]>
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> |
2765cfbb | 28 | #include <linux/pmem.h> |
289c6aed | 29 | #include <linux/sched.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> |
a254e568 CH |
34 | #include <linux/iomap.h> |
35 | #include "internal.h" | |
d475c634 | 36 | |
ac401cc7 JK |
37 | /* We choose 4096 entries - same as per-zone page wait tables */ |
38 | #define DAX_WAIT_TABLE_BITS 12 | |
39 | #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) | |
40 | ||
ce95ab0f | 41 | static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; |
ac401cc7 JK |
42 | |
43 | static int __init init_dax_wait_table(void) | |
44 | { | |
45 | int i; | |
46 | ||
47 | for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) | |
48 | init_waitqueue_head(wait_table + i); | |
49 | return 0; | |
50 | } | |
51 | fs_initcall(init_dax_wait_table); | |
52 | ||
b2e0d162 DW |
53 | static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax) |
54 | { | |
55 | struct request_queue *q = bdev->bd_queue; | |
56 | long rc = -EIO; | |
57 | ||
7a9eb206 | 58 | dax->addr = ERR_PTR(-EIO); |
b2e0d162 DW |
59 | if (blk_queue_enter(q, true) != 0) |
60 | return rc; | |
61 | ||
62 | rc = bdev_direct_access(bdev, dax); | |
63 | if (rc < 0) { | |
7a9eb206 | 64 | dax->addr = ERR_PTR(rc); |
b2e0d162 DW |
65 | blk_queue_exit(q); |
66 | return rc; | |
67 | } | |
68 | return rc; | |
69 | } | |
70 | ||
71 | static void dax_unmap_atomic(struct block_device *bdev, | |
72 | const struct blk_dax_ctl *dax) | |
73 | { | |
74 | if (IS_ERR(dax->addr)) | |
75 | return; | |
76 | blk_queue_exit(bdev->bd_queue); | |
77 | } | |
78 | ||
642261ac | 79 | static int dax_is_pmd_entry(void *entry) |
d1a5f2b4 | 80 | { |
642261ac | 81 | return (unsigned long)entry & RADIX_DAX_PMD; |
d1a5f2b4 DW |
82 | } |
83 | ||
642261ac | 84 | static int dax_is_pte_entry(void *entry) |
d475c634 | 85 | { |
642261ac | 86 | return !((unsigned long)entry & RADIX_DAX_PMD); |
d475c634 MW |
87 | } |
88 | ||
642261ac | 89 | static int dax_is_zero_entry(void *entry) |
d475c634 | 90 | { |
642261ac | 91 | return (unsigned long)entry & RADIX_DAX_HZP; |
d475c634 MW |
92 | } |
93 | ||
642261ac | 94 | static int dax_is_empty_entry(void *entry) |
b2e0d162 | 95 | { |
642261ac | 96 | return (unsigned long)entry & RADIX_DAX_EMPTY; |
b2e0d162 DW |
97 | } |
98 | ||
d1a5f2b4 | 99 | struct page *read_dax_sector(struct block_device *bdev, sector_t n) |
d475c634 | 100 | { |
d1a5f2b4 | 101 | struct page *page = alloc_pages(GFP_KERNEL, 0); |
b2e0d162 | 102 | struct blk_dax_ctl dax = { |
d1a5f2b4 DW |
103 | .size = PAGE_SIZE, |
104 | .sector = n & ~((((int) PAGE_SIZE) / 512) - 1), | |
b2e0d162 | 105 | }; |
d1a5f2b4 | 106 | long rc; |
d475c634 | 107 | |
d1a5f2b4 DW |
108 | if (!page) |
109 | return ERR_PTR(-ENOMEM); | |
d475c634 | 110 | |
d1a5f2b4 DW |
111 | rc = dax_map_atomic(bdev, &dax); |
112 | if (rc < 0) | |
113 | return ERR_PTR(rc); | |
114 | memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE); | |
b2e0d162 | 115 | dax_unmap_atomic(bdev, &dax); |
d1a5f2b4 | 116 | return page; |
d475c634 | 117 | } |
f7ca90b1 | 118 | |
ac401cc7 JK |
119 | /* |
120 | * DAX radix tree locking | |
121 | */ | |
122 | struct exceptional_entry_key { | |
123 | struct address_space *mapping; | |
63e95b5c | 124 | pgoff_t entry_start; |
ac401cc7 JK |
125 | }; |
126 | ||
127 | struct wait_exceptional_entry_queue { | |
128 | wait_queue_t wait; | |
129 | struct exceptional_entry_key key; | |
130 | }; | |
131 | ||
63e95b5c RZ |
132 | static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping, |
133 | pgoff_t index, void *entry, struct exceptional_entry_key *key) | |
134 | { | |
135 | unsigned long hash; | |
136 | ||
137 | /* | |
138 | * If 'entry' is a PMD, align the 'index' that we use for the wait | |
139 | * queue to the start of that PMD. This ensures that all offsets in | |
140 | * the range covered by the PMD map to the same bit lock. | |
141 | */ | |
642261ac | 142 | if (dax_is_pmd_entry(entry)) |
63e95b5c RZ |
143 | index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1); |
144 | ||
145 | key->mapping = mapping; | |
146 | key->entry_start = index; | |
147 | ||
148 | hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS); | |
149 | return wait_table + hash; | |
150 | } | |
151 | ||
ac401cc7 JK |
152 | static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode, |
153 | int sync, void *keyp) | |
154 | { | |
155 | struct exceptional_entry_key *key = keyp; | |
156 | struct wait_exceptional_entry_queue *ewait = | |
157 | container_of(wait, struct wait_exceptional_entry_queue, wait); | |
158 | ||
159 | if (key->mapping != ewait->key.mapping || | |
63e95b5c | 160 | key->entry_start != ewait->key.entry_start) |
ac401cc7 JK |
161 | return 0; |
162 | return autoremove_wake_function(wait, mode, sync, NULL); | |
163 | } | |
164 | ||
165 | /* | |
166 | * Check whether the given slot is locked. The function must be called with | |
167 | * mapping->tree_lock held | |
168 | */ | |
169 | static inline int slot_locked(struct address_space *mapping, void **slot) | |
170 | { | |
171 | unsigned long entry = (unsigned long) | |
172 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
173 | return entry & RADIX_DAX_ENTRY_LOCK; | |
174 | } | |
175 | ||
176 | /* | |
177 | * Mark the given slot is locked. The function must be called with | |
178 | * mapping->tree_lock held | |
179 | */ | |
180 | static inline void *lock_slot(struct address_space *mapping, void **slot) | |
181 | { | |
182 | unsigned long entry = (unsigned long) | |
183 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
184 | ||
185 | entry |= RADIX_DAX_ENTRY_LOCK; | |
6d75f366 | 186 | radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); |
ac401cc7 JK |
187 | return (void *)entry; |
188 | } | |
189 | ||
190 | /* | |
191 | * Mark the given slot is unlocked. The function must be called with | |
192 | * mapping->tree_lock held | |
193 | */ | |
194 | static inline void *unlock_slot(struct address_space *mapping, void **slot) | |
195 | { | |
196 | unsigned long entry = (unsigned long) | |
197 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
198 | ||
199 | entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK; | |
6d75f366 | 200 | radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); |
ac401cc7 JK |
201 | return (void *)entry; |
202 | } | |
203 | ||
204 | /* | |
205 | * Lookup entry in radix tree, wait for it to become unlocked if it is | |
206 | * exceptional entry and return it. The caller must call | |
207 | * put_unlocked_mapping_entry() when he decided not to lock the entry or | |
208 | * put_locked_mapping_entry() when he locked the entry and now wants to | |
209 | * unlock it. | |
210 | * | |
211 | * The function must be called with mapping->tree_lock held. | |
212 | */ | |
213 | static void *get_unlocked_mapping_entry(struct address_space *mapping, | |
214 | pgoff_t index, void ***slotp) | |
215 | { | |
e3ad61c6 | 216 | void *entry, **slot; |
ac401cc7 | 217 | struct wait_exceptional_entry_queue ewait; |
63e95b5c | 218 | wait_queue_head_t *wq; |
ac401cc7 JK |
219 | |
220 | init_wait(&ewait.wait); | |
221 | ewait.wait.func = wake_exceptional_entry_func; | |
ac401cc7 JK |
222 | |
223 | for (;;) { | |
e3ad61c6 | 224 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, |
ac401cc7 | 225 | &slot); |
e3ad61c6 | 226 | if (!entry || !radix_tree_exceptional_entry(entry) || |
ac401cc7 JK |
227 | !slot_locked(mapping, slot)) { |
228 | if (slotp) | |
229 | *slotp = slot; | |
e3ad61c6 | 230 | return entry; |
ac401cc7 | 231 | } |
63e95b5c RZ |
232 | |
233 | wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key); | |
ac401cc7 JK |
234 | prepare_to_wait_exclusive(wq, &ewait.wait, |
235 | TASK_UNINTERRUPTIBLE); | |
236 | spin_unlock_irq(&mapping->tree_lock); | |
237 | schedule(); | |
238 | finish_wait(wq, &ewait.wait); | |
239 | spin_lock_irq(&mapping->tree_lock); | |
240 | } | |
241 | } | |
242 | ||
b1aa812b JK |
243 | static void dax_unlock_mapping_entry(struct address_space *mapping, |
244 | pgoff_t index) | |
245 | { | |
246 | void *entry, **slot; | |
247 | ||
248 | spin_lock_irq(&mapping->tree_lock); | |
249 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot); | |
250 | if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) || | |
251 | !slot_locked(mapping, slot))) { | |
252 | spin_unlock_irq(&mapping->tree_lock); | |
253 | return; | |
254 | } | |
255 | unlock_slot(mapping, slot); | |
256 | spin_unlock_irq(&mapping->tree_lock); | |
257 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); | |
258 | } | |
259 | ||
422476c4 RZ |
260 | static void put_locked_mapping_entry(struct address_space *mapping, |
261 | pgoff_t index, void *entry) | |
262 | { | |
263 | if (!radix_tree_exceptional_entry(entry)) { | |
264 | unlock_page(entry); | |
265 | put_page(entry); | |
266 | } else { | |
267 | dax_unlock_mapping_entry(mapping, index); | |
268 | } | |
269 | } | |
270 | ||
271 | /* | |
272 | * Called when we are done with radix tree entry we looked up via | |
273 | * get_unlocked_mapping_entry() and which we didn't lock in the end. | |
274 | */ | |
275 | static void put_unlocked_mapping_entry(struct address_space *mapping, | |
276 | pgoff_t index, void *entry) | |
277 | { | |
278 | if (!radix_tree_exceptional_entry(entry)) | |
279 | return; | |
280 | ||
281 | /* We have to wake up next waiter for the radix tree entry lock */ | |
282 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); | |
283 | } | |
284 | ||
ac401cc7 JK |
285 | /* |
286 | * Find radix tree entry at given index. If it points to a page, return with | |
287 | * the page locked. If it points to the exceptional entry, return with the | |
288 | * radix tree entry locked. If the radix tree doesn't contain given index, | |
289 | * create empty exceptional entry for the index and return with it locked. | |
290 | * | |
642261ac RZ |
291 | * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will |
292 | * either return that locked entry or will return an error. This error will | |
293 | * happen if there are any 4k entries (either zero pages or DAX entries) | |
294 | * within the 2MiB range that we are requesting. | |
295 | * | |
296 | * We always favor 4k entries over 2MiB entries. There isn't a flow where we | |
297 | * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB | |
298 | * insertion will fail if it finds any 4k entries already in the tree, and a | |
299 | * 4k insertion will cause an existing 2MiB entry to be unmapped and | |
300 | * downgraded to 4k entries. This happens for both 2MiB huge zero pages as | |
301 | * well as 2MiB empty entries. | |
302 | * | |
303 | * The exception to this downgrade path is for 2MiB DAX PMD entries that have | |
304 | * real storage backing them. We will leave these real 2MiB DAX entries in | |
305 | * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry. | |
306 | * | |
ac401cc7 JK |
307 | * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For |
308 | * persistent memory the benefit is doubtful. We can add that later if we can | |
309 | * show it helps. | |
310 | */ | |
642261ac RZ |
311 | static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index, |
312 | unsigned long size_flag) | |
ac401cc7 | 313 | { |
642261ac | 314 | bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */ |
e3ad61c6 | 315 | void *entry, **slot; |
ac401cc7 JK |
316 | |
317 | restart: | |
318 | spin_lock_irq(&mapping->tree_lock); | |
e3ad61c6 | 319 | entry = get_unlocked_mapping_entry(mapping, index, &slot); |
642261ac RZ |
320 | |
321 | if (entry) { | |
322 | if (size_flag & RADIX_DAX_PMD) { | |
323 | if (!radix_tree_exceptional_entry(entry) || | |
324 | dax_is_pte_entry(entry)) { | |
325 | put_unlocked_mapping_entry(mapping, index, | |
326 | entry); | |
327 | entry = ERR_PTR(-EEXIST); | |
328 | goto out_unlock; | |
329 | } | |
330 | } else { /* trying to grab a PTE entry */ | |
331 | if (radix_tree_exceptional_entry(entry) && | |
332 | dax_is_pmd_entry(entry) && | |
333 | (dax_is_zero_entry(entry) || | |
334 | dax_is_empty_entry(entry))) { | |
335 | pmd_downgrade = true; | |
336 | } | |
337 | } | |
338 | } | |
339 | ||
ac401cc7 | 340 | /* No entry for given index? Make sure radix tree is big enough. */ |
642261ac | 341 | if (!entry || pmd_downgrade) { |
ac401cc7 JK |
342 | int err; |
343 | ||
642261ac RZ |
344 | if (pmd_downgrade) { |
345 | /* | |
346 | * Make sure 'entry' remains valid while we drop | |
347 | * mapping->tree_lock. | |
348 | */ | |
349 | entry = lock_slot(mapping, slot); | |
350 | } | |
351 | ||
ac401cc7 | 352 | spin_unlock_irq(&mapping->tree_lock); |
642261ac RZ |
353 | /* |
354 | * Besides huge zero pages the only other thing that gets | |
355 | * downgraded are empty entries which don't need to be | |
356 | * unmapped. | |
357 | */ | |
358 | if (pmd_downgrade && dax_is_zero_entry(entry)) | |
359 | unmap_mapping_range(mapping, | |
360 | (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0); | |
361 | ||
ac401cc7 JK |
362 | err = radix_tree_preload( |
363 | mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM); | |
0cb80b48 JK |
364 | if (err) { |
365 | if (pmd_downgrade) | |
366 | put_locked_mapping_entry(mapping, index, entry); | |
ac401cc7 | 367 | return ERR_PTR(err); |
0cb80b48 | 368 | } |
ac401cc7 | 369 | spin_lock_irq(&mapping->tree_lock); |
642261ac RZ |
370 | |
371 | if (pmd_downgrade) { | |
372 | radix_tree_delete(&mapping->page_tree, index); | |
373 | mapping->nrexceptional--; | |
374 | dax_wake_mapping_entry_waiter(mapping, index, entry, | |
375 | true); | |
376 | } | |
377 | ||
378 | entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY); | |
379 | ||
380 | err = __radix_tree_insert(&mapping->page_tree, index, | |
381 | dax_radix_order(entry), entry); | |
ac401cc7 JK |
382 | radix_tree_preload_end(); |
383 | if (err) { | |
384 | spin_unlock_irq(&mapping->tree_lock); | |
642261ac RZ |
385 | /* |
386 | * Someone already created the entry? This is a | |
387 | * normal failure when inserting PMDs in a range | |
388 | * that already contains PTEs. In that case we want | |
389 | * to return -EEXIST immediately. | |
390 | */ | |
391 | if (err == -EEXIST && !(size_flag & RADIX_DAX_PMD)) | |
ac401cc7 | 392 | goto restart; |
642261ac RZ |
393 | /* |
394 | * Our insertion of a DAX PMD entry failed, most | |
395 | * likely because it collided with a PTE sized entry | |
396 | * at a different index in the PMD range. We haven't | |
397 | * inserted anything into the radix tree and have no | |
398 | * waiters to wake. | |
399 | */ | |
ac401cc7 JK |
400 | return ERR_PTR(err); |
401 | } | |
402 | /* Good, we have inserted empty locked entry into the tree. */ | |
403 | mapping->nrexceptional++; | |
404 | spin_unlock_irq(&mapping->tree_lock); | |
e3ad61c6 | 405 | return entry; |
ac401cc7 JK |
406 | } |
407 | /* Normal page in radix tree? */ | |
e3ad61c6 RZ |
408 | if (!radix_tree_exceptional_entry(entry)) { |
409 | struct page *page = entry; | |
ac401cc7 JK |
410 | |
411 | get_page(page); | |
412 | spin_unlock_irq(&mapping->tree_lock); | |
413 | lock_page(page); | |
414 | /* Page got truncated? Retry... */ | |
415 | if (unlikely(page->mapping != mapping)) { | |
416 | unlock_page(page); | |
417 | put_page(page); | |
418 | goto restart; | |
419 | } | |
420 | return page; | |
421 | } | |
e3ad61c6 | 422 | entry = lock_slot(mapping, slot); |
642261ac | 423 | out_unlock: |
ac401cc7 | 424 | spin_unlock_irq(&mapping->tree_lock); |
e3ad61c6 | 425 | return entry; |
ac401cc7 JK |
426 | } |
427 | ||
63e95b5c RZ |
428 | /* |
429 | * We do not necessarily hold the mapping->tree_lock when we call this | |
430 | * function so it is possible that 'entry' is no longer a valid item in the | |
642261ac RZ |
431 | * radix tree. This is okay because all we really need to do is to find the |
432 | * correct waitqueue where tasks might be waiting for that old 'entry' and | |
433 | * wake them. | |
63e95b5c | 434 | */ |
ac401cc7 | 435 | void dax_wake_mapping_entry_waiter(struct address_space *mapping, |
63e95b5c | 436 | pgoff_t index, void *entry, bool wake_all) |
ac401cc7 | 437 | { |
63e95b5c RZ |
438 | struct exceptional_entry_key key; |
439 | wait_queue_head_t *wq; | |
440 | ||
441 | wq = dax_entry_waitqueue(mapping, index, entry, &key); | |
ac401cc7 JK |
442 | |
443 | /* | |
444 | * Checking for locked entry and prepare_to_wait_exclusive() happens | |
445 | * under mapping->tree_lock, ditto for entry handling in our callers. | |
446 | * So at this point all tasks that could have seen our entry locked | |
447 | * must be in the waitqueue and the following check will see them. | |
448 | */ | |
63e95b5c | 449 | if (waitqueue_active(wq)) |
ac401cc7 | 450 | __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); |
ac401cc7 JK |
451 | } |
452 | ||
ac401cc7 JK |
453 | /* |
454 | * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree | |
455 | * entry to get unlocked before deleting it. | |
456 | */ | |
457 | int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) | |
458 | { | |
459 | void *entry; | |
460 | ||
461 | spin_lock_irq(&mapping->tree_lock); | |
462 | entry = get_unlocked_mapping_entry(mapping, index, NULL); | |
463 | /* | |
464 | * This gets called from truncate / punch_hole path. As such, the caller | |
465 | * must hold locks protecting against concurrent modifications of the | |
466 | * radix tree (usually fs-private i_mmap_sem for writing). Since the | |
467 | * caller has seen exceptional entry for this index, we better find it | |
468 | * at that index as well... | |
469 | */ | |
470 | if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry))) { | |
471 | spin_unlock_irq(&mapping->tree_lock); | |
472 | return 0; | |
473 | } | |
474 | radix_tree_delete(&mapping->page_tree, index); | |
475 | mapping->nrexceptional--; | |
476 | spin_unlock_irq(&mapping->tree_lock); | |
63e95b5c | 477 | dax_wake_mapping_entry_waiter(mapping, index, entry, true); |
ac401cc7 JK |
478 | |
479 | return 1; | |
480 | } | |
481 | ||
f7ca90b1 MW |
482 | /* |
483 | * The user has performed a load from a hole in the file. Allocating | |
484 | * a new page in the file would cause excessive storage usage for | |
485 | * workloads with sparse files. We allocate a page cache page instead. | |
486 | * We'll kick it out of the page cache if it's ever written to, | |
487 | * otherwise it will simply fall out of the page cache under memory | |
488 | * pressure without ever having been dirtied. | |
489 | */ | |
ac401cc7 JK |
490 | static int dax_load_hole(struct address_space *mapping, void *entry, |
491 | struct vm_fault *vmf) | |
f7ca90b1 | 492 | { |
ac401cc7 | 493 | struct page *page; |
f7ca90b1 | 494 | |
ac401cc7 JK |
495 | /* Hole page already exists? Return it... */ |
496 | if (!radix_tree_exceptional_entry(entry)) { | |
497 | vmf->page = entry; | |
498 | return VM_FAULT_LOCKED; | |
499 | } | |
f7ca90b1 | 500 | |
ac401cc7 JK |
501 | /* This will replace locked radix tree entry with a hole page */ |
502 | page = find_or_create_page(mapping, vmf->pgoff, | |
503 | vmf->gfp_mask | __GFP_ZERO); | |
b1aa812b | 504 | if (!page) |
ac401cc7 | 505 | return VM_FAULT_OOM; |
f7ca90b1 MW |
506 | vmf->page = page; |
507 | return VM_FAULT_LOCKED; | |
508 | } | |
509 | ||
b0d5e82f CH |
510 | static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size, |
511 | struct page *to, unsigned long vaddr) | |
f7ca90b1 | 512 | { |
b2e0d162 | 513 | struct blk_dax_ctl dax = { |
b0d5e82f CH |
514 | .sector = sector, |
515 | .size = size, | |
b2e0d162 | 516 | }; |
e2e05394 RZ |
517 | void *vto; |
518 | ||
b2e0d162 DW |
519 | if (dax_map_atomic(bdev, &dax) < 0) |
520 | return PTR_ERR(dax.addr); | |
f7ca90b1 | 521 | vto = kmap_atomic(to); |
b2e0d162 | 522 | copy_user_page(vto, (void __force *)dax.addr, vaddr, to); |
f7ca90b1 | 523 | kunmap_atomic(vto); |
b2e0d162 | 524 | dax_unmap_atomic(bdev, &dax); |
f7ca90b1 MW |
525 | return 0; |
526 | } | |
527 | ||
642261ac RZ |
528 | /* |
529 | * By this point grab_mapping_entry() has ensured that we have a locked entry | |
530 | * of the appropriate size so we don't have to worry about downgrading PMDs to | |
531 | * PTEs. If we happen to be trying to insert a PTE and there is a PMD | |
532 | * already in the tree, we will skip the insertion and just dirty the PMD as | |
533 | * appropriate. | |
534 | */ | |
ac401cc7 JK |
535 | static void *dax_insert_mapping_entry(struct address_space *mapping, |
536 | struct vm_fault *vmf, | |
642261ac RZ |
537 | void *entry, sector_t sector, |
538 | unsigned long flags) | |
9973c98e RZ |
539 | { |
540 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
ac401cc7 JK |
541 | int error = 0; |
542 | bool hole_fill = false; | |
543 | void *new_entry; | |
544 | pgoff_t index = vmf->pgoff; | |
9973c98e | 545 | |
ac401cc7 | 546 | if (vmf->flags & FAULT_FLAG_WRITE) |
d2b2a28e | 547 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); |
9973c98e | 548 | |
ac401cc7 JK |
549 | /* Replacing hole page with block mapping? */ |
550 | if (!radix_tree_exceptional_entry(entry)) { | |
551 | hole_fill = true; | |
552 | /* | |
553 | * Unmap the page now before we remove it from page cache below. | |
554 | * The page is locked so it cannot be faulted in again. | |
555 | */ | |
556 | unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT, | |
557 | PAGE_SIZE, 0); | |
558 | error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM); | |
559 | if (error) | |
560 | return ERR_PTR(error); | |
642261ac RZ |
561 | } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) { |
562 | /* replacing huge zero page with PMD block mapping */ | |
563 | unmap_mapping_range(mapping, | |
564 | (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0); | |
9973c98e RZ |
565 | } |
566 | ||
ac401cc7 | 567 | spin_lock_irq(&mapping->tree_lock); |
642261ac RZ |
568 | new_entry = dax_radix_locked_entry(sector, flags); |
569 | ||
ac401cc7 JK |
570 | if (hole_fill) { |
571 | __delete_from_page_cache(entry, NULL); | |
572 | /* Drop pagecache reference */ | |
573 | put_page(entry); | |
642261ac RZ |
574 | error = __radix_tree_insert(page_tree, index, |
575 | dax_radix_order(new_entry), new_entry); | |
ac401cc7 JK |
576 | if (error) { |
577 | new_entry = ERR_PTR(error); | |
9973c98e RZ |
578 | goto unlock; |
579 | } | |
ac401cc7 | 580 | mapping->nrexceptional++; |
642261ac RZ |
581 | } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { |
582 | /* | |
583 | * Only swap our new entry into the radix tree if the current | |
584 | * entry is a zero page or an empty entry. If a normal PTE or | |
585 | * PMD entry is already in the tree, we leave it alone. This | |
586 | * means that if we are trying to insert a PTE and the | |
587 | * existing entry is a PMD, we will just leave the PMD in the | |
588 | * tree and dirty it if necessary. | |
589 | */ | |
f7942430 | 590 | struct radix_tree_node *node; |
ac401cc7 JK |
591 | void **slot; |
592 | void *ret; | |
9973c98e | 593 | |
f7942430 | 594 | ret = __radix_tree_lookup(page_tree, index, &node, &slot); |
ac401cc7 | 595 | WARN_ON_ONCE(ret != entry); |
4d693d08 JW |
596 | __radix_tree_replace(page_tree, node, slot, |
597 | new_entry, NULL, NULL); | |
9973c98e | 598 | } |
ac401cc7 | 599 | if (vmf->flags & FAULT_FLAG_WRITE) |
9973c98e RZ |
600 | radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY); |
601 | unlock: | |
602 | spin_unlock_irq(&mapping->tree_lock); | |
ac401cc7 JK |
603 | if (hole_fill) { |
604 | radix_tree_preload_end(); | |
605 | /* | |
606 | * We don't need hole page anymore, it has been replaced with | |
607 | * locked radix tree entry now. | |
608 | */ | |
609 | if (mapping->a_ops->freepage) | |
610 | mapping->a_ops->freepage(entry); | |
611 | unlock_page(entry); | |
612 | put_page(entry); | |
613 | } | |
614 | return new_entry; | |
9973c98e RZ |
615 | } |
616 | ||
617 | static int dax_writeback_one(struct block_device *bdev, | |
618 | struct address_space *mapping, pgoff_t index, void *entry) | |
619 | { | |
620 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
9973c98e RZ |
621 | struct radix_tree_node *node; |
622 | struct blk_dax_ctl dax; | |
623 | void **slot; | |
624 | int ret = 0; | |
625 | ||
626 | spin_lock_irq(&mapping->tree_lock); | |
627 | /* | |
628 | * Regular page slots are stabilized by the page lock even | |
629 | * without the tree itself locked. These unlocked entries | |
630 | * need verification under the tree lock. | |
631 | */ | |
632 | if (!__radix_tree_lookup(page_tree, index, &node, &slot)) | |
633 | goto unlock; | |
634 | if (*slot != entry) | |
635 | goto unlock; | |
636 | ||
637 | /* another fsync thread may have already written back this entry */ | |
638 | if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)) | |
639 | goto unlock; | |
640 | ||
642261ac RZ |
641 | if (WARN_ON_ONCE(dax_is_empty_entry(entry) || |
642 | dax_is_zero_entry(entry))) { | |
9973c98e RZ |
643 | ret = -EIO; |
644 | goto unlock; | |
645 | } | |
646 | ||
642261ac RZ |
647 | /* |
648 | * Even if dax_writeback_mapping_range() was given a wbc->range_start | |
649 | * in the middle of a PMD, the 'index' we are given will be aligned to | |
650 | * the start index of the PMD, as will the sector we pull from | |
651 | * 'entry'. This allows us to flush for PMD_SIZE and not have to | |
652 | * worry about partial PMD writebacks. | |
653 | */ | |
654 | dax.sector = dax_radix_sector(entry); | |
655 | dax.size = PAGE_SIZE << dax_radix_order(entry); | |
9973c98e RZ |
656 | spin_unlock_irq(&mapping->tree_lock); |
657 | ||
658 | /* | |
659 | * We cannot hold tree_lock while calling dax_map_atomic() because it | |
660 | * eventually calls cond_resched(). | |
661 | */ | |
662 | ret = dax_map_atomic(bdev, &dax); | |
663 | if (ret < 0) | |
664 | return ret; | |
665 | ||
666 | if (WARN_ON_ONCE(ret < dax.size)) { | |
667 | ret = -EIO; | |
668 | goto unmap; | |
669 | } | |
670 | ||
671 | wb_cache_pmem(dax.addr, dax.size); | |
672 | ||
673 | spin_lock_irq(&mapping->tree_lock); | |
674 | radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE); | |
675 | spin_unlock_irq(&mapping->tree_lock); | |
676 | unmap: | |
677 | dax_unmap_atomic(bdev, &dax); | |
678 | return ret; | |
679 | ||
680 | unlock: | |
681 | spin_unlock_irq(&mapping->tree_lock); | |
682 | return ret; | |
683 | } | |
684 | ||
685 | /* | |
686 | * Flush the mapping to the persistent domain within the byte range of [start, | |
687 | * end]. This is required by data integrity operations to ensure file data is | |
688 | * on persistent storage prior to completion of the operation. | |
689 | */ | |
7f6d5b52 RZ |
690 | int dax_writeback_mapping_range(struct address_space *mapping, |
691 | struct block_device *bdev, struct writeback_control *wbc) | |
9973c98e RZ |
692 | { |
693 | struct inode *inode = mapping->host; | |
642261ac | 694 | pgoff_t start_index, end_index; |
9973c98e RZ |
695 | pgoff_t indices[PAGEVEC_SIZE]; |
696 | struct pagevec pvec; | |
697 | bool done = false; | |
698 | int i, ret = 0; | |
9973c98e RZ |
699 | |
700 | if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) | |
701 | return -EIO; | |
702 | ||
7f6d5b52 RZ |
703 | if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) |
704 | return 0; | |
705 | ||
09cbfeaf KS |
706 | start_index = wbc->range_start >> PAGE_SHIFT; |
707 | end_index = wbc->range_end >> PAGE_SHIFT; | |
9973c98e RZ |
708 | |
709 | tag_pages_for_writeback(mapping, start_index, end_index); | |
710 | ||
711 | pagevec_init(&pvec, 0); | |
712 | while (!done) { | |
713 | pvec.nr = find_get_entries_tag(mapping, start_index, | |
714 | PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE, | |
715 | pvec.pages, indices); | |
716 | ||
717 | if (pvec.nr == 0) | |
718 | break; | |
719 | ||
720 | for (i = 0; i < pvec.nr; i++) { | |
721 | if (indices[i] > end_index) { | |
722 | done = true; | |
723 | break; | |
724 | } | |
725 | ||
726 | ret = dax_writeback_one(bdev, mapping, indices[i], | |
727 | pvec.pages[i]); | |
728 | if (ret < 0) | |
729 | return ret; | |
730 | } | |
731 | } | |
9973c98e RZ |
732 | return 0; |
733 | } | |
734 | EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); | |
735 | ||
ac401cc7 | 736 | static int dax_insert_mapping(struct address_space *mapping, |
1aaba095 CH |
737 | struct block_device *bdev, sector_t sector, size_t size, |
738 | void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf) | |
f7ca90b1 | 739 | { |
1a29d85e | 740 | unsigned long vaddr = vmf->address; |
b2e0d162 | 741 | struct blk_dax_ctl dax = { |
1aaba095 CH |
742 | .sector = sector, |
743 | .size = size, | |
b2e0d162 | 744 | }; |
ac401cc7 JK |
745 | void *ret; |
746 | void *entry = *entryp; | |
f7ca90b1 | 747 | |
4d9a2c87 JK |
748 | if (dax_map_atomic(bdev, &dax) < 0) |
749 | return PTR_ERR(dax.addr); | |
b2e0d162 | 750 | dax_unmap_atomic(bdev, &dax); |
f7ca90b1 | 751 | |
642261ac | 752 | ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector, 0); |
4d9a2c87 JK |
753 | if (IS_ERR(ret)) |
754 | return PTR_ERR(ret); | |
ac401cc7 | 755 | *entryp = ret; |
9973c98e | 756 | |
4d9a2c87 | 757 | return vm_insert_mixed(vma, vaddr, dax.pfn); |
f7ca90b1 MW |
758 | } |
759 | ||
0e3b210c BH |
760 | /** |
761 | * dax_pfn_mkwrite - handle first write to DAX page | |
762 | * @vma: The virtual memory area where the fault occurred | |
763 | * @vmf: The description of the fault | |
0e3b210c BH |
764 | */ |
765 | int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) | |
766 | { | |
9973c98e | 767 | struct file *file = vma->vm_file; |
ac401cc7 JK |
768 | struct address_space *mapping = file->f_mapping; |
769 | void *entry; | |
770 | pgoff_t index = vmf->pgoff; | |
30f471fd | 771 | |
ac401cc7 JK |
772 | spin_lock_irq(&mapping->tree_lock); |
773 | entry = get_unlocked_mapping_entry(mapping, index, NULL); | |
774 | if (!entry || !radix_tree_exceptional_entry(entry)) | |
775 | goto out; | |
776 | radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY); | |
777 | put_unlocked_mapping_entry(mapping, index, entry); | |
778 | out: | |
779 | spin_unlock_irq(&mapping->tree_lock); | |
0e3b210c BH |
780 | return VM_FAULT_NOPAGE; |
781 | } | |
782 | EXPORT_SYMBOL_GPL(dax_pfn_mkwrite); | |
783 | ||
4b0228fa VV |
784 | static bool dax_range_is_aligned(struct block_device *bdev, |
785 | unsigned int offset, unsigned int length) | |
786 | { | |
787 | unsigned short sector_size = bdev_logical_block_size(bdev); | |
788 | ||
789 | if (!IS_ALIGNED(offset, sector_size)) | |
790 | return false; | |
791 | if (!IS_ALIGNED(length, sector_size)) | |
792 | return false; | |
793 | ||
794 | return true; | |
795 | } | |
796 | ||
679c8bd3 CH |
797 | int __dax_zero_page_range(struct block_device *bdev, sector_t sector, |
798 | unsigned int offset, unsigned int length) | |
799 | { | |
800 | struct blk_dax_ctl dax = { | |
801 | .sector = sector, | |
802 | .size = PAGE_SIZE, | |
803 | }; | |
804 | ||
4b0228fa VV |
805 | if (dax_range_is_aligned(bdev, offset, length)) { |
806 | sector_t start_sector = dax.sector + (offset >> 9); | |
807 | ||
808 | return blkdev_issue_zeroout(bdev, start_sector, | |
809 | length >> 9, GFP_NOFS, true); | |
810 | } else { | |
811 | if (dax_map_atomic(bdev, &dax) < 0) | |
812 | return PTR_ERR(dax.addr); | |
813 | clear_pmem(dax.addr + offset, length); | |
4b0228fa VV |
814 | dax_unmap_atomic(bdev, &dax); |
815 | } | |
679c8bd3 CH |
816 | return 0; |
817 | } | |
818 | EXPORT_SYMBOL_GPL(__dax_zero_page_range); | |
819 | ||
a254e568 | 820 | #ifdef CONFIG_FS_IOMAP |
333ccc97 | 821 | static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) |
25726bc1 | 822 | { |
333ccc97 | 823 | return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9); |
25726bc1 | 824 | } |
a254e568 | 825 | |
a254e568 | 826 | static loff_t |
11c59c92 | 827 | dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, |
a254e568 CH |
828 | struct iomap *iomap) |
829 | { | |
830 | struct iov_iter *iter = data; | |
831 | loff_t end = pos + length, done = 0; | |
832 | ssize_t ret = 0; | |
833 | ||
834 | if (iov_iter_rw(iter) == READ) { | |
835 | end = min(end, i_size_read(inode)); | |
836 | if (pos >= end) | |
837 | return 0; | |
838 | ||
839 | if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) | |
840 | return iov_iter_zero(min(length, end - pos), iter); | |
841 | } | |
842 | ||
843 | if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) | |
844 | return -EIO; | |
845 | ||
846 | while (pos < end) { | |
847 | unsigned offset = pos & (PAGE_SIZE - 1); | |
848 | struct blk_dax_ctl dax = { 0 }; | |
849 | ssize_t map_len; | |
850 | ||
333ccc97 | 851 | dax.sector = dax_iomap_sector(iomap, pos); |
a254e568 CH |
852 | dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK; |
853 | map_len = dax_map_atomic(iomap->bdev, &dax); | |
854 | if (map_len < 0) { | |
855 | ret = map_len; | |
856 | break; | |
857 | } | |
858 | ||
859 | dax.addr += offset; | |
860 | map_len -= offset; | |
861 | if (map_len > end - pos) | |
862 | map_len = end - pos; | |
863 | ||
864 | if (iov_iter_rw(iter) == WRITE) | |
865 | map_len = copy_from_iter_pmem(dax.addr, map_len, iter); | |
866 | else | |
867 | map_len = copy_to_iter(dax.addr, map_len, iter); | |
868 | dax_unmap_atomic(iomap->bdev, &dax); | |
869 | if (map_len <= 0) { | |
870 | ret = map_len ? map_len : -EFAULT; | |
871 | break; | |
872 | } | |
873 | ||
874 | pos += map_len; | |
875 | length -= map_len; | |
876 | done += map_len; | |
877 | } | |
878 | ||
879 | return done ? done : ret; | |
880 | } | |
881 | ||
882 | /** | |
11c59c92 | 883 | * dax_iomap_rw - Perform I/O to a DAX file |
a254e568 CH |
884 | * @iocb: The control block for this I/O |
885 | * @iter: The addresses to do I/O from or to | |
886 | * @ops: iomap ops passed from the file system | |
887 | * | |
888 | * This function performs read and write operations to directly mapped | |
889 | * persistent memory. The callers needs to take care of read/write exclusion | |
890 | * and evicting any page cache pages in the region under I/O. | |
891 | */ | |
892 | ssize_t | |
11c59c92 | 893 | dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, |
a254e568 CH |
894 | struct iomap_ops *ops) |
895 | { | |
896 | struct address_space *mapping = iocb->ki_filp->f_mapping; | |
897 | struct inode *inode = mapping->host; | |
898 | loff_t pos = iocb->ki_pos, ret = 0, done = 0; | |
899 | unsigned flags = 0; | |
900 | ||
901 | if (iov_iter_rw(iter) == WRITE) | |
902 | flags |= IOMAP_WRITE; | |
903 | ||
904 | /* | |
905 | * Yes, even DAX files can have page cache attached to them: A zeroed | |
906 | * page is inserted into the pagecache when we have to serve a write | |
907 | * fault on a hole. It should never be dirtied and can simply be | |
908 | * dropped from the pagecache once we get real data for the page. | |
909 | * | |
910 | * XXX: This is racy against mmap, and there's nothing we can do about | |
911 | * it. We'll eventually need to shift this down even further so that | |
912 | * we can check if we allocated blocks over a hole first. | |
913 | */ | |
914 | if (mapping->nrpages) { | |
915 | ret = invalidate_inode_pages2_range(mapping, | |
916 | pos >> PAGE_SHIFT, | |
917 | (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT); | |
918 | WARN_ON_ONCE(ret); | |
919 | } | |
920 | ||
921 | while (iov_iter_count(iter)) { | |
922 | ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, | |
11c59c92 | 923 | iter, dax_iomap_actor); |
a254e568 CH |
924 | if (ret <= 0) |
925 | break; | |
926 | pos += ret; | |
927 | done += ret; | |
928 | } | |
929 | ||
930 | iocb->ki_pos += done; | |
931 | return done ? done : ret; | |
932 | } | |
11c59c92 | 933 | EXPORT_SYMBOL_GPL(dax_iomap_rw); |
a7d73fe6 CH |
934 | |
935 | /** | |
11c59c92 | 936 | * dax_iomap_fault - handle a page fault on a DAX file |
a7d73fe6 CH |
937 | * @vma: The virtual memory area where the fault occurred |
938 | * @vmf: The description of the fault | |
939 | * @ops: iomap ops passed from the file system | |
940 | * | |
941 | * When a page fault occurs, filesystems may call this helper in their fault | |
942 | * or mkwrite handler for DAX files. Assumes the caller has done all the | |
943 | * necessary locking for the page fault to proceed successfully. | |
944 | */ | |
11c59c92 | 945 | int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf, |
a7d73fe6 CH |
946 | struct iomap_ops *ops) |
947 | { | |
948 | struct address_space *mapping = vma->vm_file->f_mapping; | |
949 | struct inode *inode = mapping->host; | |
1a29d85e | 950 | unsigned long vaddr = vmf->address; |
a7d73fe6 CH |
951 | loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; |
952 | sector_t sector; | |
953 | struct iomap iomap = { 0 }; | |
9484ab1b | 954 | unsigned flags = IOMAP_FAULT; |
a7d73fe6 | 955 | int error, major = 0; |
b1aa812b | 956 | int vmf_ret = 0; |
a7d73fe6 CH |
957 | void *entry; |
958 | ||
959 | /* | |
960 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
961 | * to hold locks serializing us with truncate / punch hole so this is | |
962 | * a reliable test. | |
963 | */ | |
964 | if (pos >= i_size_read(inode)) | |
965 | return VM_FAULT_SIGBUS; | |
966 | ||
642261ac | 967 | entry = grab_mapping_entry(mapping, vmf->pgoff, 0); |
a7d73fe6 CH |
968 | if (IS_ERR(entry)) { |
969 | error = PTR_ERR(entry); | |
970 | goto out; | |
971 | } | |
972 | ||
973 | if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) | |
974 | flags |= IOMAP_WRITE; | |
975 | ||
976 | /* | |
977 | * Note that we don't bother to use iomap_apply here: DAX required | |
978 | * the file system block size to be equal the page size, which means | |
979 | * that we never have to deal with more than a single extent here. | |
980 | */ | |
981 | error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); | |
982 | if (error) | |
983 | goto unlock_entry; | |
984 | if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { | |
985 | error = -EIO; /* fs corruption? */ | |
1550290b | 986 | goto finish_iomap; |
a7d73fe6 CH |
987 | } |
988 | ||
333ccc97 | 989 | sector = dax_iomap_sector(&iomap, pos); |
a7d73fe6 CH |
990 | |
991 | if (vmf->cow_page) { | |
992 | switch (iomap.type) { | |
993 | case IOMAP_HOLE: | |
994 | case IOMAP_UNWRITTEN: | |
995 | clear_user_highpage(vmf->cow_page, vaddr); | |
996 | break; | |
997 | case IOMAP_MAPPED: | |
998 | error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE, | |
999 | vmf->cow_page, vaddr); | |
1000 | break; | |
1001 | default: | |
1002 | WARN_ON_ONCE(1); | |
1003 | error = -EIO; | |
1004 | break; | |
1005 | } | |
1006 | ||
1007 | if (error) | |
1550290b | 1008 | goto finish_iomap; |
b1aa812b JK |
1009 | |
1010 | __SetPageUptodate(vmf->cow_page); | |
1011 | vmf_ret = finish_fault(vmf); | |
1012 | if (!vmf_ret) | |
1013 | vmf_ret = VM_FAULT_DONE_COW; | |
1550290b | 1014 | goto finish_iomap; |
a7d73fe6 CH |
1015 | } |
1016 | ||
1017 | switch (iomap.type) { | |
1018 | case IOMAP_MAPPED: | |
1019 | if (iomap.flags & IOMAP_F_NEW) { | |
1020 | count_vm_event(PGMAJFAULT); | |
1021 | mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT); | |
1022 | major = VM_FAULT_MAJOR; | |
1023 | } | |
1024 | error = dax_insert_mapping(mapping, iomap.bdev, sector, | |
1025 | PAGE_SIZE, &entry, vma, vmf); | |
1026 | break; | |
1027 | case IOMAP_UNWRITTEN: | |
1028 | case IOMAP_HOLE: | |
1550290b | 1029 | if (!(vmf->flags & FAULT_FLAG_WRITE)) { |
b1aa812b | 1030 | vmf_ret = dax_load_hole(mapping, entry, vmf); |
1550290b RZ |
1031 | break; |
1032 | } | |
a7d73fe6 CH |
1033 | /*FALLTHRU*/ |
1034 | default: | |
1035 | WARN_ON_ONCE(1); | |
1036 | error = -EIO; | |
1037 | break; | |
1038 | } | |
1039 | ||
1550290b RZ |
1040 | finish_iomap: |
1041 | if (ops->iomap_end) { | |
b1aa812b | 1042 | if (error || (vmf_ret & VM_FAULT_ERROR)) { |
1550290b RZ |
1043 | /* keep previous error */ |
1044 | ops->iomap_end(inode, pos, PAGE_SIZE, 0, flags, | |
1045 | &iomap); | |
1046 | } else { | |
1047 | error = ops->iomap_end(inode, pos, PAGE_SIZE, | |
1048 | PAGE_SIZE, flags, &iomap); | |
1049 | } | |
1050 | } | |
a7d73fe6 | 1051 | unlock_entry: |
b1aa812b | 1052 | if (vmf_ret != VM_FAULT_LOCKED || error) |
1550290b | 1053 | put_locked_mapping_entry(mapping, vmf->pgoff, entry); |
a7d73fe6 CH |
1054 | out: |
1055 | if (error == -ENOMEM) | |
1056 | return VM_FAULT_OOM | major; | |
1057 | /* -EBUSY is fine, somebody else faulted on the same PTE */ | |
1058 | if (error < 0 && error != -EBUSY) | |
1059 | return VM_FAULT_SIGBUS | major; | |
b1aa812b | 1060 | if (vmf_ret) { |
1550290b | 1061 | WARN_ON_ONCE(error); /* -EBUSY from ops->iomap_end? */ |
b1aa812b | 1062 | return vmf_ret; |
1550290b | 1063 | } |
a7d73fe6 CH |
1064 | return VM_FAULT_NOPAGE | major; |
1065 | } | |
11c59c92 | 1066 | EXPORT_SYMBOL_GPL(dax_iomap_fault); |
642261ac RZ |
1067 | |
1068 | #ifdef CONFIG_FS_DAX_PMD | |
1069 | /* | |
1070 | * The 'colour' (ie low bits) within a PMD of a page offset. This comes up | |
1071 | * more often than one might expect in the below functions. | |
1072 | */ | |
1073 | #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) | |
1074 | ||
1075 | static int dax_pmd_insert_mapping(struct vm_area_struct *vma, pmd_t *pmd, | |
1076 | struct vm_fault *vmf, unsigned long address, | |
1077 | struct iomap *iomap, loff_t pos, bool write, void **entryp) | |
1078 | { | |
1079 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1080 | struct block_device *bdev = iomap->bdev; | |
1081 | struct blk_dax_ctl dax = { | |
1082 | .sector = dax_iomap_sector(iomap, pos), | |
1083 | .size = PMD_SIZE, | |
1084 | }; | |
1085 | long length = dax_map_atomic(bdev, &dax); | |
1086 | void *ret; | |
1087 | ||
1088 | if (length < 0) /* dax_map_atomic() failed */ | |
1089 | return VM_FAULT_FALLBACK; | |
1090 | if (length < PMD_SIZE) | |
1091 | goto unmap_fallback; | |
1092 | if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) | |
1093 | goto unmap_fallback; | |
1094 | if (!pfn_t_devmap(dax.pfn)) | |
1095 | goto unmap_fallback; | |
1096 | ||
1097 | dax_unmap_atomic(bdev, &dax); | |
1098 | ||
1099 | ret = dax_insert_mapping_entry(mapping, vmf, *entryp, dax.sector, | |
1100 | RADIX_DAX_PMD); | |
1101 | if (IS_ERR(ret)) | |
1102 | return VM_FAULT_FALLBACK; | |
1103 | *entryp = ret; | |
1104 | ||
1105 | return vmf_insert_pfn_pmd(vma, address, pmd, dax.pfn, write); | |
1106 | ||
1107 | unmap_fallback: | |
1108 | dax_unmap_atomic(bdev, &dax); | |
1109 | return VM_FAULT_FALLBACK; | |
1110 | } | |
1111 | ||
1112 | static int dax_pmd_load_hole(struct vm_area_struct *vma, pmd_t *pmd, | |
1113 | struct vm_fault *vmf, unsigned long address, | |
1114 | struct iomap *iomap, void **entryp) | |
1115 | { | |
1116 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1117 | unsigned long pmd_addr = address & PMD_MASK; | |
1118 | struct page *zero_page; | |
1119 | spinlock_t *ptl; | |
1120 | pmd_t pmd_entry; | |
1121 | void *ret; | |
1122 | ||
1123 | zero_page = mm_get_huge_zero_page(vma->vm_mm); | |
1124 | ||
1125 | if (unlikely(!zero_page)) | |
1126 | return VM_FAULT_FALLBACK; | |
1127 | ||
1128 | ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0, | |
1129 | RADIX_DAX_PMD | RADIX_DAX_HZP); | |
1130 | if (IS_ERR(ret)) | |
1131 | return VM_FAULT_FALLBACK; | |
1132 | *entryp = ret; | |
1133 | ||
1134 | ptl = pmd_lock(vma->vm_mm, pmd); | |
1135 | if (!pmd_none(*pmd)) { | |
1136 | spin_unlock(ptl); | |
1137 | return VM_FAULT_FALLBACK; | |
1138 | } | |
1139 | ||
1140 | pmd_entry = mk_pmd(zero_page, vma->vm_page_prot); | |
1141 | pmd_entry = pmd_mkhuge(pmd_entry); | |
1142 | set_pmd_at(vma->vm_mm, pmd_addr, pmd, pmd_entry); | |
1143 | spin_unlock(ptl); | |
1144 | return VM_FAULT_NOPAGE; | |
1145 | } | |
1146 | ||
1147 | int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address, | |
1148 | pmd_t *pmd, unsigned int flags, struct iomap_ops *ops) | |
1149 | { | |
1150 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1151 | unsigned long pmd_addr = address & PMD_MASK; | |
1152 | bool write = flags & FAULT_FLAG_WRITE; | |
9484ab1b | 1153 | unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT; |
642261ac RZ |
1154 | struct inode *inode = mapping->host; |
1155 | int result = VM_FAULT_FALLBACK; | |
1156 | struct iomap iomap = { 0 }; | |
1157 | pgoff_t max_pgoff, pgoff; | |
1158 | struct vm_fault vmf; | |
1159 | void *entry; | |
1160 | loff_t pos; | |
1161 | int error; | |
1162 | ||
1163 | /* Fall back to PTEs if we're going to COW */ | |
1164 | if (write && !(vma->vm_flags & VM_SHARED)) | |
1165 | goto fallback; | |
1166 | ||
1167 | /* If the PMD would extend outside the VMA */ | |
1168 | if (pmd_addr < vma->vm_start) | |
1169 | goto fallback; | |
1170 | if ((pmd_addr + PMD_SIZE) > vma->vm_end) | |
1171 | goto fallback; | |
1172 | ||
1173 | /* | |
1174 | * Check whether offset isn't beyond end of file now. Caller is | |
1175 | * supposed to hold locks serializing us with truncate / punch hole so | |
1176 | * this is a reliable test. | |
1177 | */ | |
1178 | pgoff = linear_page_index(vma, pmd_addr); | |
1179 | max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT; | |
1180 | ||
1181 | if (pgoff > max_pgoff) | |
1182 | return VM_FAULT_SIGBUS; | |
1183 | ||
1184 | /* If the PMD would extend beyond the file size */ | |
1185 | if ((pgoff | PG_PMD_COLOUR) > max_pgoff) | |
1186 | goto fallback; | |
1187 | ||
1188 | /* | |
1189 | * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX | |
1190 | * PMD or a HZP entry. If it can't (because a 4k page is already in | |
1191 | * the tree, for instance), it will return -EEXIST and we just fall | |
1192 | * back to 4k entries. | |
1193 | */ | |
1194 | entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD); | |
1195 | if (IS_ERR(entry)) | |
1196 | goto fallback; | |
1197 | ||
1198 | /* | |
1199 | * Note that we don't use iomap_apply here. We aren't doing I/O, only | |
1200 | * setting up a mapping, so really we're using iomap_begin() as a way | |
1201 | * to look up our filesystem block. | |
1202 | */ | |
1203 | pos = (loff_t)pgoff << PAGE_SHIFT; | |
1204 | error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap); | |
1205 | if (error) | |
1206 | goto unlock_entry; | |
1207 | if (iomap.offset + iomap.length < pos + PMD_SIZE) | |
1208 | goto finish_iomap; | |
1209 | ||
1210 | vmf.pgoff = pgoff; | |
1211 | vmf.flags = flags; | |
1212 | vmf.gfp_mask = mapping_gfp_mask(mapping) | __GFP_IO; | |
1213 | ||
1214 | switch (iomap.type) { | |
1215 | case IOMAP_MAPPED: | |
1216 | result = dax_pmd_insert_mapping(vma, pmd, &vmf, address, | |
1217 | &iomap, pos, write, &entry); | |
1218 | break; | |
1219 | case IOMAP_UNWRITTEN: | |
1220 | case IOMAP_HOLE: | |
1221 | if (WARN_ON_ONCE(write)) | |
1222 | goto finish_iomap; | |
1223 | result = dax_pmd_load_hole(vma, pmd, &vmf, address, &iomap, | |
1224 | &entry); | |
1225 | break; | |
1226 | default: | |
1227 | WARN_ON_ONCE(1); | |
1228 | break; | |
1229 | } | |
1230 | ||
1231 | finish_iomap: | |
1232 | if (ops->iomap_end) { | |
1233 | if (result == VM_FAULT_FALLBACK) { | |
1234 | ops->iomap_end(inode, pos, PMD_SIZE, 0, iomap_flags, | |
1235 | &iomap); | |
1236 | } else { | |
1237 | error = ops->iomap_end(inode, pos, PMD_SIZE, PMD_SIZE, | |
1238 | iomap_flags, &iomap); | |
1239 | if (error) | |
1240 | result = VM_FAULT_FALLBACK; | |
1241 | } | |
1242 | } | |
1243 | unlock_entry: | |
1244 | put_locked_mapping_entry(mapping, pgoff, entry); | |
1245 | fallback: | |
1246 | if (result == VM_FAULT_FALLBACK) { | |
1247 | split_huge_pmd(vma, pmd, address); | |
1248 | count_vm_event(THP_FAULT_FALLBACK); | |
1249 | } | |
1250 | return result; | |
1251 | } | |
1252 | EXPORT_SYMBOL_GPL(dax_iomap_pmd_fault); | |
1253 | #endif /* CONFIG_FS_DAX_PMD */ | |
a254e568 | 1254 | #endif /* CONFIG_FS_IOMAP */ |