]>
Commit | Line | Data |
---|---|---|
2025cf9e | 1 | // SPDX-License-Identifier: GPL-2.0-only |
d475c634 MW |
2 | /* |
3 | * fs/dax.c - Direct Access filesystem code | |
4 | * Copyright (c) 2013-2014 Intel Corporation | |
5 | * Author: Matthew Wilcox <matthew.r.wilcox@intel.com> | |
6 | * Author: Ross Zwisler <ross.zwisler@linux.intel.com> | |
d475c634 MW |
7 | */ |
8 | ||
9 | #include <linux/atomic.h> | |
10 | #include <linux/blkdev.h> | |
11 | #include <linux/buffer_head.h> | |
d77e92e2 | 12 | #include <linux/dax.h> |
d475c634 MW |
13 | #include <linux/fs.h> |
14 | #include <linux/genhd.h> | |
f7ca90b1 MW |
15 | #include <linux/highmem.h> |
16 | #include <linux/memcontrol.h> | |
17 | #include <linux/mm.h> | |
d475c634 | 18 | #include <linux/mutex.h> |
9973c98e | 19 | #include <linux/pagevec.h> |
289c6aed | 20 | #include <linux/sched.h> |
f361bf4a | 21 | #include <linux/sched/signal.h> |
d475c634 | 22 | #include <linux/uio.h> |
f7ca90b1 | 23 | #include <linux/vmstat.h> |
34c0fd54 | 24 | #include <linux/pfn_t.h> |
0e749e54 | 25 | #include <linux/sizes.h> |
4b4bb46d | 26 | #include <linux/mmu_notifier.h> |
a254e568 | 27 | #include <linux/iomap.h> |
11cf9d86 | 28 | #include <asm/pgalloc.h> |
a254e568 | 29 | #include "internal.h" |
d475c634 | 30 | |
282a8e03 RZ |
31 | #define CREATE_TRACE_POINTS |
32 | #include <trace/events/fs_dax.h> | |
33 | ||
cfc93c6c MW |
34 | static inline unsigned int pe_order(enum page_entry_size pe_size) |
35 | { | |
36 | if (pe_size == PE_SIZE_PTE) | |
37 | return PAGE_SHIFT - PAGE_SHIFT; | |
38 | if (pe_size == PE_SIZE_PMD) | |
39 | return PMD_SHIFT - PAGE_SHIFT; | |
40 | if (pe_size == PE_SIZE_PUD) | |
41 | return PUD_SHIFT - PAGE_SHIFT; | |
42 | return ~0; | |
43 | } | |
44 | ||
ac401cc7 JK |
45 | /* We choose 4096 entries - same as per-zone page wait tables */ |
46 | #define DAX_WAIT_TABLE_BITS 12 | |
47 | #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) | |
48 | ||
917f3452 RZ |
49 | /* The 'colour' (ie low bits) within a PMD of a page offset. */ |
50 | #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) | |
977fbdcd | 51 | #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT) |
917f3452 | 52 | |
cfc93c6c MW |
53 | /* The order of a PMD entry */ |
54 | #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT) | |
55 | ||
ce95ab0f | 56 | static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; |
ac401cc7 JK |
57 | |
58 | static int __init init_dax_wait_table(void) | |
59 | { | |
60 | int i; | |
61 | ||
62 | for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) | |
63 | init_waitqueue_head(wait_table + i); | |
64 | return 0; | |
65 | } | |
66 | fs_initcall(init_dax_wait_table); | |
67 | ||
527b19d0 | 68 | /* |
3159f943 MW |
69 | * DAX pagecache entries use XArray value entries so they can't be mistaken |
70 | * for pages. We use one bit for locking, one bit for the entry size (PMD) | |
71 | * and two more to tell us if the entry is a zero page or an empty entry that | |
72 | * is just used for locking. In total four special bits. | |
527b19d0 RZ |
73 | * |
74 | * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE | |
75 | * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem | |
76 | * block allocation. | |
77 | */ | |
3159f943 MW |
78 | #define DAX_SHIFT (4) |
79 | #define DAX_LOCKED (1UL << 0) | |
80 | #define DAX_PMD (1UL << 1) | |
81 | #define DAX_ZERO_PAGE (1UL << 2) | |
82 | #define DAX_EMPTY (1UL << 3) | |
527b19d0 | 83 | |
a77d19f4 | 84 | static unsigned long dax_to_pfn(void *entry) |
527b19d0 | 85 | { |
3159f943 | 86 | return xa_to_value(entry) >> DAX_SHIFT; |
527b19d0 RZ |
87 | } |
88 | ||
9f32d221 MW |
89 | static void *dax_make_entry(pfn_t pfn, unsigned long flags) |
90 | { | |
91 | return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT)); | |
92 | } | |
93 | ||
cfc93c6c MW |
94 | static bool dax_is_locked(void *entry) |
95 | { | |
96 | return xa_to_value(entry) & DAX_LOCKED; | |
97 | } | |
98 | ||
a77d19f4 | 99 | static unsigned int dax_entry_order(void *entry) |
527b19d0 | 100 | { |
3159f943 | 101 | if (xa_to_value(entry) & DAX_PMD) |
cfc93c6c | 102 | return PMD_ORDER; |
527b19d0 RZ |
103 | return 0; |
104 | } | |
105 | ||
fda490d3 | 106 | static unsigned long dax_is_pmd_entry(void *entry) |
d1a5f2b4 | 107 | { |
3159f943 | 108 | return xa_to_value(entry) & DAX_PMD; |
d1a5f2b4 DW |
109 | } |
110 | ||
fda490d3 | 111 | static bool dax_is_pte_entry(void *entry) |
d475c634 | 112 | { |
3159f943 | 113 | return !(xa_to_value(entry) & DAX_PMD); |
d475c634 MW |
114 | } |
115 | ||
642261ac | 116 | static int dax_is_zero_entry(void *entry) |
d475c634 | 117 | { |
3159f943 | 118 | return xa_to_value(entry) & DAX_ZERO_PAGE; |
d475c634 MW |
119 | } |
120 | ||
642261ac | 121 | static int dax_is_empty_entry(void *entry) |
b2e0d162 | 122 | { |
3159f943 | 123 | return xa_to_value(entry) & DAX_EMPTY; |
b2e0d162 DW |
124 | } |
125 | ||
23c84eb7 MWO |
126 | /* |
127 | * true if the entry that was found is of a smaller order than the entry | |
128 | * we were looking for | |
129 | */ | |
130 | static bool dax_is_conflict(void *entry) | |
131 | { | |
132 | return entry == XA_RETRY_ENTRY; | |
133 | } | |
134 | ||
ac401cc7 | 135 | /* |
a77d19f4 | 136 | * DAX page cache entry locking |
ac401cc7 JK |
137 | */ |
138 | struct exceptional_entry_key { | |
ec4907ff | 139 | struct xarray *xa; |
63e95b5c | 140 | pgoff_t entry_start; |
ac401cc7 JK |
141 | }; |
142 | ||
143 | struct wait_exceptional_entry_queue { | |
ac6424b9 | 144 | wait_queue_entry_t wait; |
ac401cc7 JK |
145 | struct exceptional_entry_key key; |
146 | }; | |
147 | ||
b15cd800 MW |
148 | static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas, |
149 | void *entry, struct exceptional_entry_key *key) | |
63e95b5c RZ |
150 | { |
151 | unsigned long hash; | |
b15cd800 | 152 | unsigned long index = xas->xa_index; |
63e95b5c RZ |
153 | |
154 | /* | |
155 | * If 'entry' is a PMD, align the 'index' that we use for the wait | |
156 | * queue to the start of that PMD. This ensures that all offsets in | |
157 | * the range covered by the PMD map to the same bit lock. | |
158 | */ | |
642261ac | 159 | if (dax_is_pmd_entry(entry)) |
917f3452 | 160 | index &= ~PG_PMD_COLOUR; |
b15cd800 | 161 | key->xa = xas->xa; |
63e95b5c RZ |
162 | key->entry_start = index; |
163 | ||
b15cd800 | 164 | hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS); |
63e95b5c RZ |
165 | return wait_table + hash; |
166 | } | |
167 | ||
ec4907ff MW |
168 | static int wake_exceptional_entry_func(wait_queue_entry_t *wait, |
169 | unsigned int mode, int sync, void *keyp) | |
ac401cc7 JK |
170 | { |
171 | struct exceptional_entry_key *key = keyp; | |
172 | struct wait_exceptional_entry_queue *ewait = | |
173 | container_of(wait, struct wait_exceptional_entry_queue, wait); | |
174 | ||
ec4907ff | 175 | if (key->xa != ewait->key.xa || |
63e95b5c | 176 | key->entry_start != ewait->key.entry_start) |
ac401cc7 JK |
177 | return 0; |
178 | return autoremove_wake_function(wait, mode, sync, NULL); | |
179 | } | |
180 | ||
e30331ff | 181 | /* |
b93b0163 MW |
182 | * @entry may no longer be the entry at the index in the mapping. |
183 | * The important information it's conveying is whether the entry at | |
184 | * this index used to be a PMD entry. | |
e30331ff | 185 | */ |
b15cd800 | 186 | static void dax_wake_entry(struct xa_state *xas, void *entry, bool wake_all) |
e30331ff RZ |
187 | { |
188 | struct exceptional_entry_key key; | |
189 | wait_queue_head_t *wq; | |
190 | ||
b15cd800 | 191 | wq = dax_entry_waitqueue(xas, entry, &key); |
e30331ff RZ |
192 | |
193 | /* | |
194 | * Checking for locked entry and prepare_to_wait_exclusive() happens | |
b93b0163 | 195 | * under the i_pages lock, ditto for entry handling in our callers. |
e30331ff RZ |
196 | * So at this point all tasks that could have seen our entry locked |
197 | * must be in the waitqueue and the following check will see them. | |
198 | */ | |
199 | if (waitqueue_active(wq)) | |
200 | __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); | |
201 | } | |
202 | ||
cfc93c6c MW |
203 | /* |
204 | * Look up entry in page cache, wait for it to become unlocked if it | |
205 | * is a DAX entry and return it. The caller must subsequently call | |
206 | * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry() | |
23c84eb7 MWO |
207 | * if it did. The entry returned may have a larger order than @order. |
208 | * If @order is larger than the order of the entry found in i_pages, this | |
209 | * function returns a dax_is_conflict entry. | |
cfc93c6c MW |
210 | * |
211 | * Must be called with the i_pages lock held. | |
212 | */ | |
23c84eb7 | 213 | static void *get_unlocked_entry(struct xa_state *xas, unsigned int order) |
cfc93c6c MW |
214 | { |
215 | void *entry; | |
216 | struct wait_exceptional_entry_queue ewait; | |
217 | wait_queue_head_t *wq; | |
218 | ||
219 | init_wait(&ewait.wait); | |
220 | ewait.wait.func = wake_exceptional_entry_func; | |
221 | ||
222 | for (;;) { | |
0e40de03 | 223 | entry = xas_find_conflict(xas); |
23c84eb7 MWO |
224 | if (dax_entry_order(entry) < order) |
225 | return XA_RETRY_ENTRY; | |
0e40de03 | 226 | if (!entry || WARN_ON_ONCE(!xa_is_value(entry)) || |
cfc93c6c MW |
227 | !dax_is_locked(entry)) |
228 | return entry; | |
229 | ||
b15cd800 | 230 | wq = dax_entry_waitqueue(xas, entry, &ewait.key); |
cfc93c6c MW |
231 | prepare_to_wait_exclusive(wq, &ewait.wait, |
232 | TASK_UNINTERRUPTIBLE); | |
233 | xas_unlock_irq(xas); | |
234 | xas_reset(xas); | |
235 | schedule(); | |
236 | finish_wait(wq, &ewait.wait); | |
237 | xas_lock_irq(xas); | |
238 | } | |
239 | } | |
240 | ||
55e56f06 MW |
241 | /* |
242 | * The only thing keeping the address space around is the i_pages lock | |
243 | * (it's cycled in clear_inode() after removing the entries from i_pages) | |
244 | * After we call xas_unlock_irq(), we cannot touch xas->xa. | |
245 | */ | |
246 | static void wait_entry_unlocked(struct xa_state *xas, void *entry) | |
247 | { | |
248 | struct wait_exceptional_entry_queue ewait; | |
249 | wait_queue_head_t *wq; | |
250 | ||
251 | init_wait(&ewait.wait); | |
252 | ewait.wait.func = wake_exceptional_entry_func; | |
253 | ||
254 | wq = dax_entry_waitqueue(xas, entry, &ewait.key); | |
d8a70641 DW |
255 | /* |
256 | * Unlike get_unlocked_entry() there is no guarantee that this | |
257 | * path ever successfully retrieves an unlocked entry before an | |
258 | * inode dies. Perform a non-exclusive wait in case this path | |
259 | * never successfully performs its own wake up. | |
260 | */ | |
261 | prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE); | |
55e56f06 MW |
262 | xas_unlock_irq(xas); |
263 | schedule(); | |
264 | finish_wait(wq, &ewait.wait); | |
55e56f06 MW |
265 | } |
266 | ||
cfc93c6c MW |
267 | static void put_unlocked_entry(struct xa_state *xas, void *entry) |
268 | { | |
269 | /* If we were the only waiter woken, wake the next one */ | |
23c84eb7 | 270 | if (entry && dax_is_conflict(entry)) |
cfc93c6c MW |
271 | dax_wake_entry(xas, entry, false); |
272 | } | |
273 | ||
274 | /* | |
275 | * We used the xa_state to get the entry, but then we locked the entry and | |
276 | * dropped the xa_lock, so we know the xa_state is stale and must be reset | |
277 | * before use. | |
278 | */ | |
279 | static void dax_unlock_entry(struct xa_state *xas, void *entry) | |
280 | { | |
281 | void *old; | |
282 | ||
7ae2ea7d | 283 | BUG_ON(dax_is_locked(entry)); |
cfc93c6c MW |
284 | xas_reset(xas); |
285 | xas_lock_irq(xas); | |
286 | old = xas_store(xas, entry); | |
287 | xas_unlock_irq(xas); | |
288 | BUG_ON(!dax_is_locked(old)); | |
289 | dax_wake_entry(xas, entry, false); | |
290 | } | |
291 | ||
292 | /* | |
293 | * Return: The entry stored at this location before it was locked. | |
294 | */ | |
295 | static void *dax_lock_entry(struct xa_state *xas, void *entry) | |
296 | { | |
297 | unsigned long v = xa_to_value(entry); | |
298 | return xas_store(xas, xa_mk_value(v | DAX_LOCKED)); | |
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 | ||
a77d19f4 | 313 | static unsigned long dax_end_pfn(void *entry) |
d2c997c0 | 314 | { |
a77d19f4 | 315 | return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE; |
d2c997c0 DW |
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) \ | |
a77d19f4 MW |
323 | for (pfn = dax_to_pfn(entry); \ |
324 | pfn < dax_end_pfn(entry); pfn++) | |
d2c997c0 | 325 | |
73449daf DW |
326 | /* |
327 | * TODO: for reflink+dax we need a way to associate a single page with | |
328 | * multiple address_space instances at different linear_page_index() | |
329 | * offsets. | |
330 | */ | |
331 | static void dax_associate_entry(void *entry, struct address_space *mapping, | |
332 | struct vm_area_struct *vma, unsigned long address) | |
d2c997c0 | 333 | { |
73449daf DW |
334 | unsigned long size = dax_entry_size(entry), pfn, index; |
335 | int i = 0; | |
d2c997c0 DW |
336 | |
337 | if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) | |
338 | return; | |
339 | ||
73449daf | 340 | index = linear_page_index(vma, address & ~(size - 1)); |
d2c997c0 DW |
341 | for_each_mapped_pfn(entry, pfn) { |
342 | struct page *page = pfn_to_page(pfn); | |
343 | ||
344 | WARN_ON_ONCE(page->mapping); | |
345 | page->mapping = mapping; | |
73449daf | 346 | page->index = index + i++; |
d2c997c0 DW |
347 | } |
348 | } | |
349 | ||
350 | static void dax_disassociate_entry(void *entry, struct address_space *mapping, | |
351 | bool trunc) | |
352 | { | |
353 | unsigned long pfn; | |
354 | ||
355 | if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) | |
356 | return; | |
357 | ||
358 | for_each_mapped_pfn(entry, pfn) { | |
359 | struct page *page = pfn_to_page(pfn); | |
360 | ||
361 | WARN_ON_ONCE(trunc && page_ref_count(page) > 1); | |
362 | WARN_ON_ONCE(page->mapping && page->mapping != mapping); | |
363 | page->mapping = NULL; | |
73449daf | 364 | page->index = 0; |
d2c997c0 DW |
365 | } |
366 | } | |
367 | ||
5fac7408 DW |
368 | static struct page *dax_busy_page(void *entry) |
369 | { | |
370 | unsigned long pfn; | |
371 | ||
372 | for_each_mapped_pfn(entry, pfn) { | |
373 | struct page *page = pfn_to_page(pfn); | |
374 | ||
375 | if (page_ref_count(page) > 1) | |
376 | return page; | |
377 | } | |
378 | return NULL; | |
379 | } | |
380 | ||
c5bbd451 MW |
381 | /* |
382 | * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page | |
383 | * @page: The page whose entry we want to lock | |
384 | * | |
385 | * Context: Process context. | |
27359fd6 MW |
386 | * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could |
387 | * not be locked. | |
c5bbd451 | 388 | */ |
27359fd6 | 389 | dax_entry_t dax_lock_page(struct page *page) |
c2a7d2a1 | 390 | { |
9f32d221 MW |
391 | XA_STATE(xas, NULL, 0); |
392 | void *entry; | |
c2a7d2a1 | 393 | |
c5bbd451 MW |
394 | /* Ensure page->mapping isn't freed while we look at it */ |
395 | rcu_read_lock(); | |
c2a7d2a1 | 396 | for (;;) { |
9f32d221 | 397 | struct address_space *mapping = READ_ONCE(page->mapping); |
c2a7d2a1 | 398 | |
27359fd6 | 399 | entry = NULL; |
c93db7bb | 400 | if (!mapping || !dax_mapping(mapping)) |
c5bbd451 | 401 | break; |
c2a7d2a1 DW |
402 | |
403 | /* | |
404 | * In the device-dax case there's no need to lock, a | |
405 | * struct dev_pagemap pin is sufficient to keep the | |
406 | * inode alive, and we assume we have dev_pagemap pin | |
407 | * otherwise we would not have a valid pfn_to_page() | |
408 | * translation. | |
409 | */ | |
27359fd6 | 410 | entry = (void *)~0UL; |
9f32d221 | 411 | if (S_ISCHR(mapping->host->i_mode)) |
c5bbd451 | 412 | break; |
c2a7d2a1 | 413 | |
9f32d221 MW |
414 | xas.xa = &mapping->i_pages; |
415 | xas_lock_irq(&xas); | |
c2a7d2a1 | 416 | if (mapping != page->mapping) { |
9f32d221 | 417 | xas_unlock_irq(&xas); |
c2a7d2a1 DW |
418 | continue; |
419 | } | |
9f32d221 MW |
420 | xas_set(&xas, page->index); |
421 | entry = xas_load(&xas); | |
422 | if (dax_is_locked(entry)) { | |
c5bbd451 | 423 | rcu_read_unlock(); |
55e56f06 | 424 | wait_entry_unlocked(&xas, entry); |
c5bbd451 | 425 | rcu_read_lock(); |
6d7cd8c1 | 426 | continue; |
c2a7d2a1 | 427 | } |
9f32d221 MW |
428 | dax_lock_entry(&xas, entry); |
429 | xas_unlock_irq(&xas); | |
c5bbd451 | 430 | break; |
c2a7d2a1 | 431 | } |
c5bbd451 | 432 | rcu_read_unlock(); |
27359fd6 | 433 | return (dax_entry_t)entry; |
c2a7d2a1 DW |
434 | } |
435 | ||
27359fd6 | 436 | void dax_unlock_page(struct page *page, dax_entry_t cookie) |
c2a7d2a1 DW |
437 | { |
438 | struct address_space *mapping = page->mapping; | |
9f32d221 | 439 | XA_STATE(xas, &mapping->i_pages, page->index); |
c2a7d2a1 | 440 | |
9f32d221 | 441 | if (S_ISCHR(mapping->host->i_mode)) |
c2a7d2a1 DW |
442 | return; |
443 | ||
27359fd6 | 444 | dax_unlock_entry(&xas, (void *)cookie); |
c2a7d2a1 DW |
445 | } |
446 | ||
ac401cc7 | 447 | /* |
a77d19f4 MW |
448 | * Find page cache entry at given index. If it is a DAX entry, return it |
449 | * with the entry locked. If the page cache doesn't contain an entry at | |
450 | * that index, add a locked empty entry. | |
ac401cc7 | 451 | * |
3159f943 | 452 | * When requesting an entry with size DAX_PMD, grab_mapping_entry() will |
b15cd800 MW |
453 | * either return that locked entry or will return VM_FAULT_FALLBACK. |
454 | * This will happen if there are any PTE entries within the PMD range | |
455 | * that we are requesting. | |
642261ac | 456 | * |
b15cd800 MW |
457 | * We always favor PTE entries over PMD entries. There isn't a flow where we |
458 | * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD | |
459 | * insertion will fail if it finds any PTE entries already in the tree, and a | |
460 | * PTE insertion will cause an existing PMD entry to be unmapped and | |
461 | * downgraded to PTE entries. This happens for both PMD zero pages as | |
462 | * well as PMD empty entries. | |
642261ac | 463 | * |
b15cd800 MW |
464 | * The exception to this downgrade path is for PMD entries that have |
465 | * real storage backing them. We will leave these real PMD entries in | |
466 | * the tree, and PTE writes will simply dirty the entire PMD entry. | |
642261ac | 467 | * |
ac401cc7 JK |
468 | * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For |
469 | * persistent memory the benefit is doubtful. We can add that later if we can | |
470 | * show it helps. | |
b15cd800 MW |
471 | * |
472 | * On error, this function does not return an ERR_PTR. Instead it returns | |
473 | * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values | |
474 | * overlap with xarray value entries. | |
ac401cc7 | 475 | */ |
b15cd800 | 476 | static void *grab_mapping_entry(struct xa_state *xas, |
23c84eb7 | 477 | struct address_space *mapping, unsigned int order) |
ac401cc7 | 478 | { |
b15cd800 MW |
479 | unsigned long index = xas->xa_index; |
480 | bool pmd_downgrade = false; /* splitting PMD entry into PTE entries? */ | |
481 | void *entry; | |
642261ac | 482 | |
b15cd800 MW |
483 | retry: |
484 | xas_lock_irq(xas); | |
23c84eb7 | 485 | entry = get_unlocked_entry(xas, order); |
91d25ba8 | 486 | |
642261ac | 487 | if (entry) { |
23c84eb7 MWO |
488 | if (dax_is_conflict(entry)) |
489 | goto fallback; | |
0e40de03 | 490 | if (!xa_is_value(entry)) { |
b15cd800 MW |
491 | xas_set_err(xas, EIO); |
492 | goto out_unlock; | |
493 | } | |
494 | ||
23c84eb7 | 495 | if (order == 0) { |
91d25ba8 | 496 | if (dax_is_pmd_entry(entry) && |
642261ac RZ |
497 | (dax_is_zero_entry(entry) || |
498 | dax_is_empty_entry(entry))) { | |
499 | pmd_downgrade = true; | |
500 | } | |
501 | } | |
502 | } | |
503 | ||
b15cd800 MW |
504 | if (pmd_downgrade) { |
505 | /* | |
506 | * Make sure 'entry' remains valid while we drop | |
507 | * the i_pages lock. | |
508 | */ | |
509 | dax_lock_entry(xas, entry); | |
642261ac | 510 | |
642261ac RZ |
511 | /* |
512 | * Besides huge zero pages the only other thing that gets | |
513 | * downgraded are empty entries which don't need to be | |
514 | * unmapped. | |
515 | */ | |
b15cd800 MW |
516 | if (dax_is_zero_entry(entry)) { |
517 | xas_unlock_irq(xas); | |
518 | unmap_mapping_pages(mapping, | |
519 | xas->xa_index & ~PG_PMD_COLOUR, | |
520 | PG_PMD_NR, false); | |
521 | xas_reset(xas); | |
522 | xas_lock_irq(xas); | |
e11f8b7b RZ |
523 | } |
524 | ||
b15cd800 MW |
525 | dax_disassociate_entry(entry, mapping, false); |
526 | xas_store(xas, NULL); /* undo the PMD join */ | |
527 | dax_wake_entry(xas, entry, true); | |
528 | mapping->nrexceptional--; | |
529 | entry = NULL; | |
530 | xas_set(xas, index); | |
531 | } | |
642261ac | 532 | |
b15cd800 MW |
533 | if (entry) { |
534 | dax_lock_entry(xas, entry); | |
535 | } else { | |
23c84eb7 MWO |
536 | unsigned long flags = DAX_EMPTY; |
537 | ||
538 | if (order > 0) | |
539 | flags |= DAX_PMD; | |
540 | entry = dax_make_entry(pfn_to_pfn_t(0), flags); | |
b15cd800 MW |
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))) | |
23c84eb7 | 611 | entry = get_unlocked_entry(&xas, 0); |
084a8990 MW |
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 | 637 | xas_lock_irq(&xas); |
23c84eb7 | 638 | entry = get_unlocked_entry(&xas, 0); |
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) { | |
ac46d4f3 JG |
787 | struct mmu_notifier_range range; |
788 | unsigned long address; | |
4b4bb46d JK |
789 | |
790 | cond_resched(); | |
791 | ||
792 | if (!(vma->vm_flags & VM_SHARED)) | |
793 | continue; | |
794 | ||
795 | address = pgoff_address(index, vma); | |
a4d1a885 JG |
796 | |
797 | /* | |
0cefc36b | 798 | * Note because we provide range to follow_pte_pmd it will |
a4d1a885 JG |
799 | * call mmu_notifier_invalidate_range_start() on our behalf |
800 | * before taking any lock. | |
801 | */ | |
ac46d4f3 JG |
802 | if (follow_pte_pmd(vma->vm_mm, address, &range, |
803 | &ptep, &pmdp, &ptl)) | |
4b4bb46d | 804 | continue; |
4b4bb46d | 805 | |
0f10851e JG |
806 | /* |
807 | * No need to call mmu_notifier_invalidate_range() as we are | |
808 | * downgrading page table protection not changing it to point | |
809 | * to a new page. | |
810 | * | |
ad56b738 | 811 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 812 | */ |
f729c8c9 RZ |
813 | if (pmdp) { |
814 | #ifdef CONFIG_FS_DAX_PMD | |
815 | pmd_t pmd; | |
816 | ||
817 | if (pfn != pmd_pfn(*pmdp)) | |
818 | goto unlock_pmd; | |
f6f37321 | 819 | if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp)) |
f729c8c9 RZ |
820 | goto unlock_pmd; |
821 | ||
822 | flush_cache_page(vma, address, pfn); | |
024eee0e | 823 | pmd = pmdp_invalidate(vma, address, pmdp); |
f729c8c9 RZ |
824 | pmd = pmd_wrprotect(pmd); |
825 | pmd = pmd_mkclean(pmd); | |
826 | set_pmd_at(vma->vm_mm, address, pmdp, pmd); | |
f729c8c9 | 827 | unlock_pmd: |
f729c8c9 | 828 | #endif |
ee190ca6 | 829 | spin_unlock(ptl); |
f729c8c9 RZ |
830 | } else { |
831 | if (pfn != pte_pfn(*ptep)) | |
832 | goto unlock_pte; | |
833 | if (!pte_dirty(*ptep) && !pte_write(*ptep)) | |
834 | goto unlock_pte; | |
835 | ||
836 | flush_cache_page(vma, address, pfn); | |
837 | pte = ptep_clear_flush(vma, address, ptep); | |
838 | pte = pte_wrprotect(pte); | |
839 | pte = pte_mkclean(pte); | |
840 | set_pte_at(vma->vm_mm, address, ptep, pte); | |
f729c8c9 RZ |
841 | unlock_pte: |
842 | pte_unmap_unlock(ptep, ptl); | |
843 | } | |
4b4bb46d | 844 | |
ac46d4f3 | 845 | mmu_notifier_invalidate_range_end(&range); |
4b4bb46d JK |
846 | } |
847 | i_mmap_unlock_read(mapping); | |
848 | } | |
849 | ||
9fc747f6 MW |
850 | static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev, |
851 | struct address_space *mapping, void *entry) | |
9973c98e | 852 | { |
e4b3448b | 853 | unsigned long pfn, index, count; |
3fe0791c | 854 | long ret = 0; |
9973c98e | 855 | |
9973c98e | 856 | /* |
a6abc2c0 JK |
857 | * A page got tagged dirty in DAX mapping? Something is seriously |
858 | * wrong. | |
9973c98e | 859 | */ |
3159f943 | 860 | if (WARN_ON(!xa_is_value(entry))) |
a6abc2c0 | 861 | return -EIO; |
9973c98e | 862 | |
9fc747f6 MW |
863 | if (unlikely(dax_is_locked(entry))) { |
864 | void *old_entry = entry; | |
865 | ||
23c84eb7 | 866 | entry = get_unlocked_entry(xas, 0); |
9fc747f6 MW |
867 | |
868 | /* Entry got punched out / reallocated? */ | |
869 | if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) | |
870 | goto put_unlocked; | |
871 | /* | |
872 | * Entry got reallocated elsewhere? No need to writeback. | |
873 | * We have to compare pfns as we must not bail out due to | |
874 | * difference in lockbit or entry type. | |
875 | */ | |
876 | if (dax_to_pfn(old_entry) != dax_to_pfn(entry)) | |
877 | goto put_unlocked; | |
878 | if (WARN_ON_ONCE(dax_is_empty_entry(entry) || | |
879 | dax_is_zero_entry(entry))) { | |
880 | ret = -EIO; | |
881 | goto put_unlocked; | |
882 | } | |
883 | ||
884 | /* Another fsync thread may have already done this entry */ | |
885 | if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE)) | |
886 | goto put_unlocked; | |
9973c98e RZ |
887 | } |
888 | ||
a6abc2c0 | 889 | /* Lock the entry to serialize with page faults */ |
9fc747f6 MW |
890 | dax_lock_entry(xas, entry); |
891 | ||
a6abc2c0 JK |
892 | /* |
893 | * We can clear the tag now but we have to be careful so that concurrent | |
894 | * dax_writeback_one() calls for the same index cannot finish before we | |
895 | * actually flush the caches. This is achieved as the calls will look | |
b93b0163 MW |
896 | * at the entry only under the i_pages lock and once they do that |
897 | * they will see the entry locked and wait for it to unlock. | |
a6abc2c0 | 898 | */ |
9fc747f6 MW |
899 | xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE); |
900 | xas_unlock_irq(xas); | |
a6abc2c0 | 901 | |
642261ac | 902 | /* |
e4b3448b MW |
903 | * If dax_writeback_mapping_range() was given a wbc->range_start |
904 | * in the middle of a PMD, the 'index' we use needs to be | |
905 | * aligned to the start of the PMD. | |
3fe0791c DW |
906 | * This allows us to flush for PMD_SIZE and not have to worry about |
907 | * partial PMD writebacks. | |
642261ac | 908 | */ |
a77d19f4 | 909 | pfn = dax_to_pfn(entry); |
e4b3448b MW |
910 | count = 1UL << dax_entry_order(entry); |
911 | index = xas->xa_index & ~(count - 1); | |
cccbce67 | 912 | |
e4b3448b MW |
913 | dax_entry_mkclean(mapping, index, pfn); |
914 | dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE); | |
4b4bb46d JK |
915 | /* |
916 | * After we have flushed the cache, we can clear the dirty tag. There | |
917 | * cannot be new dirty data in the pfn after the flush has completed as | |
918 | * the pfn mappings are writeprotected and fault waits for mapping | |
919 | * entry lock. | |
920 | */ | |
9fc747f6 MW |
921 | xas_reset(xas); |
922 | xas_lock_irq(xas); | |
923 | xas_store(xas, entry); | |
924 | xas_clear_mark(xas, PAGECACHE_TAG_DIRTY); | |
925 | dax_wake_entry(xas, entry, false); | |
926 | ||
e4b3448b | 927 | trace_dax_writeback_one(mapping->host, index, count); |
9973c98e RZ |
928 | return ret; |
929 | ||
a6abc2c0 | 930 | put_unlocked: |
9fc747f6 | 931 | put_unlocked_entry(xas, entry); |
9973c98e RZ |
932 | return ret; |
933 | } | |
934 | ||
935 | /* | |
936 | * Flush the mapping to the persistent domain within the byte range of [start, | |
937 | * end]. This is required by data integrity operations to ensure file data is | |
938 | * on persistent storage prior to completion of the operation. | |
939 | */ | |
7f6d5b52 RZ |
940 | int dax_writeback_mapping_range(struct address_space *mapping, |
941 | struct block_device *bdev, struct writeback_control *wbc) | |
9973c98e | 942 | { |
9fc747f6 | 943 | XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT); |
9973c98e | 944 | struct inode *inode = mapping->host; |
9fc747f6 | 945 | pgoff_t end_index = wbc->range_end >> PAGE_SHIFT; |
cccbce67 | 946 | struct dax_device *dax_dev; |
9fc747f6 MW |
947 | void *entry; |
948 | int ret = 0; | |
949 | unsigned int scanned = 0; | |
9973c98e RZ |
950 | |
951 | if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) | |
952 | return -EIO; | |
953 | ||
7f6d5b52 RZ |
954 | if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) |
955 | return 0; | |
956 | ||
cccbce67 DW |
957 | dax_dev = dax_get_by_host(bdev->bd_disk->disk_name); |
958 | if (!dax_dev) | |
959 | return -EIO; | |
960 | ||
9fc747f6 | 961 | trace_dax_writeback_range(inode, xas.xa_index, end_index); |
9973c98e | 962 | |
9fc747f6 | 963 | tag_pages_for_writeback(mapping, xas.xa_index, end_index); |
9973c98e | 964 | |
9fc747f6 MW |
965 | xas_lock_irq(&xas); |
966 | xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) { | |
967 | ret = dax_writeback_one(&xas, dax_dev, mapping, entry); | |
968 | if (ret < 0) { | |
969 | mapping_set_error(mapping, ret); | |
9973c98e | 970 | break; |
9973c98e | 971 | } |
9fc747f6 MW |
972 | if (++scanned % XA_CHECK_SCHED) |
973 | continue; | |
974 | ||
975 | xas_pause(&xas); | |
976 | xas_unlock_irq(&xas); | |
977 | cond_resched(); | |
978 | xas_lock_irq(&xas); | |
9973c98e | 979 | } |
9fc747f6 | 980 | xas_unlock_irq(&xas); |
cccbce67 | 981 | put_dax(dax_dev); |
9fc747f6 MW |
982 | trace_dax_writeback_range_done(inode, xas.xa_index, end_index); |
983 | return ret; | |
9973c98e RZ |
984 | } |
985 | EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); | |
986 | ||
31a6f1a6 | 987 | static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) |
f7ca90b1 | 988 | { |
a3841f94 | 989 | return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9; |
31a6f1a6 JK |
990 | } |
991 | ||
5e161e40 JK |
992 | static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size, |
993 | pfn_t *pfnp) | |
f7ca90b1 | 994 | { |
31a6f1a6 | 995 | const sector_t sector = dax_iomap_sector(iomap, pos); |
cccbce67 DW |
996 | pgoff_t pgoff; |
997 | int id, rc; | |
5e161e40 | 998 | long length; |
f7ca90b1 | 999 | |
5e161e40 | 1000 | rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff); |
cccbce67 DW |
1001 | if (rc) |
1002 | return rc; | |
cccbce67 | 1003 | id = dax_read_lock(); |
5e161e40 | 1004 | length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size), |
86ed913b | 1005 | NULL, pfnp); |
5e161e40 JK |
1006 | if (length < 0) { |
1007 | rc = length; | |
1008 | goto out; | |
cccbce67 | 1009 | } |
5e161e40 JK |
1010 | rc = -EINVAL; |
1011 | if (PFN_PHYS(length) < size) | |
1012 | goto out; | |
1013 | if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1)) | |
1014 | goto out; | |
1015 | /* For larger pages we need devmap */ | |
1016 | if (length > 1 && !pfn_t_devmap(*pfnp)) | |
1017 | goto out; | |
1018 | rc = 0; | |
1019 | out: | |
cccbce67 | 1020 | dax_read_unlock(id); |
5e161e40 | 1021 | return rc; |
0e3b210c | 1022 | } |
0e3b210c | 1023 | |
e30331ff | 1024 | /* |
91d25ba8 RZ |
1025 | * The user has performed a load from a hole in the file. Allocating a new |
1026 | * page in the file would cause excessive storage usage for workloads with | |
1027 | * sparse files. Instead we insert a read-only mapping of the 4k zero page. | |
1028 | * If this page is ever written to we will re-fault and change the mapping to | |
1029 | * point to real DAX storage instead. | |
e30331ff | 1030 | */ |
b15cd800 MW |
1031 | static vm_fault_t dax_load_hole(struct xa_state *xas, |
1032 | struct address_space *mapping, void **entry, | |
1033 | struct vm_fault *vmf) | |
e30331ff RZ |
1034 | { |
1035 | struct inode *inode = mapping->host; | |
91d25ba8 | 1036 | unsigned long vaddr = vmf->address; |
b90ca5cc MW |
1037 | pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr)); |
1038 | vm_fault_t ret; | |
e30331ff | 1039 | |
b15cd800 | 1040 | *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, |
3159f943 MW |
1041 | DAX_ZERO_PAGE, false); |
1042 | ||
ab77dab4 | 1043 | ret = vmf_insert_mixed(vmf->vma, vaddr, pfn); |
e30331ff RZ |
1044 | trace_dax_load_hole(inode, vmf, ret); |
1045 | return ret; | |
1046 | } | |
1047 | ||
4b0228fa VV |
1048 | static bool dax_range_is_aligned(struct block_device *bdev, |
1049 | unsigned int offset, unsigned int length) | |
1050 | { | |
1051 | unsigned short sector_size = bdev_logical_block_size(bdev); | |
1052 | ||
1053 | if (!IS_ALIGNED(offset, sector_size)) | |
1054 | return false; | |
1055 | if (!IS_ALIGNED(length, sector_size)) | |
1056 | return false; | |
1057 | ||
1058 | return true; | |
1059 | } | |
1060 | ||
cccbce67 DW |
1061 | int __dax_zero_page_range(struct block_device *bdev, |
1062 | struct dax_device *dax_dev, sector_t sector, | |
1063 | unsigned int offset, unsigned int size) | |
679c8bd3 | 1064 | { |
cccbce67 DW |
1065 | if (dax_range_is_aligned(bdev, offset, size)) { |
1066 | sector_t start_sector = sector + (offset >> 9); | |
4b0228fa VV |
1067 | |
1068 | return blkdev_issue_zeroout(bdev, start_sector, | |
53ef7d0e | 1069 | size >> 9, GFP_NOFS, 0); |
4b0228fa | 1070 | } else { |
cccbce67 DW |
1071 | pgoff_t pgoff; |
1072 | long rc, id; | |
1073 | void *kaddr; | |
cccbce67 | 1074 | |
e84b83b9 | 1075 | rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff); |
cccbce67 DW |
1076 | if (rc) |
1077 | return rc; | |
1078 | ||
1079 | id = dax_read_lock(); | |
86ed913b | 1080 | rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, NULL); |
cccbce67 DW |
1081 | if (rc < 0) { |
1082 | dax_read_unlock(id); | |
1083 | return rc; | |
1084 | } | |
81f55870 | 1085 | memset(kaddr + offset, 0, size); |
c3ca015f | 1086 | dax_flush(dax_dev, kaddr + offset, size); |
cccbce67 | 1087 | dax_read_unlock(id); |
4b0228fa | 1088 | } |
679c8bd3 CH |
1089 | return 0; |
1090 | } | |
1091 | EXPORT_SYMBOL_GPL(__dax_zero_page_range); | |
1092 | ||
a254e568 | 1093 | static loff_t |
11c59c92 | 1094 | dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, |
a254e568 CH |
1095 | struct iomap *iomap) |
1096 | { | |
cccbce67 DW |
1097 | struct block_device *bdev = iomap->bdev; |
1098 | struct dax_device *dax_dev = iomap->dax_dev; | |
a254e568 CH |
1099 | struct iov_iter *iter = data; |
1100 | loff_t end = pos + length, done = 0; | |
1101 | ssize_t ret = 0; | |
a77d4786 | 1102 | size_t xfer; |
cccbce67 | 1103 | int id; |
a254e568 CH |
1104 | |
1105 | if (iov_iter_rw(iter) == READ) { | |
1106 | end = min(end, i_size_read(inode)); | |
1107 | if (pos >= end) | |
1108 | return 0; | |
1109 | ||
1110 | if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) | |
1111 | return iov_iter_zero(min(length, end - pos), iter); | |
1112 | } | |
1113 | ||
1114 | if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) | |
1115 | return -EIO; | |
1116 | ||
e3fce68c JK |
1117 | /* |
1118 | * Write can allocate block for an area which has a hole page mapped | |
1119 | * into page tables. We have to tear down these mappings so that data | |
1120 | * written by write(2) is visible in mmap. | |
1121 | */ | |
cd656375 | 1122 | if (iomap->flags & IOMAP_F_NEW) { |
e3fce68c JK |
1123 | invalidate_inode_pages2_range(inode->i_mapping, |
1124 | pos >> PAGE_SHIFT, | |
1125 | (end - 1) >> PAGE_SHIFT); | |
1126 | } | |
1127 | ||
cccbce67 | 1128 | id = dax_read_lock(); |
a254e568 CH |
1129 | while (pos < end) { |
1130 | unsigned offset = pos & (PAGE_SIZE - 1); | |
cccbce67 DW |
1131 | const size_t size = ALIGN(length + offset, PAGE_SIZE); |
1132 | const sector_t sector = dax_iomap_sector(iomap, pos); | |
a254e568 | 1133 | ssize_t map_len; |
cccbce67 DW |
1134 | pgoff_t pgoff; |
1135 | void *kaddr; | |
a254e568 | 1136 | |
d1908f52 MH |
1137 | if (fatal_signal_pending(current)) { |
1138 | ret = -EINTR; | |
1139 | break; | |
1140 | } | |
1141 | ||
cccbce67 DW |
1142 | ret = bdev_dax_pgoff(bdev, sector, size, &pgoff); |
1143 | if (ret) | |
1144 | break; | |
1145 | ||
1146 | map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), | |
86ed913b | 1147 | &kaddr, NULL); |
a254e568 CH |
1148 | if (map_len < 0) { |
1149 | ret = map_len; | |
1150 | break; | |
1151 | } | |
1152 | ||
cccbce67 DW |
1153 | map_len = PFN_PHYS(map_len); |
1154 | kaddr += offset; | |
a254e568 CH |
1155 | map_len -= offset; |
1156 | if (map_len > end - pos) | |
1157 | map_len = end - pos; | |
1158 | ||
a2e050f5 RZ |
1159 | /* |
1160 | * The userspace address for the memory copy has already been | |
1161 | * validated via access_ok() in either vfs_read() or | |
1162 | * vfs_write(), depending on which operation we are doing. | |
1163 | */ | |
a254e568 | 1164 | if (iov_iter_rw(iter) == WRITE) |
a77d4786 | 1165 | xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr, |
fec53774 | 1166 | map_len, iter); |
a254e568 | 1167 | else |
a77d4786 | 1168 | xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr, |
b3a9a0c3 | 1169 | map_len, iter); |
a254e568 | 1170 | |
a77d4786 DW |
1171 | pos += xfer; |
1172 | length -= xfer; | |
1173 | done += xfer; | |
1174 | ||
1175 | if (xfer == 0) | |
1176 | ret = -EFAULT; | |
1177 | if (xfer < map_len) | |
1178 | break; | |
a254e568 | 1179 | } |
cccbce67 | 1180 | dax_read_unlock(id); |
a254e568 CH |
1181 | |
1182 | return done ? done : ret; | |
1183 | } | |
1184 | ||
1185 | /** | |
11c59c92 | 1186 | * dax_iomap_rw - Perform I/O to a DAX file |
a254e568 CH |
1187 | * @iocb: The control block for this I/O |
1188 | * @iter: The addresses to do I/O from or to | |
1189 | * @ops: iomap ops passed from the file system | |
1190 | * | |
1191 | * This function performs read and write operations to directly mapped | |
1192 | * persistent memory. The callers needs to take care of read/write exclusion | |
1193 | * and evicting any page cache pages in the region under I/O. | |
1194 | */ | |
1195 | ssize_t | |
11c59c92 | 1196 | dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, |
8ff6daa1 | 1197 | const struct iomap_ops *ops) |
a254e568 CH |
1198 | { |
1199 | struct address_space *mapping = iocb->ki_filp->f_mapping; | |
1200 | struct inode *inode = mapping->host; | |
1201 | loff_t pos = iocb->ki_pos, ret = 0, done = 0; | |
1202 | unsigned flags = 0; | |
1203 | ||
168316db CH |
1204 | if (iov_iter_rw(iter) == WRITE) { |
1205 | lockdep_assert_held_exclusive(&inode->i_rwsem); | |
a254e568 | 1206 | flags |= IOMAP_WRITE; |
168316db CH |
1207 | } else { |
1208 | lockdep_assert_held(&inode->i_rwsem); | |
1209 | } | |
a254e568 | 1210 | |
a254e568 CH |
1211 | while (iov_iter_count(iter)) { |
1212 | ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, | |
11c59c92 | 1213 | iter, dax_iomap_actor); |
a254e568 CH |
1214 | if (ret <= 0) |
1215 | break; | |
1216 | pos += ret; | |
1217 | done += ret; | |
1218 | } | |
1219 | ||
1220 | iocb->ki_pos += done; | |
1221 | return done ? done : ret; | |
1222 | } | |
11c59c92 | 1223 | EXPORT_SYMBOL_GPL(dax_iomap_rw); |
a7d73fe6 | 1224 | |
ab77dab4 | 1225 | static vm_fault_t dax_fault_return(int error) |
9f141d6e JK |
1226 | { |
1227 | if (error == 0) | |
1228 | return VM_FAULT_NOPAGE; | |
c9aed74e | 1229 | return vmf_error(error); |
9f141d6e JK |
1230 | } |
1231 | ||
aaa422c4 DW |
1232 | /* |
1233 | * MAP_SYNC on a dax mapping guarantees dirty metadata is | |
1234 | * flushed on write-faults (non-cow), but not read-faults. | |
1235 | */ | |
1236 | static bool dax_fault_is_synchronous(unsigned long flags, | |
1237 | struct vm_area_struct *vma, struct iomap *iomap) | |
1238 | { | |
1239 | return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) | |
1240 | && (iomap->flags & IOMAP_F_DIRTY); | |
1241 | } | |
1242 | ||
ab77dab4 | 1243 | static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp, |
c0b24625 | 1244 | int *iomap_errp, const struct iomap_ops *ops) |
a7d73fe6 | 1245 | { |
a0987ad5 JK |
1246 | struct vm_area_struct *vma = vmf->vma; |
1247 | struct address_space *mapping = vma->vm_file->f_mapping; | |
b15cd800 | 1248 | XA_STATE(xas, &mapping->i_pages, vmf->pgoff); |
a7d73fe6 | 1249 | struct inode *inode = mapping->host; |
1a29d85e | 1250 | unsigned long vaddr = vmf->address; |
a7d73fe6 | 1251 | loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; |
a7d73fe6 | 1252 | struct iomap iomap = { 0 }; |
9484ab1b | 1253 | unsigned flags = IOMAP_FAULT; |
a7d73fe6 | 1254 | int error, major = 0; |
d2c43ef1 | 1255 | bool write = vmf->flags & FAULT_FLAG_WRITE; |
caa51d26 | 1256 | bool sync; |
ab77dab4 | 1257 | vm_fault_t ret = 0; |
a7d73fe6 | 1258 | void *entry; |
1b5a1cb2 | 1259 | pfn_t pfn; |
a7d73fe6 | 1260 | |
ab77dab4 | 1261 | trace_dax_pte_fault(inode, vmf, ret); |
a7d73fe6 CH |
1262 | /* |
1263 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
1264 | * to hold locks serializing us with truncate / punch hole so this is | |
1265 | * a reliable test. | |
1266 | */ | |
a9c42b33 | 1267 | if (pos >= i_size_read(inode)) { |
ab77dab4 | 1268 | ret = VM_FAULT_SIGBUS; |
a9c42b33 RZ |
1269 | goto out; |
1270 | } | |
a7d73fe6 | 1271 | |
d2c43ef1 | 1272 | if (write && !vmf->cow_page) |
a7d73fe6 CH |
1273 | flags |= IOMAP_WRITE; |
1274 | ||
b15cd800 MW |
1275 | entry = grab_mapping_entry(&xas, mapping, 0); |
1276 | if (xa_is_internal(entry)) { | |
1277 | ret = xa_to_internal(entry); | |
13e451fd JK |
1278 | goto out; |
1279 | } | |
1280 | ||
e2093926 RZ |
1281 | /* |
1282 | * It is possible, particularly with mixed reads & writes to private | |
1283 | * mappings, that we have raced with a PMD fault that overlaps with | |
1284 | * the PTE we need to set up. If so just return and the fault will be | |
1285 | * retried. | |
1286 | */ | |
1287 | if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) { | |
ab77dab4 | 1288 | ret = VM_FAULT_NOPAGE; |
e2093926 RZ |
1289 | goto unlock_entry; |
1290 | } | |
1291 | ||
a7d73fe6 CH |
1292 | /* |
1293 | * Note that we don't bother to use iomap_apply here: DAX required | |
1294 | * the file system block size to be equal the page size, which means | |
1295 | * that we never have to deal with more than a single extent here. | |
1296 | */ | |
1297 | error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); | |
c0b24625 JK |
1298 | if (iomap_errp) |
1299 | *iomap_errp = error; | |
a9c42b33 | 1300 | if (error) { |
ab77dab4 | 1301 | ret = dax_fault_return(error); |
13e451fd | 1302 | goto unlock_entry; |
a9c42b33 | 1303 | } |
a7d73fe6 | 1304 | if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { |
13e451fd JK |
1305 | error = -EIO; /* fs corruption? */ |
1306 | goto error_finish_iomap; | |
a7d73fe6 CH |
1307 | } |
1308 | ||
a7d73fe6 | 1309 | if (vmf->cow_page) { |
31a6f1a6 JK |
1310 | sector_t sector = dax_iomap_sector(&iomap, pos); |
1311 | ||
a7d73fe6 CH |
1312 | switch (iomap.type) { |
1313 | case IOMAP_HOLE: | |
1314 | case IOMAP_UNWRITTEN: | |
1315 | clear_user_highpage(vmf->cow_page, vaddr); | |
1316 | break; | |
1317 | case IOMAP_MAPPED: | |
cccbce67 DW |
1318 | error = copy_user_dax(iomap.bdev, iomap.dax_dev, |
1319 | sector, PAGE_SIZE, vmf->cow_page, vaddr); | |
a7d73fe6 CH |
1320 | break; |
1321 | default: | |
1322 | WARN_ON_ONCE(1); | |
1323 | error = -EIO; | |
1324 | break; | |
1325 | } | |
1326 | ||
1327 | if (error) | |
13e451fd | 1328 | goto error_finish_iomap; |
b1aa812b JK |
1329 | |
1330 | __SetPageUptodate(vmf->cow_page); | |
ab77dab4 SJ |
1331 | ret = finish_fault(vmf); |
1332 | if (!ret) | |
1333 | ret = VM_FAULT_DONE_COW; | |
13e451fd | 1334 | goto finish_iomap; |
a7d73fe6 CH |
1335 | } |
1336 | ||
aaa422c4 | 1337 | sync = dax_fault_is_synchronous(flags, vma, &iomap); |
caa51d26 | 1338 | |
a7d73fe6 CH |
1339 | switch (iomap.type) { |
1340 | case IOMAP_MAPPED: | |
1341 | if (iomap.flags & IOMAP_F_NEW) { | |
1342 | count_vm_event(PGMAJFAULT); | |
a0987ad5 | 1343 | count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); |
a7d73fe6 CH |
1344 | major = VM_FAULT_MAJOR; |
1345 | } | |
1b5a1cb2 JK |
1346 | error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn); |
1347 | if (error < 0) | |
1348 | goto error_finish_iomap; | |
1349 | ||
b15cd800 | 1350 | entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn, |
caa51d26 | 1351 | 0, write && !sync); |
1b5a1cb2 | 1352 | |
caa51d26 JK |
1353 | /* |
1354 | * If we are doing synchronous page fault and inode needs fsync, | |
1355 | * we can insert PTE into page tables only after that happens. | |
1356 | * Skip insertion for now and return the pfn so that caller can | |
1357 | * insert it after fsync is done. | |
1358 | */ | |
1359 | if (sync) { | |
1360 | if (WARN_ON_ONCE(!pfnp)) { | |
1361 | error = -EIO; | |
1362 | goto error_finish_iomap; | |
1363 | } | |
1364 | *pfnp = pfn; | |
ab77dab4 | 1365 | ret = VM_FAULT_NEEDDSYNC | major; |
caa51d26 JK |
1366 | goto finish_iomap; |
1367 | } | |
1b5a1cb2 JK |
1368 | trace_dax_insert_mapping(inode, vmf, entry); |
1369 | if (write) | |
ab77dab4 | 1370 | ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn); |
1b5a1cb2 | 1371 | else |
ab77dab4 | 1372 | ret = vmf_insert_mixed(vma, vaddr, pfn); |
1b5a1cb2 | 1373 | |
ab77dab4 | 1374 | goto finish_iomap; |
a7d73fe6 CH |
1375 | case IOMAP_UNWRITTEN: |
1376 | case IOMAP_HOLE: | |
d2c43ef1 | 1377 | if (!write) { |
b15cd800 | 1378 | ret = dax_load_hole(&xas, mapping, &entry, vmf); |
13e451fd | 1379 | goto finish_iomap; |
1550290b | 1380 | } |
a7d73fe6 CH |
1381 | /*FALLTHRU*/ |
1382 | default: | |
1383 | WARN_ON_ONCE(1); | |
1384 | error = -EIO; | |
1385 | break; | |
1386 | } | |
1387 | ||
13e451fd | 1388 | error_finish_iomap: |
ab77dab4 | 1389 | ret = dax_fault_return(error); |
9f141d6e JK |
1390 | finish_iomap: |
1391 | if (ops->iomap_end) { | |
1392 | int copied = PAGE_SIZE; | |
1393 | ||
ab77dab4 | 1394 | if (ret & VM_FAULT_ERROR) |
9f141d6e JK |
1395 | copied = 0; |
1396 | /* | |
1397 | * The fault is done by now and there's no way back (other | |
1398 | * thread may be already happily using PTE we have installed). | |
1399 | * Just ignore error from ->iomap_end since we cannot do much | |
1400 | * with it. | |
1401 | */ | |
1402 | ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap); | |
1550290b | 1403 | } |
13e451fd | 1404 | unlock_entry: |
b15cd800 | 1405 | dax_unlock_entry(&xas, entry); |
13e451fd | 1406 | out: |
ab77dab4 SJ |
1407 | trace_dax_pte_fault_done(inode, vmf, ret); |
1408 | return ret | major; | |
a7d73fe6 | 1409 | } |
642261ac RZ |
1410 | |
1411 | #ifdef CONFIG_FS_DAX_PMD | |
b15cd800 MW |
1412 | static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, |
1413 | struct iomap *iomap, void **entry) | |
642261ac | 1414 | { |
f4200391 DJ |
1415 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
1416 | unsigned long pmd_addr = vmf->address & PMD_MASK; | |
11cf9d86 | 1417 | struct vm_area_struct *vma = vmf->vma; |
653b2ea3 | 1418 | struct inode *inode = mapping->host; |
11cf9d86 | 1419 | pgtable_t pgtable = NULL; |
642261ac RZ |
1420 | struct page *zero_page; |
1421 | spinlock_t *ptl; | |
1422 | pmd_t pmd_entry; | |
3fe0791c | 1423 | pfn_t pfn; |
642261ac | 1424 | |
f4200391 | 1425 | zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); |
642261ac RZ |
1426 | |
1427 | if (unlikely(!zero_page)) | |
653b2ea3 | 1428 | goto fallback; |
642261ac | 1429 | |
3fe0791c | 1430 | pfn = page_to_pfn_t(zero_page); |
b15cd800 | 1431 | *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, |
3159f943 | 1432 | DAX_PMD | DAX_ZERO_PAGE, false); |
642261ac | 1433 | |
11cf9d86 AK |
1434 | if (arch_needs_pgtable_deposit()) { |
1435 | pgtable = pte_alloc_one(vma->vm_mm); | |
1436 | if (!pgtable) | |
1437 | return VM_FAULT_OOM; | |
1438 | } | |
1439 | ||
f4200391 DJ |
1440 | ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); |
1441 | if (!pmd_none(*(vmf->pmd))) { | |
642261ac | 1442 | spin_unlock(ptl); |
653b2ea3 | 1443 | goto fallback; |
642261ac RZ |
1444 | } |
1445 | ||
11cf9d86 AK |
1446 | if (pgtable) { |
1447 | pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); | |
1448 | mm_inc_nr_ptes(vma->vm_mm); | |
1449 | } | |
f4200391 | 1450 | pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); |
642261ac | 1451 | pmd_entry = pmd_mkhuge(pmd_entry); |
f4200391 | 1452 | set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); |
642261ac | 1453 | spin_unlock(ptl); |
b15cd800 | 1454 | trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry); |
642261ac | 1455 | return VM_FAULT_NOPAGE; |
653b2ea3 RZ |
1456 | |
1457 | fallback: | |
11cf9d86 AK |
1458 | if (pgtable) |
1459 | pte_free(vma->vm_mm, pgtable); | |
b15cd800 | 1460 | trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry); |
653b2ea3 | 1461 | return VM_FAULT_FALLBACK; |
642261ac RZ |
1462 | } |
1463 | ||
ab77dab4 | 1464 | static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, |
a2d58167 | 1465 | const struct iomap_ops *ops) |
642261ac | 1466 | { |
f4200391 | 1467 | struct vm_area_struct *vma = vmf->vma; |
642261ac | 1468 | struct address_space *mapping = vma->vm_file->f_mapping; |
b15cd800 | 1469 | XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER); |
d8a849e1 DJ |
1470 | unsigned long pmd_addr = vmf->address & PMD_MASK; |
1471 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
caa51d26 | 1472 | bool sync; |
9484ab1b | 1473 | unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT; |
642261ac | 1474 | struct inode *inode = mapping->host; |
ab77dab4 | 1475 | vm_fault_t result = VM_FAULT_FALLBACK; |
642261ac | 1476 | struct iomap iomap = { 0 }; |
b15cd800 | 1477 | pgoff_t max_pgoff; |
642261ac RZ |
1478 | void *entry; |
1479 | loff_t pos; | |
1480 | int error; | |
302a5e31 | 1481 | pfn_t pfn; |
642261ac | 1482 | |
282a8e03 RZ |
1483 | /* |
1484 | * Check whether offset isn't beyond end of file now. Caller is | |
1485 | * supposed to hold locks serializing us with truncate / punch hole so | |
1486 | * this is a reliable test. | |
1487 | */ | |
957ac8c4 | 1488 | max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE); |
282a8e03 | 1489 | |
f4200391 | 1490 | trace_dax_pmd_fault(inode, vmf, max_pgoff, 0); |
282a8e03 | 1491 | |
fffa281b RZ |
1492 | /* |
1493 | * Make sure that the faulting address's PMD offset (color) matches | |
1494 | * the PMD offset from the start of the file. This is necessary so | |
1495 | * that a PMD range in the page table overlaps exactly with a PMD | |
a77d19f4 | 1496 | * range in the page cache. |
fffa281b RZ |
1497 | */ |
1498 | if ((vmf->pgoff & PG_PMD_COLOUR) != | |
1499 | ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR)) | |
1500 | goto fallback; | |
1501 | ||
642261ac RZ |
1502 | /* Fall back to PTEs if we're going to COW */ |
1503 | if (write && !(vma->vm_flags & VM_SHARED)) | |
1504 | goto fallback; | |
1505 | ||
1506 | /* If the PMD would extend outside the VMA */ | |
1507 | if (pmd_addr < vma->vm_start) | |
1508 | goto fallback; | |
1509 | if ((pmd_addr + PMD_SIZE) > vma->vm_end) | |
1510 | goto fallback; | |
1511 | ||
b15cd800 | 1512 | if (xas.xa_index >= max_pgoff) { |
282a8e03 RZ |
1513 | result = VM_FAULT_SIGBUS; |
1514 | goto out; | |
1515 | } | |
642261ac RZ |
1516 | |
1517 | /* If the PMD would extend beyond the file size */ | |
b15cd800 | 1518 | if ((xas.xa_index | PG_PMD_COLOUR) >= max_pgoff) |
642261ac RZ |
1519 | goto fallback; |
1520 | ||
876f2946 | 1521 | /* |
b15cd800 MW |
1522 | * grab_mapping_entry() will make sure we get an empty PMD entry, |
1523 | * a zero PMD entry or a DAX PMD. If it can't (because a PTE | |
1524 | * entry is already in the array, for instance), it will return | |
1525 | * VM_FAULT_FALLBACK. | |
876f2946 | 1526 | */ |
23c84eb7 | 1527 | entry = grab_mapping_entry(&xas, mapping, PMD_ORDER); |
b15cd800 MW |
1528 | if (xa_is_internal(entry)) { |
1529 | result = xa_to_internal(entry); | |
876f2946 | 1530 | goto fallback; |
b15cd800 | 1531 | } |
876f2946 | 1532 | |
e2093926 RZ |
1533 | /* |
1534 | * It is possible, particularly with mixed reads & writes to private | |
1535 | * mappings, that we have raced with a PTE fault that overlaps with | |
1536 | * the PMD we need to set up. If so just return and the fault will be | |
1537 | * retried. | |
1538 | */ | |
1539 | if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) && | |
1540 | !pmd_devmap(*vmf->pmd)) { | |
1541 | result = 0; | |
1542 | goto unlock_entry; | |
1543 | } | |
1544 | ||
642261ac RZ |
1545 | /* |
1546 | * Note that we don't use iomap_apply here. We aren't doing I/O, only | |
1547 | * setting up a mapping, so really we're using iomap_begin() as a way | |
1548 | * to look up our filesystem block. | |
1549 | */ | |
b15cd800 | 1550 | pos = (loff_t)xas.xa_index << PAGE_SHIFT; |
642261ac RZ |
1551 | error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap); |
1552 | if (error) | |
876f2946 | 1553 | goto unlock_entry; |
9f141d6e | 1554 | |
642261ac RZ |
1555 | if (iomap.offset + iomap.length < pos + PMD_SIZE) |
1556 | goto finish_iomap; | |
1557 | ||
aaa422c4 | 1558 | sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap); |
caa51d26 | 1559 | |
642261ac RZ |
1560 | switch (iomap.type) { |
1561 | case IOMAP_MAPPED: | |
302a5e31 JK |
1562 | error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn); |
1563 | if (error < 0) | |
1564 | goto finish_iomap; | |
1565 | ||
b15cd800 | 1566 | entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn, |
3159f943 | 1567 | DAX_PMD, write && !sync); |
302a5e31 | 1568 | |
caa51d26 JK |
1569 | /* |
1570 | * If we are doing synchronous page fault and inode needs fsync, | |
1571 | * we can insert PMD into page tables only after that happens. | |
1572 | * Skip insertion for now and return the pfn so that caller can | |
1573 | * insert it after fsync is done. | |
1574 | */ | |
1575 | if (sync) { | |
1576 | if (WARN_ON_ONCE(!pfnp)) | |
1577 | goto finish_iomap; | |
1578 | *pfnp = pfn; | |
1579 | result = VM_FAULT_NEEDDSYNC; | |
1580 | goto finish_iomap; | |
1581 | } | |
1582 | ||
302a5e31 | 1583 | trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry); |
fce86ff5 | 1584 | result = vmf_insert_pfn_pmd(vmf, pfn, write); |
642261ac RZ |
1585 | break; |
1586 | case IOMAP_UNWRITTEN: | |
1587 | case IOMAP_HOLE: | |
1588 | if (WARN_ON_ONCE(write)) | |
876f2946 | 1589 | break; |
b15cd800 | 1590 | result = dax_pmd_load_hole(&xas, vmf, &iomap, &entry); |
642261ac RZ |
1591 | break; |
1592 | default: | |
1593 | WARN_ON_ONCE(1); | |
1594 | break; | |
1595 | } | |
1596 | ||
1597 | finish_iomap: | |
1598 | if (ops->iomap_end) { | |
9f141d6e JK |
1599 | int copied = PMD_SIZE; |
1600 | ||
1601 | if (result == VM_FAULT_FALLBACK) | |
1602 | copied = 0; | |
1603 | /* | |
1604 | * The fault is done by now and there's no way back (other | |
1605 | * thread may be already happily using PMD we have installed). | |
1606 | * Just ignore error from ->iomap_end since we cannot do much | |
1607 | * with it. | |
1608 | */ | |
1609 | ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags, | |
1610 | &iomap); | |
642261ac | 1611 | } |
876f2946 | 1612 | unlock_entry: |
b15cd800 | 1613 | dax_unlock_entry(&xas, entry); |
642261ac RZ |
1614 | fallback: |
1615 | if (result == VM_FAULT_FALLBACK) { | |
d8a849e1 | 1616 | split_huge_pmd(vma, vmf->pmd, vmf->address); |
642261ac RZ |
1617 | count_vm_event(THP_FAULT_FALLBACK); |
1618 | } | |
282a8e03 | 1619 | out: |
f4200391 | 1620 | trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result); |
642261ac RZ |
1621 | return result; |
1622 | } | |
a2d58167 | 1623 | #else |
ab77dab4 | 1624 | static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, |
01cddfe9 | 1625 | const struct iomap_ops *ops) |
a2d58167 DJ |
1626 | { |
1627 | return VM_FAULT_FALLBACK; | |
1628 | } | |
642261ac | 1629 | #endif /* CONFIG_FS_DAX_PMD */ |
a2d58167 DJ |
1630 | |
1631 | /** | |
1632 | * dax_iomap_fault - handle a page fault on a DAX file | |
1633 | * @vmf: The description of the fault | |
cec04e8c | 1634 | * @pe_size: Size of the page to fault in |
9a0dd422 | 1635 | * @pfnp: PFN to insert for synchronous faults if fsync is required |
c0b24625 | 1636 | * @iomap_errp: Storage for detailed error code in case of error |
cec04e8c | 1637 | * @ops: Iomap ops passed from the file system |
a2d58167 DJ |
1638 | * |
1639 | * When a page fault occurs, filesystems may call this helper in | |
1640 | * their fault handler for DAX files. dax_iomap_fault() assumes the caller | |
1641 | * has done all the necessary locking for page fault to proceed | |
1642 | * successfully. | |
1643 | */ | |
ab77dab4 | 1644 | vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, |
c0b24625 | 1645 | pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops) |
a2d58167 | 1646 | { |
c791ace1 DJ |
1647 | switch (pe_size) { |
1648 | case PE_SIZE_PTE: | |
c0b24625 | 1649 | return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops); |
c791ace1 | 1650 | case PE_SIZE_PMD: |
9a0dd422 | 1651 | return dax_iomap_pmd_fault(vmf, pfnp, ops); |
a2d58167 DJ |
1652 | default: |
1653 | return VM_FAULT_FALLBACK; | |
1654 | } | |
1655 | } | |
1656 | EXPORT_SYMBOL_GPL(dax_iomap_fault); | |
71eab6df | 1657 | |
a77d19f4 | 1658 | /* |
71eab6df JK |
1659 | * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables |
1660 | * @vmf: The description of the fault | |
71eab6df | 1661 | * @pfn: PFN to insert |
cfc93c6c | 1662 | * @order: Order of entry to insert. |
71eab6df | 1663 | * |
a77d19f4 MW |
1664 | * This function inserts a writeable PTE or PMD entry into the page tables |
1665 | * for an mmaped DAX file. It also marks the page cache entry as dirty. | |
71eab6df | 1666 | */ |
cfc93c6c MW |
1667 | static vm_fault_t |
1668 | dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order) | |
71eab6df JK |
1669 | { |
1670 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; | |
cfc93c6c MW |
1671 | XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order); |
1672 | void *entry; | |
ab77dab4 | 1673 | vm_fault_t ret; |
71eab6df | 1674 | |
cfc93c6c | 1675 | xas_lock_irq(&xas); |
23c84eb7 | 1676 | entry = get_unlocked_entry(&xas, order); |
71eab6df | 1677 | /* Did we race with someone splitting entry or so? */ |
23c84eb7 MWO |
1678 | if (!entry || dax_is_conflict(entry) || |
1679 | (order == 0 && !dax_is_pte_entry(entry))) { | |
cfc93c6c MW |
1680 | put_unlocked_entry(&xas, entry); |
1681 | xas_unlock_irq(&xas); | |
71eab6df JK |
1682 | trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf, |
1683 | VM_FAULT_NOPAGE); | |
1684 | return VM_FAULT_NOPAGE; | |
1685 | } | |
cfc93c6c MW |
1686 | xas_set_mark(&xas, PAGECACHE_TAG_DIRTY); |
1687 | dax_lock_entry(&xas, entry); | |
1688 | xas_unlock_irq(&xas); | |
1689 | if (order == 0) | |
ab77dab4 | 1690 | ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); |
71eab6df | 1691 | #ifdef CONFIG_FS_DAX_PMD |
cfc93c6c | 1692 | else if (order == PMD_ORDER) |
fce86ff5 | 1693 | ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE); |
71eab6df | 1694 | #endif |
cfc93c6c | 1695 | else |
ab77dab4 | 1696 | ret = VM_FAULT_FALLBACK; |
cfc93c6c | 1697 | dax_unlock_entry(&xas, entry); |
ab77dab4 SJ |
1698 | trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret); |
1699 | return ret; | |
71eab6df JK |
1700 | } |
1701 | ||
1702 | /** | |
1703 | * dax_finish_sync_fault - finish synchronous page fault | |
1704 | * @vmf: The description of the fault | |
1705 | * @pe_size: Size of entry to be inserted | |
1706 | * @pfn: PFN to insert | |
1707 | * | |
1708 | * This function ensures that the file range touched by the page fault is | |
1709 | * stored persistently on the media and handles inserting of appropriate page | |
1710 | * table entry. | |
1711 | */ | |
ab77dab4 SJ |
1712 | vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, |
1713 | enum page_entry_size pe_size, pfn_t pfn) | |
71eab6df JK |
1714 | { |
1715 | int err; | |
1716 | loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT; | |
cfc93c6c MW |
1717 | unsigned int order = pe_order(pe_size); |
1718 | size_t len = PAGE_SIZE << order; | |
71eab6df | 1719 | |
71eab6df JK |
1720 | err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1); |
1721 | if (err) | |
1722 | return VM_FAULT_SIGBUS; | |
cfc93c6c | 1723 | return dax_insert_pfn_mkwrite(vmf, pfn, order); |
71eab6df JK |
1724 | } |
1725 | EXPORT_SYMBOL_GPL(dax_finish_sync_fault); |