]>
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 | |
e804315d JK |
37 | /* |
38 | * We use lowest available bit in exceptional entry for locking, other two | |
39 | * bits to determine entry type. In total 3 special bits. | |
40 | */ | |
41 | #define RADIX_DAX_SHIFT (RADIX_TREE_EXCEPTIONAL_SHIFT + 3) | |
42 | #define RADIX_DAX_PTE (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1)) | |
43 | #define RADIX_DAX_PMD (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2)) | |
44 | #define RADIX_DAX_TYPE_MASK (RADIX_DAX_PTE | RADIX_DAX_PMD) | |
45 | #define RADIX_DAX_TYPE(entry) ((unsigned long)entry & RADIX_DAX_TYPE_MASK) | |
78a9be0a N |
46 | #define RADIX_DAX_SECTOR(entry) (((unsigned long)entry >> RADIX_DAX_SHIFT)) |
47 | #define RADIX_DAX_ENTRY(sector, pmd) ((void *)((unsigned long)sector << \ | |
e804315d JK |
48 | RADIX_DAX_SHIFT | (pmd ? RADIX_DAX_PMD : RADIX_DAX_PTE) | \ |
49 | RADIX_TREE_EXCEPTIONAL_ENTRY)) | |
e4b27491 | 50 | |
ac401cc7 JK |
51 | /* We choose 4096 entries - same as per-zone page wait tables */ |
52 | #define DAX_WAIT_TABLE_BITS 12 | |
53 | #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) | |
54 | ||
ce95ab0f | 55 | static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; |
ac401cc7 JK |
56 | |
57 | static int __init init_dax_wait_table(void) | |
58 | { | |
59 | int i; | |
60 | ||
61 | for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) | |
62 | init_waitqueue_head(wait_table + i); | |
63 | return 0; | |
64 | } | |
65 | fs_initcall(init_dax_wait_table); | |
66 | ||
b2e0d162 DW |
67 | static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax) |
68 | { | |
69 | struct request_queue *q = bdev->bd_queue; | |
70 | long rc = -EIO; | |
71 | ||
7a9eb206 | 72 | dax->addr = ERR_PTR(-EIO); |
b2e0d162 DW |
73 | if (blk_queue_enter(q, true) != 0) |
74 | return rc; | |
75 | ||
76 | rc = bdev_direct_access(bdev, dax); | |
77 | if (rc < 0) { | |
7a9eb206 | 78 | dax->addr = ERR_PTR(rc); |
b2e0d162 DW |
79 | blk_queue_exit(q); |
80 | return rc; | |
81 | } | |
82 | return rc; | |
83 | } | |
84 | ||
85 | static void dax_unmap_atomic(struct block_device *bdev, | |
86 | const struct blk_dax_ctl *dax) | |
87 | { | |
88 | if (IS_ERR(dax->addr)) | |
89 | return; | |
90 | blk_queue_exit(bdev->bd_queue); | |
91 | } | |
92 | ||
d1a5f2b4 DW |
93 | struct page *read_dax_sector(struct block_device *bdev, sector_t n) |
94 | { | |
95 | struct page *page = alloc_pages(GFP_KERNEL, 0); | |
96 | struct blk_dax_ctl dax = { | |
97 | .size = PAGE_SIZE, | |
98 | .sector = n & ~((((int) PAGE_SIZE) / 512) - 1), | |
99 | }; | |
100 | long rc; | |
101 | ||
102 | if (!page) | |
103 | return ERR_PTR(-ENOMEM); | |
104 | ||
105 | rc = dax_map_atomic(bdev, &dax); | |
106 | if (rc < 0) | |
107 | return ERR_PTR(rc); | |
108 | memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE); | |
109 | dax_unmap_atomic(bdev, &dax); | |
110 | return page; | |
111 | } | |
112 | ||
d475c634 MW |
113 | static bool buffer_written(struct buffer_head *bh) |
114 | { | |
115 | return buffer_mapped(bh) && !buffer_unwritten(bh); | |
116 | } | |
117 | ||
b2e0d162 DW |
118 | static sector_t to_sector(const struct buffer_head *bh, |
119 | const struct inode *inode) | |
120 | { | |
121 | sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9); | |
122 | ||
123 | return sector; | |
124 | } | |
125 | ||
a95cd631 OS |
126 | static ssize_t dax_io(struct inode *inode, struct iov_iter *iter, |
127 | loff_t start, loff_t end, get_block_t get_block, | |
128 | struct buffer_head *bh) | |
d475c634 | 129 | { |
b2e0d162 | 130 | loff_t pos = start, max = start, bh_max = start; |
14df6a4e | 131 | bool hole = false; |
b2e0d162 DW |
132 | struct block_device *bdev = NULL; |
133 | int rw = iov_iter_rw(iter), rc; | |
134 | long map_len = 0; | |
135 | struct blk_dax_ctl dax = { | |
7a9eb206 | 136 | .addr = ERR_PTR(-EIO), |
b2e0d162 | 137 | }; |
069c77bc JK |
138 | unsigned blkbits = inode->i_blkbits; |
139 | sector_t file_blks = (i_size_read(inode) + (1 << blkbits) - 1) | |
140 | >> blkbits; | |
b2e0d162 DW |
141 | |
142 | if (rw == READ) | |
d475c634 MW |
143 | end = min(end, i_size_read(inode)); |
144 | ||
145 | while (pos < end) { | |
2765cfbb | 146 | size_t len; |
d475c634 | 147 | if (pos == max) { |
e94f5a22 JM |
148 | long page = pos >> PAGE_SHIFT; |
149 | sector_t block = page << (PAGE_SHIFT - blkbits); | |
d475c634 MW |
150 | unsigned first = pos - (block << blkbits); |
151 | long size; | |
152 | ||
153 | if (pos == bh_max) { | |
154 | bh->b_size = PAGE_ALIGN(end - pos); | |
155 | bh->b_state = 0; | |
b2e0d162 DW |
156 | rc = get_block(inode, block, bh, rw == WRITE); |
157 | if (rc) | |
d475c634 | 158 | break; |
d475c634 | 159 | bh_max = pos - first + bh->b_size; |
b2e0d162 | 160 | bdev = bh->b_bdev; |
069c77bc JK |
161 | /* |
162 | * We allow uninitialized buffers for writes | |
163 | * beyond EOF as those cannot race with faults | |
164 | */ | |
165 | WARN_ON_ONCE( | |
166 | (buffer_new(bh) && block < file_blks) || | |
167 | (rw == WRITE && buffer_unwritten(bh))); | |
d475c634 MW |
168 | } else { |
169 | unsigned done = bh->b_size - | |
170 | (bh_max - (pos - first)); | |
171 | bh->b_blocknr += done >> blkbits; | |
172 | bh->b_size -= done; | |
173 | } | |
174 | ||
b2e0d162 | 175 | hole = rw == READ && !buffer_written(bh); |
d475c634 | 176 | if (hole) { |
d475c634 MW |
177 | size = bh->b_size - first; |
178 | } else { | |
b2e0d162 DW |
179 | dax_unmap_atomic(bdev, &dax); |
180 | dax.sector = to_sector(bh, inode); | |
181 | dax.size = bh->b_size; | |
182 | map_len = dax_map_atomic(bdev, &dax); | |
183 | if (map_len < 0) { | |
184 | rc = map_len; | |
d475c634 | 185 | break; |
b2e0d162 | 186 | } |
b2e0d162 DW |
187 | dax.addr += first; |
188 | size = map_len - first; | |
d475c634 | 189 | } |
02395435 ES |
190 | /* |
191 | * pos + size is one past the last offset for IO, | |
192 | * so pos + size can overflow loff_t at extreme offsets. | |
193 | * Cast to u64 to catch this and get the true minimum. | |
194 | */ | |
195 | max = min_t(u64, pos + size, end); | |
d475c634 MW |
196 | } |
197 | ||
2765cfbb | 198 | if (iov_iter_rw(iter) == WRITE) { |
b2e0d162 | 199 | len = copy_from_iter_pmem(dax.addr, max - pos, iter); |
2765cfbb | 200 | } else if (!hole) |
b2e0d162 | 201 | len = copy_to_iter((void __force *) dax.addr, max - pos, |
e2e05394 | 202 | iter); |
d475c634 MW |
203 | else |
204 | len = iov_iter_zero(max - pos, iter); | |
205 | ||
cadfbb6e | 206 | if (!len) { |
b2e0d162 | 207 | rc = -EFAULT; |
d475c634 | 208 | break; |
cadfbb6e | 209 | } |
d475c634 MW |
210 | |
211 | pos += len; | |
b2e0d162 DW |
212 | if (!IS_ERR(dax.addr)) |
213 | dax.addr += len; | |
d475c634 MW |
214 | } |
215 | ||
b2e0d162 | 216 | dax_unmap_atomic(bdev, &dax); |
2765cfbb | 217 | |
b2e0d162 | 218 | return (pos == start) ? rc : pos - start; |
d475c634 MW |
219 | } |
220 | ||
221 | /** | |
222 | * dax_do_io - Perform I/O to a DAX file | |
d475c634 MW |
223 | * @iocb: The control block for this I/O |
224 | * @inode: The file which the I/O is directed at | |
225 | * @iter: The addresses to do I/O from or to | |
d475c634 MW |
226 | * @get_block: The filesystem method used to translate file offsets to blocks |
227 | * @end_io: A filesystem callback for I/O completion | |
228 | * @flags: See below | |
229 | * | |
230 | * This function uses the same locking scheme as do_blockdev_direct_IO: | |
231 | * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the | |
232 | * caller for writes. For reads, we take and release the i_mutex ourselves. | |
233 | * If DIO_LOCKING is not set, the filesystem takes care of its own locking. | |
234 | * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O | |
235 | * is in progress. | |
236 | */ | |
a95cd631 | 237 | ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode, |
c8b8e32d | 238 | struct iov_iter *iter, get_block_t get_block, |
a95cd631 | 239 | dio_iodone_t end_io, int flags) |
d475c634 MW |
240 | { |
241 | struct buffer_head bh; | |
242 | ssize_t retval = -EINVAL; | |
c8b8e32d | 243 | loff_t pos = iocb->ki_pos; |
d475c634 MW |
244 | loff_t end = pos + iov_iter_count(iter); |
245 | ||
246 | memset(&bh, 0, sizeof(bh)); | |
eab95db6 | 247 | bh.b_bdev = inode->i_sb->s_bdev; |
d475c634 | 248 | |
c3d98e39 | 249 | if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) |
5955102c | 250 | inode_lock(inode); |
d475c634 MW |
251 | |
252 | /* Protects against truncate */ | |
bbab37dd MW |
253 | if (!(flags & DIO_SKIP_DIO_COUNT)) |
254 | inode_dio_begin(inode); | |
d475c634 | 255 | |
a95cd631 | 256 | retval = dax_io(inode, iter, pos, end, get_block, &bh); |
d475c634 | 257 | |
a95cd631 | 258 | if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) |
5955102c | 259 | inode_unlock(inode); |
d475c634 | 260 | |
187372a3 CH |
261 | if (end_io) { |
262 | int err; | |
263 | ||
264 | err = end_io(iocb, pos, retval, bh.b_private); | |
265 | if (err) | |
266 | retval = err; | |
267 | } | |
d475c634 | 268 | |
bbab37dd MW |
269 | if (!(flags & DIO_SKIP_DIO_COUNT)) |
270 | inode_dio_end(inode); | |
d475c634 MW |
271 | return retval; |
272 | } | |
273 | EXPORT_SYMBOL_GPL(dax_do_io); | |
f7ca90b1 | 274 | |
ac401cc7 JK |
275 | /* |
276 | * DAX radix tree locking | |
277 | */ | |
278 | struct exceptional_entry_key { | |
279 | struct address_space *mapping; | |
63e95b5c | 280 | pgoff_t entry_start; |
ac401cc7 JK |
281 | }; |
282 | ||
283 | struct wait_exceptional_entry_queue { | |
284 | wait_queue_t wait; | |
285 | struct exceptional_entry_key key; | |
286 | }; | |
287 | ||
63e95b5c RZ |
288 | static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping, |
289 | pgoff_t index, void *entry, struct exceptional_entry_key *key) | |
290 | { | |
291 | unsigned long hash; | |
292 | ||
293 | /* | |
294 | * If 'entry' is a PMD, align the 'index' that we use for the wait | |
295 | * queue to the start of that PMD. This ensures that all offsets in | |
296 | * the range covered by the PMD map to the same bit lock. | |
297 | */ | |
298 | if (RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) | |
299 | index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1); | |
300 | ||
301 | key->mapping = mapping; | |
302 | key->entry_start = index; | |
303 | ||
304 | hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS); | |
305 | return wait_table + hash; | |
306 | } | |
307 | ||
ac401cc7 JK |
308 | static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode, |
309 | int sync, void *keyp) | |
310 | { | |
311 | struct exceptional_entry_key *key = keyp; | |
312 | struct wait_exceptional_entry_queue *ewait = | |
313 | container_of(wait, struct wait_exceptional_entry_queue, wait); | |
314 | ||
315 | if (key->mapping != ewait->key.mapping || | |
63e95b5c | 316 | key->entry_start != ewait->key.entry_start) |
ac401cc7 JK |
317 | return 0; |
318 | return autoremove_wake_function(wait, mode, sync, NULL); | |
319 | } | |
320 | ||
321 | /* | |
322 | * Check whether the given slot is locked. The function must be called with | |
323 | * mapping->tree_lock held | |
324 | */ | |
325 | static inline int slot_locked(struct address_space *mapping, void **slot) | |
326 | { | |
327 | unsigned long entry = (unsigned long) | |
328 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
329 | return entry & RADIX_DAX_ENTRY_LOCK; | |
330 | } | |
331 | ||
332 | /* | |
333 | * Mark the given slot is locked. The function must be called with | |
334 | * mapping->tree_lock held | |
335 | */ | |
336 | static inline void *lock_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 |= RADIX_DAX_ENTRY_LOCK; | |
342 | radix_tree_replace_slot(slot, (void *)entry); | |
343 | return (void *)entry; | |
344 | } | |
345 | ||
346 | /* | |
347 | * Mark the given slot is unlocked. The function must be called with | |
348 | * mapping->tree_lock held | |
349 | */ | |
350 | static inline void *unlock_slot(struct address_space *mapping, void **slot) | |
351 | { | |
352 | unsigned long entry = (unsigned long) | |
353 | radix_tree_deref_slot_protected(slot, &mapping->tree_lock); | |
354 | ||
355 | entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK; | |
356 | radix_tree_replace_slot(slot, (void *)entry); | |
357 | return (void *)entry; | |
358 | } | |
359 | ||
360 | /* | |
361 | * Lookup entry in radix tree, wait for it to become unlocked if it is | |
362 | * exceptional entry and return it. The caller must call | |
363 | * put_unlocked_mapping_entry() when he decided not to lock the entry or | |
364 | * put_locked_mapping_entry() when he locked the entry and now wants to | |
365 | * unlock it. | |
366 | * | |
367 | * The function must be called with mapping->tree_lock held. | |
368 | */ | |
369 | static void *get_unlocked_mapping_entry(struct address_space *mapping, | |
370 | pgoff_t index, void ***slotp) | |
371 | { | |
e3ad61c6 | 372 | void *entry, **slot; |
ac401cc7 | 373 | struct wait_exceptional_entry_queue ewait; |
63e95b5c | 374 | wait_queue_head_t *wq; |
ac401cc7 JK |
375 | |
376 | init_wait(&ewait.wait); | |
377 | ewait.wait.func = wake_exceptional_entry_func; | |
ac401cc7 JK |
378 | |
379 | for (;;) { | |
e3ad61c6 | 380 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, |
ac401cc7 | 381 | &slot); |
e3ad61c6 | 382 | if (!entry || !radix_tree_exceptional_entry(entry) || |
ac401cc7 JK |
383 | !slot_locked(mapping, slot)) { |
384 | if (slotp) | |
385 | *slotp = slot; | |
e3ad61c6 | 386 | return entry; |
ac401cc7 | 387 | } |
63e95b5c RZ |
388 | |
389 | wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key); | |
ac401cc7 JK |
390 | prepare_to_wait_exclusive(wq, &ewait.wait, |
391 | TASK_UNINTERRUPTIBLE); | |
392 | spin_unlock_irq(&mapping->tree_lock); | |
393 | schedule(); | |
394 | finish_wait(wq, &ewait.wait); | |
395 | spin_lock_irq(&mapping->tree_lock); | |
396 | } | |
397 | } | |
398 | ||
399 | /* | |
400 | * Find radix tree entry at given index. If it points to a page, return with | |
401 | * the page locked. If it points to the exceptional entry, return with the | |
402 | * radix tree entry locked. If the radix tree doesn't contain given index, | |
403 | * create empty exceptional entry for the index and return with it locked. | |
404 | * | |
405 | * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For | |
406 | * persistent memory the benefit is doubtful. We can add that later if we can | |
407 | * show it helps. | |
408 | */ | |
409 | static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index) | |
410 | { | |
e3ad61c6 | 411 | void *entry, **slot; |
ac401cc7 JK |
412 | |
413 | restart: | |
414 | spin_lock_irq(&mapping->tree_lock); | |
e3ad61c6 | 415 | entry = get_unlocked_mapping_entry(mapping, index, &slot); |
ac401cc7 | 416 | /* No entry for given index? Make sure radix tree is big enough. */ |
e3ad61c6 | 417 | if (!entry) { |
ac401cc7 JK |
418 | int err; |
419 | ||
420 | spin_unlock_irq(&mapping->tree_lock); | |
421 | err = radix_tree_preload( | |
422 | mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM); | |
423 | if (err) | |
424 | return ERR_PTR(err); | |
e3ad61c6 | 425 | entry = (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY | |
ac401cc7 JK |
426 | RADIX_DAX_ENTRY_LOCK); |
427 | spin_lock_irq(&mapping->tree_lock); | |
e3ad61c6 | 428 | err = radix_tree_insert(&mapping->page_tree, index, entry); |
ac401cc7 JK |
429 | radix_tree_preload_end(); |
430 | if (err) { | |
431 | spin_unlock_irq(&mapping->tree_lock); | |
432 | /* Someone already created the entry? */ | |
433 | if (err == -EEXIST) | |
434 | goto restart; | |
435 | return ERR_PTR(err); | |
436 | } | |
437 | /* Good, we have inserted empty locked entry into the tree. */ | |
438 | mapping->nrexceptional++; | |
439 | spin_unlock_irq(&mapping->tree_lock); | |
e3ad61c6 | 440 | return entry; |
ac401cc7 JK |
441 | } |
442 | /* Normal page in radix tree? */ | |
e3ad61c6 RZ |
443 | if (!radix_tree_exceptional_entry(entry)) { |
444 | struct page *page = entry; | |
ac401cc7 JK |
445 | |
446 | get_page(page); | |
447 | spin_unlock_irq(&mapping->tree_lock); | |
448 | lock_page(page); | |
449 | /* Page got truncated? Retry... */ | |
450 | if (unlikely(page->mapping != mapping)) { | |
451 | unlock_page(page); | |
452 | put_page(page); | |
453 | goto restart; | |
454 | } | |
455 | return page; | |
456 | } | |
e3ad61c6 | 457 | entry = lock_slot(mapping, slot); |
ac401cc7 | 458 | spin_unlock_irq(&mapping->tree_lock); |
e3ad61c6 | 459 | return entry; |
ac401cc7 JK |
460 | } |
461 | ||
63e95b5c RZ |
462 | /* |
463 | * We do not necessarily hold the mapping->tree_lock when we call this | |
464 | * function so it is possible that 'entry' is no longer a valid item in the | |
465 | * radix tree. This is okay, though, because all we really need to do is to | |
466 | * find the correct waitqueue where tasks might be sleeping waiting for that | |
467 | * old 'entry' and wake them. | |
468 | */ | |
ac401cc7 | 469 | void dax_wake_mapping_entry_waiter(struct address_space *mapping, |
63e95b5c | 470 | pgoff_t index, void *entry, bool wake_all) |
ac401cc7 | 471 | { |
63e95b5c RZ |
472 | struct exceptional_entry_key key; |
473 | wait_queue_head_t *wq; | |
474 | ||
475 | wq = dax_entry_waitqueue(mapping, index, entry, &key); | |
ac401cc7 JK |
476 | |
477 | /* | |
478 | * Checking for locked entry and prepare_to_wait_exclusive() happens | |
479 | * under mapping->tree_lock, ditto for entry handling in our callers. | |
480 | * So at this point all tasks that could have seen our entry locked | |
481 | * must be in the waitqueue and the following check will see them. | |
482 | */ | |
63e95b5c | 483 | if (waitqueue_active(wq)) |
ac401cc7 | 484 | __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); |
ac401cc7 JK |
485 | } |
486 | ||
bc2466e4 | 487 | void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index) |
ac401cc7 | 488 | { |
e3ad61c6 | 489 | void *entry, **slot; |
ac401cc7 JK |
490 | |
491 | spin_lock_irq(&mapping->tree_lock); | |
e3ad61c6 RZ |
492 | entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot); |
493 | if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) || | |
ac401cc7 JK |
494 | !slot_locked(mapping, slot))) { |
495 | spin_unlock_irq(&mapping->tree_lock); | |
496 | return; | |
497 | } | |
498 | unlock_slot(mapping, slot); | |
499 | spin_unlock_irq(&mapping->tree_lock); | |
63e95b5c | 500 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); |
ac401cc7 JK |
501 | } |
502 | ||
503 | static void put_locked_mapping_entry(struct address_space *mapping, | |
504 | pgoff_t index, void *entry) | |
505 | { | |
506 | if (!radix_tree_exceptional_entry(entry)) { | |
507 | unlock_page(entry); | |
508 | put_page(entry); | |
509 | } else { | |
bc2466e4 | 510 | dax_unlock_mapping_entry(mapping, index); |
ac401cc7 JK |
511 | } |
512 | } | |
513 | ||
514 | /* | |
515 | * Called when we are done with radix tree entry we looked up via | |
516 | * get_unlocked_mapping_entry() and which we didn't lock in the end. | |
517 | */ | |
518 | static void put_unlocked_mapping_entry(struct address_space *mapping, | |
519 | pgoff_t index, void *entry) | |
520 | { | |
521 | if (!radix_tree_exceptional_entry(entry)) | |
522 | return; | |
523 | ||
524 | /* We have to wake up next waiter for the radix tree entry lock */ | |
63e95b5c | 525 | dax_wake_mapping_entry_waiter(mapping, index, entry, false); |
ac401cc7 JK |
526 | } |
527 | ||
528 | /* | |
529 | * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree | |
530 | * entry to get unlocked before deleting it. | |
531 | */ | |
532 | int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) | |
533 | { | |
534 | void *entry; | |
535 | ||
536 | spin_lock_irq(&mapping->tree_lock); | |
537 | entry = get_unlocked_mapping_entry(mapping, index, NULL); | |
538 | /* | |
539 | * This gets called from truncate / punch_hole path. As such, the caller | |
540 | * must hold locks protecting against concurrent modifications of the | |
541 | * radix tree (usually fs-private i_mmap_sem for writing). Since the | |
542 | * caller has seen exceptional entry for this index, we better find it | |
543 | * at that index as well... | |
544 | */ | |
545 | if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry))) { | |
546 | spin_unlock_irq(&mapping->tree_lock); | |
547 | return 0; | |
548 | } | |
549 | radix_tree_delete(&mapping->page_tree, index); | |
550 | mapping->nrexceptional--; | |
551 | spin_unlock_irq(&mapping->tree_lock); | |
63e95b5c | 552 | dax_wake_mapping_entry_waiter(mapping, index, entry, true); |
ac401cc7 JK |
553 | |
554 | return 1; | |
555 | } | |
556 | ||
f7ca90b1 MW |
557 | /* |
558 | * The user has performed a load from a hole in the file. Allocating | |
559 | * a new page in the file would cause excessive storage usage for | |
560 | * workloads with sparse files. We allocate a page cache page instead. | |
561 | * We'll kick it out of the page cache if it's ever written to, | |
562 | * otherwise it will simply fall out of the page cache under memory | |
563 | * pressure without ever having been dirtied. | |
564 | */ | |
ac401cc7 JK |
565 | static int dax_load_hole(struct address_space *mapping, void *entry, |
566 | struct vm_fault *vmf) | |
f7ca90b1 | 567 | { |
ac401cc7 | 568 | struct page *page; |
f7ca90b1 | 569 | |
ac401cc7 JK |
570 | /* Hole page already exists? Return it... */ |
571 | if (!radix_tree_exceptional_entry(entry)) { | |
572 | vmf->page = entry; | |
573 | return VM_FAULT_LOCKED; | |
574 | } | |
f7ca90b1 | 575 | |
ac401cc7 JK |
576 | /* This will replace locked radix tree entry with a hole page */ |
577 | page = find_or_create_page(mapping, vmf->pgoff, | |
578 | vmf->gfp_mask | __GFP_ZERO); | |
579 | if (!page) { | |
580 | put_locked_mapping_entry(mapping, vmf->pgoff, entry); | |
581 | return VM_FAULT_OOM; | |
582 | } | |
f7ca90b1 MW |
583 | vmf->page = page; |
584 | return VM_FAULT_LOCKED; | |
585 | } | |
586 | ||
b0d5e82f CH |
587 | static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size, |
588 | struct page *to, unsigned long vaddr) | |
f7ca90b1 | 589 | { |
b2e0d162 | 590 | struct blk_dax_ctl dax = { |
b0d5e82f CH |
591 | .sector = sector, |
592 | .size = size, | |
b2e0d162 | 593 | }; |
e2e05394 RZ |
594 | void *vto; |
595 | ||
b2e0d162 DW |
596 | if (dax_map_atomic(bdev, &dax) < 0) |
597 | return PTR_ERR(dax.addr); | |
f7ca90b1 | 598 | vto = kmap_atomic(to); |
b2e0d162 | 599 | copy_user_page(vto, (void __force *)dax.addr, vaddr, to); |
f7ca90b1 | 600 | kunmap_atomic(vto); |
b2e0d162 | 601 | dax_unmap_atomic(bdev, &dax); |
f7ca90b1 MW |
602 | return 0; |
603 | } | |
604 | ||
09cbfeaf | 605 | #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT)) |
9973c98e | 606 | |
ac401cc7 JK |
607 | static void *dax_insert_mapping_entry(struct address_space *mapping, |
608 | struct vm_fault *vmf, | |
609 | void *entry, sector_t sector) | |
9973c98e RZ |
610 | { |
611 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
ac401cc7 JK |
612 | int error = 0; |
613 | bool hole_fill = false; | |
614 | void *new_entry; | |
615 | pgoff_t index = vmf->pgoff; | |
9973c98e | 616 | |
ac401cc7 | 617 | if (vmf->flags & FAULT_FLAG_WRITE) |
d2b2a28e | 618 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); |
9973c98e | 619 | |
ac401cc7 JK |
620 | /* Replacing hole page with block mapping? */ |
621 | if (!radix_tree_exceptional_entry(entry)) { | |
622 | hole_fill = true; | |
623 | /* | |
624 | * Unmap the page now before we remove it from page cache below. | |
625 | * The page is locked so it cannot be faulted in again. | |
626 | */ | |
627 | unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT, | |
628 | PAGE_SIZE, 0); | |
629 | error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM); | |
630 | if (error) | |
631 | return ERR_PTR(error); | |
9973c98e RZ |
632 | } |
633 | ||
ac401cc7 JK |
634 | spin_lock_irq(&mapping->tree_lock); |
635 | new_entry = (void *)((unsigned long)RADIX_DAX_ENTRY(sector, false) | | |
636 | RADIX_DAX_ENTRY_LOCK); | |
637 | if (hole_fill) { | |
638 | __delete_from_page_cache(entry, NULL); | |
639 | /* Drop pagecache reference */ | |
640 | put_page(entry); | |
641 | error = radix_tree_insert(page_tree, index, new_entry); | |
642 | if (error) { | |
643 | new_entry = ERR_PTR(error); | |
9973c98e RZ |
644 | goto unlock; |
645 | } | |
ac401cc7 JK |
646 | mapping->nrexceptional++; |
647 | } else { | |
648 | void **slot; | |
649 | void *ret; | |
9973c98e | 650 | |
ac401cc7 JK |
651 | ret = __radix_tree_lookup(page_tree, index, NULL, &slot); |
652 | WARN_ON_ONCE(ret != entry); | |
653 | radix_tree_replace_slot(slot, new_entry); | |
9973c98e | 654 | } |
ac401cc7 | 655 | if (vmf->flags & FAULT_FLAG_WRITE) |
9973c98e RZ |
656 | radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY); |
657 | unlock: | |
658 | spin_unlock_irq(&mapping->tree_lock); | |
ac401cc7 JK |
659 | if (hole_fill) { |
660 | radix_tree_preload_end(); | |
661 | /* | |
662 | * We don't need hole page anymore, it has been replaced with | |
663 | * locked radix tree entry now. | |
664 | */ | |
665 | if (mapping->a_ops->freepage) | |
666 | mapping->a_ops->freepage(entry); | |
667 | unlock_page(entry); | |
668 | put_page(entry); | |
669 | } | |
670 | return new_entry; | |
9973c98e RZ |
671 | } |
672 | ||
673 | static int dax_writeback_one(struct block_device *bdev, | |
674 | struct address_space *mapping, pgoff_t index, void *entry) | |
675 | { | |
676 | struct radix_tree_root *page_tree = &mapping->page_tree; | |
677 | int type = RADIX_DAX_TYPE(entry); | |
678 | struct radix_tree_node *node; | |
679 | struct blk_dax_ctl dax; | |
680 | void **slot; | |
681 | int ret = 0; | |
682 | ||
683 | spin_lock_irq(&mapping->tree_lock); | |
684 | /* | |
685 | * Regular page slots are stabilized by the page lock even | |
686 | * without the tree itself locked. These unlocked entries | |
687 | * need verification under the tree lock. | |
688 | */ | |
689 | if (!__radix_tree_lookup(page_tree, index, &node, &slot)) | |
690 | goto unlock; | |
691 | if (*slot != entry) | |
692 | goto unlock; | |
693 | ||
694 | /* another fsync thread may have already written back this entry */ | |
695 | if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)) | |
696 | goto unlock; | |
697 | ||
698 | if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) { | |
699 | ret = -EIO; | |
700 | goto unlock; | |
701 | } | |
702 | ||
703 | dax.sector = RADIX_DAX_SECTOR(entry); | |
704 | dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE); | |
705 | spin_unlock_irq(&mapping->tree_lock); | |
706 | ||
707 | /* | |
708 | * We cannot hold tree_lock while calling dax_map_atomic() because it | |
709 | * eventually calls cond_resched(). | |
710 | */ | |
711 | ret = dax_map_atomic(bdev, &dax); | |
712 | if (ret < 0) | |
713 | return ret; | |
714 | ||
715 | if (WARN_ON_ONCE(ret < dax.size)) { | |
716 | ret = -EIO; | |
717 | goto unmap; | |
718 | } | |
719 | ||
720 | wb_cache_pmem(dax.addr, dax.size); | |
721 | ||
722 | spin_lock_irq(&mapping->tree_lock); | |
723 | radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE); | |
724 | spin_unlock_irq(&mapping->tree_lock); | |
725 | unmap: | |
726 | dax_unmap_atomic(bdev, &dax); | |
727 | return ret; | |
728 | ||
729 | unlock: | |
730 | spin_unlock_irq(&mapping->tree_lock); | |
731 | return ret; | |
732 | } | |
733 | ||
734 | /* | |
735 | * Flush the mapping to the persistent domain within the byte range of [start, | |
736 | * end]. This is required by data integrity operations to ensure file data is | |
737 | * on persistent storage prior to completion of the operation. | |
738 | */ | |
7f6d5b52 RZ |
739 | int dax_writeback_mapping_range(struct address_space *mapping, |
740 | struct block_device *bdev, struct writeback_control *wbc) | |
9973c98e RZ |
741 | { |
742 | struct inode *inode = mapping->host; | |
9973c98e RZ |
743 | pgoff_t start_index, end_index, pmd_index; |
744 | pgoff_t indices[PAGEVEC_SIZE]; | |
745 | struct pagevec pvec; | |
746 | bool done = false; | |
747 | int i, ret = 0; | |
748 | void *entry; | |
749 | ||
750 | if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) | |
751 | return -EIO; | |
752 | ||
7f6d5b52 RZ |
753 | if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) |
754 | return 0; | |
755 | ||
09cbfeaf KS |
756 | start_index = wbc->range_start >> PAGE_SHIFT; |
757 | end_index = wbc->range_end >> PAGE_SHIFT; | |
9973c98e RZ |
758 | pmd_index = DAX_PMD_INDEX(start_index); |
759 | ||
760 | rcu_read_lock(); | |
761 | entry = radix_tree_lookup(&mapping->page_tree, pmd_index); | |
762 | rcu_read_unlock(); | |
763 | ||
764 | /* see if the start of our range is covered by a PMD entry */ | |
765 | if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) | |
766 | start_index = pmd_index; | |
767 | ||
768 | tag_pages_for_writeback(mapping, start_index, end_index); | |
769 | ||
770 | pagevec_init(&pvec, 0); | |
771 | while (!done) { | |
772 | pvec.nr = find_get_entries_tag(mapping, start_index, | |
773 | PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE, | |
774 | pvec.pages, indices); | |
775 | ||
776 | if (pvec.nr == 0) | |
777 | break; | |
778 | ||
779 | for (i = 0; i < pvec.nr; i++) { | |
780 | if (indices[i] > end_index) { | |
781 | done = true; | |
782 | break; | |
783 | } | |
784 | ||
785 | ret = dax_writeback_one(bdev, mapping, indices[i], | |
786 | pvec.pages[i]); | |
787 | if (ret < 0) | |
788 | return ret; | |
789 | } | |
790 | } | |
9973c98e RZ |
791 | return 0; |
792 | } | |
793 | EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); | |
794 | ||
ac401cc7 | 795 | static int dax_insert_mapping(struct address_space *mapping, |
1aaba095 CH |
796 | struct block_device *bdev, sector_t sector, size_t size, |
797 | void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf) | |
f7ca90b1 | 798 | { |
f7ca90b1 | 799 | unsigned long vaddr = (unsigned long)vmf->virtual_address; |
b2e0d162 | 800 | struct blk_dax_ctl dax = { |
1aaba095 CH |
801 | .sector = sector, |
802 | .size = size, | |
b2e0d162 | 803 | }; |
ac401cc7 JK |
804 | void *ret; |
805 | void *entry = *entryp; | |
f7ca90b1 | 806 | |
4d9a2c87 JK |
807 | if (dax_map_atomic(bdev, &dax) < 0) |
808 | return PTR_ERR(dax.addr); | |
b2e0d162 | 809 | dax_unmap_atomic(bdev, &dax); |
f7ca90b1 | 810 | |
ac401cc7 | 811 | ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector); |
4d9a2c87 JK |
812 | if (IS_ERR(ret)) |
813 | return PTR_ERR(ret); | |
ac401cc7 | 814 | *entryp = ret; |
9973c98e | 815 | |
4d9a2c87 | 816 | return vm_insert_mixed(vma, vaddr, dax.pfn); |
f7ca90b1 MW |
817 | } |
818 | ||
ce5c5d55 | 819 | /** |
6b524995 | 820 | * dax_fault - handle a page fault on a DAX file |
ce5c5d55 DC |
821 | * @vma: The virtual memory area where the fault occurred |
822 | * @vmf: The description of the fault | |
823 | * @get_block: The filesystem method used to translate file offsets to blocks | |
824 | * | |
825 | * When a page fault occurs, filesystems may call this helper in their | |
6b524995 | 826 | * fault handler for DAX files. dax_fault() assumes the caller has done all |
ce5c5d55 DC |
827 | * the necessary locking for the page fault to proceed successfully. |
828 | */ | |
6b524995 | 829 | int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf, |
02fbd139 | 830 | get_block_t get_block) |
f7ca90b1 MW |
831 | { |
832 | struct file *file = vma->vm_file; | |
833 | struct address_space *mapping = file->f_mapping; | |
834 | struct inode *inode = mapping->host; | |
ac401cc7 | 835 | void *entry; |
f7ca90b1 MW |
836 | struct buffer_head bh; |
837 | unsigned long vaddr = (unsigned long)vmf->virtual_address; | |
838 | unsigned blkbits = inode->i_blkbits; | |
839 | sector_t block; | |
840 | pgoff_t size; | |
841 | int error; | |
842 | int major = 0; | |
843 | ||
ac401cc7 JK |
844 | /* |
845 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
846 | * to hold locks serializing us with truncate / punch hole so this is | |
847 | * a reliable test. | |
848 | */ | |
f7ca90b1 MW |
849 | size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT; |
850 | if (vmf->pgoff >= size) | |
851 | return VM_FAULT_SIGBUS; | |
852 | ||
853 | memset(&bh, 0, sizeof(bh)); | |
854 | block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits); | |
eab95db6 | 855 | bh.b_bdev = inode->i_sb->s_bdev; |
f7ca90b1 MW |
856 | bh.b_size = PAGE_SIZE; |
857 | ||
ac401cc7 JK |
858 | entry = grab_mapping_entry(mapping, vmf->pgoff); |
859 | if (IS_ERR(entry)) { | |
860 | error = PTR_ERR(entry); | |
861 | goto out; | |
f7ca90b1 MW |
862 | } |
863 | ||
864 | error = get_block(inode, block, &bh, 0); | |
865 | if (!error && (bh.b_size < PAGE_SIZE)) | |
866 | error = -EIO; /* fs corruption? */ | |
867 | if (error) | |
ac401cc7 | 868 | goto unlock_entry; |
f7ca90b1 MW |
869 | |
870 | if (vmf->cow_page) { | |
871 | struct page *new_page = vmf->cow_page; | |
872 | if (buffer_written(&bh)) | |
b0d5e82f CH |
873 | error = copy_user_dax(bh.b_bdev, to_sector(&bh, inode), |
874 | bh.b_size, new_page, vaddr); | |
f7ca90b1 MW |
875 | else |
876 | clear_user_highpage(new_page, vaddr); | |
877 | if (error) | |
ac401cc7 JK |
878 | goto unlock_entry; |
879 | if (!radix_tree_exceptional_entry(entry)) { | |
880 | vmf->page = entry; | |
bc2466e4 | 881 | return VM_FAULT_LOCKED; |
ac401cc7 | 882 | } |
bc2466e4 JK |
883 | vmf->entry = entry; |
884 | return VM_FAULT_DAX_LOCKED; | |
f7ca90b1 | 885 | } |
f7ca90b1 | 886 | |
ac401cc7 | 887 | if (!buffer_mapped(&bh)) { |
f7ca90b1 MW |
888 | if (vmf->flags & FAULT_FLAG_WRITE) { |
889 | error = get_block(inode, block, &bh, 1); | |
890 | count_vm_event(PGMAJFAULT); | |
891 | mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT); | |
892 | major = VM_FAULT_MAJOR; | |
893 | if (!error && (bh.b_size < PAGE_SIZE)) | |
894 | error = -EIO; | |
895 | if (error) | |
ac401cc7 | 896 | goto unlock_entry; |
f7ca90b1 | 897 | } else { |
ac401cc7 | 898 | return dax_load_hole(mapping, entry, vmf); |
f7ca90b1 MW |
899 | } |
900 | } | |
901 | ||
02fbd139 | 902 | /* Filesystem should not return unwritten buffers to us! */ |
2b10945c | 903 | WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh)); |
1aaba095 CH |
904 | error = dax_insert_mapping(mapping, bh.b_bdev, to_sector(&bh, inode), |
905 | bh.b_size, &entry, vma, vmf); | |
ac401cc7 JK |
906 | unlock_entry: |
907 | put_locked_mapping_entry(mapping, vmf->pgoff, entry); | |
f7ca90b1 MW |
908 | out: |
909 | if (error == -ENOMEM) | |
910 | return VM_FAULT_OOM | major; | |
911 | /* -EBUSY is fine, somebody else faulted on the same PTE */ | |
912 | if ((error < 0) && (error != -EBUSY)) | |
913 | return VM_FAULT_SIGBUS | major; | |
914 | return VM_FAULT_NOPAGE | major; | |
f7ca90b1 | 915 | } |
f7ca90b1 | 916 | EXPORT_SYMBOL_GPL(dax_fault); |
4c0ccfef | 917 | |
0e3b210c BH |
918 | /** |
919 | * dax_pfn_mkwrite - handle first write to DAX page | |
920 | * @vma: The virtual memory area where the fault occurred | |
921 | * @vmf: The description of the fault | |
0e3b210c BH |
922 | */ |
923 | int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) | |
924 | { | |
9973c98e | 925 | struct file *file = vma->vm_file; |
ac401cc7 JK |
926 | struct address_space *mapping = file->f_mapping; |
927 | void *entry; | |
928 | pgoff_t index = vmf->pgoff; | |
30f471fd | 929 | |
ac401cc7 JK |
930 | spin_lock_irq(&mapping->tree_lock); |
931 | entry = get_unlocked_mapping_entry(mapping, index, NULL); | |
932 | if (!entry || !radix_tree_exceptional_entry(entry)) | |
933 | goto out; | |
934 | radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY); | |
935 | put_unlocked_mapping_entry(mapping, index, entry); | |
936 | out: | |
937 | spin_unlock_irq(&mapping->tree_lock); | |
0e3b210c BH |
938 | return VM_FAULT_NOPAGE; |
939 | } | |
940 | EXPORT_SYMBOL_GPL(dax_pfn_mkwrite); | |
941 | ||
4b0228fa VV |
942 | static bool dax_range_is_aligned(struct block_device *bdev, |
943 | unsigned int offset, unsigned int length) | |
944 | { | |
945 | unsigned short sector_size = bdev_logical_block_size(bdev); | |
946 | ||
947 | if (!IS_ALIGNED(offset, sector_size)) | |
948 | return false; | |
949 | if (!IS_ALIGNED(length, sector_size)) | |
950 | return false; | |
951 | ||
952 | return true; | |
953 | } | |
954 | ||
679c8bd3 CH |
955 | int __dax_zero_page_range(struct block_device *bdev, sector_t sector, |
956 | unsigned int offset, unsigned int length) | |
957 | { | |
958 | struct blk_dax_ctl dax = { | |
959 | .sector = sector, | |
960 | .size = PAGE_SIZE, | |
961 | }; | |
962 | ||
4b0228fa VV |
963 | if (dax_range_is_aligned(bdev, offset, length)) { |
964 | sector_t start_sector = dax.sector + (offset >> 9); | |
965 | ||
966 | return blkdev_issue_zeroout(bdev, start_sector, | |
967 | length >> 9, GFP_NOFS, true); | |
968 | } else { | |
969 | if (dax_map_atomic(bdev, &dax) < 0) | |
970 | return PTR_ERR(dax.addr); | |
971 | clear_pmem(dax.addr + offset, length); | |
4b0228fa VV |
972 | dax_unmap_atomic(bdev, &dax); |
973 | } | |
679c8bd3 CH |
974 | return 0; |
975 | } | |
976 | EXPORT_SYMBOL_GPL(__dax_zero_page_range); | |
977 | ||
4c0ccfef | 978 | /** |
25726bc1 | 979 | * dax_zero_page_range - zero a range within a page of a DAX file |
4c0ccfef MW |
980 | * @inode: The file being truncated |
981 | * @from: The file offset that is being truncated to | |
25726bc1 | 982 | * @length: The number of bytes to zero |
4c0ccfef MW |
983 | * @get_block: The filesystem method used to translate file offsets to blocks |
984 | * | |
25726bc1 MW |
985 | * This function can be called by a filesystem when it is zeroing part of a |
986 | * page in a DAX file. This is intended for hole-punch operations. If | |
987 | * you are truncating a file, the helper function dax_truncate_page() may be | |
988 | * more convenient. | |
4c0ccfef | 989 | */ |
25726bc1 MW |
990 | int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length, |
991 | get_block_t get_block) | |
4c0ccfef MW |
992 | { |
993 | struct buffer_head bh; | |
09cbfeaf KS |
994 | pgoff_t index = from >> PAGE_SHIFT; |
995 | unsigned offset = from & (PAGE_SIZE-1); | |
4c0ccfef MW |
996 | int err; |
997 | ||
998 | /* Block boundary? Nothing to do */ | |
999 | if (!length) | |
1000 | return 0; | |
aada54f9 RZ |
1001 | if (WARN_ON_ONCE((offset + length) > PAGE_SIZE)) |
1002 | return -EINVAL; | |
4c0ccfef MW |
1003 | |
1004 | memset(&bh, 0, sizeof(bh)); | |
eab95db6 | 1005 | bh.b_bdev = inode->i_sb->s_bdev; |
09cbfeaf | 1006 | bh.b_size = PAGE_SIZE; |
4c0ccfef | 1007 | err = get_block(inode, index, &bh, 0); |
679c8bd3 | 1008 | if (err < 0 || !buffer_written(&bh)) |
4c0ccfef | 1009 | return err; |
4c0ccfef | 1010 | |
679c8bd3 CH |
1011 | return __dax_zero_page_range(bh.b_bdev, to_sector(&bh, inode), |
1012 | offset, length); | |
4c0ccfef | 1013 | } |
25726bc1 MW |
1014 | EXPORT_SYMBOL_GPL(dax_zero_page_range); |
1015 | ||
1016 | /** | |
1017 | * dax_truncate_page - handle a partial page being truncated in a DAX file | |
1018 | * @inode: The file being truncated | |
1019 | * @from: The file offset that is being truncated to | |
1020 | * @get_block: The filesystem method used to translate file offsets to blocks | |
1021 | * | |
1022 | * Similar to block_truncate_page(), this function can be called by a | |
1023 | * filesystem when it is truncating a DAX file to handle the partial page. | |
25726bc1 MW |
1024 | */ |
1025 | int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block) | |
1026 | { | |
09cbfeaf | 1027 | unsigned length = PAGE_ALIGN(from) - from; |
25726bc1 MW |
1028 | return dax_zero_page_range(inode, from, length, get_block); |
1029 | } | |
4c0ccfef | 1030 | EXPORT_SYMBOL_GPL(dax_truncate_page); |
a254e568 CH |
1031 | |
1032 | #ifdef CONFIG_FS_IOMAP | |
1033 | static loff_t | |
1034 | iomap_dax_actor(struct inode *inode, loff_t pos, loff_t length, void *data, | |
1035 | struct iomap *iomap) | |
1036 | { | |
1037 | struct iov_iter *iter = data; | |
1038 | loff_t end = pos + length, done = 0; | |
1039 | ssize_t ret = 0; | |
1040 | ||
1041 | if (iov_iter_rw(iter) == READ) { | |
1042 | end = min(end, i_size_read(inode)); | |
1043 | if (pos >= end) | |
1044 | return 0; | |
1045 | ||
1046 | if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) | |
1047 | return iov_iter_zero(min(length, end - pos), iter); | |
1048 | } | |
1049 | ||
1050 | if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) | |
1051 | return -EIO; | |
1052 | ||
1053 | while (pos < end) { | |
1054 | unsigned offset = pos & (PAGE_SIZE - 1); | |
1055 | struct blk_dax_ctl dax = { 0 }; | |
1056 | ssize_t map_len; | |
1057 | ||
1058 | dax.sector = iomap->blkno + | |
1059 | (((pos & PAGE_MASK) - iomap->offset) >> 9); | |
1060 | dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK; | |
1061 | map_len = dax_map_atomic(iomap->bdev, &dax); | |
1062 | if (map_len < 0) { | |
1063 | ret = map_len; | |
1064 | break; | |
1065 | } | |
1066 | ||
1067 | dax.addr += offset; | |
1068 | map_len -= offset; | |
1069 | if (map_len > end - pos) | |
1070 | map_len = end - pos; | |
1071 | ||
1072 | if (iov_iter_rw(iter) == WRITE) | |
1073 | map_len = copy_from_iter_pmem(dax.addr, map_len, iter); | |
1074 | else | |
1075 | map_len = copy_to_iter(dax.addr, map_len, iter); | |
1076 | dax_unmap_atomic(iomap->bdev, &dax); | |
1077 | if (map_len <= 0) { | |
1078 | ret = map_len ? map_len : -EFAULT; | |
1079 | break; | |
1080 | } | |
1081 | ||
1082 | pos += map_len; | |
1083 | length -= map_len; | |
1084 | done += map_len; | |
1085 | } | |
1086 | ||
1087 | return done ? done : ret; | |
1088 | } | |
1089 | ||
1090 | /** | |
1091 | * iomap_dax_rw - Perform I/O to a DAX file | |
1092 | * @iocb: The control block for this I/O | |
1093 | * @iter: The addresses to do I/O from or to | |
1094 | * @ops: iomap ops passed from the file system | |
1095 | * | |
1096 | * This function performs read and write operations to directly mapped | |
1097 | * persistent memory. The callers needs to take care of read/write exclusion | |
1098 | * and evicting any page cache pages in the region under I/O. | |
1099 | */ | |
1100 | ssize_t | |
1101 | iomap_dax_rw(struct kiocb *iocb, struct iov_iter *iter, | |
1102 | struct iomap_ops *ops) | |
1103 | { | |
1104 | struct address_space *mapping = iocb->ki_filp->f_mapping; | |
1105 | struct inode *inode = mapping->host; | |
1106 | loff_t pos = iocb->ki_pos, ret = 0, done = 0; | |
1107 | unsigned flags = 0; | |
1108 | ||
1109 | if (iov_iter_rw(iter) == WRITE) | |
1110 | flags |= IOMAP_WRITE; | |
1111 | ||
1112 | /* | |
1113 | * Yes, even DAX files can have page cache attached to them: A zeroed | |
1114 | * page is inserted into the pagecache when we have to serve a write | |
1115 | * fault on a hole. It should never be dirtied and can simply be | |
1116 | * dropped from the pagecache once we get real data for the page. | |
1117 | * | |
1118 | * XXX: This is racy against mmap, and there's nothing we can do about | |
1119 | * it. We'll eventually need to shift this down even further so that | |
1120 | * we can check if we allocated blocks over a hole first. | |
1121 | */ | |
1122 | if (mapping->nrpages) { | |
1123 | ret = invalidate_inode_pages2_range(mapping, | |
1124 | pos >> PAGE_SHIFT, | |
1125 | (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT); | |
1126 | WARN_ON_ONCE(ret); | |
1127 | } | |
1128 | ||
1129 | while (iov_iter_count(iter)) { | |
1130 | ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, | |
1131 | iter, iomap_dax_actor); | |
1132 | if (ret <= 0) | |
1133 | break; | |
1134 | pos += ret; | |
1135 | done += ret; | |
1136 | } | |
1137 | ||
1138 | iocb->ki_pos += done; | |
1139 | return done ? done : ret; | |
1140 | } | |
1141 | EXPORT_SYMBOL_GPL(iomap_dax_rw); | |
a7d73fe6 CH |
1142 | |
1143 | /** | |
1144 | * iomap_dax_fault - handle a page fault on a DAX file | |
1145 | * @vma: The virtual memory area where the fault occurred | |
1146 | * @vmf: The description of the fault | |
1147 | * @ops: iomap ops passed from the file system | |
1148 | * | |
1149 | * When a page fault occurs, filesystems may call this helper in their fault | |
1150 | * or mkwrite handler for DAX files. Assumes the caller has done all the | |
1151 | * necessary locking for the page fault to proceed successfully. | |
1152 | */ | |
1153 | int iomap_dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf, | |
1154 | struct iomap_ops *ops) | |
1155 | { | |
1156 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1157 | struct inode *inode = mapping->host; | |
1158 | unsigned long vaddr = (unsigned long)vmf->virtual_address; | |
1159 | loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; | |
1160 | sector_t sector; | |
1161 | struct iomap iomap = { 0 }; | |
1162 | unsigned flags = 0; | |
1163 | int error, major = 0; | |
1164 | void *entry; | |
1165 | ||
1166 | /* | |
1167 | * Check whether offset isn't beyond end of file now. Caller is supposed | |
1168 | * to hold locks serializing us with truncate / punch hole so this is | |
1169 | * a reliable test. | |
1170 | */ | |
1171 | if (pos >= i_size_read(inode)) | |
1172 | return VM_FAULT_SIGBUS; | |
1173 | ||
1174 | entry = grab_mapping_entry(mapping, vmf->pgoff); | |
1175 | if (IS_ERR(entry)) { | |
1176 | error = PTR_ERR(entry); | |
1177 | goto out; | |
1178 | } | |
1179 | ||
1180 | if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) | |
1181 | flags |= IOMAP_WRITE; | |
1182 | ||
1183 | /* | |
1184 | * Note that we don't bother to use iomap_apply here: DAX required | |
1185 | * the file system block size to be equal the page size, which means | |
1186 | * that we never have to deal with more than a single extent here. | |
1187 | */ | |
1188 | error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); | |
1189 | if (error) | |
1190 | goto unlock_entry; | |
1191 | if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { | |
1192 | error = -EIO; /* fs corruption? */ | |
1193 | goto unlock_entry; | |
1194 | } | |
1195 | ||
1196 | sector = iomap.blkno + (((pos & PAGE_MASK) - iomap.offset) >> 9); | |
1197 | ||
1198 | if (vmf->cow_page) { | |
1199 | switch (iomap.type) { | |
1200 | case IOMAP_HOLE: | |
1201 | case IOMAP_UNWRITTEN: | |
1202 | clear_user_highpage(vmf->cow_page, vaddr); | |
1203 | break; | |
1204 | case IOMAP_MAPPED: | |
1205 | error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE, | |
1206 | vmf->cow_page, vaddr); | |
1207 | break; | |
1208 | default: | |
1209 | WARN_ON_ONCE(1); | |
1210 | error = -EIO; | |
1211 | break; | |
1212 | } | |
1213 | ||
1214 | if (error) | |
1215 | goto unlock_entry; | |
1216 | if (!radix_tree_exceptional_entry(entry)) { | |
1217 | vmf->page = entry; | |
1218 | return VM_FAULT_LOCKED; | |
1219 | } | |
1220 | vmf->entry = entry; | |
1221 | return VM_FAULT_DAX_LOCKED; | |
1222 | } | |
1223 | ||
1224 | switch (iomap.type) { | |
1225 | case IOMAP_MAPPED: | |
1226 | if (iomap.flags & IOMAP_F_NEW) { | |
1227 | count_vm_event(PGMAJFAULT); | |
1228 | mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT); | |
1229 | major = VM_FAULT_MAJOR; | |
1230 | } | |
1231 | error = dax_insert_mapping(mapping, iomap.bdev, sector, | |
1232 | PAGE_SIZE, &entry, vma, vmf); | |
1233 | break; | |
1234 | case IOMAP_UNWRITTEN: | |
1235 | case IOMAP_HOLE: | |
1236 | if (!(vmf->flags & FAULT_FLAG_WRITE)) | |
1237 | return dax_load_hole(mapping, entry, vmf); | |
1238 | /*FALLTHRU*/ | |
1239 | default: | |
1240 | WARN_ON_ONCE(1); | |
1241 | error = -EIO; | |
1242 | break; | |
1243 | } | |
1244 | ||
1245 | unlock_entry: | |
1246 | put_locked_mapping_entry(mapping, vmf->pgoff, entry); | |
1247 | out: | |
1248 | if (error == -ENOMEM) | |
1249 | return VM_FAULT_OOM | major; | |
1250 | /* -EBUSY is fine, somebody else faulted on the same PTE */ | |
1251 | if (error < 0 && error != -EBUSY) | |
1252 | return VM_FAULT_SIGBUS | major; | |
1253 | return VM_FAULT_NOPAGE | major; | |
1254 | } | |
1255 | EXPORT_SYMBOL_GPL(iomap_dax_fault); | |
a254e568 | 1256 | #endif /* CONFIG_FS_IOMAP */ |