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