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 | ||
cfc93c6c MW |
41 | static inline unsigned int pe_order(enum page_entry_size pe_size) |
42 | { | |
43 | if (pe_size == PE_SIZE_PTE) | |
44 | return PAGE_SHIFT - PAGE_SHIFT; | |
45 | if (pe_size == PE_SIZE_PMD) | |
46 | return PMD_SHIFT - PAGE_SHIFT; | |
47 | if (pe_size == PE_SIZE_PUD) | |
48 | return PUD_SHIFT - PAGE_SHIFT; | |
49 | return ~0; | |
50 | } | |
51 | ||
ac401cc7 JK |
52 | /* We choose 4096 entries - same as per-zone page wait tables */ |
53 | #define DAX_WAIT_TABLE_BITS 12 | |
54 | #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) | |
55 | ||
917f3452 RZ |
56 | /* The 'colour' (ie low bits) within a PMD of a page offset. */ |
57 | #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) | |
977fbdcd | 58 | #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT) |
917f3452 | 59 | |
cfc93c6c MW |
60 | /* The order of a PMD entry */ |
61 | #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT) | |
62 | ||
ce95ab0f | 63 | static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; |
ac401cc7 JK |
64 | |
65 | static int __init init_dax_wait_table(void) | |
66 | { | |
67 | int i; | |
68 | ||
69 | for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) | |
70 | init_waitqueue_head(wait_table + i); | |
71 | return 0; | |
72 | } | |
73 | fs_initcall(init_dax_wait_table); | |
74 | ||
527b19d0 | 75 | /* |
3159f943 MW |
76 | * DAX pagecache entries use XArray value entries so they can't be mistaken |
77 | * for pages. We use one bit for locking, one bit for the entry size (PMD) | |
78 | * and two more to tell us if the entry is a zero page or an empty entry that | |
79 | * is just used for locking. In total four special bits. | |
527b19d0 RZ |
80 | * |
81 | * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE | |
82 | * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem | |
83 | * block allocation. | |
84 | */ | |
3159f943 MW |
85 | #define DAX_SHIFT (4) |
86 | #define DAX_LOCKED (1UL << 0) | |
87 | #define DAX_PMD (1UL << 1) | |
88 | #define DAX_ZERO_PAGE (1UL << 2) | |
89 | #define DAX_EMPTY (1UL << 3) | |
527b19d0 | 90 | |
a77d19f4 | 91 | static unsigned long dax_to_pfn(void *entry) |
527b19d0 | 92 | { |
3159f943 | 93 | return xa_to_value(entry) >> DAX_SHIFT; |
527b19d0 RZ |
94 | } |
95 | ||
9f32d221 MW |
96 | static void *dax_make_entry(pfn_t pfn, unsigned long flags) |
97 | { | |
98 | return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT)); | |
99 | } | |
100 | ||
cfc93c6c MW |
101 | static bool dax_is_locked(void *entry) |
102 | { | |
103 | return xa_to_value(entry) & DAX_LOCKED; | |
104 | } | |
105 | ||
a77d19f4 | 106 | static unsigned int dax_entry_order(void *entry) |
527b19d0 | 107 | { |
3159f943 | 108 | if (xa_to_value(entry) & DAX_PMD) |
cfc93c6c | 109 | return PMD_ORDER; |
527b19d0 RZ |
110 | return 0; |
111 | } | |
112 | ||
fda490d3 | 113 | static unsigned long dax_is_pmd_entry(void *entry) |
d1a5f2b4 | 114 | { |
3159f943 | 115 | return xa_to_value(entry) & DAX_PMD; |
d1a5f2b4 DW |
116 | } |
117 | ||
fda490d3 | 118 | static bool dax_is_pte_entry(void *entry) |
d475c634 | 119 | { |
3159f943 | 120 | return !(xa_to_value(entry) & DAX_PMD); |
d475c634 MW |
121 | } |
122 | ||
642261ac | 123 | static int dax_is_zero_entry(void *entry) |
d475c634 | 124 | { |
3159f943 | 125 | return xa_to_value(entry) & DAX_ZERO_PAGE; |
d475c634 MW |
126 | } |
127 | ||
642261ac | 128 | static int dax_is_empty_entry(void *entry) |
b2e0d162 | 129 | { |
3159f943 | 130 | return xa_to_value(entry) & DAX_EMPTY; |
b2e0d162 DW |
131 | } |
132 | ||
ac401cc7 | 133 | /* |
a77d19f4 | 134 | * DAX page cache entry locking |
ac401cc7 JK |
135 | */ |
136 | struct exceptional_entry_key { | |
ec4907ff | 137 | struct xarray *xa; |
63e95b5c | 138 | pgoff_t entry_start; |
ac401cc7 JK |
139 | }; |
140 | ||
141 | struct wait_exceptional_entry_queue { | |
ac6424b9 | 142 | wait_queue_entry_t wait; |
ac401cc7 JK |
143 | struct exceptional_entry_key key; |
144 | }; | |
145 | ||
b15cd800 MW |
146 | static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas, |
147 | void *entry, struct exceptional_entry_key *key) | |
63e95b5c RZ |
148 | { |
149 | unsigned long hash; | |
b15cd800 | 150 | unsigned long index = xas->xa_index; |
63e95b5c RZ |
151 | |
152 | /* | |
153 | * If 'entry' is a PMD, align the 'index' that we use for the wait | |
154 | * queue to the start of that PMD. This ensures that all offsets in | |
155 | * the range covered by the PMD map to the same bit lock. | |
156 | */ | |
642261ac | 157 | if (dax_is_pmd_entry(entry)) |
917f3452 | 158 | index &= ~PG_PMD_COLOUR; |
b15cd800 | 159 | key->xa = xas->xa; |
63e95b5c RZ |
160 | key->entry_start = index; |
161 | ||
b15cd800 | 162 | hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS); |
63e95b5c RZ |
163 | return wait_table + hash; |
164 | } | |
165 | ||
ec4907ff MW |
166 | static int wake_exceptional_entry_func(wait_queue_entry_t *wait, |
167 | unsigned int mode, int sync, void *keyp) | |
ac401cc7 JK |
168 | { |
169 | struct exceptional_entry_key *key = keyp; | |
170 | struct wait_exceptional_entry_queue *ewait = | |
171 | container_of(wait, struct wait_exceptional_entry_queue, wait); | |
172 | ||
ec4907ff | 173 | if (key->xa != ewait->key.xa || |
63e95b5c | 174 | key->entry_start != ewait->key.entry_start) |
ac401cc7 JK |
175 | return 0; |
176 | return autoremove_wake_function(wait, mode, sync, NULL); | |
177 | } | |
178 | ||
e30331ff | 179 | /* |
b93b0163 MW |
180 | * @entry may no longer be the entry at the index in the mapping. |
181 | * The important information it's conveying is whether the entry at | |
182 | * this index used to be a PMD entry. | |
e30331ff | 183 | */ |
b15cd800 | 184 | static void dax_wake_entry(struct xa_state *xas, void *entry, bool wake_all) |
e30331ff RZ |
185 | { |
186 | struct exceptional_entry_key key; | |
187 | wait_queue_head_t *wq; | |
188 | ||
b15cd800 | 189 | wq = dax_entry_waitqueue(xas, entry, &key); |
e30331ff RZ |
190 | |
191 | /* | |
192 | * Checking for locked entry and prepare_to_wait_exclusive() happens | |
b93b0163 | 193 | * under the i_pages lock, ditto for entry handling in our callers. |
e30331ff RZ |
194 | * So at this point all tasks that could have seen our entry locked |
195 | * must be in the waitqueue and the following check will see them. | |
196 | */ | |
197 | if (waitqueue_active(wq)) | |
198 | __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); | |
199 | } | |
200 | ||
cfc93c6c MW |
201 | /* |
202 | * Look up entry in page cache, wait for it to become unlocked if it | |
203 | * is a DAX entry and return it. The caller must subsequently call | |
204 | * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry() | |
205 | * if it did. | |
206 | * | |
207 | * Must be called with the i_pages lock held. | |
208 | */ | |
209 | static void *get_unlocked_entry(struct xa_state *xas) | |
210 | { | |
211 | void *entry; | |
212 | struct wait_exceptional_entry_queue ewait; | |
213 | wait_queue_head_t *wq; | |
214 | ||
215 | init_wait(&ewait.wait); | |
216 | ewait.wait.func = wake_exceptional_entry_func; | |
217 | ||
218 | for (;;) { | |
0e40de03 MW |
219 | entry = xas_find_conflict(xas); |
220 | if (!entry || WARN_ON_ONCE(!xa_is_value(entry)) || | |
cfc93c6c MW |
221 | !dax_is_locked(entry)) |
222 | return entry; | |
223 | ||
b15cd800 | 224 | wq = dax_entry_waitqueue(xas, entry, &ewait.key); |
cfc93c6c MW |
225 | prepare_to_wait_exclusive(wq, &ewait.wait, |
226 | TASK_UNINTERRUPTIBLE); | |
227 | xas_unlock_irq(xas); | |
228 | xas_reset(xas); | |
229 | schedule(); | |
230 | finish_wait(wq, &ewait.wait); | |
231 | xas_lock_irq(xas); | |
232 | } | |
233 | } | |
234 | ||
55e56f06 MW |
235 | /* |
236 | * The only thing keeping the address space around is the i_pages lock | |
237 | * (it's cycled in clear_inode() after removing the entries from i_pages) | |
238 | * After we call xas_unlock_irq(), we cannot touch xas->xa. | |
239 | */ | |
240 | static void wait_entry_unlocked(struct xa_state *xas, void *entry) | |
241 | { | |
242 | struct wait_exceptional_entry_queue ewait; | |
243 | wait_queue_head_t *wq; | |
244 | ||
245 | init_wait(&ewait.wait); | |
246 | ewait.wait.func = wake_exceptional_entry_func; | |
247 | ||
248 | wq = dax_entry_waitqueue(xas, entry, &ewait.key); | |
249 | prepare_to_wait_exclusive(wq, &ewait.wait, TASK_UNINTERRUPTIBLE); | |
250 | xas_unlock_irq(xas); | |
251 | schedule(); | |
252 | finish_wait(wq, &ewait.wait); | |
253 | ||
254 | /* | |
255 | * Entry lock waits are exclusive. Wake up the next waiter since | |
256 | * we aren't sure we will acquire the entry lock and thus wake | |
257 | * the next waiter up on unlock. | |
258 | */ | |
259 | if (waitqueue_active(wq)) | |
260 | __wake_up(wq, TASK_NORMAL, 1, &ewait.key); | |
261 | } | |
262 | ||
cfc93c6c MW |
263 | static void put_unlocked_entry(struct xa_state *xas, void *entry) |
264 | { | |
265 | /* If we were the only waiter woken, wake the next one */ | |
266 | if (entry) | |
267 | dax_wake_entry(xas, entry, false); | |
268 | } | |
269 | ||
270 | /* | |
271 | * We used the xa_state to get the entry, but then we locked the entry and | |
272 | * dropped the xa_lock, so we know the xa_state is stale and must be reset | |
273 | * before use. | |
274 | */ | |
275 | static void dax_unlock_entry(struct xa_state *xas, void *entry) | |
276 | { | |
277 | void *old; | |
278 | ||
7ae2ea7d | 279 | BUG_ON(dax_is_locked(entry)); |
cfc93c6c MW |
280 | xas_reset(xas); |
281 | xas_lock_irq(xas); | |
282 | old = xas_store(xas, entry); | |
283 | xas_unlock_irq(xas); | |
284 | BUG_ON(!dax_is_locked(old)); | |
285 | dax_wake_entry(xas, entry, false); | |
286 | } | |
287 | ||
288 | /* | |
289 | * Return: The entry stored at this location before it was locked. | |
290 | */ | |
291 | static void *dax_lock_entry(struct xa_state *xas, void *entry) | |
292 | { | |
293 | unsigned long v = xa_to_value(entry); | |
294 | return xas_store(xas, xa_mk_value(v | DAX_LOCKED)); | |
295 | } | |
296 | ||
d2c997c0 DW |
297 | static unsigned long dax_entry_size(void *entry) |
298 | { | |
299 | if (dax_is_zero_entry(entry)) | |
300 | return 0; | |
301 | else if (dax_is_empty_entry(entry)) | |
302 | return 0; | |
303 | else if (dax_is_pmd_entry(entry)) | |
304 | return PMD_SIZE; | |
305 | else | |
306 | return PAGE_SIZE; | |
307 | } | |
308 | ||
a77d19f4 | 309 | static unsigned long dax_end_pfn(void *entry) |
d2c997c0 | 310 | { |
a77d19f4 | 311 | return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE; |
d2c997c0 DW |
312 | } |
313 | ||
314 | /* | |
315 | * Iterate through all mapped pfns represented by an entry, i.e. skip | |
316 | * 'empty' and 'zero' entries. | |
317 | */ | |
318 | #define for_each_mapped_pfn(entry, pfn) \ | |
a77d19f4 MW |
319 | for (pfn = dax_to_pfn(entry); \ |
320 | pfn < dax_end_pfn(entry); pfn++) | |
d2c997c0 | 321 | |
73449daf DW |
322 | /* |
323 | * TODO: for reflink+dax we need a way to associate a single page with | |
324 | * multiple address_space instances at different linear_page_index() | |
325 | * offsets. | |
326 | */ | |
327 | static void dax_associate_entry(void *entry, struct address_space *mapping, | |
328 | struct vm_area_struct *vma, unsigned long address) | |
d2c997c0 | 329 | { |
73449daf DW |
330 | unsigned long size = dax_entry_size(entry), pfn, index; |
331 | int i = 0; | |
d2c997c0 DW |
332 | |
333 | if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) | |
334 | return; | |
335 | ||
73449daf | 336 | index = linear_page_index(vma, address & ~(size - 1)); |
d2c997c0 DW |
337 | for_each_mapped_pfn(entry, pfn) { |
338 | struct page *page = pfn_to_page(pfn); | |
339 | ||
340 | WARN_ON_ONCE(page->mapping); | |
341 | page->mapping = mapping; | |
73449daf | 342 | page->index = index + i++; |
d2c997c0 DW |
343 | } |
344 | } | |
345 | ||
346 | static void dax_disassociate_entry(void *entry, struct address_space *mapping, | |
347 | bool trunc) | |
348 | { | |
349 | unsigned long pfn; | |
350 | ||
351 | if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) | |
352 | return; | |
353 | ||
354 | for_each_mapped_pfn(entry, pfn) { | |
355 | struct page *page = pfn_to_page(pfn); | |
356 | ||
357 | WARN_ON_ONCE(trunc && page_ref_count(page) > 1); | |
358 | WARN_ON_ONCE(page->mapping && page->mapping != mapping); | |
359 | page->mapping = NULL; | |
73449daf | 360 | page->index = 0; |
d2c997c0 DW |
361 | } |
362 | } | |
363 | ||
5fac7408 DW |
364 | static struct page *dax_busy_page(void *entry) |
365 | { | |
366 | unsigned long pfn; | |
367 | ||
368 | for_each_mapped_pfn(entry, pfn) { | |
369 | struct page *page = pfn_to_page(pfn); | |
370 | ||
371 | if (page_ref_count(page) > 1) | |
372 | return page; | |
373 | } | |
374 | return NULL; | |
375 | } | |
376 | ||
c5bbd451 MW |
377 | /* |
378 | * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page | |
379 | * @page: The page whose entry we want to lock | |
380 | * | |
381 | * Context: Process context. | |
382 | * Return: %true if the entry was locked or does not need to be locked. | |
383 | */ | |
c2a7d2a1 DW |
384 | bool dax_lock_mapping_entry(struct page *page) |
385 | { | |
9f32d221 MW |
386 | XA_STATE(xas, NULL, 0); |
387 | void *entry; | |
c5bbd451 | 388 | bool locked; |
c2a7d2a1 | 389 | |
c5bbd451 MW |
390 | /* Ensure page->mapping isn't freed while we look at it */ |
391 | rcu_read_lock(); | |
c2a7d2a1 | 392 | for (;;) { |
9f32d221 | 393 | struct address_space *mapping = READ_ONCE(page->mapping); |
c2a7d2a1 | 394 | |
c5bbd451 | 395 | locked = false; |
c93db7bb | 396 | if (!mapping || !dax_mapping(mapping)) |
c5bbd451 | 397 | break; |
c2a7d2a1 DW |
398 | |
399 | /* | |
400 | * In the device-dax case there's no need to lock, a | |
401 | * struct dev_pagemap pin is sufficient to keep the | |
402 | * inode alive, and we assume we have dev_pagemap pin | |
403 | * otherwise we would not have a valid pfn_to_page() | |
404 | * translation. | |
405 | */ | |
c5bbd451 | 406 | locked = true; |
9f32d221 | 407 | if (S_ISCHR(mapping->host->i_mode)) |
c5bbd451 | 408 | break; |
c2a7d2a1 | 409 | |
9f32d221 MW |
410 | xas.xa = &mapping->i_pages; |
411 | xas_lock_irq(&xas); | |
c2a7d2a1 | 412 | if (mapping != page->mapping) { |
9f32d221 | 413 | xas_unlock_irq(&xas); |
c2a7d2a1 DW |
414 | continue; |
415 | } | |
9f32d221 MW |
416 | xas_set(&xas, page->index); |
417 | entry = xas_load(&xas); | |
418 | if (dax_is_locked(entry)) { | |
c5bbd451 | 419 | rcu_read_unlock(); |
55e56f06 | 420 | wait_entry_unlocked(&xas, entry); |
c5bbd451 | 421 | rcu_read_lock(); |
6d7cd8c1 | 422 | continue; |
c2a7d2a1 | 423 | } |
9f32d221 MW |
424 | dax_lock_entry(&xas, entry); |
425 | xas_unlock_irq(&xas); | |
c5bbd451 | 426 | break; |
c2a7d2a1 | 427 | } |
c5bbd451 MW |
428 | rcu_read_unlock(); |
429 | return locked; | |
c2a7d2a1 DW |
430 | } |
431 | ||
432 | void dax_unlock_mapping_entry(struct page *page) | |
433 | { | |
434 | struct address_space *mapping = page->mapping; | |
9f32d221 | 435 | XA_STATE(xas, &mapping->i_pages, page->index); |
fda490d3 | 436 | void *entry; |
c2a7d2a1 | 437 | |
9f32d221 | 438 | if (S_ISCHR(mapping->host->i_mode)) |
c2a7d2a1 DW |
439 | return; |
440 | ||
fda490d3 MW |
441 | rcu_read_lock(); |
442 | entry = xas_load(&xas); | |
443 | rcu_read_unlock(); | |
444 | entry = dax_make_entry(page_to_pfn_t(page), dax_is_pmd_entry(entry)); | |
445 | dax_unlock_entry(&xas, entry); | |
c2a7d2a1 DW |
446 | } |
447 | ||
ac401cc7 | 448 | /* |
a77d19f4 MW |
449 | * Find page cache entry at given index. If it is a DAX entry, return it |
450 | * with the entry locked. If the page cache doesn't contain an entry at | |
451 | * that index, add a locked empty entry. | |
ac401cc7 | 452 | * |
3159f943 | 453 | * When requesting an entry with size DAX_PMD, grab_mapping_entry() will |
b15cd800 MW |
454 | * either return that locked entry or will return VM_FAULT_FALLBACK. |
455 | * This will happen if there are any PTE entries within the PMD range | |
456 | * that we are requesting. | |
642261ac | 457 | * |
b15cd800 MW |
458 | * We always favor PTE entries over PMD entries. There isn't a flow where we |
459 | * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD | |
460 | * insertion will fail if it finds any PTE entries already in the tree, and a | |
461 | * PTE insertion will cause an existing PMD entry to be unmapped and | |
462 | * downgraded to PTE entries. This happens for both PMD zero pages as | |
463 | * well as PMD empty entries. | |
642261ac | 464 | * |
b15cd800 MW |
465 | * The exception to this downgrade path is for PMD entries that have |
466 | * real storage backing them. We will leave these real PMD entries in | |
467 | * the tree, and PTE writes will simply dirty the entire PMD entry. | |
642261ac | 468 | * |
ac401cc7 JK |
469 | * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For |
470 | * persistent memory the benefit is doubtful. We can add that later if we can | |
471 | * show it helps. | |
b15cd800 MW |
472 | * |
473 | * On error, this function does not return an ERR_PTR. Instead it returns | |
474 | * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values | |
475 | * overlap with xarray value entries. | |
ac401cc7 | 476 | */ |
b15cd800 MW |
477 | static void *grab_mapping_entry(struct xa_state *xas, |
478 | struct address_space *mapping, unsigned long size_flag) | |
ac401cc7 | 479 | { |
b15cd800 MW |
480 | unsigned long index = xas->xa_index; |
481 | bool pmd_downgrade = false; /* splitting PMD entry into PTE entries? */ | |
482 | void *entry; | |
642261ac | 483 | |
b15cd800 MW |
484 | retry: |
485 | xas_lock_irq(xas); | |
486 | entry = get_unlocked_entry(xas); | |
91d25ba8 | 487 | |
642261ac | 488 | if (entry) { |
0e40de03 | 489 | if (!xa_is_value(entry)) { |
b15cd800 MW |
490 | xas_set_err(xas, EIO); |
491 | goto out_unlock; | |
492 | } | |
493 | ||
3159f943 | 494 | if (size_flag & DAX_PMD) { |
91d25ba8 | 495 | if (dax_is_pte_entry(entry)) { |
b15cd800 MW |
496 | put_unlocked_entry(xas, entry); |
497 | goto fallback; | |
642261ac RZ |
498 | } |
499 | } else { /* trying to grab a PTE entry */ | |
91d25ba8 | 500 | if (dax_is_pmd_entry(entry) && |
642261ac RZ |
501 | (dax_is_zero_entry(entry) || |
502 | dax_is_empty_entry(entry))) { | |
503 | pmd_downgrade = true; | |
504 | } | |
505 | } | |
506 | } | |
507 | ||
b15cd800 MW |
508 | if (pmd_downgrade) { |
509 | /* | |
510 | * Make sure 'entry' remains valid while we drop | |
511 | * the i_pages lock. | |
512 | */ | |
513 | dax_lock_entry(xas, entry); | |
642261ac | 514 | |
642261ac RZ |
515 | /* |
516 | * Besides huge zero pages the only other thing that gets | |
517 | * downgraded are empty entries which don't need to be | |
518 | * unmapped. | |
519 | */ | |
b15cd800 MW |
520 | if (dax_is_zero_entry(entry)) { |
521 | xas_unlock_irq(xas); | |
522 | unmap_mapping_pages(mapping, | |
523 | xas->xa_index & ~PG_PMD_COLOUR, | |
524 | PG_PMD_NR, false); | |
525 | xas_reset(xas); | |
526 | xas_lock_irq(xas); | |
e11f8b7b RZ |
527 | } |
528 | ||
b15cd800 MW |
529 | dax_disassociate_entry(entry, mapping, false); |
530 | xas_store(xas, NULL); /* undo the PMD join */ | |
531 | dax_wake_entry(xas, entry, true); | |
532 | mapping->nrexceptional--; | |
533 | entry = NULL; | |
534 | xas_set(xas, index); | |
535 | } | |
642261ac | 536 | |
b15cd800 MW |
537 | if (entry) { |
538 | dax_lock_entry(xas, entry); | |
539 | } else { | |
540 | entry = dax_make_entry(pfn_to_pfn_t(0), size_flag | DAX_EMPTY); | |
541 | dax_lock_entry(xas, entry); | |
542 | if (xas_error(xas)) | |
543 | goto out_unlock; | |
ac401cc7 | 544 | mapping->nrexceptional++; |
ac401cc7 | 545 | } |
b15cd800 MW |
546 | |
547 | out_unlock: | |
548 | xas_unlock_irq(xas); | |
549 | if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM)) | |
550 | goto retry; | |
551 | if (xas->xa_node == XA_ERROR(-ENOMEM)) | |
552 | return xa_mk_internal(VM_FAULT_OOM); | |
553 | if (xas_error(xas)) | |
554 | return xa_mk_internal(VM_FAULT_SIGBUS); | |
e3ad61c6 | 555 | return entry; |
b15cd800 MW |
556 | fallback: |
557 | xas_unlock_irq(xas); | |
558 | return xa_mk_internal(VM_FAULT_FALLBACK); | |
ac401cc7 JK |
559 | } |
560 | ||
5fac7408 DW |
561 | /** |
562 | * dax_layout_busy_page - find first pinned page in @mapping | |
563 | * @mapping: address space to scan for a page with ref count > 1 | |
564 | * | |
565 | * DAX requires ZONE_DEVICE mapped pages. These pages are never | |
566 | * 'onlined' to the page allocator so they are considered idle when | |
567 | * page->count == 1. A filesystem uses this interface to determine if | |
568 | * any page in the mapping is busy, i.e. for DMA, or other | |
569 | * get_user_pages() usages. | |
570 | * | |
571 | * It is expected that the filesystem is holding locks to block the | |
572 | * establishment of new mappings in this address_space. I.e. it expects | |
573 | * to be able to run unmap_mapping_range() and subsequently not race | |
574 | * mapping_mapped() becoming true. | |
575 | */ | |
576 | struct page *dax_layout_busy_page(struct address_space *mapping) | |
577 | { | |
084a8990 MW |
578 | XA_STATE(xas, &mapping->i_pages, 0); |
579 | void *entry; | |
580 | unsigned int scanned = 0; | |
5fac7408 | 581 | struct page *page = NULL; |
5fac7408 DW |
582 | |
583 | /* | |
584 | * In the 'limited' case get_user_pages() for dax is disabled. | |
585 | */ | |
586 | if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) | |
587 | return NULL; | |
588 | ||
589 | if (!dax_mapping(mapping) || !mapping_mapped(mapping)) | |
590 | return NULL; | |
591 | ||
5fac7408 DW |
592 | /* |
593 | * If we race get_user_pages_fast() here either we'll see the | |
084a8990 | 594 | * elevated page count in the iteration and wait, or |
5fac7408 DW |
595 | * get_user_pages_fast() will see that the page it took a reference |
596 | * against is no longer mapped in the page tables and bail to the | |
597 | * get_user_pages() slow path. The slow path is protected by | |
598 | * pte_lock() and pmd_lock(). New references are not taken without | |
599 | * holding those locks, and unmap_mapping_range() will not zero the | |
600 | * pte or pmd without holding the respective lock, so we are | |
601 | * guaranteed to either see new references or prevent new | |
602 | * references from being established. | |
603 | */ | |
604 | unmap_mapping_range(mapping, 0, 0, 1); | |
605 | ||
084a8990 MW |
606 | xas_lock_irq(&xas); |
607 | xas_for_each(&xas, entry, ULONG_MAX) { | |
608 | if (WARN_ON_ONCE(!xa_is_value(entry))) | |
609 | continue; | |
610 | if (unlikely(dax_is_locked(entry))) | |
611 | entry = get_unlocked_entry(&xas); | |
612 | if (entry) | |
613 | page = dax_busy_page(entry); | |
614 | put_unlocked_entry(&xas, entry); | |
5fac7408 DW |
615 | if (page) |
616 | break; | |
084a8990 MW |
617 | if (++scanned % XA_CHECK_SCHED) |
618 | continue; | |
619 | ||
620 | xas_pause(&xas); | |
621 | xas_unlock_irq(&xas); | |
622 | cond_resched(); | |
623 | xas_lock_irq(&xas); | |
5fac7408 | 624 | } |
084a8990 | 625 | xas_unlock_irq(&xas); |
5fac7408 DW |
626 | return page; |
627 | } | |
628 | EXPORT_SYMBOL_GPL(dax_layout_busy_page); | |
629 | ||
a77d19f4 | 630 | static int __dax_invalidate_entry(struct address_space *mapping, |
c6dcf52c JK |
631 | pgoff_t index, bool trunc) |
632 | { | |
07f2d89c | 633 | XA_STATE(xas, &mapping->i_pages, index); |
c6dcf52c JK |
634 | int ret = 0; |
635 | void *entry; | |
c6dcf52c | 636 | |
07f2d89c MW |
637 | xas_lock_irq(&xas); |
638 | entry = get_unlocked_entry(&xas); | |
3159f943 | 639 | if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) |
c6dcf52c JK |
640 | goto out; |
641 | if (!trunc && | |
07f2d89c MW |
642 | (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) || |
643 | xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE))) | |
c6dcf52c | 644 | goto out; |
d2c997c0 | 645 | dax_disassociate_entry(entry, mapping, trunc); |
07f2d89c | 646 | xas_store(&xas, NULL); |
c6dcf52c JK |
647 | mapping->nrexceptional--; |
648 | ret = 1; | |
649 | out: | |
07f2d89c MW |
650 | put_unlocked_entry(&xas, entry); |
651 | xas_unlock_irq(&xas); | |
c6dcf52c JK |
652 | return ret; |
653 | } | |
07f2d89c | 654 | |
ac401cc7 | 655 | /* |
3159f943 MW |
656 | * Delete DAX entry at @index from @mapping. Wait for it |
657 | * to be unlocked before deleting it. | |
ac401cc7 JK |
658 | */ |
659 | int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) | |
660 | { | |
a77d19f4 | 661 | int ret = __dax_invalidate_entry(mapping, index, true); |
ac401cc7 | 662 | |
ac401cc7 JK |
663 | /* |
664 | * This gets called from truncate / punch_hole path. As such, the caller | |
665 | * must hold locks protecting against concurrent modifications of the | |
a77d19f4 | 666 | * page cache (usually fs-private i_mmap_sem for writing). Since the |
3159f943 | 667 | * caller has seen a DAX entry for this index, we better find it |
ac401cc7 JK |
668 | * at that index as well... |
669 | */ | |
c6dcf52c JK |
670 | WARN_ON_ONCE(!ret); |
671 | return ret; | |
672 | } | |
673 | ||
c6dcf52c | 674 | /* |
3159f943 | 675 | * Invalidate DAX entry if it is clean. |
c6dcf52c JK |
676 | */ |
677 | int dax_invalidate_mapping_entry_sync(struct address_space *mapping, | |
678 | pgoff_t index) | |
679 | { | |
a77d19f4 | 680 | return __dax_invalidate_entry(mapping, index, false); |
ac401cc7 JK |
681 | } |
682 | ||
cccbce67 DW |
683 | static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev, |
684 | sector_t sector, size_t size, struct page *to, | |
685 | unsigned long vaddr) | |
f7ca90b1 | 686 | { |
cccbce67 DW |
687 | void *vto, *kaddr; |
688 | pgoff_t pgoff; | |
cccbce67 DW |
689 | long rc; |
690 | int id; | |
691 | ||
692 | rc = bdev_dax_pgoff(bdev, sector, size, &pgoff); | |
693 | if (rc) | |
694 | return rc; | |
695 | ||
696 | id = dax_read_lock(); | |
86ed913b | 697 | rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, NULL); |
cccbce67 DW |
698 | if (rc < 0) { |
699 | dax_read_unlock(id); | |
700 | return rc; | |
701 | } | |
f7ca90b1 | 702 | vto = kmap_atomic(to); |
cccbce67 | 703 | copy_user_page(vto, (void __force *)kaddr, vaddr, to); |
f7ca90b1 | 704 | kunmap_atomic(vto); |
cccbce67 | 705 | dax_read_unlock(id); |
f7ca90b1 MW |
706 | return 0; |
707 | } | |
708 | ||
642261ac RZ |
709 | /* |
710 | * By this point grab_mapping_entry() has ensured that we have a locked entry | |
711 | * of the appropriate size so we don't have to worry about downgrading PMDs to | |
712 | * PTEs. If we happen to be trying to insert a PTE and there is a PMD | |
713 | * already in the tree, we will skip the insertion and just dirty the PMD as | |
714 | * appropriate. | |
715 | */ | |
b15cd800 MW |
716 | static void *dax_insert_entry(struct xa_state *xas, |
717 | struct address_space *mapping, struct vm_fault *vmf, | |
718 | void *entry, pfn_t pfn, unsigned long flags, bool dirty) | |
9973c98e | 719 | { |
b15cd800 | 720 | void *new_entry = dax_make_entry(pfn, flags); |
9973c98e | 721 | |
f5b7b748 | 722 | if (dirty) |
d2b2a28e | 723 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); |
9973c98e | 724 | |
3159f943 | 725 | if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) { |
b15cd800 | 726 | unsigned long index = xas->xa_index; |
91d25ba8 RZ |
727 | /* we are replacing a zero page with block mapping */ |
728 | if (dax_is_pmd_entry(entry)) | |
977fbdcd | 729 | unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR, |
b15cd800 | 730 | PG_PMD_NR, false); |
91d25ba8 | 731 | else /* pte entry */ |
b15cd800 | 732 | unmap_mapping_pages(mapping, index, 1, false); |
9973c98e RZ |
733 | } |
734 | ||
b15cd800 MW |
735 | xas_reset(xas); |
736 | xas_lock_irq(xas); | |
d2c997c0 DW |
737 | if (dax_entry_size(entry) != dax_entry_size(new_entry)) { |
738 | dax_disassociate_entry(entry, mapping, false); | |
73449daf | 739 | dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address); |
d2c997c0 | 740 | } |
642261ac | 741 | |
91d25ba8 | 742 | if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { |
642261ac | 743 | /* |
a77d19f4 | 744 | * Only swap our new entry into the page cache if the current |
642261ac | 745 | * entry is a zero page or an empty entry. If a normal PTE or |
a77d19f4 | 746 | * PMD entry is already in the cache, we leave it alone. This |
642261ac RZ |
747 | * means that if we are trying to insert a PTE and the |
748 | * existing entry is a PMD, we will just leave the PMD in the | |
749 | * tree and dirty it if necessary. | |
750 | */ | |
b15cd800 MW |
751 | void *old = dax_lock_entry(xas, new_entry); |
752 | WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) | | |
753 | DAX_LOCKED)); | |
91d25ba8 | 754 | entry = new_entry; |
b15cd800 MW |
755 | } else { |
756 | xas_load(xas); /* Walk the xa_state */ | |
9973c98e | 757 | } |
91d25ba8 | 758 | |
f5b7b748 | 759 | if (dirty) |
b15cd800 | 760 | xas_set_mark(xas, PAGECACHE_TAG_DIRTY); |
91d25ba8 | 761 | |
b15cd800 | 762 | xas_unlock_irq(xas); |
91d25ba8 | 763 | return entry; |
9973c98e RZ |
764 | } |
765 | ||
a77d19f4 MW |
766 | static inline |
767 | unsigned long pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma) | |
4b4bb46d JK |
768 | { |
769 | unsigned long address; | |
770 | ||
771 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
772 | VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); | |
773 | return address; | |
774 | } | |
775 | ||
776 | /* Walk all mappings of a given index of a file and writeprotect them */ | |
a77d19f4 MW |
777 | static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index, |
778 | unsigned long pfn) | |
4b4bb46d JK |
779 | { |
780 | struct vm_area_struct *vma; | |
f729c8c9 RZ |
781 | pte_t pte, *ptep = NULL; |
782 | pmd_t *pmdp = NULL; | |
4b4bb46d | 783 | spinlock_t *ptl; |
4b4bb46d JK |
784 | |
785 | i_mmap_lock_read(mapping); | |
786 | vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) { | |
a4d1a885 | 787 | unsigned long address, start, end; |
4b4bb46d JK |
788 | |
789 | cond_resched(); | |
790 | ||
791 | if (!(vma->vm_flags & VM_SHARED)) | |
792 | continue; | |
793 | ||
794 | address = pgoff_address(index, vma); | |
a4d1a885 JG |
795 | |
796 | /* | |
797 | * Note because we provide start/end to follow_pte_pmd it will | |
798 | * call mmu_notifier_invalidate_range_start() on our behalf | |
799 | * before taking any lock. | |
800 | */ | |
801 | if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl)) | |
4b4bb46d | 802 | continue; |
4b4bb46d | 803 | |
0f10851e JG |
804 | /* |
805 | * No need to call mmu_notifier_invalidate_range() as we are | |
806 | * downgrading page table protection not changing it to point | |
807 | * to a new page. | |
808 | * | |
ad56b738 | 809 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 810 | */ |
f729c8c9 RZ |
811 | if (pmdp) { |
812 | #ifdef CONFIG_FS_DAX_PMD | |
813 | pmd_t pmd; | |
814 | ||
815 | if (pfn != pmd_pfn(*pmdp)) | |
816 | goto unlock_pmd; | |
f6f37321 | 817 | if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp)) |
f729c8c9 RZ |
818 | goto unlock_pmd; |
819 | ||
820 | flush_cache_page(vma, address, pfn); | |
821 | pmd = pmdp_huge_clear_flush(vma, address, pmdp); | |
822 | pmd = pmd_wrprotect(pmd); | |
823 | pmd = pmd_mkclean(pmd); | |
824 | set_pmd_at(vma->vm_mm, address, pmdp, pmd); | |
f729c8c9 | 825 | unlock_pmd: |
f729c8c9 | 826 | #endif |
ee190ca6 | 827 | spin_unlock(ptl); |
f729c8c9 RZ |
828 | } else { |
829 | if (pfn != pte_pfn(*ptep)) | |
830 | goto unlock_pte; | |
831 | if (!pte_dirty(*ptep) && !pte_write(*ptep)) | |
832 | goto unlock_pte; | |
833 | ||
834 | flush_cache_page(vma, address, pfn); | |
835 | pte = ptep_clear_flush(vma, address, ptep); | |
836 | pte = pte_wrprotect(pte); | |
837 | pte = pte_mkclean(pte); | |
838 | set_pte_at(vma->vm_mm, address, ptep, pte); | |
f729c8c9 RZ |
839 | unlock_pte: |
840 | pte_unmap_unlock(ptep, ptl); | |
841 | } | |
4b4bb46d | 842 | |
a4d1a885 | 843 | mmu_notifier_invalidate_range_end(vma->vm_mm, start, end); |
4b4bb46d JK |
844 | } |
845 | i_mmap_unlock_read(mapping); | |
846 | } | |
847 | ||
9fc747f6 MW |
848 | static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev, |
849 | struct address_space *mapping, void *entry) | |
9973c98e | 850 | { |
3fe0791c DW |
851 | unsigned long pfn; |
852 | long ret = 0; | |
cccbce67 | 853 | size_t size; |
9973c98e | 854 | |
9973c98e | 855 | /* |
a6abc2c0 JK |
856 | * A page got tagged dirty in DAX mapping? Something is seriously |
857 | * wrong. | |
9973c98e | 858 | */ |
3159f943 | 859 | if (WARN_ON(!xa_is_value(entry))) |
a6abc2c0 | 860 | return -EIO; |
9973c98e | 861 | |
9fc747f6 MW |
862 | if (unlikely(dax_is_locked(entry))) { |
863 | void *old_entry = entry; | |
864 | ||
865 | entry = get_unlocked_entry(xas); | |
866 | ||
867 | /* Entry got punched out / reallocated? */ | |
868 | if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) | |
869 | goto put_unlocked; | |
870 | /* | |
871 | * Entry got reallocated elsewhere? No need to writeback. | |
872 | * We have to compare pfns as we must not bail out due to | |
873 | * difference in lockbit or entry type. | |
874 | */ | |
875 | if (dax_to_pfn(old_entry) != dax_to_pfn(entry)) | |
876 | goto put_unlocked; | |
877 | if (WARN_ON_ONCE(dax_is_empty_entry(entry) || | |
878 | dax_is_zero_entry(entry))) { | |
879 | ret = -EIO; | |
880 | goto put_unlocked; | |
881 | } | |
882 | ||
883 | /* Another fsync thread may have already done this entry */ | |
884 | if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE)) | |
885 | goto put_unlocked; | |
9973c98e RZ |
886 | } |
887 | ||
a6abc2c0 | 888 | /* Lock the entry to serialize with page faults */ |
9fc747f6 MW |
889 | dax_lock_entry(xas, entry); |
890 | ||
a6abc2c0 JK |
891 | /* |
892 | * We can clear the tag now but we have to be careful so that concurrent | |
893 | * dax_writeback_one() calls for the same index cannot finish before we | |
894 | * actually flush the caches. This is achieved as the calls will look | |
b93b0163 MW |
895 | * at the entry only under the i_pages lock and once they do that |
896 | * they will see the entry locked and wait for it to unlock. | |
a6abc2c0 | 897 | */ |
9fc747f6 MW |
898 | xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE); |
899 | xas_unlock_irq(xas); | |
a6abc2c0 | 900 | |
642261ac RZ |
901 | /* |
902 | * Even if dax_writeback_mapping_range() was given a wbc->range_start | |
903 | * in the middle of a PMD, the 'index' we are given will be aligned to | |
3fe0791c DW |
904 | * the start index of the PMD, as will the pfn we pull from 'entry'. |
905 | * This allows us to flush for PMD_SIZE and not have to worry about | |
906 | * partial PMD writebacks. | |
642261ac | 907 | */ |
a77d19f4 MW |
908 | pfn = dax_to_pfn(entry); |
909 | size = PAGE_SIZE << dax_entry_order(entry); | |
cccbce67 | 910 | |
9fc747f6 | 911 | dax_entry_mkclean(mapping, xas->xa_index, pfn); |
3fe0791c | 912 | dax_flush(dax_dev, page_address(pfn_to_page(pfn)), size); |
4b4bb46d JK |
913 | /* |
914 | * After we have flushed the cache, we can clear the dirty tag. There | |
915 | * cannot be new dirty data in the pfn after the flush has completed as | |
916 | * the pfn mappings are writeprotected and fault waits for mapping | |
917 | * entry lock. | |
918 | */ | |
9fc747f6 MW |
919 | xas_reset(xas); |
920 | xas_lock_irq(xas); | |
921 | xas_store(xas, entry); | |
922 | xas_clear_mark(xas, PAGECACHE_TAG_DIRTY); | |
923 | dax_wake_entry(xas, entry, false); | |
924 | ||
925 | trace_dax_writeback_one(mapping->host, xas->xa_index, | |
926 | size >> PAGE_SHIFT); | |
9973c98e RZ |
927 | return ret; |
928 | ||
a6abc2c0 | 929 | put_unlocked: |
9fc747f6 | 930 | put_unlocked_entry(xas, entry); |
9973c98e RZ |
931 | return ret; |
932 | } | |
933 | ||
934 | /* | |
935 | * Flush the mapping to the persistent domain within the byte range of [start, | |
936 | * end]. This is required by data integrity operations to ensure file data is | |
937 | * on persistent storage prior to completion of the operation. | |
938 | */ | |
7f6d5b52 RZ |
939 | int dax_writeback_mapping_range(struct address_space *mapping, |
940 | struct block_device *bdev, struct writeback_control *wbc) | |
9973c98e | 941 | { |
9fc747f6 | 942 | XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT); |
9973c98e | 943 | struct inode *inode = mapping->host; |
9fc747f6 | 944 | pgoff_t end_index = wbc->range_end >> PAGE_SHIFT; |
cccbce67 | 945 | struct dax_device *dax_dev; |
9fc747f6 MW |
946 | void *entry; |
947 | int ret = 0; | |
948 | unsigned int scanned = 0; | |
9973c98e RZ |
949 | |
950 | if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) | |
951 | return -EIO; | |
952 | ||
7f6d5b52 RZ |
953 | if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) |
954 | return 0; | |
955 | ||
cccbce67 DW |
956 | dax_dev = dax_get_by_host(bdev->bd_disk->disk_name); |
957 | if (!dax_dev) | |
958 | return -EIO; | |
959 | ||
9fc747f6 | 960 | trace_dax_writeback_range(inode, xas.xa_index, end_index); |
9973c98e | 961 | |
9fc747f6 | 962 | tag_pages_for_writeback(mapping, xas.xa_index, end_index); |
9973c98e | 963 | |
9fc747f6 MW |
964 | xas_lock_irq(&xas); |
965 | xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) { | |
966 | ret = dax_writeback_one(&xas, dax_dev, mapping, entry); | |
967 | if (ret < 0) { | |
968 | mapping_set_error(mapping, ret); | |
9973c98e | 969 | break; |
9973c98e | 970 | } |
9fc747f6 MW |
971 | if (++scanned % XA_CHECK_SCHED) |
972 | continue; | |
973 | ||
974 | xas_pause(&xas); | |
975 | xas_unlock_irq(&xas); | |
976 | cond_resched(); | |
977 | xas_lock_irq(&xas); | |
9973c98e | 978 | } |
9fc747f6 | 979 | xas_unlock_irq(&xas); |
cccbce67 | 980 | put_dax(dax_dev); |
9fc747f6 MW |
981 | trace_dax_writeback_range_done(inode, xas.xa_index, end_index); |
982 | return ret; | |
9973c98e RZ |
983 | } |
984 | EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); | |
985 | ||
31a6f1a6 | 986 | static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) |
f7ca90b1 | 987 | { |
a3841f94 | 988 | return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9; |
31a6f1a6 JK |
989 | } |
990 | ||
5e161e40 JK |
991 | static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size, |
992 | pfn_t *pfnp) | |
f7ca90b1 | 993 | { |
31a6f1a6 | 994 | const sector_t sector = dax_iomap_sector(iomap, pos); |
cccbce67 DW |
995 | pgoff_t pgoff; |
996 | int id, rc; | |
5e161e40 | 997 | long length; |
f7ca90b1 | 998 | |
5e161e40 | 999 | rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff); |
cccbce67 DW |
1000 | if (rc) |
1001 | return rc; | |
cccbce67 | 1002 | id = dax_read_lock(); |
5e161e40 | 1003 | length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size), |
86ed913b | 1004 | NULL, pfnp); |
5e161e40 JK |
1005 | if (length < 0) { |
1006 | rc = length; | |
1007 | goto out; | |
cccbce67 | 1008 | } |
5e161e40 JK |
1009 | rc = -EINVAL; |
1010 | if (PFN_PHYS(length) < size) | |
1011 | goto out; | |
1012 | if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1)) | |
1013 | goto out; | |
1014 | /* For larger pages we need devmap */ | |
1015 | if (length > 1 && !pfn_t_devmap(*pfnp)) | |
1016 | goto out; | |
1017 | rc = 0; | |
1018 | out: | |
cccbce67 | 1019 | dax_read_unlock(id); |
5e161e40 | 1020 | return rc; |
0e3b210c | 1021 | } |
0e3b210c | 1022 | |
e30331ff | 1023 | /* |
91d25ba8 RZ |
1024 | * The user has performed a load from a hole in the file. Allocating a new |
1025 | * page in the file would cause excessive storage usage for workloads with | |
1026 | * sparse files. Instead we insert a read-only mapping of the 4k zero page. | |
1027 | * If this page is ever written to we will re-fault and change the mapping to | |
1028 | * point to real DAX storage instead. | |
e30331ff | 1029 | */ |
b15cd800 MW |
1030 | static vm_fault_t dax_load_hole(struct xa_state *xas, |
1031 | struct address_space *mapping, void **entry, | |
1032 | struct vm_fault *vmf) | |
e30331ff RZ |
1033 | { |
1034 | struct inode *inode = mapping->host; | |
91d25ba8 | 1035 | unsigned long vaddr = vmf->address; |
b90ca5cc MW |
1036 | pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr)); |
1037 | vm_fault_t ret; | |
e30331ff | 1038 | |
b15cd800 | 1039 | *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, |
3159f943 MW |
1040 | DAX_ZERO_PAGE, false); |
1041 | ||
ab77dab4 | 1042 | ret = vmf_insert_mixed(vmf->vma, vaddr, pfn); |
e30331ff RZ |
1043 | trace_dax_load_hole(inode, vmf, ret); |
1044 | return ret; | |
1045 | } | |
1046 | ||
4b0228fa VV |
1047 | static bool dax_range_is_aligned(struct block_device *bdev, |
1048 | unsigned int offset, unsigned int length) | |
1049 | { | |
1050 | unsigned short sector_size = bdev_logical_block_size(bdev); | |
1051 | ||
1052 | if (!IS_ALIGNED(offset, sector_size)) | |
1053 | return false; | |
1054 | if (!IS_ALIGNED(length, sector_size)) | |
1055 | return false; | |
1056 | ||
1057 | return true; | |
1058 | } | |
1059 | ||
cccbce67 DW |
1060 | int __dax_zero_page_range(struct block_device *bdev, |
1061 | struct dax_device *dax_dev, sector_t sector, | |
1062 | unsigned int offset, unsigned int size) | |
679c8bd3 | 1063 | { |
cccbce67 DW |
1064 | if (dax_range_is_aligned(bdev, offset, size)) { |
1065 | sector_t start_sector = sector + (offset >> 9); | |
4b0228fa VV |
1066 | |
1067 | return blkdev_issue_zeroout(bdev, start_sector, | |
53ef7d0e | 1068 | size >> 9, GFP_NOFS, 0); |
4b0228fa | 1069 | } else { |
cccbce67 DW |
1070 | pgoff_t pgoff; |
1071 | long rc, id; | |
1072 | void *kaddr; | |
cccbce67 | 1073 | |
e84b83b9 | 1074 | rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff); |
cccbce67 DW |
1075 | if (rc) |
1076 | return rc; | |
1077 | ||
1078 | id = dax_read_lock(); | |
86ed913b | 1079 | rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, NULL); |
cccbce67 DW |
1080 | if (rc < 0) { |
1081 | dax_read_unlock(id); | |
1082 | return rc; | |
1083 | } | |
81f55870 | 1084 | memset(kaddr + offset, 0, size); |
c3ca015f | 1085 | dax_flush(dax_dev, kaddr + offset, size); |
cccbce67 | 1086 | dax_read_unlock(id); |
4b0228fa | 1087 | } |
679c8bd3 CH |
1088 | return 0; |
1089 | } | |
1090 | EXPORT_SYMBOL_GPL(__dax_zero_page_range); | |
1091 | ||
a254e568 | 1092 | static loff_t |
11c59c92 | 1093 | dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, |
a254e568 CH |
1094 | struct iomap *iomap) |
1095 | { | |
cccbce67 DW |
1096 | struct block_device *bdev = iomap->bdev; |
1097 | struct dax_device *dax_dev = iomap->dax_dev; | |
a254e568 CH |
1098 | struct iov_iter *iter = data; |
1099 | loff_t end = pos + length, done = 0; | |
1100 | ssize_t ret = 0; | |
a77d4786 | 1101 | size_t xfer; |
cccbce67 | 1102 | int id; |
a254e568 CH |
1103 | |
1104 | if (iov_iter_rw(iter) == READ) { | |
1105 | end = min(end, i_size_read(inode)); | |
1106 | if (pos >= end) | |
1107 | return 0; | |
1108 | ||
1109 | if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) | |
1110 | return iov_iter_zero(min(length, end - pos), iter); | |
1111 | } | |
1112 | ||
1113 | if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) | |
1114 | return -EIO; | |
1115 | ||
e3fce68c JK |
1116 | /* |
1117 | * Write can allocate block for an area which has a hole page mapped | |
1118 | * into page tables. We have to tear down these mappings so that data | |
1119 | * written by write(2) is visible in mmap. | |
1120 | */ | |
cd656375 | 1121 | if (iomap->flags & IOMAP_F_NEW) { |
e3fce68c JK |
1122 | invalidate_inode_pages2_range(inode->i_mapping, |
1123 | pos >> PAGE_SHIFT, | |
1124 | (end - 1) >> PAGE_SHIFT); | |
1125 | } | |
1126 | ||
cccbce67 | 1127 | id = dax_read_lock(); |
a254e568 CH |
1128 | while (pos < end) { |
1129 | unsigned offset = pos & (PAGE_SIZE - 1); | |
cccbce67 DW |
1130 | const size_t size = ALIGN(length + offset, PAGE_SIZE); |
1131 | const sector_t sector = dax_iomap_sector(iomap, pos); | |
a254e568 | 1132 | ssize_t map_len; |
cccbce67 DW |
1133 | pgoff_t pgoff; |
1134 | void *kaddr; | |
a254e568 | 1135 | |
d1908f52 MH |
1136 | if (fatal_signal_pending(current)) { |
1137 | ret = -EINTR; | |
1138 | break; | |
1139 | } | |
1140 | ||
cccbce67 DW |
1141 | ret = bdev_dax_pgoff(bdev, sector, size, &pgoff); |
1142 | if (ret) | |
1143 | break; | |
1144 | ||
1145 | map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), | |
86ed913b | 1146 | &kaddr, NULL); |
a254e568 CH |
1147 | if (map_len < 0) { |
1148 | ret = map_len; | |
1149 | break; | |
1150 | } | |
1151 | ||
cccbce67 DW |
1152 | map_len = PFN_PHYS(map_len); |
1153 | kaddr += offset; | |
a254e568 CH |
1154 | map_len -= offset; |
1155 | if (map_len > end - pos) | |
1156 | map_len = end - pos; | |
1157 | ||
a2e050f5 RZ |
1158 | /* |
1159 | * The userspace address for the memory copy has already been | |
1160 | * validated via access_ok() in either vfs_read() or | |
1161 | * vfs_write(), depending on which operation we are doing. | |
1162 | */ | |
a254e568 | 1163 | if (iov_iter_rw(iter) == WRITE) |
a77d4786 | 1164 | xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr, |
fec53774 | 1165 | map_len, iter); |
a254e568 | 1166 | else |
a77d4786 | 1167 | xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr, |
b3a9a0c3 | 1168 | map_len, iter); |
a254e568 | 1169 | |
a77d4786 DW |
1170 | pos += xfer; |
1171 | length -= xfer; | |
1172 | done += xfer; | |
1173 | ||
1174 | if (xfer == 0) | |
1175 | ret = -EFAULT; | |
1176 | if (xfer < map_len) | |
1177 | break; | |
a254e568 | 1178 | } |
cccbce67 | 1179 | dax_read_unlock(id); |
a254e568 CH |
1180 | |
1181 | return done ? done : ret; | |
1182 | } | |
1183 | ||
1184 | /** | |
11c59c92 | 1185 | * dax_iomap_rw - Perform I/O to a DAX file |
a254e568 CH |
1186 | * @iocb: The control block for this I/O |
1187 | * @iter: The addresses to do I/O from or to | |
1188 | * @ops: iomap ops passed from the file system | |
1189 | * | |
1190 | * This function performs read and write operations to directly mapped | |
1191 | * persistent memory. The callers needs to take care of read/write exclusion | |
1192 | * and evicting any page cache pages in the region under I/O. | |
1193 | */ | |
1194 | ssize_t | |
11c59c92 | 1195 | dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, |
8ff6daa1 | 1196 | const struct iomap_ops *ops) |
a254e568 CH |
1197 | { |
1198 | struct address_space *mapping = iocb->ki_filp->f_mapping; | |
1199 | struct inode *inode = mapping->host; | |
1200 | loff_t pos = iocb->ki_pos, ret = 0, done = 0; | |
1201 | unsigned flags = 0; | |
1202 | ||
168316db CH |
1203 | if (iov_iter_rw(iter) == WRITE) { |
1204 | lockdep_assert_held_exclusive(&inode->i_rwsem); | |
a254e568 | 1205 | flags |= IOMAP_WRITE; |
168316db CH |
1206 | } else { |
1207 | lockdep_assert_held(&inode->i_rwsem); | |
1208 | } | |
a254e568 | 1209 | |
a254e568 CH |
1210 | while (iov_iter_count(iter)) { |
1211 | ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, | |
11c59c92 | 1212 | iter, dax_iomap_actor); |
a254e568 CH |
1213 | if (ret <= 0) |
1214 | break; | |
1215 | pos += ret; | |
1216 | done += ret; | |
1217 | } | |
1218 | ||
1219 | iocb->ki_pos += done; | |
1220 | return done ? done : ret; | |
1221 | } | |
11c59c92 | 1222 | EXPORT_SYMBOL_GPL(dax_iomap_rw); |
a7d73fe6 | 1223 | |
ab77dab4 | 1224 | static vm_fault_t dax_fault_return(int error) |
9f141d6e JK |
1225 | { |
1226 | if (error == 0) | |
1227 | return VM_FAULT_NOPAGE; | |
1228 | if (error == -ENOMEM) | |
1229 | return VM_FAULT_OOM; | |
1230 | return VM_FAULT_SIGBUS; | |
1231 | } | |
1232 | ||
aaa422c4 DW |
1233 | /* |
1234 | * MAP_SYNC on a dax mapping guarantees dirty metadata is | |
1235 | * flushed on write-faults (non-cow), but not read-faults. | |
1236 | */ | |
1237 | static bool dax_fault_is_synchronous(unsigned long flags, | |
1238 | struct vm_area_struct *vma, struct iomap *iomap) | |
1239 | { | |
1240 | return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) | |
1241 | && (iomap->flags & IOMAP_F_DIRTY); | |
1242 | } | |
1243 | ||
ab77dab4 | 1244 | static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp, |
c0b24625 | 1245 | int *iomap_errp, const struct iomap_ops *ops) |
a7d73fe6 | 1246 | { |
a0987ad5 JK |
1247 | struct vm_area_struct *vma = vmf->vma; |
1248 | struct address_space *mapping = vma->vm_file->f_mapping; | |
b15cd800 | 1249 | XA_STATE(xas, &mapping->i_pages, vmf->pgoff); |
a7d73fe6 | 1250 | struct inode *inode = mapping->host; |
1a29d85e | 1251 | unsigned long vaddr = vmf->address; |
a7d73fe6 | 1252 | loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; |
a7d73fe6 | 1253 | struct iomap iomap = { 0 }; |
9484ab1b | 1254 | unsigned flags = IOMAP_FAULT; |
a7d73fe6 | 1255 | int error, major = 0; |
d2c43ef1 | 1256 | bool write = vmf->flags & FAULT_FLAG_WRITE; |
caa51d26 | 1257 | bool sync; |
ab77dab4 | 1258 | vm_fault_t ret = 0; |
a7d73fe6 | 1259 | void *entry; |
1b5a1cb2 | 1260 | pfn_t pfn; |
a7d73fe6 | 1261 | |
ab77dab4 | 1262 | trace_dax_pte_fault(inode, vmf, ret); |
a7d73fe6 CH |
1263 | /* |
1264 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
1265 | * to hold locks serializing us with truncate / punch hole so this is | |
1266 | * a reliable test. | |
1267 | */ | |
a9c42b33 | 1268 | if (pos >= i_size_read(inode)) { |
ab77dab4 | 1269 | ret = VM_FAULT_SIGBUS; |
a9c42b33 RZ |
1270 | goto out; |
1271 | } | |
a7d73fe6 | 1272 | |
d2c43ef1 | 1273 | if (write && !vmf->cow_page) |
a7d73fe6 CH |
1274 | flags |= IOMAP_WRITE; |
1275 | ||
b15cd800 MW |
1276 | entry = grab_mapping_entry(&xas, mapping, 0); |
1277 | if (xa_is_internal(entry)) { | |
1278 | ret = xa_to_internal(entry); | |
13e451fd JK |
1279 | goto out; |
1280 | } | |
1281 | ||
e2093926 RZ |
1282 | /* |
1283 | * It is possible, particularly with mixed reads & writes to private | |
1284 | * mappings, that we have raced with a PMD fault that overlaps with | |
1285 | * the PTE we need to set up. If so just return and the fault will be | |
1286 | * retried. | |
1287 | */ | |
1288 | if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) { | |
ab77dab4 | 1289 | ret = VM_FAULT_NOPAGE; |
e2093926 RZ |
1290 | goto unlock_entry; |
1291 | } | |
1292 | ||
a7d73fe6 CH |
1293 | /* |
1294 | * Note that we don't bother to use iomap_apply here: DAX required | |
1295 | * the file system block size to be equal the page size, which means | |
1296 | * that we never have to deal with more than a single extent here. | |
1297 | */ | |
1298 | error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); | |
c0b24625 JK |
1299 | if (iomap_errp) |
1300 | *iomap_errp = error; | |
a9c42b33 | 1301 | if (error) { |
ab77dab4 | 1302 | ret = dax_fault_return(error); |
13e451fd | 1303 | goto unlock_entry; |
a9c42b33 | 1304 | } |
a7d73fe6 | 1305 | if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { |
13e451fd JK |
1306 | error = -EIO; /* fs corruption? */ |
1307 | goto error_finish_iomap; | |
a7d73fe6 CH |
1308 | } |
1309 | ||
a7d73fe6 | 1310 | if (vmf->cow_page) { |
31a6f1a6 JK |
1311 | sector_t sector = dax_iomap_sector(&iomap, pos); |
1312 | ||
a7d73fe6 CH |
1313 | switch (iomap.type) { |
1314 | case IOMAP_HOLE: | |
1315 | case IOMAP_UNWRITTEN: | |
1316 | clear_user_highpage(vmf->cow_page, vaddr); | |
1317 | break; | |
1318 | case IOMAP_MAPPED: | |
cccbce67 DW |
1319 | error = copy_user_dax(iomap.bdev, iomap.dax_dev, |
1320 | sector, PAGE_SIZE, vmf->cow_page, vaddr); | |
a7d73fe6 CH |
1321 | break; |
1322 | default: | |
1323 | WARN_ON_ONCE(1); | |
1324 | error = -EIO; | |
1325 | break; | |
1326 | } | |
1327 | ||
1328 | if (error) | |
13e451fd | 1329 | goto error_finish_iomap; |
b1aa812b JK |
1330 | |
1331 | __SetPageUptodate(vmf->cow_page); | |
ab77dab4 SJ |
1332 | ret = finish_fault(vmf); |
1333 | if (!ret) | |
1334 | ret = VM_FAULT_DONE_COW; | |
13e451fd | 1335 | goto finish_iomap; |
a7d73fe6 CH |
1336 | } |
1337 | ||
aaa422c4 | 1338 | sync = dax_fault_is_synchronous(flags, vma, &iomap); |
caa51d26 | 1339 | |
a7d73fe6 CH |
1340 | switch (iomap.type) { |
1341 | case IOMAP_MAPPED: | |
1342 | if (iomap.flags & IOMAP_F_NEW) { | |
1343 | count_vm_event(PGMAJFAULT); | |
a0987ad5 | 1344 | count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); |
a7d73fe6 CH |
1345 | major = VM_FAULT_MAJOR; |
1346 | } | |
1b5a1cb2 JK |
1347 | error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn); |
1348 | if (error < 0) | |
1349 | goto error_finish_iomap; | |
1350 | ||
b15cd800 | 1351 | entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn, |
caa51d26 | 1352 | 0, write && !sync); |
1b5a1cb2 | 1353 | |
caa51d26 JK |
1354 | /* |
1355 | * If we are doing synchronous page fault and inode needs fsync, | |
1356 | * we can insert PTE into page tables only after that happens. | |
1357 | * Skip insertion for now and return the pfn so that caller can | |
1358 | * insert it after fsync is done. | |
1359 | */ | |
1360 | if (sync) { | |
1361 | if (WARN_ON_ONCE(!pfnp)) { | |
1362 | error = -EIO; | |
1363 | goto error_finish_iomap; | |
1364 | } | |
1365 | *pfnp = pfn; | |
ab77dab4 | 1366 | ret = VM_FAULT_NEEDDSYNC | major; |
caa51d26 JK |
1367 | goto finish_iomap; |
1368 | } | |
1b5a1cb2 JK |
1369 | trace_dax_insert_mapping(inode, vmf, entry); |
1370 | if (write) | |
ab77dab4 | 1371 | ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn); |
1b5a1cb2 | 1372 | else |
ab77dab4 | 1373 | ret = vmf_insert_mixed(vma, vaddr, pfn); |
1b5a1cb2 | 1374 | |
ab77dab4 | 1375 | goto finish_iomap; |
a7d73fe6 CH |
1376 | case IOMAP_UNWRITTEN: |
1377 | case IOMAP_HOLE: | |
d2c43ef1 | 1378 | if (!write) { |
b15cd800 | 1379 | ret = dax_load_hole(&xas, mapping, &entry, vmf); |
13e451fd | 1380 | goto finish_iomap; |
1550290b | 1381 | } |
a7d73fe6 CH |
1382 | /*FALLTHRU*/ |
1383 | default: | |
1384 | WARN_ON_ONCE(1); | |
1385 | error = -EIO; | |
1386 | break; | |
1387 | } | |
1388 | ||
13e451fd | 1389 | error_finish_iomap: |
ab77dab4 | 1390 | ret = dax_fault_return(error); |
9f141d6e JK |
1391 | finish_iomap: |
1392 | if (ops->iomap_end) { | |
1393 | int copied = PAGE_SIZE; | |
1394 | ||
ab77dab4 | 1395 | if (ret & VM_FAULT_ERROR) |
9f141d6e JK |
1396 | copied = 0; |
1397 | /* | |
1398 | * The fault is done by now and there's no way back (other | |
1399 | * thread may be already happily using PTE we have installed). | |
1400 | * Just ignore error from ->iomap_end since we cannot do much | |
1401 | * with it. | |
1402 | */ | |
1403 | ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap); | |
1550290b | 1404 | } |
13e451fd | 1405 | unlock_entry: |
b15cd800 | 1406 | dax_unlock_entry(&xas, entry); |
13e451fd | 1407 | out: |
ab77dab4 SJ |
1408 | trace_dax_pte_fault_done(inode, vmf, ret); |
1409 | return ret | major; | |
a7d73fe6 | 1410 | } |
642261ac RZ |
1411 | |
1412 | #ifdef CONFIG_FS_DAX_PMD | |
b15cd800 MW |
1413 | static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, |
1414 | struct iomap *iomap, void **entry) | |
642261ac | 1415 | { |
f4200391 DJ |
1416 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
1417 | unsigned long pmd_addr = vmf->address & PMD_MASK; | |
653b2ea3 | 1418 | struct inode *inode = mapping->host; |
642261ac RZ |
1419 | struct page *zero_page; |
1420 | spinlock_t *ptl; | |
1421 | pmd_t pmd_entry; | |
3fe0791c | 1422 | pfn_t pfn; |
642261ac | 1423 | |
f4200391 | 1424 | zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); |
642261ac RZ |
1425 | |
1426 | if (unlikely(!zero_page)) | |
653b2ea3 | 1427 | goto fallback; |
642261ac | 1428 | |
3fe0791c | 1429 | pfn = page_to_pfn_t(zero_page); |
b15cd800 | 1430 | *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, |
3159f943 | 1431 | DAX_PMD | DAX_ZERO_PAGE, false); |
642261ac | 1432 | |
f4200391 DJ |
1433 | ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); |
1434 | if (!pmd_none(*(vmf->pmd))) { | |
642261ac | 1435 | spin_unlock(ptl); |
653b2ea3 | 1436 | goto fallback; |
642261ac RZ |
1437 | } |
1438 | ||
f4200391 | 1439 | pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); |
642261ac | 1440 | pmd_entry = pmd_mkhuge(pmd_entry); |
f4200391 | 1441 | set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); |
642261ac | 1442 | spin_unlock(ptl); |
b15cd800 | 1443 | trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry); |
642261ac | 1444 | return VM_FAULT_NOPAGE; |
653b2ea3 RZ |
1445 | |
1446 | fallback: | |
b15cd800 | 1447 | trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry); |
653b2ea3 | 1448 | return VM_FAULT_FALLBACK; |
642261ac RZ |
1449 | } |
1450 | ||
ab77dab4 | 1451 | static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, |
a2d58167 | 1452 | const struct iomap_ops *ops) |
642261ac | 1453 | { |
f4200391 | 1454 | struct vm_area_struct *vma = vmf->vma; |
642261ac | 1455 | struct address_space *mapping = vma->vm_file->f_mapping; |
b15cd800 | 1456 | XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER); |
d8a849e1 DJ |
1457 | unsigned long pmd_addr = vmf->address & PMD_MASK; |
1458 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
caa51d26 | 1459 | bool sync; |
9484ab1b | 1460 | unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT; |
642261ac | 1461 | struct inode *inode = mapping->host; |
ab77dab4 | 1462 | vm_fault_t result = VM_FAULT_FALLBACK; |
642261ac | 1463 | struct iomap iomap = { 0 }; |
b15cd800 | 1464 | pgoff_t max_pgoff; |
642261ac RZ |
1465 | void *entry; |
1466 | loff_t pos; | |
1467 | int error; | |
302a5e31 | 1468 | pfn_t pfn; |
642261ac | 1469 | |
282a8e03 RZ |
1470 | /* |
1471 | * Check whether offset isn't beyond end of file now. Caller is | |
1472 | * supposed to hold locks serializing us with truncate / punch hole so | |
1473 | * this is a reliable test. | |
1474 | */ | |
957ac8c4 | 1475 | max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE); |
282a8e03 | 1476 | |
f4200391 | 1477 | trace_dax_pmd_fault(inode, vmf, max_pgoff, 0); |
282a8e03 | 1478 | |
fffa281b RZ |
1479 | /* |
1480 | * Make sure that the faulting address's PMD offset (color) matches | |
1481 | * the PMD offset from the start of the file. This is necessary so | |
1482 | * that a PMD range in the page table overlaps exactly with a PMD | |
a77d19f4 | 1483 | * range in the page cache. |
fffa281b RZ |
1484 | */ |
1485 | if ((vmf->pgoff & PG_PMD_COLOUR) != | |
1486 | ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR)) | |
1487 | goto fallback; | |
1488 | ||
642261ac RZ |
1489 | /* Fall back to PTEs if we're going to COW */ |
1490 | if (write && !(vma->vm_flags & VM_SHARED)) | |
1491 | goto fallback; | |
1492 | ||
1493 | /* If the PMD would extend outside the VMA */ | |
1494 | if (pmd_addr < vma->vm_start) | |
1495 | goto fallback; | |
1496 | if ((pmd_addr + PMD_SIZE) > vma->vm_end) | |
1497 | goto fallback; | |
1498 | ||
b15cd800 | 1499 | if (xas.xa_index >= max_pgoff) { |
282a8e03 RZ |
1500 | result = VM_FAULT_SIGBUS; |
1501 | goto out; | |
1502 | } | |
642261ac RZ |
1503 | |
1504 | /* If the PMD would extend beyond the file size */ | |
b15cd800 | 1505 | if ((xas.xa_index | PG_PMD_COLOUR) >= max_pgoff) |
642261ac RZ |
1506 | goto fallback; |
1507 | ||
876f2946 | 1508 | /* |
b15cd800 MW |
1509 | * grab_mapping_entry() will make sure we get an empty PMD entry, |
1510 | * a zero PMD entry or a DAX PMD. If it can't (because a PTE | |
1511 | * entry is already in the array, for instance), it will return | |
1512 | * VM_FAULT_FALLBACK. | |
876f2946 | 1513 | */ |
b15cd800 MW |
1514 | entry = grab_mapping_entry(&xas, mapping, DAX_PMD); |
1515 | if (xa_is_internal(entry)) { | |
1516 | result = xa_to_internal(entry); | |
876f2946 | 1517 | goto fallback; |
b15cd800 | 1518 | } |
876f2946 | 1519 | |
e2093926 RZ |
1520 | /* |
1521 | * It is possible, particularly with mixed reads & writes to private | |
1522 | * mappings, that we have raced with a PTE fault that overlaps with | |
1523 | * the PMD we need to set up. If so just return and the fault will be | |
1524 | * retried. | |
1525 | */ | |
1526 | if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) && | |
1527 | !pmd_devmap(*vmf->pmd)) { | |
1528 | result = 0; | |
1529 | goto unlock_entry; | |
1530 | } | |
1531 | ||
642261ac RZ |
1532 | /* |
1533 | * Note that we don't use iomap_apply here. We aren't doing I/O, only | |
1534 | * setting up a mapping, so really we're using iomap_begin() as a way | |
1535 | * to look up our filesystem block. | |
1536 | */ | |
b15cd800 | 1537 | pos = (loff_t)xas.xa_index << PAGE_SHIFT; |
642261ac RZ |
1538 | error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap); |
1539 | if (error) | |
876f2946 | 1540 | goto unlock_entry; |
9f141d6e | 1541 | |
642261ac RZ |
1542 | if (iomap.offset + iomap.length < pos + PMD_SIZE) |
1543 | goto finish_iomap; | |
1544 | ||
aaa422c4 | 1545 | sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap); |
caa51d26 | 1546 | |
642261ac RZ |
1547 | switch (iomap.type) { |
1548 | case IOMAP_MAPPED: | |
302a5e31 JK |
1549 | error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn); |
1550 | if (error < 0) | |
1551 | goto finish_iomap; | |
1552 | ||
b15cd800 | 1553 | entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn, |
3159f943 | 1554 | DAX_PMD, write && !sync); |
302a5e31 | 1555 | |
caa51d26 JK |
1556 | /* |
1557 | * If we are doing synchronous page fault and inode needs fsync, | |
1558 | * we can insert PMD into page tables only after that happens. | |
1559 | * Skip insertion for now and return the pfn so that caller can | |
1560 | * insert it after fsync is done. | |
1561 | */ | |
1562 | if (sync) { | |
1563 | if (WARN_ON_ONCE(!pfnp)) | |
1564 | goto finish_iomap; | |
1565 | *pfnp = pfn; | |
1566 | result = VM_FAULT_NEEDDSYNC; | |
1567 | goto finish_iomap; | |
1568 | } | |
1569 | ||
302a5e31 JK |
1570 | trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry); |
1571 | result = vmf_insert_pfn_pmd(vma, vmf->address, vmf->pmd, pfn, | |
1572 | write); | |
642261ac RZ |
1573 | break; |
1574 | case IOMAP_UNWRITTEN: | |
1575 | case IOMAP_HOLE: | |
1576 | if (WARN_ON_ONCE(write)) | |
876f2946 | 1577 | break; |
b15cd800 | 1578 | result = dax_pmd_load_hole(&xas, vmf, &iomap, &entry); |
642261ac RZ |
1579 | break; |
1580 | default: | |
1581 | WARN_ON_ONCE(1); | |
1582 | break; | |
1583 | } | |
1584 | ||
1585 | finish_iomap: | |
1586 | if (ops->iomap_end) { | |
9f141d6e JK |
1587 | int copied = PMD_SIZE; |
1588 | ||
1589 | if (result == VM_FAULT_FALLBACK) | |
1590 | copied = 0; | |
1591 | /* | |
1592 | * The fault is done by now and there's no way back (other | |
1593 | * thread may be already happily using PMD we have installed). | |
1594 | * Just ignore error from ->iomap_end since we cannot do much | |
1595 | * with it. | |
1596 | */ | |
1597 | ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags, | |
1598 | &iomap); | |
642261ac | 1599 | } |
876f2946 | 1600 | unlock_entry: |
b15cd800 | 1601 | dax_unlock_entry(&xas, entry); |
642261ac RZ |
1602 | fallback: |
1603 | if (result == VM_FAULT_FALLBACK) { | |
d8a849e1 | 1604 | split_huge_pmd(vma, vmf->pmd, vmf->address); |
642261ac RZ |
1605 | count_vm_event(THP_FAULT_FALLBACK); |
1606 | } | |
282a8e03 | 1607 | out: |
f4200391 | 1608 | trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result); |
642261ac RZ |
1609 | return result; |
1610 | } | |
a2d58167 | 1611 | #else |
ab77dab4 | 1612 | static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, |
01cddfe9 | 1613 | const struct iomap_ops *ops) |
a2d58167 DJ |
1614 | { |
1615 | return VM_FAULT_FALLBACK; | |
1616 | } | |
642261ac | 1617 | #endif /* CONFIG_FS_DAX_PMD */ |
a2d58167 DJ |
1618 | |
1619 | /** | |
1620 | * dax_iomap_fault - handle a page fault on a DAX file | |
1621 | * @vmf: The description of the fault | |
cec04e8c | 1622 | * @pe_size: Size of the page to fault in |
9a0dd422 | 1623 | * @pfnp: PFN to insert for synchronous faults if fsync is required |
c0b24625 | 1624 | * @iomap_errp: Storage for detailed error code in case of error |
cec04e8c | 1625 | * @ops: Iomap ops passed from the file system |
a2d58167 DJ |
1626 | * |
1627 | * When a page fault occurs, filesystems may call this helper in | |
1628 | * their fault handler for DAX files. dax_iomap_fault() assumes the caller | |
1629 | * has done all the necessary locking for page fault to proceed | |
1630 | * successfully. | |
1631 | */ | |
ab77dab4 | 1632 | vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, |
c0b24625 | 1633 | pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops) |
a2d58167 | 1634 | { |
c791ace1 DJ |
1635 | switch (pe_size) { |
1636 | case PE_SIZE_PTE: | |
c0b24625 | 1637 | return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops); |
c791ace1 | 1638 | case PE_SIZE_PMD: |
9a0dd422 | 1639 | return dax_iomap_pmd_fault(vmf, pfnp, ops); |
a2d58167 DJ |
1640 | default: |
1641 | return VM_FAULT_FALLBACK; | |
1642 | } | |
1643 | } | |
1644 | EXPORT_SYMBOL_GPL(dax_iomap_fault); | |
71eab6df | 1645 | |
a77d19f4 | 1646 | /* |
71eab6df JK |
1647 | * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables |
1648 | * @vmf: The description of the fault | |
71eab6df | 1649 | * @pfn: PFN to insert |
cfc93c6c | 1650 | * @order: Order of entry to insert. |
71eab6df | 1651 | * |
a77d19f4 MW |
1652 | * This function inserts a writeable PTE or PMD entry into the page tables |
1653 | * for an mmaped DAX file. It also marks the page cache entry as dirty. | |
71eab6df | 1654 | */ |
cfc93c6c MW |
1655 | static vm_fault_t |
1656 | dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order) | |
71eab6df JK |
1657 | { |
1658 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; | |
cfc93c6c MW |
1659 | XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order); |
1660 | void *entry; | |
ab77dab4 | 1661 | vm_fault_t ret; |
71eab6df | 1662 | |
cfc93c6c MW |
1663 | xas_lock_irq(&xas); |
1664 | entry = get_unlocked_entry(&xas); | |
71eab6df JK |
1665 | /* Did we race with someone splitting entry or so? */ |
1666 | if (!entry || | |
cfc93c6c | 1667 | (order == 0 && !dax_is_pte_entry(entry)) || |
0e40de03 | 1668 | (order == PMD_ORDER && !dax_is_pmd_entry(entry))) { |
cfc93c6c MW |
1669 | put_unlocked_entry(&xas, entry); |
1670 | xas_unlock_irq(&xas); | |
71eab6df JK |
1671 | trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf, |
1672 | VM_FAULT_NOPAGE); | |
1673 | return VM_FAULT_NOPAGE; | |
1674 | } | |
cfc93c6c MW |
1675 | xas_set_mark(&xas, PAGECACHE_TAG_DIRTY); |
1676 | dax_lock_entry(&xas, entry); | |
1677 | xas_unlock_irq(&xas); | |
1678 | if (order == 0) | |
ab77dab4 | 1679 | ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); |
71eab6df | 1680 | #ifdef CONFIG_FS_DAX_PMD |
cfc93c6c | 1681 | else if (order == PMD_ORDER) |
ab77dab4 | 1682 | ret = vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd, |
71eab6df | 1683 | pfn, true); |
71eab6df | 1684 | #endif |
cfc93c6c | 1685 | else |
ab77dab4 | 1686 | ret = VM_FAULT_FALLBACK; |
cfc93c6c | 1687 | dax_unlock_entry(&xas, entry); |
ab77dab4 SJ |
1688 | trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret); |
1689 | return ret; | |
71eab6df JK |
1690 | } |
1691 | ||
1692 | /** | |
1693 | * dax_finish_sync_fault - finish synchronous page fault | |
1694 | * @vmf: The description of the fault | |
1695 | * @pe_size: Size of entry to be inserted | |
1696 | * @pfn: PFN to insert | |
1697 | * | |
1698 | * This function ensures that the file range touched by the page fault is | |
1699 | * stored persistently on the media and handles inserting of appropriate page | |
1700 | * table entry. | |
1701 | */ | |
ab77dab4 SJ |
1702 | vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, |
1703 | enum page_entry_size pe_size, pfn_t pfn) | |
71eab6df JK |
1704 | { |
1705 | int err; | |
1706 | loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT; | |
cfc93c6c MW |
1707 | unsigned int order = pe_order(pe_size); |
1708 | size_t len = PAGE_SIZE << order; | |
71eab6df | 1709 | |
71eab6df JK |
1710 | err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1); |
1711 | if (err) | |
1712 | return VM_FAULT_SIGBUS; | |
cfc93c6c | 1713 | return dax_insert_pfn_mkwrite(vmf, pfn, order); |
71eab6df JK |
1714 | } |
1715 | EXPORT_SYMBOL_GPL(dax_finish_sync_fault); |