]>
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 | ||
d1a5f2b4 DW |
79 | struct page *read_dax_sector(struct block_device *bdev, sector_t n) |
80 | { | |
81 | struct page *page = alloc_pages(GFP_KERNEL, 0); | |
82 | struct blk_dax_ctl dax = { | |
83 | .size = PAGE_SIZE, | |
84 | .sector = n & ~((((int) PAGE_SIZE) / 512) - 1), | |
85 | }; | |
86 | long rc; | |
87 | ||
88 | if (!page) | |
89 | return ERR_PTR(-ENOMEM); | |
90 | ||
91 | rc = dax_map_atomic(bdev, &dax); | |
92 | if (rc < 0) | |
93 | return ERR_PTR(rc); | |
94 | memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE); | |
95 | dax_unmap_atomic(bdev, &dax); | |
96 | return page; | |
97 | } | |
98 | ||
d475c634 MW |
99 | static bool buffer_written(struct buffer_head *bh) |
100 | { | |
101 | return buffer_mapped(bh) && !buffer_unwritten(bh); | |
102 | } | |
103 | ||
b2e0d162 DW |
104 | static sector_t to_sector(const struct buffer_head *bh, |
105 | const struct inode *inode) | |
106 | { | |
107 | sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9); | |
108 | ||
109 | return sector; | |
110 | } | |
111 | ||
a95cd631 OS |
112 | static ssize_t dax_io(struct inode *inode, struct iov_iter *iter, |
113 | loff_t start, loff_t end, get_block_t get_block, | |
114 | struct buffer_head *bh) | |
d475c634 | 115 | { |
b2e0d162 | 116 | loff_t pos = start, max = start, bh_max = start; |
14df6a4e | 117 | bool hole = false; |
b2e0d162 DW |
118 | struct block_device *bdev = NULL; |
119 | int rw = iov_iter_rw(iter), rc; | |
120 | long map_len = 0; | |
121 | struct blk_dax_ctl dax = { | |
7a9eb206 | 122 | .addr = ERR_PTR(-EIO), |
b2e0d162 | 123 | }; |
069c77bc JK |
124 | unsigned blkbits = inode->i_blkbits; |
125 | sector_t file_blks = (i_size_read(inode) + (1 << blkbits) - 1) | |
126 | >> blkbits; | |
b2e0d162 DW |
127 | |
128 | if (rw == READ) | |
d475c634 MW |
129 | end = min(end, i_size_read(inode)); |
130 | ||
131 | while (pos < end) { | |
2765cfbb | 132 | size_t len; |
d475c634 | 133 | if (pos == max) { |
e94f5a22 JM |
134 | long page = pos >> PAGE_SHIFT; |
135 | sector_t block = page << (PAGE_SHIFT - blkbits); | |
d475c634 MW |
136 | unsigned first = pos - (block << blkbits); |
137 | long size; | |
138 | ||
139 | if (pos == bh_max) { | |
140 | bh->b_size = PAGE_ALIGN(end - pos); | |
141 | bh->b_state = 0; | |
b2e0d162 DW |
142 | rc = get_block(inode, block, bh, rw == WRITE); |
143 | if (rc) | |
d475c634 | 144 | break; |
d475c634 | 145 | bh_max = pos - first + bh->b_size; |
b2e0d162 | 146 | bdev = bh->b_bdev; |
069c77bc JK |
147 | /* |
148 | * We allow uninitialized buffers for writes | |
149 | * beyond EOF as those cannot race with faults | |
150 | */ | |
151 | WARN_ON_ONCE( | |
152 | (buffer_new(bh) && block < file_blks) || | |
153 | (rw == WRITE && buffer_unwritten(bh))); | |
d475c634 MW |
154 | } else { |
155 | unsigned done = bh->b_size - | |
156 | (bh_max - (pos - first)); | |
157 | bh->b_blocknr += done >> blkbits; | |
158 | bh->b_size -= done; | |
159 | } | |
160 | ||
b2e0d162 | 161 | hole = rw == READ && !buffer_written(bh); |
d475c634 | 162 | if (hole) { |
d475c634 MW |
163 | size = bh->b_size - first; |
164 | } else { | |
b2e0d162 DW |
165 | dax_unmap_atomic(bdev, &dax); |
166 | dax.sector = to_sector(bh, inode); | |
167 | dax.size = bh->b_size; | |
168 | map_len = dax_map_atomic(bdev, &dax); | |
169 | if (map_len < 0) { | |
170 | rc = map_len; | |
d475c634 | 171 | break; |
b2e0d162 | 172 | } |
b2e0d162 DW |
173 | dax.addr += first; |
174 | size = map_len - first; | |
d475c634 | 175 | } |
02395435 ES |
176 | /* |
177 | * pos + size is one past the last offset for IO, | |
178 | * so pos + size can overflow loff_t at extreme offsets. | |
179 | * Cast to u64 to catch this and get the true minimum. | |
180 | */ | |
181 | max = min_t(u64, pos + size, end); | |
d475c634 MW |
182 | } |
183 | ||
2765cfbb | 184 | if (iov_iter_rw(iter) == WRITE) { |
b2e0d162 | 185 | len = copy_from_iter_pmem(dax.addr, max - pos, iter); |
2765cfbb | 186 | } else if (!hole) |
b2e0d162 | 187 | len = copy_to_iter((void __force *) dax.addr, max - pos, |
e2e05394 | 188 | iter); |
d475c634 MW |
189 | else |
190 | len = iov_iter_zero(max - pos, iter); | |
191 | ||
cadfbb6e | 192 | if (!len) { |
b2e0d162 | 193 | rc = -EFAULT; |
d475c634 | 194 | break; |
cadfbb6e | 195 | } |
d475c634 MW |
196 | |
197 | pos += len; | |
b2e0d162 DW |
198 | if (!IS_ERR(dax.addr)) |
199 | dax.addr += len; | |
d475c634 MW |
200 | } |
201 | ||
b2e0d162 | 202 | dax_unmap_atomic(bdev, &dax); |
2765cfbb | 203 | |
b2e0d162 | 204 | return (pos == start) ? rc : pos - start; |
d475c634 MW |
205 | } |
206 | ||
207 | /** | |
208 | * dax_do_io - Perform I/O to a DAX file | |
d475c634 MW |
209 | * @iocb: The control block for this I/O |
210 | * @inode: The file which the I/O is directed at | |
211 | * @iter: The addresses to do I/O from or to | |
d475c634 MW |
212 | * @get_block: The filesystem method used to translate file offsets to blocks |
213 | * @end_io: A filesystem callback for I/O completion | |
214 | * @flags: See below | |
215 | * | |
216 | * This function uses the same locking scheme as do_blockdev_direct_IO: | |
217 | * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the | |
218 | * caller for writes. For reads, we take and release the i_mutex ourselves. | |
219 | * If DIO_LOCKING is not set, the filesystem takes care of its own locking. | |
220 | * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O | |
221 | * is in progress. | |
222 | */ | |
a95cd631 | 223 | ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode, |
c8b8e32d | 224 | struct iov_iter *iter, get_block_t get_block, |
a95cd631 | 225 | dio_iodone_t end_io, int flags) |
d475c634 MW |
226 | { |
227 | struct buffer_head bh; | |
228 | ssize_t retval = -EINVAL; | |
c8b8e32d | 229 | loff_t pos = iocb->ki_pos; |
d475c634 MW |
230 | loff_t end = pos + iov_iter_count(iter); |
231 | ||
232 | memset(&bh, 0, sizeof(bh)); | |
eab95db6 | 233 | bh.b_bdev = inode->i_sb->s_bdev; |
d475c634 | 234 | |
c3d98e39 | 235 | if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) |
5955102c | 236 | inode_lock(inode); |
d475c634 MW |
237 | |
238 | /* Protects against truncate */ | |
bbab37dd MW |
239 | if (!(flags & DIO_SKIP_DIO_COUNT)) |
240 | inode_dio_begin(inode); | |
d475c634 | 241 | |
a95cd631 | 242 | retval = dax_io(inode, iter, pos, end, get_block, &bh); |
d475c634 | 243 | |
a95cd631 | 244 | if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) |
5955102c | 245 | inode_unlock(inode); |
d475c634 | 246 | |
187372a3 CH |
247 | if (end_io) { |
248 | int err; | |
249 | ||
250 | err = end_io(iocb, pos, retval, bh.b_private); | |
251 | if (err) | |
252 | retval = err; | |
253 | } | |
d475c634 | 254 | |
bbab37dd MW |
255 | if (!(flags & DIO_SKIP_DIO_COUNT)) |
256 | inode_dio_end(inode); | |
d475c634 MW |
257 | return retval; |
258 | } | |
259 | EXPORT_SYMBOL_GPL(dax_do_io); | |
f7ca90b1 | 260 | |
ac401cc7 JK |
261 | /* |
262 | * DAX radix tree locking | |
263 | */ | |
264 | struct exceptional_entry_key { | |
265 | struct address_space *mapping; | |
63e95b5c | 266 | pgoff_t entry_start; |
ac401cc7 JK |
267 | }; |
268 | ||
269 | struct wait_exceptional_entry_queue { | |
270 | wait_queue_t wait; | |
271 | struct exceptional_entry_key key; | |
272 | }; | |
273 | ||
63e95b5c RZ |
274 | static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping, |
275 | pgoff_t index, void *entry, struct exceptional_entry_key *key) | |
276 | { | |
277 | unsigned long hash; | |
278 | ||
279 | /* | |
280 | * If 'entry' is a PMD, align the 'index' that we use for the wait | |
281 | * queue to the start of that PMD. This ensures that all offsets in | |
282 | * the range covered by the PMD map to the same bit lock. | |
283 | */ | |
284 | if (RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) | |
285 | index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1); | |
286 | ||
287 | key->mapping = mapping; | |
288 | key->entry_start = index; | |
289 | ||
290 | hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS); | |
291 | return wait_table + hash; | |
292 | } | |
293 | ||
ac401cc7 JK |
294 | static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode, |
295 | int sync, void *keyp) | |
296 | { | |
297 | struct exceptional_entry_key *key = keyp; | |
298 | struct wait_exceptional_entry_queue *ewait = | |
299 | container_of(wait, struct wait_exceptional_entry_queue, wait); | |
300 | ||
301 | if (key->mapping != ewait->key.mapping || | |
63e95b5c | 302 | key->entry_start != ewait->key.entry_start) |
ac401cc7 JK |
303 | return 0; |
304 | return autoremove_wake_function(wait, mode, sync, NULL); | |
305 | } | |
306 | ||
307 | /* | |
308 | * Check whether the given slot is locked. The function must be called with | |
309 | * mapping->tree_lock held | |
310 | */ | |
311 | static inline int slot_locked(struct address_space *mapping, void **slot) | |
312 | { | |
313 | unsigned long entry = (unsigned long) | |
314 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
315 | return entry & RADIX_DAX_ENTRY_LOCK; | |
316 | } | |
317 | ||
318 | /* | |
319 | * Mark the given slot is locked. The function must be called with | |
320 | * mapping->tree_lock held | |
321 | */ | |
322 | static inline void *lock_slot(struct address_space *mapping, void **slot) | |
323 | { | |
324 | unsigned long entry = (unsigned long) | |
325 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
326 | ||
327 | entry |= RADIX_DAX_ENTRY_LOCK; | |
328 | radix_tree_replace_slot(slot, (void *)entry); | |
329 | return (void *)entry; | |
330 | } | |
331 | ||
332 | /* | |
333 | * Mark the given slot is unlocked. The function must be called with | |
334 | * mapping->tree_lock held | |
335 | */ | |
336 | static inline void *unlock_slot(struct address_space *mapping, void **slot) | |
337 | { | |
338 | unsigned long entry = (unsigned long) | |
339 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
340 | ||
341 | entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK; | |
342 | radix_tree_replace_slot(slot, (void *)entry); | |
343 | return (void *)entry; | |
344 | } | |
345 | ||
346 | /* | |
347 | * Lookup entry in radix tree, wait for it to become unlocked if it is | |
348 | * exceptional entry and return it. The caller must call | |
349 | * put_unlocked_mapping_entry() when he decided not to lock the entry or | |
350 | * put_locked_mapping_entry() when he locked the entry and now wants to | |
351 | * unlock it. | |
352 | * | |
353 | * The function must be called with mapping->tree_lock held. | |
354 | */ | |
355 | static void *get_unlocked_mapping_entry(struct address_space *mapping, | |
356 | pgoff_t index, void ***slotp) | |
357 | { | |
e3ad61c6 | 358 | void *entry, **slot; |
ac401cc7 | 359 | struct wait_exceptional_entry_queue ewait; |
63e95b5c | 360 | wait_queue_head_t *wq; |
ac401cc7 JK |
361 | |
362 | init_wait(&ewait.wait); | |
363 | ewait.wait.func = wake_exceptional_entry_func; | |
ac401cc7 JK |
364 | |
365 | for (;;) { | |
e3ad61c6 | 366 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, |
ac401cc7 | 367 | &slot); |
e3ad61c6 | 368 | if (!entry || !radix_tree_exceptional_entry(entry) || |
ac401cc7 JK |
369 | !slot_locked(mapping, slot)) { |
370 | if (slotp) | |
371 | *slotp = slot; | |
e3ad61c6 | 372 | return entry; |
ac401cc7 | 373 | } |
63e95b5c RZ |
374 | |
375 | wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key); | |
ac401cc7 JK |
376 | prepare_to_wait_exclusive(wq, &ewait.wait, |
377 | TASK_UNINTERRUPTIBLE); | |
378 | spin_unlock_irq(&mapping->tree_lock); | |
379 | schedule(); | |
380 | finish_wait(wq, &ewait.wait); | |
381 | spin_lock_irq(&mapping->tree_lock); | |
382 | } | |
383 | } | |
384 | ||
385 | /* | |
386 | * Find radix tree entry at given index. If it points to a page, return with | |
387 | * the page locked. If it points to the exceptional entry, return with the | |
388 | * radix tree entry locked. If the radix tree doesn't contain given index, | |
389 | * create empty exceptional entry for the index and return with it locked. | |
390 | * | |
391 | * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For | |
392 | * persistent memory the benefit is doubtful. We can add that later if we can | |
393 | * show it helps. | |
394 | */ | |
395 | static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index) | |
396 | { | |
e3ad61c6 | 397 | void *entry, **slot; |
ac401cc7 JK |
398 | |
399 | restart: | |
400 | spin_lock_irq(&mapping->tree_lock); | |
e3ad61c6 | 401 | entry = get_unlocked_mapping_entry(mapping, index, &slot); |
ac401cc7 | 402 | /* No entry for given index? Make sure radix tree is big enough. */ |
e3ad61c6 | 403 | if (!entry) { |
ac401cc7 JK |
404 | int err; |
405 | ||
406 | spin_unlock_irq(&mapping->tree_lock); | |
407 | err = radix_tree_preload( | |
408 | mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM); | |
409 | if (err) | |
410 | return ERR_PTR(err); | |
e3ad61c6 | 411 | entry = (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY | |
ac401cc7 JK |
412 | RADIX_DAX_ENTRY_LOCK); |
413 | spin_lock_irq(&mapping->tree_lock); | |
e3ad61c6 | 414 | err = radix_tree_insert(&mapping->page_tree, index, entry); |
ac401cc7 JK |
415 | radix_tree_preload_end(); |
416 | if (err) { | |
417 | spin_unlock_irq(&mapping->tree_lock); | |
418 | /* Someone already created the entry? */ | |
419 | if (err == -EEXIST) | |
420 | goto restart; | |
421 | return ERR_PTR(err); | |
422 | } | |
423 | /* Good, we have inserted empty locked entry into the tree. */ | |
424 | mapping->nrexceptional++; | |
425 | spin_unlock_irq(&mapping->tree_lock); | |
e3ad61c6 | 426 | return entry; |
ac401cc7 JK |
427 | } |
428 | /* Normal page in radix tree? */ | |
e3ad61c6 RZ |
429 | if (!radix_tree_exceptional_entry(entry)) { |
430 | struct page *page = entry; | |
ac401cc7 JK |
431 | |
432 | get_page(page); | |
433 | spin_unlock_irq(&mapping->tree_lock); | |
434 | lock_page(page); | |
435 | /* Page got truncated? Retry... */ | |
436 | if (unlikely(page->mapping != mapping)) { | |
437 | unlock_page(page); | |
438 | put_page(page); | |
439 | goto restart; | |
440 | } | |
441 | return page; | |
442 | } | |
e3ad61c6 | 443 | entry = lock_slot(mapping, slot); |
ac401cc7 | 444 | spin_unlock_irq(&mapping->tree_lock); |
e3ad61c6 | 445 | return entry; |
ac401cc7 JK |
446 | } |
447 | ||
63e95b5c RZ |
448 | /* |
449 | * We do not necessarily hold the mapping->tree_lock when we call this | |
450 | * function so it is possible that 'entry' is no longer a valid item in the | |
451 | * radix tree. This is okay, though, because all we really need to do is to | |
452 | * find the correct waitqueue where tasks might be sleeping waiting for that | |
453 | * old 'entry' and wake them. | |
454 | */ | |
ac401cc7 | 455 | void dax_wake_mapping_entry_waiter(struct address_space *mapping, |
63e95b5c | 456 | pgoff_t index, void *entry, bool wake_all) |
ac401cc7 | 457 | { |
63e95b5c RZ |
458 | struct exceptional_entry_key key; |
459 | wait_queue_head_t *wq; | |
460 | ||
461 | wq = dax_entry_waitqueue(mapping, index, entry, &key); | |
ac401cc7 JK |
462 | |
463 | /* | |
464 | * Checking for locked entry and prepare_to_wait_exclusive() happens | |
465 | * under mapping->tree_lock, ditto for entry handling in our callers. | |
466 | * So at this point all tasks that could have seen our entry locked | |
467 | * must be in the waitqueue and the following check will see them. | |
468 | */ | |
63e95b5c | 469 | if (waitqueue_active(wq)) |
ac401cc7 | 470 | __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); |
ac401cc7 JK |
471 | } |
472 | ||
bc2466e4 | 473 | void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index) |
ac401cc7 | 474 | { |
e3ad61c6 | 475 | void *entry, **slot; |
ac401cc7 JK |
476 | |
477 | spin_lock_irq(&mapping->tree_lock); | |
e3ad61c6 RZ |
478 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot); |
479 | if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) || | |
ac401cc7 JK |
480 | !slot_locked(mapping, slot))) { |
481 | spin_unlock_irq(&mapping->tree_lock); | |
482 | return; | |
483 | } | |
484 | unlock_slot(mapping, slot); | |
485 | spin_unlock_irq(&mapping->tree_lock); | |
63e95b5c | 486 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); |
ac401cc7 JK |
487 | } |
488 | ||
489 | static void put_locked_mapping_entry(struct address_space *mapping, | |
490 | pgoff_t index, void *entry) | |
491 | { | |
492 | if (!radix_tree_exceptional_entry(entry)) { | |
493 | unlock_page(entry); | |
494 | put_page(entry); | |
495 | } else { | |
bc2466e4 | 496 | dax_unlock_mapping_entry(mapping, index); |
ac401cc7 JK |
497 | } |
498 | } | |
499 | ||
500 | /* | |
501 | * Called when we are done with radix tree entry we looked up via | |
502 | * get_unlocked_mapping_entry() and which we didn't lock in the end. | |
503 | */ | |
504 | static void put_unlocked_mapping_entry(struct address_space *mapping, | |
505 | pgoff_t index, void *entry) | |
506 | { | |
507 | if (!radix_tree_exceptional_entry(entry)) | |
508 | return; | |
509 | ||
510 | /* We have to wake up next waiter for the radix tree entry lock */ | |
63e95b5c | 511 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); |
ac401cc7 JK |
512 | } |
513 | ||
514 | /* | |
515 | * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree | |
516 | * entry to get unlocked before deleting it. | |
517 | */ | |
518 | int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) | |
519 | { | |
520 | void *entry; | |
521 | ||
522 | spin_lock_irq(&mapping->tree_lock); | |
523 | entry = get_unlocked_mapping_entry(mapping, index, NULL); | |
524 | /* | |
525 | * This gets called from truncate / punch_hole path. As such, the caller | |
526 | * must hold locks protecting against concurrent modifications of the | |
527 | * radix tree (usually fs-private i_mmap_sem for writing). Since the | |
528 | * caller has seen exceptional entry for this index, we better find it | |
529 | * at that index as well... | |
530 | */ | |
531 | if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry))) { | |
532 | spin_unlock_irq(&mapping->tree_lock); | |
533 | return 0; | |
534 | } | |
535 | radix_tree_delete(&mapping->page_tree, index); | |
536 | mapping->nrexceptional--; | |
537 | spin_unlock_irq(&mapping->tree_lock); | |
63e95b5c | 538 | dax_wake_mapping_entry_waiter(mapping, index, entry, true); |
ac401cc7 JK |
539 | |
540 | return 1; | |
541 | } | |
542 | ||
f7ca90b1 MW |
543 | /* |
544 | * The user has performed a load from a hole in the file. Allocating | |
545 | * a new page in the file would cause excessive storage usage for | |
546 | * workloads with sparse files. We allocate a page cache page instead. | |
547 | * We'll kick it out of the page cache if it's ever written to, | |
548 | * otherwise it will simply fall out of the page cache under memory | |
549 | * pressure without ever having been dirtied. | |
550 | */ | |
ac401cc7 JK |
551 | static int dax_load_hole(struct address_space *mapping, void *entry, |
552 | struct vm_fault *vmf) | |
f7ca90b1 | 553 | { |
ac401cc7 | 554 | struct page *page; |
f7ca90b1 | 555 | |
ac401cc7 JK |
556 | /* Hole page already exists? Return it... */ |
557 | if (!radix_tree_exceptional_entry(entry)) { | |
558 | vmf->page = entry; | |
559 | return VM_FAULT_LOCKED; | |
560 | } | |
f7ca90b1 | 561 | |
ac401cc7 JK |
562 | /* This will replace locked radix tree entry with a hole page */ |
563 | page = find_or_create_page(mapping, vmf->pgoff, | |
564 | vmf->gfp_mask | __GFP_ZERO); | |
565 | if (!page) { | |
566 | put_locked_mapping_entry(mapping, vmf->pgoff, entry); | |
567 | return VM_FAULT_OOM; | |
568 | } | |
f7ca90b1 MW |
569 | vmf->page = page; |
570 | return VM_FAULT_LOCKED; | |
571 | } | |
572 | ||
b0d5e82f CH |
573 | static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size, |
574 | struct page *to, unsigned long vaddr) | |
f7ca90b1 | 575 | { |
b2e0d162 | 576 | struct blk_dax_ctl dax = { |
b0d5e82f CH |
577 | .sector = sector, |
578 | .size = size, | |
b2e0d162 | 579 | }; |
e2e05394 RZ |
580 | void *vto; |
581 | ||
b2e0d162 DW |
582 | if (dax_map_atomic(bdev, &dax) < 0) |
583 | return PTR_ERR(dax.addr); | |
f7ca90b1 | 584 | vto = kmap_atomic(to); |
b2e0d162 | 585 | copy_user_page(vto, (void __force *)dax.addr, vaddr, to); |
f7ca90b1 | 586 | kunmap_atomic(vto); |
b2e0d162 | 587 | dax_unmap_atomic(bdev, &dax); |
f7ca90b1 MW |
588 | return 0; |
589 | } | |
590 | ||
09cbfeaf | 591 | #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT)) |
9973c98e | 592 | |
ac401cc7 JK |
593 | static void *dax_insert_mapping_entry(struct address_space *mapping, |
594 | struct vm_fault *vmf, | |
595 | void *entry, sector_t sector) | |
9973c98e RZ |
596 | { |
597 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
ac401cc7 JK |
598 | int error = 0; |
599 | bool hole_fill = false; | |
600 | void *new_entry; | |
601 | pgoff_t index = vmf->pgoff; | |
9973c98e | 602 | |
ac401cc7 | 603 | if (vmf->flags & FAULT_FLAG_WRITE) |
d2b2a28e | 604 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); |
9973c98e | 605 | |
ac401cc7 JK |
606 | /* Replacing hole page with block mapping? */ |
607 | if (!radix_tree_exceptional_entry(entry)) { | |
608 | hole_fill = true; | |
609 | /* | |
610 | * Unmap the page now before we remove it from page cache below. | |
611 | * The page is locked so it cannot be faulted in again. | |
612 | */ | |
613 | unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT, | |
614 | PAGE_SIZE, 0); | |
615 | error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM); | |
616 | if (error) | |
617 | return ERR_PTR(error); | |
9973c98e RZ |
618 | } |
619 | ||
ac401cc7 JK |
620 | spin_lock_irq(&mapping->tree_lock); |
621 | new_entry = (void *)((unsigned long)RADIX_DAX_ENTRY(sector, false) | | |
622 | RADIX_DAX_ENTRY_LOCK); | |
623 | if (hole_fill) { | |
624 | __delete_from_page_cache(entry, NULL); | |
625 | /* Drop pagecache reference */ | |
626 | put_page(entry); | |
627 | error = radix_tree_insert(page_tree, index, new_entry); | |
628 | if (error) { | |
629 | new_entry = ERR_PTR(error); | |
9973c98e RZ |
630 | goto unlock; |
631 | } | |
ac401cc7 JK |
632 | mapping->nrexceptional++; |
633 | } else { | |
634 | void **slot; | |
635 | void *ret; | |
9973c98e | 636 | |
ac401cc7 JK |
637 | ret = __radix_tree_lookup(page_tree, index, NULL, &slot); |
638 | WARN_ON_ONCE(ret != entry); | |
639 | radix_tree_replace_slot(slot, new_entry); | |
9973c98e | 640 | } |
ac401cc7 | 641 | if (vmf->flags & FAULT_FLAG_WRITE) |
9973c98e RZ |
642 | radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY); |
643 | unlock: | |
644 | spin_unlock_irq(&mapping->tree_lock); | |
ac401cc7 JK |
645 | if (hole_fill) { |
646 | radix_tree_preload_end(); | |
647 | /* | |
648 | * We don't need hole page anymore, it has been replaced with | |
649 | * locked radix tree entry now. | |
650 | */ | |
651 | if (mapping->a_ops->freepage) | |
652 | mapping->a_ops->freepage(entry); | |
653 | unlock_page(entry); | |
654 | put_page(entry); | |
655 | } | |
656 | return new_entry; | |
9973c98e RZ |
657 | } |
658 | ||
659 | static int dax_writeback_one(struct block_device *bdev, | |
660 | struct address_space *mapping, pgoff_t index, void *entry) | |
661 | { | |
662 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
663 | int type = RADIX_DAX_TYPE(entry); | |
664 | struct radix_tree_node *node; | |
665 | struct blk_dax_ctl dax; | |
666 | void **slot; | |
667 | int ret = 0; | |
668 | ||
669 | spin_lock_irq(&mapping->tree_lock); | |
670 | /* | |
671 | * Regular page slots are stabilized by the page lock even | |
672 | * without the tree itself locked. These unlocked entries | |
673 | * need verification under the tree lock. | |
674 | */ | |
675 | if (!__radix_tree_lookup(page_tree, index, &node, &slot)) | |
676 | goto unlock; | |
677 | if (*slot != entry) | |
678 | goto unlock; | |
679 | ||
680 | /* another fsync thread may have already written back this entry */ | |
681 | if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)) | |
682 | goto unlock; | |
683 | ||
684 | if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) { | |
685 | ret = -EIO; | |
686 | goto unlock; | |
687 | } | |
688 | ||
689 | dax.sector = RADIX_DAX_SECTOR(entry); | |
690 | dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE); | |
691 | spin_unlock_irq(&mapping->tree_lock); | |
692 | ||
693 | /* | |
694 | * We cannot hold tree_lock while calling dax_map_atomic() because it | |
695 | * eventually calls cond_resched(). | |
696 | */ | |
697 | ret = dax_map_atomic(bdev, &dax); | |
698 | if (ret < 0) | |
699 | return ret; | |
700 | ||
701 | if (WARN_ON_ONCE(ret < dax.size)) { | |
702 | ret = -EIO; | |
703 | goto unmap; | |
704 | } | |
705 | ||
706 | wb_cache_pmem(dax.addr, dax.size); | |
707 | ||
708 | spin_lock_irq(&mapping->tree_lock); | |
709 | radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE); | |
710 | spin_unlock_irq(&mapping->tree_lock); | |
711 | unmap: | |
712 | dax_unmap_atomic(bdev, &dax); | |
713 | return ret; | |
714 | ||
715 | unlock: | |
716 | spin_unlock_irq(&mapping->tree_lock); | |
717 | return ret; | |
718 | } | |
719 | ||
720 | /* | |
721 | * Flush the mapping to the persistent domain within the byte range of [start, | |
722 | * end]. This is required by data integrity operations to ensure file data is | |
723 | * on persistent storage prior to completion of the operation. | |
724 | */ | |
7f6d5b52 RZ |
725 | int dax_writeback_mapping_range(struct address_space *mapping, |
726 | struct block_device *bdev, struct writeback_control *wbc) | |
9973c98e RZ |
727 | { |
728 | struct inode *inode = mapping->host; | |
9973c98e RZ |
729 | pgoff_t start_index, end_index, pmd_index; |
730 | pgoff_t indices[PAGEVEC_SIZE]; | |
731 | struct pagevec pvec; | |
732 | bool done = false; | |
733 | int i, ret = 0; | |
734 | void *entry; | |
735 | ||
736 | if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) | |
737 | return -EIO; | |
738 | ||
7f6d5b52 RZ |
739 | if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) |
740 | return 0; | |
741 | ||
09cbfeaf KS |
742 | start_index = wbc->range_start >> PAGE_SHIFT; |
743 | end_index = wbc->range_end >> PAGE_SHIFT; | |
9973c98e RZ |
744 | pmd_index = DAX_PMD_INDEX(start_index); |
745 | ||
746 | rcu_read_lock(); | |
747 | entry = radix_tree_lookup(&mapping->page_tree, pmd_index); | |
748 | rcu_read_unlock(); | |
749 | ||
750 | /* see if the start of our range is covered by a PMD entry */ | |
751 | if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) | |
752 | start_index = pmd_index; | |
753 | ||
754 | tag_pages_for_writeback(mapping, start_index, end_index); | |
755 | ||
756 | pagevec_init(&pvec, 0); | |
757 | while (!done) { | |
758 | pvec.nr = find_get_entries_tag(mapping, start_index, | |
759 | PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE, | |
760 | pvec.pages, indices); | |
761 | ||
762 | if (pvec.nr == 0) | |
763 | break; | |
764 | ||
765 | for (i = 0; i < pvec.nr; i++) { | |
766 | if (indices[i] > end_index) { | |
767 | done = true; | |
768 | break; | |
769 | } | |
770 | ||
771 | ret = dax_writeback_one(bdev, mapping, indices[i], | |
772 | pvec.pages[i]); | |
773 | if (ret < 0) | |
774 | return ret; | |
775 | } | |
776 | } | |
9973c98e RZ |
777 | return 0; |
778 | } | |
779 | EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); | |
780 | ||
ac401cc7 | 781 | static int dax_insert_mapping(struct address_space *mapping, |
1aaba095 CH |
782 | struct block_device *bdev, sector_t sector, size_t size, |
783 | void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf) | |
f7ca90b1 | 784 | { |
f7ca90b1 | 785 | unsigned long vaddr = (unsigned long)vmf->virtual_address; |
b2e0d162 | 786 | struct blk_dax_ctl dax = { |
1aaba095 CH |
787 | .sector = sector, |
788 | .size = size, | |
b2e0d162 | 789 | }; |
ac401cc7 JK |
790 | void *ret; |
791 | void *entry = *entryp; | |
f7ca90b1 | 792 | |
4d9a2c87 JK |
793 | if (dax_map_atomic(bdev, &dax) < 0) |
794 | return PTR_ERR(dax.addr); | |
b2e0d162 | 795 | dax_unmap_atomic(bdev, &dax); |
f7ca90b1 | 796 | |
ac401cc7 | 797 | ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector); |
4d9a2c87 JK |
798 | if (IS_ERR(ret)) |
799 | return PTR_ERR(ret); | |
ac401cc7 | 800 | *entryp = ret; |
9973c98e | 801 | |
4d9a2c87 | 802 | return vm_insert_mixed(vma, vaddr, dax.pfn); |
f7ca90b1 MW |
803 | } |
804 | ||
ce5c5d55 | 805 | /** |
6b524995 | 806 | * dax_fault - handle a page fault on a DAX file |
ce5c5d55 DC |
807 | * @vma: The virtual memory area where the fault occurred |
808 | * @vmf: The description of the fault | |
809 | * @get_block: The filesystem method used to translate file offsets to blocks | |
810 | * | |
811 | * When a page fault occurs, filesystems may call this helper in their | |
6b524995 | 812 | * fault handler for DAX files. dax_fault() assumes the caller has done all |
ce5c5d55 DC |
813 | * the necessary locking for the page fault to proceed successfully. |
814 | */ | |
6b524995 | 815 | int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf, |
02fbd139 | 816 | get_block_t get_block) |
f7ca90b1 MW |
817 | { |
818 | struct file *file = vma->vm_file; | |
819 | struct address_space *mapping = file->f_mapping; | |
820 | struct inode *inode = mapping->host; | |
ac401cc7 | 821 | void *entry; |
f7ca90b1 MW |
822 | struct buffer_head bh; |
823 | unsigned long vaddr = (unsigned long)vmf->virtual_address; | |
824 | unsigned blkbits = inode->i_blkbits; | |
825 | sector_t block; | |
826 | pgoff_t size; | |
827 | int error; | |
828 | int major = 0; | |
829 | ||
ac401cc7 JK |
830 | /* |
831 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
832 | * to hold locks serializing us with truncate / punch hole so this is | |
833 | * a reliable test. | |
834 | */ | |
f7ca90b1 MW |
835 | size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT; |
836 | if (vmf->pgoff >= size) | |
837 | return VM_FAULT_SIGBUS; | |
838 | ||
839 | memset(&bh, 0, sizeof(bh)); | |
840 | block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits); | |
eab95db6 | 841 | bh.b_bdev = inode->i_sb->s_bdev; |
f7ca90b1 MW |
842 | bh.b_size = PAGE_SIZE; |
843 | ||
ac401cc7 JK |
844 | entry = grab_mapping_entry(mapping, vmf->pgoff); |
845 | if (IS_ERR(entry)) { | |
846 | error = PTR_ERR(entry); | |
847 | goto out; | |
f7ca90b1 MW |
848 | } |
849 | ||
850 | error = get_block(inode, block, &bh, 0); | |
851 | if (!error && (bh.b_size < PAGE_SIZE)) | |
852 | error = -EIO; /* fs corruption? */ | |
853 | if (error) | |
ac401cc7 | 854 | goto unlock_entry; |
f7ca90b1 MW |
855 | |
856 | if (vmf->cow_page) { | |
857 | struct page *new_page = vmf->cow_page; | |
858 | if (buffer_written(&bh)) | |
b0d5e82f CH |
859 | error = copy_user_dax(bh.b_bdev, to_sector(&bh, inode), |
860 | bh.b_size, new_page, vaddr); | |
f7ca90b1 MW |
861 | else |
862 | clear_user_highpage(new_page, vaddr); | |
863 | if (error) | |
ac401cc7 JK |
864 | goto unlock_entry; |
865 | if (!radix_tree_exceptional_entry(entry)) { | |
866 | vmf->page = entry; | |
bc2466e4 | 867 | return VM_FAULT_LOCKED; |
ac401cc7 | 868 | } |
bc2466e4 JK |
869 | vmf->entry = entry; |
870 | return VM_FAULT_DAX_LOCKED; | |
f7ca90b1 | 871 | } |
f7ca90b1 | 872 | |
ac401cc7 | 873 | if (!buffer_mapped(&bh)) { |
f7ca90b1 MW |
874 | if (vmf->flags & FAULT_FLAG_WRITE) { |
875 | error = get_block(inode, block, &bh, 1); | |
876 | count_vm_event(PGMAJFAULT); | |
877 | mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT); | |
878 | major = VM_FAULT_MAJOR; | |
879 | if (!error && (bh.b_size < PAGE_SIZE)) | |
880 | error = -EIO; | |
881 | if (error) | |
ac401cc7 | 882 | goto unlock_entry; |
f7ca90b1 | 883 | } else { |
ac401cc7 | 884 | return dax_load_hole(mapping, entry, vmf); |
f7ca90b1 MW |
885 | } |
886 | } | |
887 | ||
02fbd139 | 888 | /* Filesystem should not return unwritten buffers to us! */ |
2b10945c | 889 | WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh)); |
1aaba095 CH |
890 | error = dax_insert_mapping(mapping, bh.b_bdev, to_sector(&bh, inode), |
891 | bh.b_size, &entry, vma, vmf); | |
ac401cc7 JK |
892 | unlock_entry: |
893 | put_locked_mapping_entry(mapping, vmf->pgoff, entry); | |
f7ca90b1 MW |
894 | out: |
895 | if (error == -ENOMEM) | |
896 | return VM_FAULT_OOM | major; | |
897 | /* -EBUSY is fine, somebody else faulted on the same PTE */ | |
898 | if ((error < 0) && (error != -EBUSY)) | |
899 | return VM_FAULT_SIGBUS | major; | |
900 | return VM_FAULT_NOPAGE | major; | |
f7ca90b1 | 901 | } |
f7ca90b1 | 902 | EXPORT_SYMBOL_GPL(dax_fault); |
4c0ccfef | 903 | |
0e3b210c BH |
904 | /** |
905 | * dax_pfn_mkwrite - handle first write to DAX page | |
906 | * @vma: The virtual memory area where the fault occurred | |
907 | * @vmf: The description of the fault | |
0e3b210c BH |
908 | */ |
909 | int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) | |
910 | { | |
9973c98e | 911 | struct file *file = vma->vm_file; |
ac401cc7 JK |
912 | struct address_space *mapping = file->f_mapping; |
913 | void *entry; | |
914 | pgoff_t index = vmf->pgoff; | |
30f471fd | 915 | |
ac401cc7 JK |
916 | spin_lock_irq(&mapping->tree_lock); |
917 | entry = get_unlocked_mapping_entry(mapping, index, NULL); | |
918 | if (!entry || !radix_tree_exceptional_entry(entry)) | |
919 | goto out; | |
920 | radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY); | |
921 | put_unlocked_mapping_entry(mapping, index, entry); | |
922 | out: | |
923 | spin_unlock_irq(&mapping->tree_lock); | |
0e3b210c BH |
924 | return VM_FAULT_NOPAGE; |
925 | } | |
926 | EXPORT_SYMBOL_GPL(dax_pfn_mkwrite); | |
927 | ||
4b0228fa VV |
928 | static bool dax_range_is_aligned(struct block_device *bdev, |
929 | unsigned int offset, unsigned int length) | |
930 | { | |
931 | unsigned short sector_size = bdev_logical_block_size(bdev); | |
932 | ||
933 | if (!IS_ALIGNED(offset, sector_size)) | |
934 | return false; | |
935 | if (!IS_ALIGNED(length, sector_size)) | |
936 | return false; | |
937 | ||
938 | return true; | |
939 | } | |
940 | ||
679c8bd3 CH |
941 | int __dax_zero_page_range(struct block_device *bdev, sector_t sector, |
942 | unsigned int offset, unsigned int length) | |
943 | { | |
944 | struct blk_dax_ctl dax = { | |
945 | .sector = sector, | |
946 | .size = PAGE_SIZE, | |
947 | }; | |
948 | ||
4b0228fa VV |
949 | if (dax_range_is_aligned(bdev, offset, length)) { |
950 | sector_t start_sector = dax.sector + (offset >> 9); | |
951 | ||
952 | return blkdev_issue_zeroout(bdev, start_sector, | |
953 | length >> 9, GFP_NOFS, true); | |
954 | } else { | |
955 | if (dax_map_atomic(bdev, &dax) < 0) | |
956 | return PTR_ERR(dax.addr); | |
957 | clear_pmem(dax.addr + offset, length); | |
4b0228fa VV |
958 | dax_unmap_atomic(bdev, &dax); |
959 | } | |
679c8bd3 CH |
960 | return 0; |
961 | } | |
962 | EXPORT_SYMBOL_GPL(__dax_zero_page_range); | |
963 | ||
4c0ccfef | 964 | /** |
25726bc1 | 965 | * dax_zero_page_range - zero a range within a page of a DAX file |
4c0ccfef MW |
966 | * @inode: The file being truncated |
967 | * @from: The file offset that is being truncated to | |
25726bc1 | 968 | * @length: The number of bytes to zero |
4c0ccfef MW |
969 | * @get_block: The filesystem method used to translate file offsets to blocks |
970 | * | |
25726bc1 MW |
971 | * This function can be called by a filesystem when it is zeroing part of a |
972 | * page in a DAX file. This is intended for hole-punch operations. If | |
973 | * you are truncating a file, the helper function dax_truncate_page() may be | |
974 | * more convenient. | |
4c0ccfef | 975 | */ |
25726bc1 MW |
976 | int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length, |
977 | get_block_t get_block) | |
4c0ccfef MW |
978 | { |
979 | struct buffer_head bh; | |
09cbfeaf KS |
980 | pgoff_t index = from >> PAGE_SHIFT; |
981 | unsigned offset = from & (PAGE_SIZE-1); | |
4c0ccfef MW |
982 | int err; |
983 | ||
984 | /* Block boundary? Nothing to do */ | |
985 | if (!length) | |
986 | return 0; | |
aada54f9 RZ |
987 | if (WARN_ON_ONCE((offset + length) > PAGE_SIZE)) |
988 | return -EINVAL; | |
4c0ccfef MW |
989 | |
990 | memset(&bh, 0, sizeof(bh)); | |
eab95db6 | 991 | bh.b_bdev = inode->i_sb->s_bdev; |
09cbfeaf | 992 | bh.b_size = PAGE_SIZE; |
4c0ccfef | 993 | err = get_block(inode, index, &bh, 0); |
679c8bd3 | 994 | if (err < 0 || !buffer_written(&bh)) |
4c0ccfef | 995 | return err; |
4c0ccfef | 996 | |
679c8bd3 CH |
997 | return __dax_zero_page_range(bh.b_bdev, to_sector(&bh, inode), |
998 | offset, length); | |
4c0ccfef | 999 | } |
25726bc1 MW |
1000 | EXPORT_SYMBOL_GPL(dax_zero_page_range); |
1001 | ||
1002 | /** | |
1003 | * dax_truncate_page - handle a partial page being truncated in a DAX file | |
1004 | * @inode: The file being truncated | |
1005 | * @from: The file offset that is being truncated to | |
1006 | * @get_block: The filesystem method used to translate file offsets to blocks | |
1007 | * | |
1008 | * Similar to block_truncate_page(), this function can be called by a | |
1009 | * filesystem when it is truncating a DAX file to handle the partial page. | |
25726bc1 MW |
1010 | */ |
1011 | int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block) | |
1012 | { | |
09cbfeaf | 1013 | unsigned length = PAGE_ALIGN(from) - from; |
25726bc1 MW |
1014 | return dax_zero_page_range(inode, from, length, get_block); |
1015 | } | |
4c0ccfef | 1016 | EXPORT_SYMBOL_GPL(dax_truncate_page); |
a254e568 CH |
1017 | |
1018 | #ifdef CONFIG_FS_IOMAP | |
333ccc97 RZ |
1019 | static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) |
1020 | { | |
1021 | return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9); | |
1022 | } | |
1023 | ||
a254e568 | 1024 | static loff_t |
11c59c92 | 1025 | dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, |
a254e568 CH |
1026 | struct iomap *iomap) |
1027 | { | |
1028 | struct iov_iter *iter = data; | |
1029 | loff_t end = pos + length, done = 0; | |
1030 | ssize_t ret = 0; | |
1031 | ||
1032 | if (iov_iter_rw(iter) == READ) { | |
1033 | end = min(end, i_size_read(inode)); | |
1034 | if (pos >= end) | |
1035 | return 0; | |
1036 | ||
1037 | if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) | |
1038 | return iov_iter_zero(min(length, end - pos), iter); | |
1039 | } | |
1040 | ||
1041 | if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) | |
1042 | return -EIO; | |
1043 | ||
1044 | while (pos < end) { | |
1045 | unsigned offset = pos & (PAGE_SIZE - 1); | |
1046 | struct blk_dax_ctl dax = { 0 }; | |
1047 | ssize_t map_len; | |
1048 | ||
333ccc97 | 1049 | dax.sector = dax_iomap_sector(iomap, pos); |
a254e568 CH |
1050 | dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK; |
1051 | map_len = dax_map_atomic(iomap->bdev, &dax); | |
1052 | if (map_len < 0) { | |
1053 | ret = map_len; | |
1054 | break; | |
1055 | } | |
1056 | ||
1057 | dax.addr += offset; | |
1058 | map_len -= offset; | |
1059 | if (map_len > end - pos) | |
1060 | map_len = end - pos; | |
1061 | ||
1062 | if (iov_iter_rw(iter) == WRITE) | |
1063 | map_len = copy_from_iter_pmem(dax.addr, map_len, iter); | |
1064 | else | |
1065 | map_len = copy_to_iter(dax.addr, map_len, iter); | |
1066 | dax_unmap_atomic(iomap->bdev, &dax); | |
1067 | if (map_len <= 0) { | |
1068 | ret = map_len ? map_len : -EFAULT; | |
1069 | break; | |
1070 | } | |
1071 | ||
1072 | pos += map_len; | |
1073 | length -= map_len; | |
1074 | done += map_len; | |
1075 | } | |
1076 | ||
1077 | return done ? done : ret; | |
1078 | } | |
1079 | ||
1080 | /** | |
11c59c92 | 1081 | * dax_iomap_rw - Perform I/O to a DAX file |
a254e568 CH |
1082 | * @iocb: The control block for this I/O |
1083 | * @iter: The addresses to do I/O from or to | |
1084 | * @ops: iomap ops passed from the file system | |
1085 | * | |
1086 | * This function performs read and write operations to directly mapped | |
1087 | * persistent memory. The callers needs to take care of read/write exclusion | |
1088 | * and evicting any page cache pages in the region under I/O. | |
1089 | */ | |
1090 | ssize_t | |
11c59c92 | 1091 | dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, |
a254e568 CH |
1092 | struct iomap_ops *ops) |
1093 | { | |
1094 | struct address_space *mapping = iocb->ki_filp->f_mapping; | |
1095 | struct inode *inode = mapping->host; | |
1096 | loff_t pos = iocb->ki_pos, ret = 0, done = 0; | |
1097 | unsigned flags = 0; | |
1098 | ||
1099 | if (iov_iter_rw(iter) == WRITE) | |
1100 | flags |= IOMAP_WRITE; | |
1101 | ||
1102 | /* | |
1103 | * Yes, even DAX files can have page cache attached to them: A zeroed | |
1104 | * page is inserted into the pagecache when we have to serve a write | |
1105 | * fault on a hole. It should never be dirtied and can simply be | |
1106 | * dropped from the pagecache once we get real data for the page. | |
1107 | * | |
1108 | * XXX: This is racy against mmap, and there's nothing we can do about | |
1109 | * it. We'll eventually need to shift this down even further so that | |
1110 | * we can check if we allocated blocks over a hole first. | |
1111 | */ | |
1112 | if (mapping->nrpages) { | |
1113 | ret = invalidate_inode_pages2_range(mapping, | |
1114 | pos >> PAGE_SHIFT, | |
1115 | (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT); | |
1116 | WARN_ON_ONCE(ret); | |
1117 | } | |
1118 | ||
1119 | while (iov_iter_count(iter)) { | |
1120 | ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, | |
11c59c92 | 1121 | iter, dax_iomap_actor); |
a254e568 CH |
1122 | if (ret <= 0) |
1123 | break; | |
1124 | pos += ret; | |
1125 | done += ret; | |
1126 | } | |
1127 | ||
1128 | iocb->ki_pos += done; | |
1129 | return done ? done : ret; | |
1130 | } | |
11c59c92 | 1131 | EXPORT_SYMBOL_GPL(dax_iomap_rw); |
a7d73fe6 CH |
1132 | |
1133 | /** | |
11c59c92 | 1134 | * dax_iomap_fault - handle a page fault on a DAX file |
a7d73fe6 CH |
1135 | * @vma: The virtual memory area where the fault occurred |
1136 | * @vmf: The description of the fault | |
1137 | * @ops: iomap ops passed from the file system | |
1138 | * | |
1139 | * When a page fault occurs, filesystems may call this helper in their fault | |
1140 | * or mkwrite handler for DAX files. Assumes the caller has done all the | |
1141 | * necessary locking for the page fault to proceed successfully. | |
1142 | */ | |
11c59c92 | 1143 | int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf, |
a7d73fe6 CH |
1144 | struct iomap_ops *ops) |
1145 | { | |
1146 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1147 | struct inode *inode = mapping->host; | |
1148 | unsigned long vaddr = (unsigned long)vmf->virtual_address; | |
1149 | loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; | |
1150 | sector_t sector; | |
1151 | struct iomap iomap = { 0 }; | |
1152 | unsigned flags = 0; | |
1153 | int error, major = 0; | |
1550290b | 1154 | int locked_status = 0; |
a7d73fe6 CH |
1155 | void *entry; |
1156 | ||
1157 | /* | |
1158 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
1159 | * to hold locks serializing us with truncate / punch hole so this is | |
1160 | * a reliable test. | |
1161 | */ | |
1162 | if (pos >= i_size_read(inode)) | |
1163 | return VM_FAULT_SIGBUS; | |
1164 | ||
1165 | entry = grab_mapping_entry(mapping, vmf->pgoff); | |
1166 | if (IS_ERR(entry)) { | |
1167 | error = PTR_ERR(entry); | |
1168 | goto out; | |
1169 | } | |
1170 | ||
1171 | if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) | |
1172 | flags |= IOMAP_WRITE; | |
1173 | ||
1174 | /* | |
1175 | * Note that we don't bother to use iomap_apply here: DAX required | |
1176 | * the file system block size to be equal the page size, which means | |
1177 | * that we never have to deal with more than a single extent here. | |
1178 | */ | |
1179 | error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); | |
1180 | if (error) | |
1181 | goto unlock_entry; | |
1182 | if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { | |
1183 | error = -EIO; /* fs corruption? */ | |
1550290b | 1184 | goto finish_iomap; |
a7d73fe6 CH |
1185 | } |
1186 | ||
333ccc97 | 1187 | sector = dax_iomap_sector(&iomap, pos); |
a7d73fe6 CH |
1188 | |
1189 | if (vmf->cow_page) { | |
1190 | switch (iomap.type) { | |
1191 | case IOMAP_HOLE: | |
1192 | case IOMAP_UNWRITTEN: | |
1193 | clear_user_highpage(vmf->cow_page, vaddr); | |
1194 | break; | |
1195 | case IOMAP_MAPPED: | |
1196 | error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE, | |
1197 | vmf->cow_page, vaddr); | |
1198 | break; | |
1199 | default: | |
1200 | WARN_ON_ONCE(1); | |
1201 | error = -EIO; | |
1202 | break; | |
1203 | } | |
1204 | ||
1205 | if (error) | |
1550290b | 1206 | goto finish_iomap; |
a7d73fe6 CH |
1207 | if (!radix_tree_exceptional_entry(entry)) { |
1208 | vmf->page = entry; | |
1550290b RZ |
1209 | locked_status = VM_FAULT_LOCKED; |
1210 | } else { | |
1211 | vmf->entry = entry; | |
1212 | locked_status = VM_FAULT_DAX_LOCKED; | |
a7d73fe6 | 1213 | } |
1550290b | 1214 | goto finish_iomap; |
a7d73fe6 CH |
1215 | } |
1216 | ||
1217 | switch (iomap.type) { | |
1218 | case IOMAP_MAPPED: | |
1219 | if (iomap.flags & IOMAP_F_NEW) { | |
1220 | count_vm_event(PGMAJFAULT); | |
1221 | mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT); | |
1222 | major = VM_FAULT_MAJOR; | |
1223 | } | |
1224 | error = dax_insert_mapping(mapping, iomap.bdev, sector, | |
1225 | PAGE_SIZE, &entry, vma, vmf); | |
1226 | break; | |
1227 | case IOMAP_UNWRITTEN: | |
1228 | case IOMAP_HOLE: | |
1550290b RZ |
1229 | if (!(vmf->flags & FAULT_FLAG_WRITE)) { |
1230 | locked_status = dax_load_hole(mapping, entry, vmf); | |
1231 | break; | |
1232 | } | |
a7d73fe6 CH |
1233 | /*FALLTHRU*/ |
1234 | default: | |
1235 | WARN_ON_ONCE(1); | |
1236 | error = -EIO; | |
1237 | break; | |
1238 | } | |
1239 | ||
1550290b RZ |
1240 | finish_iomap: |
1241 | if (ops->iomap_end) { | |
1242 | if (error) { | |
1243 | /* keep previous error */ | |
1244 | ops->iomap_end(inode, pos, PAGE_SIZE, 0, flags, | |
1245 | &iomap); | |
1246 | } else { | |
1247 | error = ops->iomap_end(inode, pos, PAGE_SIZE, | |
1248 | PAGE_SIZE, flags, &iomap); | |
1249 | } | |
1250 | } | |
a7d73fe6 | 1251 | unlock_entry: |
1550290b RZ |
1252 | if (!locked_status || error) |
1253 | put_locked_mapping_entry(mapping, vmf->pgoff, entry); | |
a7d73fe6 CH |
1254 | out: |
1255 | if (error == -ENOMEM) | |
1256 | return VM_FAULT_OOM | major; | |
1257 | /* -EBUSY is fine, somebody else faulted on the same PTE */ | |
1258 | if (error < 0 && error != -EBUSY) | |
1259 | return VM_FAULT_SIGBUS | major; | |
1550290b RZ |
1260 | if (locked_status) { |
1261 | WARN_ON_ONCE(error); /* -EBUSY from ops->iomap_end? */ | |
1262 | return locked_status; | |
1263 | } | |
a7d73fe6 CH |
1264 | return VM_FAULT_NOPAGE | major; |
1265 | } | |
11c59c92 | 1266 | EXPORT_SYMBOL_GPL(dax_iomap_fault); |
a254e568 | 1267 | #endif /* CONFIG_FS_IOMAP */ |