out; |
510 | if (!trunc && | |
511 | (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) || | |
512 | radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))) | |
513 | goto out; | |
d2c997c0 | 514 | dax_disassociate_entry(entry, mapping, trunc); |
c6dcf52c JK |
515 | radix_tree_delete(page_tree, index); |
516 | mapping->nrexceptional--; | |
517 | ret = 1; | |
518 | out: | |
519 | put_unlocked_mapping_entry(mapping, index, entry); | |
520 | spin_unlock_irq(&mapping->tree_lock); | |
521 | return ret; | |
522 | } | |
ac401cc7 JK |
523 | /* |
524 | * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree | |
525 | * entry to get unlocked before deleting it. | |
526 | */ | |
527 | int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) | |
528 | { | |
c6dcf52c | 529 | int ret = __dax_invalidate_mapping_entry(mapping, index, true); |
ac401cc7 | 530 | |
ac401cc7 JK |
531 | /* |
532 | * This gets called from truncate / punch_hole path. As such, the caller | |
533 | * must hold locks protecting against concurrent modifications of the | |
534 | * radix tree (usually fs-private i_mmap_sem for writing). Since the | |
535 | * caller has seen exceptional entry for this index, we better find it | |
536 | * at that index as well... | |
537 | */ | |
c6dcf52c JK |
538 | WARN_ON_ONCE(!ret); |
539 | return ret; | |
540 | } | |
541 | ||
c6dcf52c JK |
542 | /* |
543 | * Invalidate exceptional DAX entry if it is clean. | |
544 | */ | |
545 | int dax_invalidate_mapping_entry_sync(struct address_space *mapping, | |
546 | pgoff_t index) | |
547 | { | |
548 | return __dax_invalidate_mapping_entry(mapping, index, false); | |
ac401cc7 JK |
549 | } |
550 | ||
cccbce67 DW |
551 | static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev, |
552 | sector_t sector, size_t size, struct page *to, | |
553 | unsigned long vaddr) | |
f7ca90b1 | 554 | { |
cccbce67 DW |
555 | void *vto, *kaddr; |
556 | pgoff_t pgoff; | |
557 | pfn_t pfn; | |
558 | long rc; | |
559 | int id; | |
560 | ||
561 | rc = bdev_dax_pgoff(bdev, sector, size, &pgoff); | |
562 | if (rc) | |
563 | return rc; | |
564 | ||
565 | id = dax_read_lock(); | |
566 | rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn); | |
567 | if (rc < 0) { | |
568 | dax_read_unlock(id); | |
569 | return rc; | |
570 | } | |
f7ca90b1 | 571 | vto = kmap_atomic(to); |
cccbce67 | 572 | copy_user_page(vto, (void __force *)kaddr, vaddr, to); |
f7ca90b1 | 573 | kunmap_atomic(vto); |
cccbce67 | 574 | dax_read_unlock(id); |
f7ca90b1 MW |
575 | return 0; |
576 | } | |
577 | ||
642261ac RZ |
578 | /* |
579 | * By this point grab_mapping_entry() has ensured that we have a locked entry | |
580 | * of the appropriate size so we don't have to worry about downgrading PMDs to | |
581 | * PTEs. If we happen to be trying to insert a PTE and there is a PMD | |
582 | * already in the tree, we will skip the insertion and just dirty the PMD as | |
583 | * appropriate. | |
584 | */ | |
ac401cc7 JK |
585 | static void *dax_insert_mapping_entry(struct address_space *mapping, |
586 | struct vm_fault *vmf, | |
3fe0791c | 587 | void *entry, pfn_t pfn_t, |
f5b7b748 | 588 | unsigned long flags, bool dirty) |
9973c98e RZ |
589 | { |
590 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
3fe0791c | 591 | unsigned long pfn = pfn_t_to_pfn(pfn_t); |
ac401cc7 | 592 | pgoff_t index = vmf->pgoff; |
3fe0791c | 593 | void *new_entry; |
9973c98e | 594 | |
f5b7b748 | 595 | if (dirty) |
d2b2a28e | 596 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); |
9973c98e | 597 | |
91d25ba8 RZ |
598 | if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_ZERO_PAGE)) { |
599 | /* we are replacing a zero page with block mapping */ | |
600 | if (dax_is_pmd_entry(entry)) | |
977fbdcd MW |
601 | unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR, |
602 | PG_PMD_NR, false); | |
91d25ba8 | 603 | else /* pte entry */ |
977fbdcd | 604 | unmap_mapping_pages(mapping, vmf->pgoff, 1, false); |
9973c98e RZ |
605 | } |
606 | ||
ac401cc7 | 607 | spin_lock_irq(&mapping->tree_lock); |
3fe0791c | 608 | new_entry = dax_radix_locked_entry(pfn, flags); |
d2c997c0 DW |
609 | if (dax_entry_size(entry) != dax_entry_size(new_entry)) { |
610 | dax_disassociate_entry(entry, mapping, false); | |
611 | dax_associate_entry(new_entry, mapping); | |
612 | } | |
642261ac | 613 | |
91d25ba8 | 614 | if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { |
642261ac RZ |
615 | /* |
616 | * Only swap our new entry into the radix tree if the current | |
617 | * entry is a zero page or an empty entry. If a normal PTE or | |
618 | * PMD entry is already in the tree, we leave it alone. This | |
619 | * means that if we are trying to insert a PTE and the | |
620 | * existing entry is a PMD, we will just leave the PMD in the | |
621 | * tree and dirty it if necessary. | |
622 | */ | |
f7942430 | 623 | struct radix_tree_node *node; |
ac401cc7 JK |
624 | void **slot; |
625 | void *ret; | |
9973c98e | 626 | |
f7942430 | 627 | ret = __radix_tree_lookup(page_tree, index, &node, &slot); |
ac401cc7 | 628 | WARN_ON_ONCE(ret != entry); |
4d693d08 | 629 | __radix_tree_replace(page_tree, node, slot, |
c7df8ad2 | 630 | new_entry, NULL); |
91d25ba8 | 631 | entry = new_entry; |
9973c98e | 632 | } |
91d25ba8 | 633 | |
f5b7b748 | 634 | if (dirty) |
9973c98e | 635 | radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY); |
91d25ba8 | 636 | |
9973c98e | 637 | spin_unlock_irq(&mapping->tree_lock); |
91d25ba8 | 638 | return entry; |
9973c98e RZ |
639 | } |
640 | ||
4b4bb46d JK |
641 | static inline unsigned long |
642 | pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma) | |
643 | { | |
644 | unsigned long address; | |
645 | ||
646 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
647 | VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); | |
648 | return address; | |
649 | } | |
650 | ||
651 | /* Walk all mappings of a given index of a file and writeprotect them */ | |
652 | static void dax_mapping_entry_mkclean(struct address_space *mapping, | |
653 | pgoff_t index, unsigned long pfn) | |
654 | { | |
655 | struct vm_area_struct *vma; | |
f729c8c9 RZ |
656 | pte_t pte, *ptep = NULL; |
657 | pmd_t *pmdp = NULL; | |
4b4bb46d | 658 | spinlock_t *ptl; |
4b4bb46d JK |
659 | |
660 | i_mmap_lock_read(mapping); | |
661 | vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) { | |
a4d1a885 | 662 | unsigned long address, start, end; |
4b4bb46d JK |
663 | |
664 | cond_resched(); | |
665 | ||
666 | if (!(vma->vm_flags & VM_SHARED)) | |
667 | continue; | |
668 | ||
669 | address = pgoff_address(index, vma); | |
a4d1a885 JG |
670 | |
671 | /* | |
672 | * Note because we provide start/end to follow_pte_pmd it will | |
673 | * call mmu_notifier_invalidate_range_start() on our behalf | |
674 | * before taking any lock. | |
675 | */ | |
676 | if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl)) | |
4b4bb46d | 677 | continue; |
4b4bb46d | 678 | |
0f10851e JG |
679 | /* |
680 | * No need to call mmu_notifier_invalidate_range() as we are | |
681 | * downgrading page table protection not changing it to point | |
682 | * to a new page. | |
683 | * | |
684 | * See Documentation/vm/mmu_notifier.txt | |
685 | */ | |
f729c8c9 RZ |
686 | if (pmdp) { |
687 | #ifdef CONFIG_FS_DAX_PMD | |
688 | pmd_t pmd; | |
689 | ||
690 | if (pfn != pmd_pfn(*pmdp)) | |
691 | goto unlock_pmd; | |
f6f37321 | 692 | if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp)) |
f729c8c9 RZ |
693 | goto unlock_pmd; |
694 | ||
695 | flush_cache_page(vma, address, pfn); | |
696 | pmd = pmdp_huge_clear_flush(vma, address, pmdp); | |
697 | pmd = pmd_wrprotect(pmd); | |
698 | pmd = pmd_mkclean(pmd); | |
699 | set_pmd_at(vma->vm_mm, address, pmdp, pmd); | |
f729c8c9 | 700 | unlock_pmd: |
f729c8c9 | 701 | #endif |
ee190ca6 | 702 | spin_unlock(ptl); |
f729c8c9 RZ |
703 | } else { |
704 | if (pfn != pte_pfn(*ptep)) | |
705 | goto unlock_pte; | |
706 | if (!pte_dirty(*ptep) && !pte_write(*ptep)) | |
707 | goto unlock_pte; | |
708 | ||
709 | flush_cache_page(vma, address, pfn); | |
710 | pte = ptep_clear_flush(vma, address, ptep); | |
711 | pte = pte_wrprotect(pte); | |
712 | pte = pte_mkclean(pte); | |
713 | set_pte_at(vma->vm_mm, address, ptep, pte); | |
f729c8c9 RZ |
714 | unlock_pte: |
715 | pte_unmap_unlock(ptep, ptl); | |
716 | } | |
4b4bb46d | 717 | |
a4d1a885 | 718 | mmu_notifier_invalidate_range_end(vma->vm_mm, start, end); |
4b4bb46d JK |
719 | } |
720 | i_mmap_unlock_read(mapping); | |
721 | } | |
722 | ||
3fe0791c DW |
723 | static int dax_writeback_one(struct dax_device *dax_dev, |
724 | struct address_space *mapping, pgoff_t index, void *entry) | |
9973c98e RZ |
725 | { |
726 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
3fe0791c DW |
727 | void *entry2, **slot; |
728 | unsigned long pfn; | |
729 | long ret = 0; | |
cccbce67 | 730 | size_t size; |
9973c98e | 731 | |
9973c98e | 732 | /* |
a6abc2c0 JK |
733 | * A page got tagged dirty in DAX mapping? Something is seriously |
734 | * wrong. | |
9973c98e | 735 | */ |
a6abc2c0 JK |
736 | if (WARN_ON(!radix_tree_exceptional_entry(entry))) |
737 | return -EIO; | |
9973c98e | 738 | |
a6abc2c0 JK |
739 | spin_lock_irq(&mapping->tree_lock); |
740 | entry2 = get_unlocked_mapping_entry(mapping, index, &slot); | |
741 | /* Entry got punched out / reallocated? */ | |
91d25ba8 | 742 | if (!entry2 || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry2))) |
a6abc2c0 JK |
743 | goto put_unlocked; |
744 | /* | |
745 | * Entry got reallocated elsewhere? No need to writeback. We have to | |
3fe0791c | 746 | * compare pfns as we must not bail out due to difference in lockbit |
a6abc2c0 JK |
747 | * or entry type. |
748 | */ | |
3fe0791c | 749 | if (dax_radix_pfn(entry2) != dax_radix_pfn(entry)) |
a6abc2c0 | 750 | goto put_unlocked; |
642261ac RZ |
751 | if (WARN_ON_ONCE(dax_is_empty_entry(entry) || |
752 | dax_is_zero_entry(entry))) { | |
9973c98e | 753 | ret = -EIO; |
a6abc2c0 | 754 | goto put_unlocked; |
9973c98e RZ |
755 | } |
756 | ||
a6abc2c0 JK |
757 | /* Another fsync thread may have already written back this entry */ |
758 | if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)) | |
759 | goto put_unlocked; | |
760 | /* Lock the entry to serialize with page faults */ | |
761 | entry = lock_slot(mapping, slot); | |
762 | /* | |
763 | * We can clear the tag now but we have to be careful so that concurrent | |
764 | * dax_writeback_one() calls for the same index cannot finish before we | |
765 | * actually flush the caches. This is achieved as the calls will look | |
766 | * at the entry only under tree_lock and once they do that they will | |
767 | * see the entry locked and wait for it to unlock. | |
768 | */ | |
769 | radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE); | |
770 | spin_unlock_irq(&mapping->tree_lock); | |
771 | ||
642261ac RZ |
772 | /* |
773 | * Even if dax_writeback_mapping_range() was given a wbc->range_start | |
774 | * in the middle of a PMD, the 'index' we are given will be aligned to | |
3fe0791c DW |
775 | * the start index of the PMD, as will the pfn we pull from 'entry'. |
776 | * This allows us to flush for PMD_SIZE and not have to worry about | |
777 | * partial PMD writebacks. | |
642261ac | 778 | */ |
3fe0791c | 779 | pfn = dax_radix_pfn(entry); |
cccbce67 DW |
780 | size = PAGE_SIZE << dax_radix_order(entry); |
781 | ||
3fe0791c DW |
782 | dax_mapping_entry_mkclean(mapping, index, pfn); |
783 | dax_flush(dax_dev, page_address(pfn_to_page(pfn)), size); | |
4b4bb46d JK |
784 | /* |
785 | * After we have flushed the cache, we can clear the dirty tag. There | |
786 | * cannot be new dirty data in the pfn after the flush has completed as | |
787 | * the pfn mappings are writeprotected and fault waits for mapping | |
788 | * entry lock. | |
789 | */ | |
790 | spin_lock_irq(&mapping->tree_lock); | |
791 | radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY); | |
792 | spin_unlock_irq(&mapping->tree_lock); | |
f9bc3a07 | 793 | trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT); |
91d25ba8 | 794 | put_locked_mapping_entry(mapping, index); |
9973c98e RZ |
795 | return ret; |
796 | ||
a6abc2c0 JK |
797 | put_unlocked: |
798 | put_unlocked_mapping_entry(mapping, index, entry2); | |
9973c98e RZ |
799 | spin_unlock_irq(&mapping->tree_lock); |
800 | return ret; | |
801 | } | |
802 | ||
803 | /* | |
804 | * Flush the mapping to the persistent domain within the byte range of [start, | |
805 | * end]. This is required by data integrity operations to ensure file data is | |
806 | * on persistent storage prior to completion of the operation. | |
807 | */ | |
7f6d5b52 RZ |
808 | int dax_writeback_mapping_range(struct address_space *mapping, |
809 | struct block_device *bdev, struct writeback_control *wbc) | |
9973c98e RZ |
810 | { |
811 | struct inode *inode = mapping->host; | |
642261ac | 812 | pgoff_t start_index, end_index; |
9973c98e | 813 | pgoff_t indices[PAGEVEC_SIZE]; |
cccbce67 | 814 | struct dax_device *dax_dev; |
9973c98e RZ |
815 | struct pagevec pvec; |
816 | bool done = false; | |
817 | int i, ret = 0; | |
9973c98e RZ |
818 | |
819 | if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) | |
820 | return -EIO; | |
821 | ||
7f6d5b52 RZ |
822 | if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) |
823 | return 0; | |
824 | ||
cccbce67 DW |
825 | dax_dev = dax_get_by_host(bdev->bd_disk->disk_name); |
826 | if (!dax_dev) | |
827 | return -EIO; | |
828 | ||
09cbfeaf KS |
829 | start_index = wbc->range_start >> PAGE_SHIFT; |
830 | end_index = wbc->range_end >> PAGE_SHIFT; | |
9973c98e | 831 | |
d14a3f48 RZ |
832 | trace_dax_writeback_range(inode, start_index, end_index); |
833 | ||
9973c98e RZ |
834 | tag_pages_for_writeback(mapping, start_index, end_index); |
835 | ||
86679820 | 836 | pagevec_init(&pvec); |
9973c98e RZ |
837 | while (!done) { |
838 | pvec.nr = find_get_entries_tag(mapping, start_index, | |
839 | PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE, | |
840 | pvec.pages, indices); | |
841 | ||
842 | if (pvec.nr == 0) | |
843 | break; | |
844 | ||
845 | for (i = 0; i < pvec.nr; i++) { | |
846 | if (indices[i] > end_index) { | |
847 | done = true; | |
848 | break; | |
849 | } | |
850 | ||
3fe0791c DW |
851 | ret = dax_writeback_one(dax_dev, mapping, indices[i], |
852 | pvec.pages[i]); | |
819ec6b9 JL |
853 | if (ret < 0) { |
854 | mapping_set_error(mapping, ret); | |
d14a3f48 | 855 | goto out; |
819ec6b9 | 856 | } |
9973c98e | 857 | } |
1eb643d0 | 858 | start_index = indices[pvec.nr - 1] + 1; |
9973c98e | 859 | } |
d14a3f48 | 860 | out: |
cccbce67 | 861 | put_dax(dax_dev); |
d14a3f48 RZ |
862 | trace_dax_writeback_range_done(inode, start_index, end_index); |
863 | return (ret < 0 ? ret : 0); | |
9973c98e RZ |
864 | } |
865 | EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); | |
866 | ||
31a6f1a6 | 867 | static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) |
f7ca90b1 | 868 | { |
a3841f94 | 869 | return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9; |
31a6f1a6 JK |
870 | } |
871 | ||
5e161e40 JK |
872 | static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size, |
873 | pfn_t *pfnp) | |
f7ca90b1 | 874 | { |
31a6f1a6 | 875 | const sector_t sector = dax_iomap_sector(iomap, pos); |
cccbce67 | 876 | pgoff_t pgoff; |
5e161e40 | 877 | void *kaddr; |
cccbce67 | 878 | int id, rc; |
5e161e40 | 879 | long length; |
f7ca90b1 | 880 | |
5e161e40 | 881 | rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff); |
cccbce67 DW |
882 | if (rc) |
883 | return rc; | |
cccbce67 | 884 | id = dax_read_lock(); |
5e161e40 JK |
885 | length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size), |
886 | &kaddr, pfnp); | |
887 | if (length < 0) { | |
888 | rc = length; | |
889 | goto out; | |
cccbce67 | 890 | } |
5e161e40 JK |
891 | rc = -EINVAL; |
892 | if (PFN_PHYS(length) < size) | |
893 | goto out; | |
894 | if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1)) | |
895 | goto out; | |
896 | /* For larger pages we need devmap */ | |
897 | if (length > 1 && !pfn_t_devmap(*pfnp)) | |
898 | goto out; | |
899 | rc = 0; | |
900 | out: | |
cccbce67 | 901 | dax_read_unlock(id); |
5e161e40 | 902 | return rc; |
0e3b210c | 903 | } |
0e3b210c | 904 | |
e30331ff | 905 | /* |
91d25ba8 RZ |
906 | * The user has performed a load from a hole in the file. Allocating a new |
907 | * page in the file would cause excessive storage usage for workloads with | |
908 | * sparse files. Instead we insert a read-only mapping of the 4k zero page. | |
909 | * If this page is ever written to we will re-fault and change the mapping to | |
910 | * point to real DAX storage instead. | |
e30331ff | 911 | */ |
91d25ba8 | 912 | static int dax_load_hole(struct address_space *mapping, void *entry, |
e30331ff RZ |
913 | struct vm_fault *vmf) |
914 | { | |
915 | struct inode *inode = mapping->host; | |
91d25ba8 RZ |
916 | unsigned long vaddr = vmf->address; |
917 | int ret = VM_FAULT_NOPAGE; | |
918 | struct page *zero_page; | |
919 | void *entry2; | |
3fe0791c | 920 | pfn_t pfn; |
e30331ff | 921 | |
91d25ba8 RZ |
922 | zero_page = ZERO_PAGE(0); |
923 | if (unlikely(!zero_page)) { | |
e30331ff RZ |
924 | ret = VM_FAULT_OOM; |
925 | goto out; | |
926 | } | |
927 | ||
3fe0791c DW |
928 | pfn = page_to_pfn_t(zero_page); |
929 | entry2 = dax_insert_mapping_entry(mapping, vmf, entry, pfn, | |
f5b7b748 | 930 | RADIX_DAX_ZERO_PAGE, false); |
91d25ba8 RZ |
931 | if (IS_ERR(entry2)) { |
932 | ret = VM_FAULT_SIGBUS; | |
933 | goto out; | |
e30331ff | 934 | } |
91d25ba8 | 935 | |
3fe0791c | 936 | vm_insert_mixed(vmf->vma, vaddr, pfn); |
e30331ff RZ |
937 | out: |
938 | trace_dax_load_hole(inode, vmf, ret); | |
939 | return ret; | |
940 | } | |
941 | ||
4b0228fa VV |
942 | static bool dax_range_is_aligned(struct block_device *bdev, |
943 | unsigned int offset, unsigned int length) | |
944 | { | |
945 | unsigned short sector_size = bdev_logical_block_size(bdev); | |
946 | ||
947 | if (!IS_ALIGNED(offset, sector_size)) | |
948 | return false; | |
949 | if (!IS_ALIGNED(length, sector_size)) | |
950 | return false; | |
951 | ||
952 | return true; | |
953 | } | |
954 | ||
cccbce67 DW |
955 | int __dax_zero_page_range(struct block_device *bdev, |
956 | struct dax_device *dax_dev, sector_t sector, | |
957 | unsigned int offset, unsigned int size) | |
679c8bd3 | 958 | { |
cccbce67 DW |
959 | if (dax_range_is_aligned(bdev, offset, size)) { |
960 | sector_t start_sector = sector + (offset >> 9); | |
4b0228fa VV |
961 | |
962 | return blkdev_issue_zeroout(bdev, start_sector, | |
53ef7d0e | 963 | size >> 9, GFP_NOFS, 0); |
4b0228fa | 964 | } else { |
cccbce67 DW |
965 | pgoff_t pgoff; |
966 | long rc, id; | |
967 | void *kaddr; | |
968 | pfn_t pfn; | |
969 | ||
e84b83b9 | 970 | rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff); |
cccbce67 DW |
971 | if (rc) |
972 | return rc; | |
973 | ||
974 | id = dax_read_lock(); | |
e84b83b9 | 975 | rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, |
cccbce67 DW |
976 | &pfn); |
977 | if (rc < 0) { | |
978 | dax_read_unlock(id); | |
979 | return rc; | |
980 | } | |
81f55870 | 981 | memset(kaddr + offset, 0, size); |
c3ca015f | 982 | dax_flush(dax_dev, kaddr + offset, size); |
cccbce67 | 983 | dax_read_unlock(id); |
4b0228fa | 984 | } |
679c8bd3 CH |
985 | return 0; |
986 | } | |
987 | EXPORT_SYMBOL_GPL(__dax_zero_page_range); | |
988 | ||
a254e568 | 989 | static loff_t |
11c59c92 | 990 | dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, |
a254e568 CH |
991 | struct iomap *iomap) |
992 | { | |
cccbce67 DW |
993 | struct block_device *bdev = iomap->bdev; |
994 | struct dax_device *dax_dev = iomap->dax_dev; | |
a254e568 CH |
995 | struct iov_iter *iter = data; |
996 | loff_t end = pos + length, done = 0; | |
997 | ssize_t ret = 0; | |
cccbce67 | 998 | int id; |
a254e568 CH |
999 | |
1000 | if (iov_iter_rw(iter) == READ) { | |
1001 | end = min(end, i_size_read(inode)); | |
1002 | if (pos >= end) | |
1003 | return 0; | |
1004 | ||
1005 | if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) | |
1006 | return iov_iter_zero(min(length, end - pos), iter); | |
1007 | } | |
1008 | ||
1009 | if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) | |
1010 | return -EIO; | |
1011 | ||
e3fce68c JK |
1012 | /* |
1013 | * Write can allocate block for an area which has a hole page mapped | |
1014 | * into page tables. We have to tear down these mappings so that data | |
1015 | * written by write(2) is visible in mmap. | |
1016 | */ | |
cd656375 | 1017 | if (iomap->flags & IOMAP_F_NEW) { |
e3fce68c JK |
1018 | invalidate_inode_pages2_range(inode->i_mapping, |
1019 | pos >> PAGE_SHIFT, | |
1020 | (end - 1) >> PAGE_SHIFT); | |
1021 | } | |
1022 | ||
cccbce67 | 1023 | id = dax_read_lock(); |
a254e568 CH |
1024 | while (pos < end) { |
1025 | unsigned offset = pos & (PAGE_SIZE - 1); | |
cccbce67 DW |
1026 | const size_t size = ALIGN(length + offset, PAGE_SIZE); |
1027 | const sector_t sector = dax_iomap_sector(iomap, pos); | |
a254e568 | 1028 | ssize_t map_len; |
cccbce67 DW |
1029 | pgoff_t pgoff; |
1030 | void *kaddr; | |
1031 | pfn_t pfn; | |
a254e568 | 1032 | |
d1908f52 MH |
1033 | if (fatal_signal_pending(current)) { |
1034 | ret = -EINTR; | |
1035 | break; | |
1036 | } | |
1037 | ||
cccbce67 DW |
1038 | ret = bdev_dax_pgoff(bdev, sector, size, &pgoff); |
1039 | if (ret) | |
1040 | break; | |
1041 | ||
1042 | map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), | |
1043 | &kaddr, &pfn); | |
a254e568 CH |
1044 | if (map_len < 0) { |
1045 | ret = map_len; | |
1046 | break; | |
1047 | } | |
1048 | ||
cccbce67 DW |
1049 | map_len = PFN_PHYS(map_len); |
1050 | kaddr += offset; | |
a254e568 CH |
1051 | map_len -= offset; |
1052 | if (map_len > end - pos) | |
1053 | map_len = end - pos; | |
1054 | ||
a2e050f5 RZ |
1055 | /* |
1056 | * The userspace address for the memory copy has already been | |
1057 | * validated via access_ok() in either vfs_read() or | |
1058 | * vfs_write(), depending on which operation we are doing. | |
1059 | */ | |
a254e568 | 1060 | if (iov_iter_rw(iter) == WRITE) |
fec53774 DW |
1061 | map_len = dax_copy_from_iter(dax_dev, pgoff, kaddr, |
1062 | map_len, iter); | |
a254e568 | 1063 | else |
cccbce67 | 1064 | map_len = copy_to_iter(kaddr, map_len, iter); |
a254e568 CH |
1065 | if (map_len <= 0) { |
1066 | ret = map_len ? map_len : -EFAULT; | |
1067 | break; | |
1068 | } | |
1069 | ||
1070 | pos += map_len; | |
1071 | length -= map_len; | |
1072 | done += map_len; | |
1073 | } | |
cccbce67 | 1074 | dax_read_unlock(id); |
a254e568 CH |
1075 | |
1076 | return done ? done : ret; | |
1077 | } | |
1078 | ||
1079 | /** | |
11c59c92 | 1080 | * dax_iomap_rw - Perform I/O to a DAX file |
a254e568 CH |
1081 | * @iocb: The control block for this I/O |
1082 | * @iter: The addresses to do I/O from or to | |
1083 | * @ops: iomap ops passed from the file system | |
1084 | * | |
1085 | * This function performs read and write operations to directly mapped | |
1086 | * persistent memory. The callers needs to take care of read/write exclusion | |
1087 | * and evicting any page cache pages in the region under I/O. | |
1088 | */ | |
1089 | ssize_t | |
11c59c92 | 1090 | dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, |
8ff6daa1 | 1091 | const struct iomap_ops *ops) |
a254e568 CH |
1092 | { |
1093 | struct address_space *mapping = iocb->ki_filp->f_mapping; | |
1094 | struct inode *inode = mapping->host; | |
1095 | loff_t pos = iocb->ki_pos, ret = 0, done = 0; | |
1096 | unsigned flags = 0; | |
1097 | ||
168316db CH |
1098 | if (iov_iter_rw(iter) == WRITE) { |
1099 | lockdep_assert_held_exclusive(&inode->i_rwsem); | |
a254e568 | 1100 | flags |= IOMAP_WRITE; |
168316db CH |
1101 | } else { |
1102 | lockdep_assert_held(&inode->i_rwsem); | |
1103 | } | |
a254e568 | 1104 | |
a254e568 CH |
1105 | while (iov_iter_count(iter)) { |
1106 | ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, | |
11c59c92 | 1107 | iter, dax_iomap_actor); |
a254e568 CH |
1108 | if (ret <= 0) |
1109 | break; | |
1110 | pos += ret; | |
1111 | done += ret; | |
1112 | } | |
1113 | ||
1114 | iocb->ki_pos += done; | |
1115 | return done ? done : ret; | |
1116 | } | |
11c59c92 | 1117 | EXPORT_SYMBOL_GPL(dax_iomap_rw); |
a7d73fe6 | 1118 | |
9f141d6e JK |
1119 | static int dax_fault_return(int error) |
1120 | { | |
1121 | if (error == 0) | |
1122 | return VM_FAULT_NOPAGE; | |
1123 | if (error == -ENOMEM) | |
1124 | return VM_FAULT_OOM; | |
1125 | return VM_FAULT_SIGBUS; | |
1126 | } | |
1127 | ||
aaa422c4 DW |
1128 | /* |
1129 | * MAP_SYNC on a dax mapping guarantees dirty metadata is | |
1130 | * flushed on write-faults (non-cow), but not read-faults. | |
1131 | */ | |
1132 | static bool dax_fault_is_synchronous(unsigned long flags, | |
1133 | struct vm_area_struct *vma, struct iomap *iomap) | |
1134 | { | |
1135 | return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) | |
1136 | && (iomap->flags & IOMAP_F_DIRTY); | |
1137 | } | |
1138 | ||
9a0dd422 | 1139 | static int dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp, |
c0b24625 | 1140 | int *iomap_errp, const struct iomap_ops *ops) |
a7d73fe6 | 1141 | { |
a0987ad5 JK |
1142 | struct vm_area_struct *vma = vmf->vma; |
1143 | struct address_space *mapping = vma->vm_file->f_mapping; | |
a7d73fe6 | 1144 | struct inode *inode = mapping->host; |
1a29d85e | 1145 | unsigned long vaddr = vmf->address; |
a7d73fe6 | 1146 | loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; |
a7d73fe6 | 1147 | struct iomap iomap = { 0 }; |
9484ab1b | 1148 | unsigned flags = IOMAP_FAULT; |
a7d73fe6 | 1149 | int error, major = 0; |
d2c43ef1 | 1150 | bool write = vmf->flags & FAULT_FLAG_WRITE; |
caa51d26 | 1151 | bool sync; |
b1aa812b | 1152 | int vmf_ret = 0; |
a7d73fe6 | 1153 | void *entry; |
1b5a1cb2 | 1154 | pfn_t pfn; |
a7d73fe6 | 1155 | |
a9c42b33 | 1156 | trace_dax_pte_fault(inode, vmf, vmf_ret); |
a7d73fe6 CH |
1157 | /* |
1158 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
1159 | * to hold locks serializing us with truncate / punch hole so this is | |
1160 | * a reliable test. | |
1161 | */ | |
a9c42b33 RZ |
1162 | if (pos >= i_size_read(inode)) { |
1163 | vmf_ret = VM_FAULT_SIGBUS; | |
1164 | goto out; | |
1165 | } | |
a7d73fe6 | 1166 | |
d2c43ef1 | 1167 | if (write && !vmf->cow_page) |
a7d73fe6 CH |
1168 | flags |= IOMAP_WRITE; |
1169 | ||
13e451fd JK |
1170 | entry = grab_mapping_entry(mapping, vmf->pgoff, 0); |
1171 | if (IS_ERR(entry)) { | |
1172 | vmf_ret = dax_fault_return(PTR_ERR(entry)); | |
1173 | goto out; | |
1174 | } | |
1175 | ||
e2093926 RZ |
1176 | /* |
1177 | * It is possible, particularly with mixed reads & writes to private | |
1178 | * mappings, that we have raced with a PMD fault that overlaps with | |
1179 | * the PTE we need to set up. If so just return and the fault will be | |
1180 | * retried. | |
1181 | */ | |
1182 | if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) { | |
1183 | vmf_ret = VM_FAULT_NOPAGE; | |
1184 | goto unlock_entry; | |
1185 | } | |
1186 | ||
a7d73fe6 CH |
1187 | /* |
1188 | * Note that we don't bother to use iomap_apply here: DAX required | |
1189 | * the file system block size to be equal the page size, which means | |
1190 | * that we never have to deal with more than a single extent here. | |
1191 | */ | |
1192 | error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); | |
c0b24625 JK |
1193 | if (iomap_errp) |
1194 | *iomap_errp = error; | |
a9c42b33 RZ |
1195 | if (error) { |
1196 | vmf_ret = dax_fault_return(error); | |
13e451fd | 1197 | goto unlock_entry; |
a9c42b33 | 1198 | } |
a7d73fe6 | 1199 | if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { |
13e451fd JK |
1200 | error = -EIO; /* fs corruption? */ |
1201 | goto error_finish_iomap; | |
a7d73fe6 CH |
1202 | } |
1203 | ||
a7d73fe6 | 1204 | if (vmf->cow_page) { |
31a6f1a6 JK |
1205 | sector_t sector = dax_iomap_sector(&iomap, pos); |
1206 | ||
a7d73fe6 CH |
1207 | switch (iomap.type) { |
1208 | case IOMAP_HOLE: | |
1209 | case IOMAP_UNWRITTEN: | |
1210 | clear_user_highpage(vmf->cow_page, vaddr); | |
1211 | break; | |
1212 | case IOMAP_MAPPED: | |
cccbce67 DW |
1213 | error = copy_user_dax(iomap.bdev, iomap.dax_dev, |
1214 | sector, PAGE_SIZE, vmf->cow_page, vaddr); | |
a7d73fe6 CH |
1215 | break; |
1216 | default: | |
1217 | WARN_ON_ONCE(1); | |
1218 | error = -EIO; | |
1219 | break; | |
1220 | } | |
1221 | ||
1222 | if (error) | |
13e451fd | 1223 | goto error_finish_iomap; |
b1aa812b JK |
1224 | |
1225 | __SetPageUptodate(vmf->cow_page); | |
1226 | vmf_ret = finish_fault(vmf); | |
1227 | if (!vmf_ret) | |
1228 | vmf_ret = VM_FAULT_DONE_COW; | |
13e451fd | 1229 | goto finish_iomap; |
a7d73fe6 CH |
1230 | } |
1231 | ||
aaa422c4 | 1232 | sync = dax_fault_is_synchronous(flags, vma, &iomap); |
caa51d26 | 1233 | |
a7d73fe6 CH |
1234 | switch (iomap.type) { |
1235 | case IOMAP_MAPPED: | |
1236 | if (iomap.flags & IOMAP_F_NEW) { | |
1237 | count_vm_event(PGMAJFAULT); | |
a0987ad5 | 1238 | count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); |
a7d73fe6 CH |
1239 | major = VM_FAULT_MAJOR; |
1240 | } | |
1b5a1cb2 JK |
1241 | error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn); |
1242 | if (error < 0) | |
1243 | goto error_finish_iomap; | |
1244 | ||
3fe0791c | 1245 | entry = dax_insert_mapping_entry(mapping, vmf, entry, pfn, |
caa51d26 | 1246 | 0, write && !sync); |
1b5a1cb2 JK |
1247 | if (IS_ERR(entry)) { |
1248 | error = PTR_ERR(entry); | |
1249 | goto error_finish_iomap; | |
1250 | } | |
1251 | ||
caa51d26 JK |
1252 | /* |
1253 | * If we are doing synchronous page fault and inode needs fsync, | |
1254 | * we can insert PTE into page tables only after that happens. | |
1255 | * Skip insertion for now and return the pfn so that caller can | |
1256 | * insert it after fsync is done. | |
1257 | */ | |
1258 | if (sync) { | |
1259 | if (WARN_ON_ONCE(!pfnp)) { | |
1260 | error = -EIO; | |
1261 | goto error_finish_iomap; | |
1262 | } | |
1263 | *pfnp = pfn; | |
1264 | vmf_ret = VM_FAULT_NEEDDSYNC | major; | |
1265 | goto finish_iomap; | |
1266 | } | |
1b5a1cb2 JK |
1267 | trace_dax_insert_mapping(inode, vmf, entry); |
1268 | if (write) | |
1269 | error = vm_insert_mixed_mkwrite(vma, vaddr, pfn); | |
1270 | else | |
1271 | error = vm_insert_mixed(vma, vaddr, pfn); | |
1272 | ||
9f141d6e JK |
1273 | /* -EBUSY is fine, somebody else faulted on the same PTE */ |
1274 | if (error == -EBUSY) | |
1275 | error = 0; | |
a7d73fe6 CH |
1276 | break; |
1277 | case IOMAP_UNWRITTEN: | |
1278 | case IOMAP_HOLE: | |
d2c43ef1 | 1279 | if (!write) { |
91d25ba8 | 1280 | vmf_ret = dax_load_hole(mapping, entry, vmf); |
13e451fd | 1281 | goto finish_iomap; |
1550290b | 1282 | } |
a7d73fe6 CH |
1283 | /*FALLTHRU*/ |
1284 | default: | |
1285 | WARN_ON_ONCE(1); | |
1286 | error = -EIO; | |
1287 | break; | |
1288 | } | |
1289 | ||
13e451fd | 1290 | error_finish_iomap: |
9f141d6e | 1291 | vmf_ret = dax_fault_return(error) | major; |
9f141d6e JK |
1292 | finish_iomap: |
1293 | if (ops->iomap_end) { | |
1294 | int copied = PAGE_SIZE; | |
1295 | ||
1296 | if (vmf_ret & VM_FAULT_ERROR) | |
1297 | copied = 0; | |
1298 | /* | |
1299 | * The fault is done by now and there's no way back (other | |
1300 | * thread may be already happily using PTE we have installed). | |
1301 | * Just ignore error from ->iomap_end since we cannot do much | |
1302 | * with it. | |
1303 | */ | |
1304 | ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap); | |
1550290b | 1305 | } |
13e451fd | 1306 | unlock_entry: |
91d25ba8 | 1307 | put_locked_mapping_entry(mapping, vmf->pgoff); |
13e451fd | 1308 | out: |
a9c42b33 | 1309 | trace_dax_pte_fault_done(inode, vmf, vmf_ret); |
9f141d6e | 1310 | return vmf_ret; |
a7d73fe6 | 1311 | } |
642261ac RZ |
1312 | |
1313 | #ifdef CONFIG_FS_DAX_PMD | |
f4200391 | 1314 | static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap, |
91d25ba8 | 1315 | void *entry) |
642261ac | 1316 | { |
f4200391 DJ |
1317 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
1318 | unsigned long pmd_addr = vmf->address & PMD_MASK; | |
653b2ea3 | 1319 | struct inode *inode = mapping->host; |
642261ac | 1320 | struct page *zero_page; |
653b2ea3 | 1321 | void *ret = NULL; |
642261ac RZ |
1322 | spinlock_t *ptl; |
1323 | pmd_t pmd_entry; | |
3fe0791c | 1324 | pfn_t pfn; |
642261ac | 1325 | |
f4200391 | 1326 | zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); |
642261ac RZ |
1327 | |
1328 | if (unlikely(!zero_page)) | |
653b2ea3 | 1329 | goto fallback; |
642261ac | 1330 | |
3fe0791c DW |
1331 | pfn = page_to_pfn_t(zero_page); |
1332 | ret = dax_insert_mapping_entry(mapping, vmf, entry, pfn, | |
f5b7b748 | 1333 | RADIX_DAX_PMD | RADIX_DAX_ZERO_PAGE, false); |
642261ac | 1334 | if (IS_ERR(ret)) |
653b2ea3 | 1335 | goto fallback; |
642261ac | 1336 | |
f4200391 DJ |
1337 | ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); |
1338 | if (!pmd_none(*(vmf->pmd))) { | |
642261ac | 1339 | spin_unlock(ptl); |
653b2ea3 | 1340 | goto fallback; |
642261ac RZ |
1341 | } |
1342 | ||
f4200391 | 1343 | pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); |
642261ac | 1344 | pmd_entry = pmd_mkhuge(pmd_entry); |
f4200391 | 1345 | set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); |
642261ac | 1346 | spin_unlock(ptl); |
f4200391 | 1347 | trace_dax_pmd_load_hole(inode, vmf, zero_page, ret); |
642261ac | 1348 | return VM_FAULT_NOPAGE; |
653b2ea3 RZ |
1349 | |
1350 | fallback: | |
f4200391 | 1351 | trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret); |
653b2ea3 | 1352 | return VM_FAULT_FALLBACK; |
642261ac RZ |
1353 | } |
1354 | ||
9a0dd422 | 1355 | static int dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, |
a2d58167 | 1356 | const struct iomap_ops *ops) |
642261ac | 1357 | { |
f4200391 | 1358 | struct vm_area_struct *vma = vmf->vma; |
642261ac | 1359 | struct address_space *mapping = vma->vm_file->f_mapping; |
d8a849e1 DJ |
1360 | unsigned long pmd_addr = vmf->address & PMD_MASK; |
1361 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
caa51d26 | 1362 | bool sync; |
9484ab1b | 1363 | unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT; |
642261ac RZ |
1364 | struct inode *inode = mapping->host; |
1365 | int result = VM_FAULT_FALLBACK; | |
1366 | struct iomap iomap = { 0 }; | |
1367 | pgoff_t max_pgoff, pgoff; | |
642261ac RZ |
1368 | void *entry; |
1369 | loff_t pos; | |
1370 | int error; | |
302a5e31 | 1371 | pfn_t pfn; |
642261ac | 1372 | |
282a8e03 RZ |
1373 | /* |
1374 | * Check whether offset isn't beyond end of file now. Caller is | |
1375 | * supposed to hold locks serializing us with truncate / punch hole so | |
1376 | * this is a reliable test. | |
1377 | */ | |
1378 | pgoff = linear_page_index(vma, pmd_addr); | |
957ac8c4 | 1379 | max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE); |
282a8e03 | 1380 | |
f4200391 | 1381 | trace_dax_pmd_fault(inode, vmf, max_pgoff, 0); |
282a8e03 | 1382 | |
fffa281b RZ |
1383 | /* |
1384 | * Make sure that the faulting address's PMD offset (color) matches | |
1385 | * the PMD offset from the start of the file. This is necessary so | |
1386 | * that a PMD range in the page table overlaps exactly with a PMD | |
1387 | * range in the radix tree. | |
1388 | */ | |
1389 | if ((vmf->pgoff & PG_PMD_COLOUR) != | |
1390 | ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR)) | |
1391 | goto fallback; | |
1392 | ||
642261ac RZ |
1393 | /* Fall back to PTEs if we're going to COW */ |
1394 | if (write && !(vma->vm_flags & VM_SHARED)) | |
1395 | goto fallback; | |
1396 | ||
1397 | /* If the PMD would extend outside the VMA */ | |
1398 | if (pmd_addr < vma->vm_start) | |
1399 | goto fallback; | |
1400 | if ((pmd_addr + PMD_SIZE) > vma->vm_end) | |
1401 | goto fallback; | |
1402 | ||
957ac8c4 | 1403 | if (pgoff >= max_pgoff) { |
282a8e03 RZ |
1404 | result = VM_FAULT_SIGBUS; |
1405 | goto out; | |
1406 | } | |
642261ac RZ |
1407 | |
1408 | /* If the PMD would extend beyond the file size */ | |
957ac8c4 | 1409 | if ((pgoff | PG_PMD_COLOUR) >= max_pgoff) |
642261ac RZ |
1410 | goto fallback; |
1411 | ||
876f2946 | 1412 | /* |
91d25ba8 RZ |
1413 | * grab_mapping_entry() will make sure we get a 2MiB empty entry, a |
1414 | * 2MiB zero page entry or a DAX PMD. If it can't (because a 4k page | |
1415 | * is already in the tree, for instance), it will return -EEXIST and | |
1416 | * we just fall back to 4k entries. | |
876f2946 RZ |
1417 | */ |
1418 | entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD); | |
1419 | if (IS_ERR(entry)) | |
1420 | goto fallback; | |
1421 | ||
e2093926 RZ |
1422 | /* |
1423 | * It is possible, particularly with mixed reads & writes to private | |
1424 | * mappings, that we have raced with a PTE fault that overlaps with | |
1425 | * the PMD we need to set up. If so just return and the fault will be | |
1426 | * retried. | |
1427 | */ | |
1428 | if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) && | |
1429 | !pmd_devmap(*vmf->pmd)) { | |
1430 | result = 0; | |
1431 | goto unlock_entry; | |
1432 | } | |
1433 | ||
642261ac RZ |
1434 | /* |
1435 | * Note that we don't use iomap_apply here. We aren't doing I/O, only | |
1436 | * setting up a mapping, so really we're using iomap_begin() as a way | |
1437 | * to look up our filesystem block. | |
1438 | */ | |
1439 | pos = (loff_t)pgoff << PAGE_SHIFT; | |
1440 | error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap); | |
1441 | if (error) | |
876f2946 | 1442 | goto unlock_entry; |
9f141d6e | 1443 | |
642261ac RZ |
1444 | if (iomap.offset + iomap.length < pos + PMD_SIZE) |
1445 | goto finish_iomap; | |
1446 | ||
aaa422c4 | 1447 | sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap); |
caa51d26 | 1448 | |
642261ac RZ |
1449 | switch (iomap.type) { |
1450 | case IOMAP_MAPPED: | |
302a5e31 JK |
1451 | error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn); |
1452 | if (error < 0) | |
1453 | goto finish_iomap; | |
1454 | ||
3fe0791c | 1455 | entry = dax_insert_mapping_entry(mapping, vmf, entry, pfn, |
caa51d26 | 1456 | RADIX_DAX_PMD, write && !sync); |
302a5e31 JK |
1457 | if (IS_ERR(entry)) |
1458 | goto finish_iomap; | |
1459 | ||
caa51d26 JK |
1460 | /* |
1461 | * If we are doing synchronous page fault and inode needs fsync, | |
1462 | * we can insert PMD into page tables only after that happens. | |
1463 | * Skip insertion for now and return the pfn so that caller can | |
1464 | * insert it after fsync is done. | |
1465 | */ | |
1466 | if (sync) { | |
1467 | if (WARN_ON_ONCE(!pfnp)) | |
1468 | goto finish_iomap; | |
1469 | *pfnp = pfn; | |
1470 | result = VM_FAULT_NEEDDSYNC; | |
1471 | goto finish_iomap; | |
1472 | } | |
1473 | ||
302a5e31 JK |
1474 | trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry); |
1475 | result = vmf_insert_pfn_pmd(vma, vmf->address, vmf->pmd, pfn, | |
1476 | write); | |
642261ac RZ |
1477 | break; |
1478 | case IOMAP_UNWRITTEN: | |
1479 | case IOMAP_HOLE: | |
1480 | if (WARN_ON_ONCE(write)) | |
876f2946 | 1481 | break; |
91d25ba8 | 1482 | result = dax_pmd_load_hole(vmf, &iomap, entry); |
642261ac RZ |
1483 | break; |
1484 | default: | |
1485 | WARN_ON_ONCE(1); | |
1486 | break; | |
1487 | } | |
1488 | ||
1489 | finish_iomap: | |
1490 | if (ops->iomap_end) { | |
9f141d6e JK |
1491 | int copied = PMD_SIZE; |
1492 | ||
1493 | if (result == VM_FAULT_FALLBACK) | |
1494 | copied = 0; | |
1495 | /* | |
1496 | * The fault is done by now and there's no way back (other | |
1497 | * thread may be already happily using PMD we have installed). | |
1498 | * Just ignore error from ->iomap_end since we cannot do much | |
1499 | * with it. | |
1500 | */ | |
1501 | ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags, | |
1502 | &iomap); | |
642261ac | 1503 | } |
876f2946 | 1504 | unlock_entry: |
91d25ba8 | 1505 | put_locked_mapping_entry(mapping, pgoff); |
642261ac RZ |
1506 | fallback: |
1507 | if (result == VM_FAULT_FALLBACK) { | |
d8a849e1 | 1508 | split_huge_pmd(vma, vmf->pmd, vmf->address); |
642261ac RZ |
1509 | count_vm_event(THP_FAULT_FALLBACK); |
1510 | } | |
282a8e03 | 1511 | out: |
f4200391 | 1512 | trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result); |
642261ac RZ |
1513 | return result; |
1514 | } | |
a2d58167 | 1515 | #else |
9a0dd422 | 1516 | static int dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, |
01cddfe9 | 1517 | const struct iomap_ops *ops) |
a2d58167 DJ |
1518 | { |
1519 | return VM_FAULT_FALLBACK; | |
1520 | } | |
642261ac | 1521 | #endif /* CONFIG_FS_DAX_PMD */ |
a2d58167 DJ |
1522 | |
1523 | /** | |
1524 | * dax_iomap_fault - handle a page fault on a DAX file | |
1525 | * @vmf: The description of the fault | |
cec04e8c | 1526 | * @pe_size: Size of the page to fault in |
9a0dd422 | 1527 | * @pfnp: PFN to insert for synchronous faults if fsync is required |
c0b24625 | 1528 | * @iomap_errp: Storage for detailed error code in case of error |
cec04e8c | 1529 | * @ops: Iomap ops passed from the file system |
a2d58167 DJ |
1530 | * |
1531 | * When a page fault occurs, filesystems may call this helper in | |
1532 | * their fault handler for DAX files. dax_iomap_fault() assumes the caller | |
1533 | * has done all the necessary locking for page fault to proceed | |
1534 | * successfully. | |
1535 | */ | |
c791ace1 | 1536 | int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, |
c0b24625 | 1537 | pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops) |
a2d58167 | 1538 | { |
c791ace1 DJ |
1539 | switch (pe_size) { |
1540 | case PE_SIZE_PTE: | |
c0b24625 | 1541 | return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops); |
c791ace1 | 1542 | case PE_SIZE_PMD: |
9a0dd422 | 1543 | return dax_iomap_pmd_fault(vmf, pfnp, ops); |
a2d58167 DJ |
1544 | default: |
1545 | return VM_FAULT_FALLBACK; | |
1546 | } | |
1547 | } | |
1548 | EXPORT_SYMBOL_GPL(dax_iomap_fault); | |
71eab6df JK |
1549 | |
1550 | /** | |
1551 | * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables | |
1552 | * @vmf: The description of the fault | |
1553 | * @pe_size: Size of entry to be inserted | |
1554 | * @pfn: PFN to insert | |
1555 | * | |
1556 | * This function inserts writeable PTE or PMD entry into page tables for mmaped | |
1557 | * DAX file. It takes care of marking corresponding radix tree entry as dirty | |
1558 | * as well. | |
1559 | */ | |
1560 | static int dax_insert_pfn_mkwrite(struct vm_fault *vmf, | |
1561 | enum page_entry_size pe_size, | |
1562 | pfn_t pfn) | |
1563 | { | |
1564 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; | |
1565 | void *entry, **slot; | |
1566 | pgoff_t index = vmf->pgoff; | |
1567 | int vmf_ret, error; | |
1568 | ||
1569 | spin_lock_irq(&mapping->tree_lock); | |
1570 | entry = get_unlocked_mapping_entry(mapping, index, &slot); | |
1571 | /* Did we race with someone splitting entry or so? */ | |
1572 | if (!entry || | |
1573 | (pe_size == PE_SIZE_PTE && !dax_is_pte_entry(entry)) || | |
1574 | (pe_size == PE_SIZE_PMD && !dax_is_pmd_entry(entry))) { | |
1575 | put_unlocked_mapping_entry(mapping, index, entry); | |
1576 | spin_unlock_irq(&mapping->tree_lock); | |
1577 | trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf, | |
1578 | VM_FAULT_NOPAGE); | |
1579 | return VM_FAULT_NOPAGE; | |
1580 | } | |
1581 | radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY); | |
1582 | entry = lock_slot(mapping, slot); | |
1583 | spin_unlock_irq(&mapping->tree_lock); | |
1584 | switch (pe_size) { | |
1585 | case PE_SIZE_PTE: | |
1586 | error = vm_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); | |
1587 | vmf_ret = dax_fault_return(error); | |
1588 | break; | |
1589 | #ifdef CONFIG_FS_DAX_PMD | |
1590 | case PE_SIZE_PMD: | |
1591 | vmf_ret = vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd, | |
1592 | pfn, true); | |
1593 | break; | |
1594 | #endif | |
1595 | default: | |
1596 | vmf_ret = VM_FAULT_FALLBACK; | |
1597 | } | |
1598 | put_locked_mapping_entry(mapping, index); | |
1599 | trace_dax_insert_pfn_mkwrite(mapping->host, vmf, vmf_ret); | |
1600 | return vmf_ret; | |
1601 | } | |
1602 | ||
1603 | /** | |
1604 | * dax_finish_sync_fault - finish synchronous page fault | |
1605 | * @vmf: The description of the fault | |
1606 | * @pe_size: Size of entry to be inserted | |
1607 | * @pfn: PFN to insert | |
1608 | * | |
1609 | * This function ensures that the file range touched by the page fault is | |
1610 | * stored persistently on the media and handles inserting of appropriate page | |
1611 | * table entry. | |
1612 | */ | |
1613 | int dax_finish_sync_fault(struct vm_fault *vmf, enum page_entry_size pe_size, | |
1614 | pfn_t pfn) | |
1615 | { | |
1616 | int err; | |
1617 | loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT; | |
1618 | size_t len = 0; | |
1619 | ||
1620 | if (pe_size == PE_SIZE_PTE) | |
1621 | len = PAGE_SIZE; | |
1622 | else if (pe_size == PE_SIZE_PMD) | |
1623 | len = PMD_SIZE; | |
1624 | else | |
1625 | WARN_ON_ONCE(1); | |
1626 | err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1); | |
1627 | if (err) | |
1628 | return VM_FAULT_SIGBUS; | |
1629 | return dax_insert_pfn_mkwrite(vmf, pe_size, pfn); | |
1630 | } | |
1631 | EXPORT_SYMBOL_GPL(dax_finish_sync_fault); |