]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Copyright (c) 2000-2006 Silicon Graphics, Inc. | |
3 | * All Rights Reserved. | |
4 | * | |
5 | * This program is free software; you can redistribute it and/or | |
6 | * modify it under the terms of the GNU General Public License as | |
7 | * published by the Free Software Foundation. | |
8 | * | |
9 | * This program is distributed in the hope that it would be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public License | |
15 | * along with this program; if not, write the Free Software Foundation, | |
16 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA | |
17 | */ | |
18 | #include "xfs.h" | |
19 | #include <linux/stddef.h> | |
20 | #include <linux/errno.h> | |
21 | #include <linux/gfp.h> | |
22 | #include <linux/pagemap.h> | |
23 | #include <linux/init.h> | |
24 | #include <linux/vmalloc.h> | |
25 | #include <linux/bio.h> | |
26 | #include <linux/sysctl.h> | |
27 | #include <linux/proc_fs.h> | |
28 | #include <linux/workqueue.h> | |
29 | #include <linux/percpu.h> | |
30 | #include <linux/blkdev.h> | |
31 | #include <linux/hash.h> | |
32 | #include <linux/kthread.h> | |
33 | #include <linux/migrate.h> | |
34 | #include <linux/backing-dev.h> | |
35 | #include <linux/freezer.h> | |
36 | #include <linux/sched/mm.h> | |
37 | ||
38 | #include "xfs_format.h" | |
39 | #include "xfs_log_format.h" | |
40 | #include "xfs_trans_resv.h" | |
41 | #include "xfs_sb.h" | |
42 | #include "xfs_mount.h" | |
43 | #include "xfs_trace.h" | |
44 | #include "xfs_log.h" | |
45 | #include "xfs_errortag.h" | |
46 | #include "xfs_error.h" | |
47 | ||
48 | static kmem_zone_t *xfs_buf_zone; | |
49 | ||
50 | #ifdef XFS_BUF_LOCK_TRACKING | |
51 | # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid) | |
52 | # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1) | |
53 | # define XB_GET_OWNER(bp) ((bp)->b_last_holder) | |
54 | #else | |
55 | # define XB_SET_OWNER(bp) do { } while (0) | |
56 | # define XB_CLEAR_OWNER(bp) do { } while (0) | |
57 | # define XB_GET_OWNER(bp) do { } while (0) | |
58 | #endif | |
59 | ||
60 | #define xb_to_gfp(flags) \ | |
61 | ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN) | |
62 | ||
63 | ||
64 | static inline int | |
65 | xfs_buf_is_vmapped( | |
66 | struct xfs_buf *bp) | |
67 | { | |
68 | /* | |
69 | * Return true if the buffer is vmapped. | |
70 | * | |
71 | * b_addr is null if the buffer is not mapped, but the code is clever | |
72 | * enough to know it doesn't have to map a single page, so the check has | |
73 | * to be both for b_addr and bp->b_page_count > 1. | |
74 | */ | |
75 | return bp->b_addr && bp->b_page_count > 1; | |
76 | } | |
77 | ||
78 | static inline int | |
79 | xfs_buf_vmap_len( | |
80 | struct xfs_buf *bp) | |
81 | { | |
82 | return (bp->b_page_count * PAGE_SIZE) - bp->b_offset; | |
83 | } | |
84 | ||
85 | /* | |
86 | * Bump the I/O in flight count on the buftarg if we haven't yet done so for | |
87 | * this buffer. The count is incremented once per buffer (per hold cycle) | |
88 | * because the corresponding decrement is deferred to buffer release. Buffers | |
89 | * can undergo I/O multiple times in a hold-release cycle and per buffer I/O | |
90 | * tracking adds unnecessary overhead. This is used for sychronization purposes | |
91 | * with unmount (see xfs_wait_buftarg()), so all we really need is a count of | |
92 | * in-flight buffers. | |
93 | * | |
94 | * Buffers that are never released (e.g., superblock, iclog buffers) must set | |
95 | * the XBF_NO_IOACCT flag before I/O submission. Otherwise, the buftarg count | |
96 | * never reaches zero and unmount hangs indefinitely. | |
97 | */ | |
98 | static inline void | |
99 | xfs_buf_ioacct_inc( | |
100 | struct xfs_buf *bp) | |
101 | { | |
102 | if (bp->b_flags & XBF_NO_IOACCT) | |
103 | return; | |
104 | ||
105 | ASSERT(bp->b_flags & XBF_ASYNC); | |
106 | spin_lock(&bp->b_lock); | |
107 | if (!(bp->b_state & XFS_BSTATE_IN_FLIGHT)) { | |
108 | bp->b_state |= XFS_BSTATE_IN_FLIGHT; | |
109 | percpu_counter_inc(&bp->b_target->bt_io_count); | |
110 | } | |
111 | spin_unlock(&bp->b_lock); | |
112 | } | |
113 | ||
114 | /* | |
115 | * Clear the in-flight state on a buffer about to be released to the LRU or | |
116 | * freed and unaccount from the buftarg. | |
117 | */ | |
118 | static inline void | |
119 | __xfs_buf_ioacct_dec( | |
120 | struct xfs_buf *bp) | |
121 | { | |
122 | lockdep_assert_held(&bp->b_lock); | |
123 | ||
124 | if (bp->b_state & XFS_BSTATE_IN_FLIGHT) { | |
125 | bp->b_state &= ~XFS_BSTATE_IN_FLIGHT; | |
126 | percpu_counter_dec(&bp->b_target->bt_io_count); | |
127 | } | |
128 | } | |
129 | ||
130 | static inline void | |
131 | xfs_buf_ioacct_dec( | |
132 | struct xfs_buf *bp) | |
133 | { | |
134 | spin_lock(&bp->b_lock); | |
135 | __xfs_buf_ioacct_dec(bp); | |
136 | spin_unlock(&bp->b_lock); | |
137 | } | |
138 | ||
139 | /* | |
140 | * When we mark a buffer stale, we remove the buffer from the LRU and clear the | |
141 | * b_lru_ref count so that the buffer is freed immediately when the buffer | |
142 | * reference count falls to zero. If the buffer is already on the LRU, we need | |
143 | * to remove the reference that LRU holds on the buffer. | |
144 | * | |
145 | * This prevents build-up of stale buffers on the LRU. | |
146 | */ | |
147 | void | |
148 | xfs_buf_stale( | |
149 | struct xfs_buf *bp) | |
150 | { | |
151 | ASSERT(xfs_buf_islocked(bp)); | |
152 | ||
153 | bp->b_flags |= XBF_STALE; | |
154 | ||
155 | /* | |
156 | * Clear the delwri status so that a delwri queue walker will not | |
157 | * flush this buffer to disk now that it is stale. The delwri queue has | |
158 | * a reference to the buffer, so this is safe to do. | |
159 | */ | |
160 | bp->b_flags &= ~_XBF_DELWRI_Q; | |
161 | ||
162 | /* | |
163 | * Once the buffer is marked stale and unlocked, a subsequent lookup | |
164 | * could reset b_flags. There is no guarantee that the buffer is | |
165 | * unaccounted (released to LRU) before that occurs. Drop in-flight | |
166 | * status now to preserve accounting consistency. | |
167 | */ | |
168 | spin_lock(&bp->b_lock); | |
169 | __xfs_buf_ioacct_dec(bp); | |
170 | ||
171 | atomic_set(&bp->b_lru_ref, 0); | |
172 | if (!(bp->b_state & XFS_BSTATE_DISPOSE) && | |
173 | (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru))) | |
174 | atomic_dec(&bp->b_hold); | |
175 | ||
176 | ASSERT(atomic_read(&bp->b_hold) >= 1); | |
177 | spin_unlock(&bp->b_lock); | |
178 | } | |
179 | ||
180 | static int | |
181 | xfs_buf_get_maps( | |
182 | struct xfs_buf *bp, | |
183 | int map_count) | |
184 | { | |
185 | ASSERT(bp->b_maps == NULL); | |
186 | bp->b_map_count = map_count; | |
187 | ||
188 | if (map_count == 1) { | |
189 | bp->b_maps = &bp->__b_map; | |
190 | return 0; | |
191 | } | |
192 | ||
193 | bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map), | |
194 | KM_NOFS); | |
195 | if (!bp->b_maps) | |
196 | return -ENOMEM; | |
197 | return 0; | |
198 | } | |
199 | ||
200 | /* | |
201 | * Frees b_pages if it was allocated. | |
202 | */ | |
203 | static void | |
204 | xfs_buf_free_maps( | |
205 | struct xfs_buf *bp) | |
206 | { | |
207 | if (bp->b_maps != &bp->__b_map) { | |
208 | kmem_free(bp->b_maps); | |
209 | bp->b_maps = NULL; | |
210 | } | |
211 | } | |
212 | ||
213 | struct xfs_buf * | |
214 | _xfs_buf_alloc( | |
215 | struct xfs_buftarg *target, | |
216 | struct xfs_buf_map *map, | |
217 | int nmaps, | |
218 | xfs_buf_flags_t flags) | |
219 | { | |
220 | struct xfs_buf *bp; | |
221 | int error; | |
222 | int i; | |
223 | ||
224 | bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS); | |
225 | if (unlikely(!bp)) | |
226 | return NULL; | |
227 | ||
228 | /* | |
229 | * We don't want certain flags to appear in b_flags unless they are | |
230 | * specifically set by later operations on the buffer. | |
231 | */ | |
232 | flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD); | |
233 | ||
234 | atomic_set(&bp->b_hold, 1); | |
235 | atomic_set(&bp->b_lru_ref, 1); | |
236 | init_completion(&bp->b_iowait); | |
237 | INIT_LIST_HEAD(&bp->b_lru); | |
238 | INIT_LIST_HEAD(&bp->b_list); | |
239 | sema_init(&bp->b_sema, 0); /* held, no waiters */ | |
240 | spin_lock_init(&bp->b_lock); | |
241 | XB_SET_OWNER(bp); | |
242 | bp->b_target = target; | |
243 | bp->b_flags = flags; | |
244 | ||
245 | /* | |
246 | * Set length and io_length to the same value initially. | |
247 | * I/O routines should use io_length, which will be the same in | |
248 | * most cases but may be reset (e.g. XFS recovery). | |
249 | */ | |
250 | error = xfs_buf_get_maps(bp, nmaps); | |
251 | if (error) { | |
252 | kmem_zone_free(xfs_buf_zone, bp); | |
253 | return NULL; | |
254 | } | |
255 | ||
256 | bp->b_bn = map[0].bm_bn; | |
257 | bp->b_length = 0; | |
258 | for (i = 0; i < nmaps; i++) { | |
259 | bp->b_maps[i].bm_bn = map[i].bm_bn; | |
260 | bp->b_maps[i].bm_len = map[i].bm_len; | |
261 | bp->b_length += map[i].bm_len; | |
262 | } | |
263 | bp->b_io_length = bp->b_length; | |
264 | ||
265 | atomic_set(&bp->b_pin_count, 0); | |
266 | init_waitqueue_head(&bp->b_waiters); | |
267 | ||
268 | XFS_STATS_INC(target->bt_mount, xb_create); | |
269 | trace_xfs_buf_init(bp, _RET_IP_); | |
270 | ||
271 | return bp; | |
272 | } | |
273 | ||
274 | /* | |
275 | * Allocate a page array capable of holding a specified number | |
276 | * of pages, and point the page buf at it. | |
277 | */ | |
278 | STATIC int | |
279 | _xfs_buf_get_pages( | |
280 | xfs_buf_t *bp, | |
281 | int page_count) | |
282 | { | |
283 | /* Make sure that we have a page list */ | |
284 | if (bp->b_pages == NULL) { | |
285 | bp->b_page_count = page_count; | |
286 | if (page_count <= XB_PAGES) { | |
287 | bp->b_pages = bp->b_page_array; | |
288 | } else { | |
289 | bp->b_pages = kmem_alloc(sizeof(struct page *) * | |
290 | page_count, KM_NOFS); | |
291 | if (bp->b_pages == NULL) | |
292 | return -ENOMEM; | |
293 | } | |
294 | memset(bp->b_pages, 0, sizeof(struct page *) * page_count); | |
295 | } | |
296 | return 0; | |
297 | } | |
298 | ||
299 | /* | |
300 | * Frees b_pages if it was allocated. | |
301 | */ | |
302 | STATIC void | |
303 | _xfs_buf_free_pages( | |
304 | xfs_buf_t *bp) | |
305 | { | |
306 | if (bp->b_pages != bp->b_page_array) { | |
307 | kmem_free(bp->b_pages); | |
308 | bp->b_pages = NULL; | |
309 | } | |
310 | } | |
311 | ||
312 | /* | |
313 | * Releases the specified buffer. | |
314 | * | |
315 | * The modification state of any associated pages is left unchanged. | |
316 | * The buffer must not be on any hash - use xfs_buf_rele instead for | |
317 | * hashed and refcounted buffers | |
318 | */ | |
319 | void | |
320 | xfs_buf_free( | |
321 | xfs_buf_t *bp) | |
322 | { | |
323 | trace_xfs_buf_free(bp, _RET_IP_); | |
324 | ||
325 | ASSERT(list_empty(&bp->b_lru)); | |
326 | ||
327 | if (bp->b_flags & _XBF_PAGES) { | |
328 | uint i; | |
329 | ||
330 | if (xfs_buf_is_vmapped(bp)) | |
331 | vm_unmap_ram(bp->b_addr - bp->b_offset, | |
332 | bp->b_page_count); | |
333 | ||
334 | for (i = 0; i < bp->b_page_count; i++) { | |
335 | struct page *page = bp->b_pages[i]; | |
336 | ||
337 | __free_page(page); | |
338 | } | |
339 | } else if (bp->b_flags & _XBF_KMEM) | |
340 | kmem_free(bp->b_addr); | |
341 | _xfs_buf_free_pages(bp); | |
342 | xfs_buf_free_maps(bp); | |
343 | kmem_zone_free(xfs_buf_zone, bp); | |
344 | } | |
345 | ||
346 | /* | |
347 | * Allocates all the pages for buffer in question and builds it's page list. | |
348 | */ | |
349 | STATIC int | |
350 | xfs_buf_allocate_memory( | |
351 | xfs_buf_t *bp, | |
352 | uint flags) | |
353 | { | |
354 | size_t size; | |
355 | size_t nbytes, offset; | |
356 | gfp_t gfp_mask = xb_to_gfp(flags); | |
357 | unsigned short page_count, i; | |
358 | xfs_off_t start, end; | |
359 | int error; | |
360 | ||
361 | /* | |
362 | * for buffers that are contained within a single page, just allocate | |
363 | * the memory from the heap - there's no need for the complexity of | |
364 | * page arrays to keep allocation down to order 0. | |
365 | */ | |
366 | size = BBTOB(bp->b_length); | |
367 | if (size < PAGE_SIZE) { | |
368 | bp->b_addr = kmem_alloc(size, KM_NOFS); | |
369 | if (!bp->b_addr) { | |
370 | /* low memory - use alloc_page loop instead */ | |
371 | goto use_alloc_page; | |
372 | } | |
373 | ||
374 | if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) != | |
375 | ((unsigned long)bp->b_addr & PAGE_MASK)) { | |
376 | /* b_addr spans two pages - use alloc_page instead */ | |
377 | kmem_free(bp->b_addr); | |
378 | bp->b_addr = NULL; | |
379 | goto use_alloc_page; | |
380 | } | |
381 | bp->b_offset = offset_in_page(bp->b_addr); | |
382 | bp->b_pages = bp->b_page_array; | |
383 | bp->b_pages[0] = virt_to_page(bp->b_addr); | |
384 | bp->b_page_count = 1; | |
385 | bp->b_flags |= _XBF_KMEM; | |
386 | return 0; | |
387 | } | |
388 | ||
389 | use_alloc_page: | |
390 | start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT; | |
391 | end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1) | |
392 | >> PAGE_SHIFT; | |
393 | page_count = end - start; | |
394 | error = _xfs_buf_get_pages(bp, page_count); | |
395 | if (unlikely(error)) | |
396 | return error; | |
397 | ||
398 | offset = bp->b_offset; | |
399 | bp->b_flags |= _XBF_PAGES; | |
400 | ||
401 | for (i = 0; i < bp->b_page_count; i++) { | |
402 | struct page *page; | |
403 | uint retries = 0; | |
404 | retry: | |
405 | page = alloc_page(gfp_mask); | |
406 | if (unlikely(page == NULL)) { | |
407 | if (flags & XBF_READ_AHEAD) { | |
408 | bp->b_page_count = i; | |
409 | error = -ENOMEM; | |
410 | goto out_free_pages; | |
411 | } | |
412 | ||
413 | /* | |
414 | * This could deadlock. | |
415 | * | |
416 | * But until all the XFS lowlevel code is revamped to | |
417 | * handle buffer allocation failures we can't do much. | |
418 | */ | |
419 | if (!(++retries % 100)) | |
420 | xfs_err(NULL, | |
421 | "%s(%u) possible memory allocation deadlock in %s (mode:0x%x)", | |
422 | current->comm, current->pid, | |
423 | __func__, gfp_mask); | |
424 | ||
425 | XFS_STATS_INC(bp->b_target->bt_mount, xb_page_retries); | |
426 | congestion_wait(BLK_RW_ASYNC, HZ/50); | |
427 | goto retry; | |
428 | } | |
429 | ||
430 | XFS_STATS_INC(bp->b_target->bt_mount, xb_page_found); | |
431 | ||
432 | nbytes = min_t(size_t, size, PAGE_SIZE - offset); | |
433 | size -= nbytes; | |
434 | bp->b_pages[i] = page; | |
435 | offset = 0; | |
436 | } | |
437 | return 0; | |
438 | ||
439 | out_free_pages: | |
440 | for (i = 0; i < bp->b_page_count; i++) | |
441 | __free_page(bp->b_pages[i]); | |
442 | bp->b_flags &= ~_XBF_PAGES; | |
443 | return error; | |
444 | } | |
445 | ||
446 | /* | |
447 | * Map buffer into kernel address-space if necessary. | |
448 | */ | |
449 | STATIC int | |
450 | _xfs_buf_map_pages( | |
451 | xfs_buf_t *bp, | |
452 | uint flags) | |
453 | { | |
454 | ASSERT(bp->b_flags & _XBF_PAGES); | |
455 | if (bp->b_page_count == 1) { | |
456 | /* A single page buffer is always mappable */ | |
457 | bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset; | |
458 | } else if (flags & XBF_UNMAPPED) { | |
459 | bp->b_addr = NULL; | |
460 | } else { | |
461 | int retried = 0; | |
462 | unsigned nofs_flag; | |
463 | ||
464 | /* | |
465 | * vm_map_ram() will allocate auxillary structures (e.g. | |
466 | * pagetables) with GFP_KERNEL, yet we are likely to be under | |
467 | * GFP_NOFS context here. Hence we need to tell memory reclaim | |
468 | * that we are in such a context via PF_MEMALLOC_NOFS to prevent | |
469 | * memory reclaim re-entering the filesystem here and | |
470 | * potentially deadlocking. | |
471 | */ | |
472 | nofs_flag = memalloc_nofs_save(); | |
473 | do { | |
474 | bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count, | |
475 | -1, PAGE_KERNEL); | |
476 | if (bp->b_addr) | |
477 | break; | |
478 | vm_unmap_aliases(); | |
479 | } while (retried++ <= 1); | |
480 | memalloc_nofs_restore(nofs_flag); | |
481 | ||
482 | if (!bp->b_addr) | |
483 | return -ENOMEM; | |
484 | bp->b_addr += bp->b_offset; | |
485 | } | |
486 | ||
487 | return 0; | |
488 | } | |
489 | ||
490 | /* | |
491 | * Finding and Reading Buffers | |
492 | */ | |
493 | static int | |
494 | _xfs_buf_obj_cmp( | |
495 | struct rhashtable_compare_arg *arg, | |
496 | const void *obj) | |
497 | { | |
498 | const struct xfs_buf_map *map = arg->key; | |
499 | const struct xfs_buf *bp = obj; | |
500 | ||
501 | /* | |
502 | * The key hashing in the lookup path depends on the key being the | |
503 | * first element of the compare_arg, make sure to assert this. | |
504 | */ | |
505 | BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0); | |
506 | ||
507 | if (bp->b_bn != map->bm_bn) | |
508 | return 1; | |
509 | ||
510 | if (unlikely(bp->b_length != map->bm_len)) { | |
511 | /* | |
512 | * found a block number match. If the range doesn't | |
513 | * match, the only way this is allowed is if the buffer | |
514 | * in the cache is stale and the transaction that made | |
515 | * it stale has not yet committed. i.e. we are | |
516 | * reallocating a busy extent. Skip this buffer and | |
517 | * continue searching for an exact match. | |
518 | */ | |
519 | ASSERT(bp->b_flags & XBF_STALE); | |
520 | return 1; | |
521 | } | |
522 | return 0; | |
523 | } | |
524 | ||
525 | static const struct rhashtable_params xfs_buf_hash_params = { | |
526 | .min_size = 32, /* empty AGs have minimal footprint */ | |
527 | .nelem_hint = 16, | |
528 | .key_len = sizeof(xfs_daddr_t), | |
529 | .key_offset = offsetof(struct xfs_buf, b_bn), | |
530 | .head_offset = offsetof(struct xfs_buf, b_rhash_head), | |
531 | .automatic_shrinking = true, | |
532 | .obj_cmpfn = _xfs_buf_obj_cmp, | |
533 | }; | |
534 | ||
535 | int | |
536 | xfs_buf_hash_init( | |
537 | struct xfs_perag *pag) | |
538 | { | |
539 | spin_lock_init(&pag->pag_buf_lock); | |
540 | return rhashtable_init(&pag->pag_buf_hash, &xfs_buf_hash_params); | |
541 | } | |
542 | ||
543 | void | |
544 | xfs_buf_hash_destroy( | |
545 | struct xfs_perag *pag) | |
546 | { | |
547 | rhashtable_destroy(&pag->pag_buf_hash); | |
548 | } | |
549 | ||
550 | /* | |
551 | * Look up, and creates if absent, a lockable buffer for | |
552 | * a given range of an inode. The buffer is returned | |
553 | * locked. No I/O is implied by this call. | |
554 | */ | |
555 | xfs_buf_t * | |
556 | _xfs_buf_find( | |
557 | struct xfs_buftarg *btp, | |
558 | struct xfs_buf_map *map, | |
559 | int nmaps, | |
560 | xfs_buf_flags_t flags, | |
561 | xfs_buf_t *new_bp) | |
562 | { | |
563 | struct xfs_perag *pag; | |
564 | xfs_buf_t *bp; | |
565 | struct xfs_buf_map cmap = { .bm_bn = map[0].bm_bn }; | |
566 | xfs_daddr_t eofs; | |
567 | int i; | |
568 | ||
569 | for (i = 0; i < nmaps; i++) | |
570 | cmap.bm_len += map[i].bm_len; | |
571 | ||
572 | /* Check for IOs smaller than the sector size / not sector aligned */ | |
573 | ASSERT(!(BBTOB(cmap.bm_len) < btp->bt_meta_sectorsize)); | |
574 | ASSERT(!(BBTOB(cmap.bm_bn) & (xfs_off_t)btp->bt_meta_sectormask)); | |
575 | ||
576 | /* | |
577 | * Corrupted block numbers can get through to here, unfortunately, so we | |
578 | * have to check that the buffer falls within the filesystem bounds. | |
579 | */ | |
580 | eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks); | |
581 | if (cmap.bm_bn < 0 || cmap.bm_bn >= eofs) { | |
582 | /* | |
583 | * XXX (dgc): we should really be returning -EFSCORRUPTED here, | |
584 | * but none of the higher level infrastructure supports | |
585 | * returning a specific error on buffer lookup failures. | |
586 | */ | |
587 | xfs_alert(btp->bt_mount, | |
588 | "%s: Block out of range: block 0x%llx, EOFS 0x%llx ", | |
589 | __func__, cmap.bm_bn, eofs); | |
590 | WARN_ON(1); | |
591 | return NULL; | |
592 | } | |
593 | ||
594 | pag = xfs_perag_get(btp->bt_mount, | |
595 | xfs_daddr_to_agno(btp->bt_mount, cmap.bm_bn)); | |
596 | ||
597 | spin_lock(&pag->pag_buf_lock); | |
598 | bp = rhashtable_lookup_fast(&pag->pag_buf_hash, &cmap, | |
599 | xfs_buf_hash_params); | |
600 | if (bp) { | |
601 | atomic_inc(&bp->b_hold); | |
602 | goto found; | |
603 | } | |
604 | ||
605 | /* No match found */ | |
606 | if (new_bp) { | |
607 | /* the buffer keeps the perag reference until it is freed */ | |
608 | new_bp->b_pag = pag; | |
609 | rhashtable_insert_fast(&pag->pag_buf_hash, | |
610 | &new_bp->b_rhash_head, | |
611 | xfs_buf_hash_params); | |
612 | spin_unlock(&pag->pag_buf_lock); | |
613 | } else { | |
614 | XFS_STATS_INC(btp->bt_mount, xb_miss_locked); | |
615 | spin_unlock(&pag->pag_buf_lock); | |
616 | xfs_perag_put(pag); | |
617 | } | |
618 | return new_bp; | |
619 | ||
620 | found: | |
621 | spin_unlock(&pag->pag_buf_lock); | |
622 | xfs_perag_put(pag); | |
623 | ||
624 | if (!xfs_buf_trylock(bp)) { | |
625 | if (flags & XBF_TRYLOCK) { | |
626 | xfs_buf_rele(bp); | |
627 | XFS_STATS_INC(btp->bt_mount, xb_busy_locked); | |
628 | return NULL; | |
629 | } | |
630 | xfs_buf_lock(bp); | |
631 | XFS_STATS_INC(btp->bt_mount, xb_get_locked_waited); | |
632 | } | |
633 | ||
634 | /* | |
635 | * if the buffer is stale, clear all the external state associated with | |
636 | * it. We need to keep flags such as how we allocated the buffer memory | |
637 | * intact here. | |
638 | */ | |
639 | if (bp->b_flags & XBF_STALE) { | |
640 | ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0); | |
641 | ASSERT(bp->b_iodone == NULL); | |
642 | bp->b_flags &= _XBF_KMEM | _XBF_PAGES; | |
643 | bp->b_ops = NULL; | |
644 | } | |
645 | ||
646 | trace_xfs_buf_find(bp, flags, _RET_IP_); | |
647 | XFS_STATS_INC(btp->bt_mount, xb_get_locked); | |
648 | return bp; | |
649 | } | |
650 | ||
651 | /* | |
652 | * Assembles a buffer covering the specified range. The code is optimised for | |
653 | * cache hits, as metadata intensive workloads will see 3 orders of magnitude | |
654 | * more hits than misses. | |
655 | */ | |
656 | struct xfs_buf * | |
657 | xfs_buf_get_map( | |
658 | struct xfs_buftarg *target, | |
659 | struct xfs_buf_map *map, | |
660 | int nmaps, | |
661 | xfs_buf_flags_t flags) | |
662 | { | |
663 | struct xfs_buf *bp; | |
664 | struct xfs_buf *new_bp; | |
665 | int error = 0; | |
666 | ||
667 | bp = _xfs_buf_find(target, map, nmaps, flags, NULL); | |
668 | if (likely(bp)) | |
669 | goto found; | |
670 | ||
671 | new_bp = _xfs_buf_alloc(target, map, nmaps, flags); | |
672 | if (unlikely(!new_bp)) | |
673 | return NULL; | |
674 | ||
675 | error = xfs_buf_allocate_memory(new_bp, flags); | |
676 | if (error) { | |
677 | xfs_buf_free(new_bp); | |
678 | return NULL; | |
679 | } | |
680 | ||
681 | bp = _xfs_buf_find(target, map, nmaps, flags, new_bp); | |
682 | if (!bp) { | |
683 | xfs_buf_free(new_bp); | |
684 | return NULL; | |
685 | } | |
686 | ||
687 | if (bp != new_bp) | |
688 | xfs_buf_free(new_bp); | |
689 | ||
690 | found: | |
691 | if (!bp->b_addr) { | |
692 | error = _xfs_buf_map_pages(bp, flags); | |
693 | if (unlikely(error)) { | |
694 | xfs_warn(target->bt_mount, | |
695 | "%s: failed to map pagesn", __func__); | |
696 | xfs_buf_relse(bp); | |
697 | return NULL; | |
698 | } | |
699 | } | |
700 | ||
701 | /* | |
702 | * Clear b_error if this is a lookup from a caller that doesn't expect | |
703 | * valid data to be found in the buffer. | |
704 | */ | |
705 | if (!(flags & XBF_READ)) | |
706 | xfs_buf_ioerror(bp, 0); | |
707 | ||
708 | XFS_STATS_INC(target->bt_mount, xb_get); | |
709 | trace_xfs_buf_get(bp, flags, _RET_IP_); | |
710 | return bp; | |
711 | } | |
712 | ||
713 | STATIC int | |
714 | _xfs_buf_read( | |
715 | xfs_buf_t *bp, | |
716 | xfs_buf_flags_t flags) | |
717 | { | |
718 | ASSERT(!(flags & XBF_WRITE)); | |
719 | ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL); | |
720 | ||
721 | bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD); | |
722 | bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD); | |
723 | ||
724 | if (flags & XBF_ASYNC) { | |
725 | xfs_buf_submit(bp); | |
726 | return 0; | |
727 | } | |
728 | return xfs_buf_submit_wait(bp); | |
729 | } | |
730 | ||
731 | xfs_buf_t * | |
732 | xfs_buf_read_map( | |
733 | struct xfs_buftarg *target, | |
734 | struct xfs_buf_map *map, | |
735 | int nmaps, | |
736 | xfs_buf_flags_t flags, | |
737 | const struct xfs_buf_ops *ops) | |
738 | { | |
739 | struct xfs_buf *bp; | |
740 | ||
741 | flags |= XBF_READ; | |
742 | ||
743 | bp = xfs_buf_get_map(target, map, nmaps, flags); | |
744 | if (bp) { | |
745 | trace_xfs_buf_read(bp, flags, _RET_IP_); | |
746 | ||
747 | if (!(bp->b_flags & XBF_DONE)) { | |
748 | XFS_STATS_INC(target->bt_mount, xb_get_read); | |
749 | bp->b_ops = ops; | |
750 | _xfs_buf_read(bp, flags); | |
751 | } else if (flags & XBF_ASYNC) { | |
752 | /* | |
753 | * Read ahead call which is already satisfied, | |
754 | * drop the buffer | |
755 | */ | |
756 | xfs_buf_relse(bp); | |
757 | return NULL; | |
758 | } else { | |
759 | /* We do not want read in the flags */ | |
760 | bp->b_flags &= ~XBF_READ; | |
761 | } | |
762 | } | |
763 | ||
764 | return bp; | |
765 | } | |
766 | ||
767 | /* | |
768 | * If we are not low on memory then do the readahead in a deadlock | |
769 | * safe manner. | |
770 | */ | |
771 | void | |
772 | xfs_buf_readahead_map( | |
773 | struct xfs_buftarg *target, | |
774 | struct xfs_buf_map *map, | |
775 | int nmaps, | |
776 | const struct xfs_buf_ops *ops) | |
777 | { | |
778 | if (bdi_read_congested(target->bt_bdev->bd_bdi)) | |
779 | return; | |
780 | ||
781 | xfs_buf_read_map(target, map, nmaps, | |
782 | XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops); | |
783 | } | |
784 | ||
785 | /* | |
786 | * Read an uncached buffer from disk. Allocates and returns a locked | |
787 | * buffer containing the disk contents or nothing. | |
788 | */ | |
789 | int | |
790 | xfs_buf_read_uncached( | |
791 | struct xfs_buftarg *target, | |
792 | xfs_daddr_t daddr, | |
793 | size_t numblks, | |
794 | int flags, | |
795 | struct xfs_buf **bpp, | |
796 | const struct xfs_buf_ops *ops) | |
797 | { | |
798 | struct xfs_buf *bp; | |
799 | ||
800 | *bpp = NULL; | |
801 | ||
802 | bp = xfs_buf_get_uncached(target, numblks, flags); | |
803 | if (!bp) | |
804 | return -ENOMEM; | |
805 | ||
806 | /* set up the buffer for a read IO */ | |
807 | ASSERT(bp->b_map_count == 1); | |
808 | bp->b_bn = XFS_BUF_DADDR_NULL; /* always null for uncached buffers */ | |
809 | bp->b_maps[0].bm_bn = daddr; | |
810 | bp->b_flags |= XBF_READ; | |
811 | bp->b_ops = ops; | |
812 | ||
813 | xfs_buf_submit_wait(bp); | |
814 | if (bp->b_error) { | |
815 | int error = bp->b_error; | |
816 | xfs_buf_relse(bp); | |
817 | return error; | |
818 | } | |
819 | ||
820 | *bpp = bp; | |
821 | return 0; | |
822 | } | |
823 | ||
824 | /* | |
825 | * Return a buffer allocated as an empty buffer and associated to external | |
826 | * memory via xfs_buf_associate_memory() back to it's empty state. | |
827 | */ | |
828 | void | |
829 | xfs_buf_set_empty( | |
830 | struct xfs_buf *bp, | |
831 | size_t numblks) | |
832 | { | |
833 | if (bp->b_pages) | |
834 | _xfs_buf_free_pages(bp); | |
835 | ||
836 | bp->b_pages = NULL; | |
837 | bp->b_page_count = 0; | |
838 | bp->b_addr = NULL; | |
839 | bp->b_length = numblks; | |
840 | bp->b_io_length = numblks; | |
841 | ||
842 | ASSERT(bp->b_map_count == 1); | |
843 | bp->b_bn = XFS_BUF_DADDR_NULL; | |
844 | bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL; | |
845 | bp->b_maps[0].bm_len = bp->b_length; | |
846 | } | |
847 | ||
848 | static inline struct page * | |
849 | mem_to_page( | |
850 | void *addr) | |
851 | { | |
852 | if ((!is_vmalloc_addr(addr))) { | |
853 | return virt_to_page(addr); | |
854 | } else { | |
855 | return vmalloc_to_page(addr); | |
856 | } | |
857 | } | |
858 | ||
859 | int | |
860 | xfs_buf_associate_memory( | |
861 | xfs_buf_t *bp, | |
862 | void *mem, | |
863 | size_t len) | |
864 | { | |
865 | int rval; | |
866 | int i = 0; | |
867 | unsigned long pageaddr; | |
868 | unsigned long offset; | |
869 | size_t buflen; | |
870 | int page_count; | |
871 | ||
872 | pageaddr = (unsigned long)mem & PAGE_MASK; | |
873 | offset = (unsigned long)mem - pageaddr; | |
874 | buflen = PAGE_ALIGN(len + offset); | |
875 | page_count = buflen >> PAGE_SHIFT; | |
876 | ||
877 | /* Free any previous set of page pointers */ | |
878 | if (bp->b_pages) | |
879 | _xfs_buf_free_pages(bp); | |
880 | ||
881 | bp->b_pages = NULL; | |
882 | bp->b_addr = mem; | |
883 | ||
884 | rval = _xfs_buf_get_pages(bp, page_count); | |
885 | if (rval) | |
886 | return rval; | |
887 | ||
888 | bp->b_offset = offset; | |
889 | ||
890 | for (i = 0; i < bp->b_page_count; i++) { | |
891 | bp->b_pages[i] = mem_to_page((void *)pageaddr); | |
892 | pageaddr += PAGE_SIZE; | |
893 | } | |
894 | ||
895 | bp->b_io_length = BTOBB(len); | |
896 | bp->b_length = BTOBB(buflen); | |
897 | ||
898 | return 0; | |
899 | } | |
900 | ||
901 | xfs_buf_t * | |
902 | xfs_buf_get_uncached( | |
903 | struct xfs_buftarg *target, | |
904 | size_t numblks, | |
905 | int flags) | |
906 | { | |
907 | unsigned long page_count; | |
908 | int error, i; | |
909 | struct xfs_buf *bp; | |
910 | DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks); | |
911 | ||
912 | /* flags might contain irrelevant bits, pass only what we care about */ | |
913 | bp = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT); | |
914 | if (unlikely(bp == NULL)) | |
915 | goto fail; | |
916 | ||
917 | page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT; | |
918 | error = _xfs_buf_get_pages(bp, page_count); | |
919 | if (error) | |
920 | goto fail_free_buf; | |
921 | ||
922 | for (i = 0; i < page_count; i++) { | |
923 | bp->b_pages[i] = alloc_page(xb_to_gfp(flags)); | |
924 | if (!bp->b_pages[i]) | |
925 | goto fail_free_mem; | |
926 | } | |
927 | bp->b_flags |= _XBF_PAGES; | |
928 | ||
929 | error = _xfs_buf_map_pages(bp, 0); | |
930 | if (unlikely(error)) { | |
931 | xfs_warn(target->bt_mount, | |
932 | "%s: failed to map pages", __func__); | |
933 | goto fail_free_mem; | |
934 | } | |
935 | ||
936 | trace_xfs_buf_get_uncached(bp, _RET_IP_); | |
937 | return bp; | |
938 | ||
939 | fail_free_mem: | |
940 | while (--i >= 0) | |
941 | __free_page(bp->b_pages[i]); | |
942 | _xfs_buf_free_pages(bp); | |
943 | fail_free_buf: | |
944 | xfs_buf_free_maps(bp); | |
945 | kmem_zone_free(xfs_buf_zone, bp); | |
946 | fail: | |
947 | return NULL; | |
948 | } | |
949 | ||
950 | /* | |
951 | * Increment reference count on buffer, to hold the buffer concurrently | |
952 | * with another thread which may release (free) the buffer asynchronously. | |
953 | * Must hold the buffer already to call this function. | |
954 | */ | |
955 | void | |
956 | xfs_buf_hold( | |
957 | xfs_buf_t *bp) | |
958 | { | |
959 | trace_xfs_buf_hold(bp, _RET_IP_); | |
960 | atomic_inc(&bp->b_hold); | |
961 | } | |
962 | ||
963 | /* | |
964 | * Release a hold on the specified buffer. If the hold count is 1, the buffer is | |
965 | * placed on LRU or freed (depending on b_lru_ref). | |
966 | */ | |
967 | void | |
968 | xfs_buf_rele( | |
969 | xfs_buf_t *bp) | |
970 | { | |
971 | struct xfs_perag *pag = bp->b_pag; | |
972 | bool release; | |
973 | bool freebuf = false; | |
974 | ||
975 | trace_xfs_buf_rele(bp, _RET_IP_); | |
976 | ||
977 | if (!pag) { | |
978 | ASSERT(list_empty(&bp->b_lru)); | |
979 | if (atomic_dec_and_test(&bp->b_hold)) { | |
980 | xfs_buf_ioacct_dec(bp); | |
981 | xfs_buf_free(bp); | |
982 | } | |
983 | return; | |
984 | } | |
985 | ||
986 | ASSERT(atomic_read(&bp->b_hold) > 0); | |
987 | ||
988 | release = atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock); | |
989 | spin_lock(&bp->b_lock); | |
990 | if (!release) { | |
991 | /* | |
992 | * Drop the in-flight state if the buffer is already on the LRU | |
993 | * and it holds the only reference. This is racy because we | |
994 | * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT | |
995 | * ensures the decrement occurs only once per-buf. | |
996 | */ | |
997 | if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru)) | |
998 | __xfs_buf_ioacct_dec(bp); | |
999 | goto out_unlock; | |
1000 | } | |
1001 | ||
1002 | /* the last reference has been dropped ... */ | |
1003 | __xfs_buf_ioacct_dec(bp); | |
1004 | if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) { | |
1005 | /* | |
1006 | * If the buffer is added to the LRU take a new reference to the | |
1007 | * buffer for the LRU and clear the (now stale) dispose list | |
1008 | * state flag | |
1009 | */ | |
1010 | if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) { | |
1011 | bp->b_state &= ~XFS_BSTATE_DISPOSE; | |
1012 | atomic_inc(&bp->b_hold); | |
1013 | } | |
1014 | spin_unlock(&pag->pag_buf_lock); | |
1015 | } else { | |
1016 | /* | |
1017 | * most of the time buffers will already be removed from the | |
1018 | * LRU, so optimise that case by checking for the | |
1019 | * XFS_BSTATE_DISPOSE flag indicating the last list the buffer | |
1020 | * was on was the disposal list | |
1021 | */ | |
1022 | if (!(bp->b_state & XFS_BSTATE_DISPOSE)) { | |
1023 | list_lru_del(&bp->b_target->bt_lru, &bp->b_lru); | |
1024 | } else { | |
1025 | ASSERT(list_empty(&bp->b_lru)); | |
1026 | } | |
1027 | ||
1028 | ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); | |
1029 | rhashtable_remove_fast(&pag->pag_buf_hash, &bp->b_rhash_head, | |
1030 | xfs_buf_hash_params); | |
1031 | spin_unlock(&pag->pag_buf_lock); | |
1032 | xfs_perag_put(pag); | |
1033 | freebuf = true; | |
1034 | } | |
1035 | ||
1036 | out_unlock: | |
1037 | spin_unlock(&bp->b_lock); | |
1038 | ||
1039 | if (freebuf) | |
1040 | xfs_buf_free(bp); | |
1041 | } | |
1042 | ||
1043 | ||
1044 | /* | |
1045 | * Lock a buffer object, if it is not already locked. | |
1046 | * | |
1047 | * If we come across a stale, pinned, locked buffer, we know that we are | |
1048 | * being asked to lock a buffer that has been reallocated. Because it is | |
1049 | * pinned, we know that the log has not been pushed to disk and hence it | |
1050 | * will still be locked. Rather than continuing to have trylock attempts | |
1051 | * fail until someone else pushes the log, push it ourselves before | |
1052 | * returning. This means that the xfsaild will not get stuck trying | |
1053 | * to push on stale inode buffers. | |
1054 | */ | |
1055 | int | |
1056 | xfs_buf_trylock( | |
1057 | struct xfs_buf *bp) | |
1058 | { | |
1059 | int locked; | |
1060 | ||
1061 | locked = down_trylock(&bp->b_sema) == 0; | |
1062 | if (locked) { | |
1063 | XB_SET_OWNER(bp); | |
1064 | trace_xfs_buf_trylock(bp, _RET_IP_); | |
1065 | } else { | |
1066 | trace_xfs_buf_trylock_fail(bp, _RET_IP_); | |
1067 | } | |
1068 | return locked; | |
1069 | } | |
1070 | ||
1071 | /* | |
1072 | * Lock a buffer object. | |
1073 | * | |
1074 | * If we come across a stale, pinned, locked buffer, we know that we | |
1075 | * are being asked to lock a buffer that has been reallocated. Because | |
1076 | * it is pinned, we know that the log has not been pushed to disk and | |
1077 | * hence it will still be locked. Rather than sleeping until someone | |
1078 | * else pushes the log, push it ourselves before trying to get the lock. | |
1079 | */ | |
1080 | void | |
1081 | xfs_buf_lock( | |
1082 | struct xfs_buf *bp) | |
1083 | { | |
1084 | trace_xfs_buf_lock(bp, _RET_IP_); | |
1085 | ||
1086 | if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE)) | |
1087 | xfs_log_force(bp->b_target->bt_mount, 0); | |
1088 | down(&bp->b_sema); | |
1089 | XB_SET_OWNER(bp); | |
1090 | ||
1091 | trace_xfs_buf_lock_done(bp, _RET_IP_); | |
1092 | } | |
1093 | ||
1094 | void | |
1095 | xfs_buf_unlock( | |
1096 | struct xfs_buf *bp) | |
1097 | { | |
1098 | ASSERT(xfs_buf_islocked(bp)); | |
1099 | ||
1100 | XB_CLEAR_OWNER(bp); | |
1101 | up(&bp->b_sema); | |
1102 | ||
1103 | trace_xfs_buf_unlock(bp, _RET_IP_); | |
1104 | } | |
1105 | ||
1106 | STATIC void | |
1107 | xfs_buf_wait_unpin( | |
1108 | xfs_buf_t *bp) | |
1109 | { | |
1110 | DECLARE_WAITQUEUE (wait, current); | |
1111 | ||
1112 | if (atomic_read(&bp->b_pin_count) == 0) | |
1113 | return; | |
1114 | ||
1115 | add_wait_queue(&bp->b_waiters, &wait); | |
1116 | for (;;) { | |
1117 | set_current_state(TASK_UNINTERRUPTIBLE); | |
1118 | if (atomic_read(&bp->b_pin_count) == 0) | |
1119 | break; | |
1120 | io_schedule(); | |
1121 | } | |
1122 | remove_wait_queue(&bp->b_waiters, &wait); | |
1123 | set_current_state(TASK_RUNNING); | |
1124 | } | |
1125 | ||
1126 | /* | |
1127 | * Buffer Utility Routines | |
1128 | */ | |
1129 | ||
1130 | void | |
1131 | xfs_buf_ioend( | |
1132 | struct xfs_buf *bp) | |
1133 | { | |
1134 | bool read = bp->b_flags & XBF_READ; | |
1135 | ||
1136 | trace_xfs_buf_iodone(bp, _RET_IP_); | |
1137 | ||
1138 | bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); | |
1139 | ||
1140 | /* | |
1141 | * Pull in IO completion errors now. We are guaranteed to be running | |
1142 | * single threaded, so we don't need the lock to read b_io_error. | |
1143 | */ | |
1144 | if (!bp->b_error && bp->b_io_error) | |
1145 | xfs_buf_ioerror(bp, bp->b_io_error); | |
1146 | ||
1147 | /* Only validate buffers that were read without errors */ | |
1148 | if (read && !bp->b_error && bp->b_ops) { | |
1149 | ASSERT(!bp->b_iodone); | |
1150 | bp->b_ops->verify_read(bp); | |
1151 | } | |
1152 | ||
1153 | if (!bp->b_error) | |
1154 | bp->b_flags |= XBF_DONE; | |
1155 | ||
1156 | if (bp->b_iodone) | |
1157 | (*(bp->b_iodone))(bp); | |
1158 | else if (bp->b_flags & XBF_ASYNC) | |
1159 | xfs_buf_relse(bp); | |
1160 | else | |
1161 | complete(&bp->b_iowait); | |
1162 | } | |
1163 | ||
1164 | static void | |
1165 | xfs_buf_ioend_work( | |
1166 | struct work_struct *work) | |
1167 | { | |
1168 | struct xfs_buf *bp = | |
1169 | container_of(work, xfs_buf_t, b_ioend_work); | |
1170 | ||
1171 | xfs_buf_ioend(bp); | |
1172 | } | |
1173 | ||
1174 | static void | |
1175 | xfs_buf_ioend_async( | |
1176 | struct xfs_buf *bp) | |
1177 | { | |
1178 | INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work); | |
1179 | queue_work(bp->b_ioend_wq, &bp->b_ioend_work); | |
1180 | } | |
1181 | ||
1182 | void | |
1183 | __xfs_buf_ioerror( | |
1184 | xfs_buf_t *bp, | |
1185 | int error, | |
1186 | xfs_failaddr_t failaddr) | |
1187 | { | |
1188 | ASSERT(error <= 0 && error >= -1000); | |
1189 | bp->b_error = error; | |
1190 | trace_xfs_buf_ioerror(bp, error, failaddr); | |
1191 | } | |
1192 | ||
1193 | void | |
1194 | xfs_buf_ioerror_alert( | |
1195 | struct xfs_buf *bp, | |
1196 | const char *func) | |
1197 | { | |
1198 | xfs_alert(bp->b_target->bt_mount, | |
1199 | "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d", | |
1200 | (uint64_t)XFS_BUF_ADDR(bp), func, -bp->b_error, bp->b_length); | |
1201 | } | |
1202 | ||
1203 | int | |
1204 | xfs_bwrite( | |
1205 | struct xfs_buf *bp) | |
1206 | { | |
1207 | int error; | |
1208 | ||
1209 | ASSERT(xfs_buf_islocked(bp)); | |
1210 | ||
1211 | bp->b_flags |= XBF_WRITE; | |
1212 | bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q | | |
1213 | XBF_WRITE_FAIL | XBF_DONE); | |
1214 | ||
1215 | error = xfs_buf_submit_wait(bp); | |
1216 | if (error) { | |
1217 | xfs_force_shutdown(bp->b_target->bt_mount, | |
1218 | SHUTDOWN_META_IO_ERROR); | |
1219 | } | |
1220 | return error; | |
1221 | } | |
1222 | ||
1223 | static void | |
1224 | xfs_buf_bio_end_io( | |
1225 | struct bio *bio) | |
1226 | { | |
1227 | struct xfs_buf *bp = (struct xfs_buf *)bio->bi_private; | |
1228 | ||
1229 | /* | |
1230 | * don't overwrite existing errors - otherwise we can lose errors on | |
1231 | * buffers that require multiple bios to complete. | |
1232 | */ | |
1233 | if (bio->bi_status) { | |
1234 | int error = blk_status_to_errno(bio->bi_status); | |
1235 | ||
1236 | cmpxchg(&bp->b_io_error, 0, error); | |
1237 | } | |
1238 | ||
1239 | if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ)) | |
1240 | invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp)); | |
1241 | ||
1242 | if (atomic_dec_and_test(&bp->b_io_remaining) == 1) | |
1243 | xfs_buf_ioend_async(bp); | |
1244 | bio_put(bio); | |
1245 | } | |
1246 | ||
1247 | static void | |
1248 | xfs_buf_ioapply_map( | |
1249 | struct xfs_buf *bp, | |
1250 | int map, | |
1251 | int *buf_offset, | |
1252 | int *count, | |
1253 | int op, | |
1254 | int op_flags) | |
1255 | { | |
1256 | int page_index; | |
1257 | int total_nr_pages = bp->b_page_count; | |
1258 | int nr_pages; | |
1259 | struct bio *bio; | |
1260 | sector_t sector = bp->b_maps[map].bm_bn; | |
1261 | int size; | |
1262 | int offset; | |
1263 | ||
1264 | /* skip the pages in the buffer before the start offset */ | |
1265 | page_index = 0; | |
1266 | offset = *buf_offset; | |
1267 | while (offset >= PAGE_SIZE) { | |
1268 | page_index++; | |
1269 | offset -= PAGE_SIZE; | |
1270 | } | |
1271 | ||
1272 | /* | |
1273 | * Limit the IO size to the length of the current vector, and update the | |
1274 | * remaining IO count for the next time around. | |
1275 | */ | |
1276 | size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count); | |
1277 | *count -= size; | |
1278 | *buf_offset += size; | |
1279 | ||
1280 | next_chunk: | |
1281 | atomic_inc(&bp->b_io_remaining); | |
1282 | nr_pages = min(total_nr_pages, BIO_MAX_PAGES); | |
1283 | ||
1284 | bio = bio_alloc(GFP_NOIO, nr_pages); | |
1285 | bio_set_dev(bio, bp->b_target->bt_bdev); | |
1286 | bio->bi_iter.bi_sector = sector; | |
1287 | bio->bi_end_io = xfs_buf_bio_end_io; | |
1288 | bio->bi_private = bp; | |
1289 | bio_set_op_attrs(bio, op, op_flags); | |
1290 | ||
1291 | for (; size && nr_pages; nr_pages--, page_index++) { | |
1292 | int rbytes, nbytes = PAGE_SIZE - offset; | |
1293 | ||
1294 | if (nbytes > size) | |
1295 | nbytes = size; | |
1296 | ||
1297 | rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes, | |
1298 | offset); | |
1299 | if (rbytes < nbytes) | |
1300 | break; | |
1301 | ||
1302 | offset = 0; | |
1303 | sector += BTOBB(nbytes); | |
1304 | size -= nbytes; | |
1305 | total_nr_pages--; | |
1306 | } | |
1307 | ||
1308 | if (likely(bio->bi_iter.bi_size)) { | |
1309 | if (xfs_buf_is_vmapped(bp)) { | |
1310 | flush_kernel_vmap_range(bp->b_addr, | |
1311 | xfs_buf_vmap_len(bp)); | |
1312 | } | |
1313 | submit_bio(bio); | |
1314 | if (size) | |
1315 | goto next_chunk; | |
1316 | } else { | |
1317 | /* | |
1318 | * This is guaranteed not to be the last io reference count | |
1319 | * because the caller (xfs_buf_submit) holds a count itself. | |
1320 | */ | |
1321 | atomic_dec(&bp->b_io_remaining); | |
1322 | xfs_buf_ioerror(bp, -EIO); | |
1323 | bio_put(bio); | |
1324 | } | |
1325 | ||
1326 | } | |
1327 | ||
1328 | STATIC void | |
1329 | _xfs_buf_ioapply( | |
1330 | struct xfs_buf *bp) | |
1331 | { | |
1332 | struct blk_plug plug; | |
1333 | int op; | |
1334 | int op_flags = 0; | |
1335 | int offset; | |
1336 | int size; | |
1337 | int i; | |
1338 | ||
1339 | /* | |
1340 | * Make sure we capture only current IO errors rather than stale errors | |
1341 | * left over from previous use of the buffer (e.g. failed readahead). | |
1342 | */ | |
1343 | bp->b_error = 0; | |
1344 | ||
1345 | /* | |
1346 | * Initialize the I/O completion workqueue if we haven't yet or the | |
1347 | * submitter has not opted to specify a custom one. | |
1348 | */ | |
1349 | if (!bp->b_ioend_wq) | |
1350 | bp->b_ioend_wq = bp->b_target->bt_mount->m_buf_workqueue; | |
1351 | ||
1352 | if (bp->b_flags & XBF_WRITE) { | |
1353 | op = REQ_OP_WRITE; | |
1354 | if (bp->b_flags & XBF_SYNCIO) | |
1355 | op_flags = REQ_SYNC; | |
1356 | if (bp->b_flags & XBF_FUA) | |
1357 | op_flags |= REQ_FUA; | |
1358 | if (bp->b_flags & XBF_FLUSH) | |
1359 | op_flags |= REQ_PREFLUSH; | |
1360 | ||
1361 | /* | |
1362 | * Run the write verifier callback function if it exists. If | |
1363 | * this function fails it will mark the buffer with an error and | |
1364 | * the IO should not be dispatched. | |
1365 | */ | |
1366 | if (bp->b_ops) { | |
1367 | bp->b_ops->verify_write(bp); | |
1368 | if (bp->b_error) { | |
1369 | xfs_force_shutdown(bp->b_target->bt_mount, | |
1370 | SHUTDOWN_CORRUPT_INCORE); | |
1371 | return; | |
1372 | } | |
1373 | } else if (bp->b_bn != XFS_BUF_DADDR_NULL) { | |
1374 | struct xfs_mount *mp = bp->b_target->bt_mount; | |
1375 | ||
1376 | /* | |
1377 | * non-crc filesystems don't attach verifiers during | |
1378 | * log recovery, so don't warn for such filesystems. | |
1379 | */ | |
1380 | if (xfs_sb_version_hascrc(&mp->m_sb)) { | |
1381 | xfs_warn(mp, | |
1382 | "%s: no ops on block 0x%llx/0x%x", | |
1383 | __func__, bp->b_bn, bp->b_length); | |
1384 | xfs_hex_dump(bp->b_addr, | |
1385 | XFS_CORRUPTION_DUMP_LEN); | |
1386 | dump_stack(); | |
1387 | } | |
1388 | } | |
1389 | } else if (bp->b_flags & XBF_READ_AHEAD) { | |
1390 | op = REQ_OP_READ; | |
1391 | op_flags = REQ_RAHEAD; | |
1392 | } else { | |
1393 | op = REQ_OP_READ; | |
1394 | } | |
1395 | ||
1396 | /* we only use the buffer cache for meta-data */ | |
1397 | op_flags |= REQ_META; | |
1398 | ||
1399 | /* | |
1400 | * Walk all the vectors issuing IO on them. Set up the initial offset | |
1401 | * into the buffer and the desired IO size before we start - | |
1402 | * _xfs_buf_ioapply_vec() will modify them appropriately for each | |
1403 | * subsequent call. | |
1404 | */ | |
1405 | offset = bp->b_offset; | |
1406 | size = BBTOB(bp->b_io_length); | |
1407 | blk_start_plug(&plug); | |
1408 | for (i = 0; i < bp->b_map_count; i++) { | |
1409 | xfs_buf_ioapply_map(bp, i, &offset, &size, op, op_flags); | |
1410 | if (bp->b_error) | |
1411 | break; | |
1412 | if (size <= 0) | |
1413 | break; /* all done */ | |
1414 | } | |
1415 | blk_finish_plug(&plug); | |
1416 | } | |
1417 | ||
1418 | /* | |
1419 | * Asynchronous IO submission path. This transfers the buffer lock ownership and | |
1420 | * the current reference to the IO. It is not safe to reference the buffer after | |
1421 | * a call to this function unless the caller holds an additional reference | |
1422 | * itself. | |
1423 | */ | |
1424 | void | |
1425 | xfs_buf_submit( | |
1426 | struct xfs_buf *bp) | |
1427 | { | |
1428 | trace_xfs_buf_submit(bp, _RET_IP_); | |
1429 | ||
1430 | ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); | |
1431 | ASSERT(bp->b_flags & XBF_ASYNC); | |
1432 | ||
1433 | /* on shutdown we stale and complete the buffer immediately */ | |
1434 | if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) { | |
1435 | xfs_buf_ioerror(bp, -EIO); | |
1436 | bp->b_flags &= ~XBF_DONE; | |
1437 | xfs_buf_stale(bp); | |
1438 | xfs_buf_ioend(bp); | |
1439 | return; | |
1440 | } | |
1441 | ||
1442 | if (bp->b_flags & XBF_WRITE) | |
1443 | xfs_buf_wait_unpin(bp); | |
1444 | ||
1445 | /* clear the internal error state to avoid spurious errors */ | |
1446 | bp->b_io_error = 0; | |
1447 | ||
1448 | /* | |
1449 | * The caller's reference is released during I/O completion. | |
1450 | * This occurs some time after the last b_io_remaining reference is | |
1451 | * released, so after we drop our Io reference we have to have some | |
1452 | * other reference to ensure the buffer doesn't go away from underneath | |
1453 | * us. Take a direct reference to ensure we have safe access to the | |
1454 | * buffer until we are finished with it. | |
1455 | */ | |
1456 | xfs_buf_hold(bp); | |
1457 | ||
1458 | /* | |
1459 | * Set the count to 1 initially, this will stop an I/O completion | |
1460 | * callout which happens before we have started all the I/O from calling | |
1461 | * xfs_buf_ioend too early. | |
1462 | */ | |
1463 | atomic_set(&bp->b_io_remaining, 1); | |
1464 | xfs_buf_ioacct_inc(bp); | |
1465 | _xfs_buf_ioapply(bp); | |
1466 | ||
1467 | /* | |
1468 | * If _xfs_buf_ioapply failed, we can get back here with only the IO | |
1469 | * reference we took above. If we drop it to zero, run completion so | |
1470 | * that we don't return to the caller with completion still pending. | |
1471 | */ | |
1472 | if (atomic_dec_and_test(&bp->b_io_remaining) == 1) { | |
1473 | if (bp->b_error) | |
1474 | xfs_buf_ioend(bp); | |
1475 | else | |
1476 | xfs_buf_ioend_async(bp); | |
1477 | } | |
1478 | ||
1479 | xfs_buf_rele(bp); | |
1480 | /* Note: it is not safe to reference bp now we've dropped our ref */ | |
1481 | } | |
1482 | ||
1483 | /* | |
1484 | * Synchronous buffer IO submission path, read or write. | |
1485 | */ | |
1486 | int | |
1487 | xfs_buf_submit_wait( | |
1488 | struct xfs_buf *bp) | |
1489 | { | |
1490 | int error; | |
1491 | ||
1492 | trace_xfs_buf_submit_wait(bp, _RET_IP_); | |
1493 | ||
1494 | ASSERT(!(bp->b_flags & (_XBF_DELWRI_Q | XBF_ASYNC))); | |
1495 | ||
1496 | if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) { | |
1497 | xfs_buf_ioerror(bp, -EIO); | |
1498 | xfs_buf_stale(bp); | |
1499 | bp->b_flags &= ~XBF_DONE; | |
1500 | return -EIO; | |
1501 | } | |
1502 | ||
1503 | if (bp->b_flags & XBF_WRITE) | |
1504 | xfs_buf_wait_unpin(bp); | |
1505 | ||
1506 | /* clear the internal error state to avoid spurious errors */ | |
1507 | bp->b_io_error = 0; | |
1508 | ||
1509 | /* | |
1510 | * For synchronous IO, the IO does not inherit the submitters reference | |
1511 | * count, nor the buffer lock. Hence we cannot release the reference we | |
1512 | * are about to take until we've waited for all IO completion to occur, | |
1513 | * including any xfs_buf_ioend_async() work that may be pending. | |
1514 | */ | |
1515 | xfs_buf_hold(bp); | |
1516 | ||
1517 | /* | |
1518 | * Set the count to 1 initially, this will stop an I/O completion | |
1519 | * callout which happens before we have started all the I/O from calling | |
1520 | * xfs_buf_ioend too early. | |
1521 | */ | |
1522 | atomic_set(&bp->b_io_remaining, 1); | |
1523 | _xfs_buf_ioapply(bp); | |
1524 | ||
1525 | /* | |
1526 | * make sure we run completion synchronously if it raced with us and is | |
1527 | * already complete. | |
1528 | */ | |
1529 | if (atomic_dec_and_test(&bp->b_io_remaining) == 1) | |
1530 | xfs_buf_ioend(bp); | |
1531 | ||
1532 | /* wait for completion before gathering the error from the buffer */ | |
1533 | trace_xfs_buf_iowait(bp, _RET_IP_); | |
1534 | wait_for_completion(&bp->b_iowait); | |
1535 | trace_xfs_buf_iowait_done(bp, _RET_IP_); | |
1536 | error = bp->b_error; | |
1537 | ||
1538 | /* | |
1539 | * all done now, we can release the hold that keeps the buffer | |
1540 | * referenced for the entire IO. | |
1541 | */ | |
1542 | xfs_buf_rele(bp); | |
1543 | return error; | |
1544 | } | |
1545 | ||
1546 | void * | |
1547 | xfs_buf_offset( | |
1548 | struct xfs_buf *bp, | |
1549 | size_t offset) | |
1550 | { | |
1551 | struct page *page; | |
1552 | ||
1553 | if (bp->b_addr) | |
1554 | return bp->b_addr + offset; | |
1555 | ||
1556 | offset += bp->b_offset; | |
1557 | page = bp->b_pages[offset >> PAGE_SHIFT]; | |
1558 | return page_address(page) + (offset & (PAGE_SIZE-1)); | |
1559 | } | |
1560 | ||
1561 | /* | |
1562 | * Move data into or out of a buffer. | |
1563 | */ | |
1564 | void | |
1565 | xfs_buf_iomove( | |
1566 | xfs_buf_t *bp, /* buffer to process */ | |
1567 | size_t boff, /* starting buffer offset */ | |
1568 | size_t bsize, /* length to copy */ | |
1569 | void *data, /* data address */ | |
1570 | xfs_buf_rw_t mode) /* read/write/zero flag */ | |
1571 | { | |
1572 | size_t bend; | |
1573 | ||
1574 | bend = boff + bsize; | |
1575 | while (boff < bend) { | |
1576 | struct page *page; | |
1577 | int page_index, page_offset, csize; | |
1578 | ||
1579 | page_index = (boff + bp->b_offset) >> PAGE_SHIFT; | |
1580 | page_offset = (boff + bp->b_offset) & ~PAGE_MASK; | |
1581 | page = bp->b_pages[page_index]; | |
1582 | csize = min_t(size_t, PAGE_SIZE - page_offset, | |
1583 | BBTOB(bp->b_io_length) - boff); | |
1584 | ||
1585 | ASSERT((csize + page_offset) <= PAGE_SIZE); | |
1586 | ||
1587 | switch (mode) { | |
1588 | case XBRW_ZERO: | |
1589 | memset(page_address(page) + page_offset, 0, csize); | |
1590 | break; | |
1591 | case XBRW_READ: | |
1592 | memcpy(data, page_address(page) + page_offset, csize); | |
1593 | break; | |
1594 | case XBRW_WRITE: | |
1595 | memcpy(page_address(page) + page_offset, data, csize); | |
1596 | } | |
1597 | ||
1598 | boff += csize; | |
1599 | data += csize; | |
1600 | } | |
1601 | } | |
1602 | ||
1603 | /* | |
1604 | * Handling of buffer targets (buftargs). | |
1605 | */ | |
1606 | ||
1607 | /* | |
1608 | * Wait for any bufs with callbacks that have been submitted but have not yet | |
1609 | * returned. These buffers will have an elevated hold count, so wait on those | |
1610 | * while freeing all the buffers only held by the LRU. | |
1611 | */ | |
1612 | static enum lru_status | |
1613 | xfs_buftarg_wait_rele( | |
1614 | struct list_head *item, | |
1615 | struct list_lru_one *lru, | |
1616 | spinlock_t *lru_lock, | |
1617 | void *arg) | |
1618 | ||
1619 | { | |
1620 | struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru); | |
1621 | struct list_head *dispose = arg; | |
1622 | ||
1623 | if (atomic_read(&bp->b_hold) > 1) { | |
1624 | /* need to wait, so skip it this pass */ | |
1625 | trace_xfs_buf_wait_buftarg(bp, _RET_IP_); | |
1626 | return LRU_SKIP; | |
1627 | } | |
1628 | if (!spin_trylock(&bp->b_lock)) | |
1629 | return LRU_SKIP; | |
1630 | ||
1631 | /* | |
1632 | * clear the LRU reference count so the buffer doesn't get | |
1633 | * ignored in xfs_buf_rele(). | |
1634 | */ | |
1635 | atomic_set(&bp->b_lru_ref, 0); | |
1636 | bp->b_state |= XFS_BSTATE_DISPOSE; | |
1637 | list_lru_isolate_move(lru, item, dispose); | |
1638 | spin_unlock(&bp->b_lock); | |
1639 | return LRU_REMOVED; | |
1640 | } | |
1641 | ||
1642 | void | |
1643 | xfs_wait_buftarg( | |
1644 | struct xfs_buftarg *btp) | |
1645 | { | |
1646 | LIST_HEAD(dispose); | |
1647 | int loop = 0; | |
1648 | ||
1649 | /* | |
1650 | * First wait on the buftarg I/O count for all in-flight buffers to be | |
1651 | * released. This is critical as new buffers do not make the LRU until | |
1652 | * they are released. | |
1653 | * | |
1654 | * Next, flush the buffer workqueue to ensure all completion processing | |
1655 | * has finished. Just waiting on buffer locks is not sufficient for | |
1656 | * async IO as the reference count held over IO is not released until | |
1657 | * after the buffer lock is dropped. Hence we need to ensure here that | |
1658 | * all reference counts have been dropped before we start walking the | |
1659 | * LRU list. | |
1660 | */ | |
1661 | while (percpu_counter_sum(&btp->bt_io_count)) | |
1662 | delay(100); | |
1663 | flush_workqueue(btp->bt_mount->m_buf_workqueue); | |
1664 | ||
1665 | /* loop until there is nothing left on the lru list. */ | |
1666 | while (list_lru_count(&btp->bt_lru)) { | |
1667 | list_lru_walk(&btp->bt_lru, xfs_buftarg_wait_rele, | |
1668 | &dispose, LONG_MAX); | |
1669 | ||
1670 | while (!list_empty(&dispose)) { | |
1671 | struct xfs_buf *bp; | |
1672 | bp = list_first_entry(&dispose, struct xfs_buf, b_lru); | |
1673 | list_del_init(&bp->b_lru); | |
1674 | if (bp->b_flags & XBF_WRITE_FAIL) { | |
1675 | xfs_alert(btp->bt_mount, | |
1676 | "Corruption Alert: Buffer at block 0x%llx had permanent write failures!", | |
1677 | (long long)bp->b_bn); | |
1678 | xfs_alert(btp->bt_mount, | |
1679 | "Please run xfs_repair to determine the extent of the problem."); | |
1680 | } | |
1681 | xfs_buf_rele(bp); | |
1682 | } | |
1683 | if (loop++ != 0) | |
1684 | delay(100); | |
1685 | } | |
1686 | } | |
1687 | ||
1688 | static enum lru_status | |
1689 | xfs_buftarg_isolate( | |
1690 | struct list_head *item, | |
1691 | struct list_lru_one *lru, | |
1692 | spinlock_t *lru_lock, | |
1693 | void *arg) | |
1694 | { | |
1695 | struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru); | |
1696 | struct list_head *dispose = arg; | |
1697 | ||
1698 | /* | |
1699 | * we are inverting the lru lock/bp->b_lock here, so use a trylock. | |
1700 | * If we fail to get the lock, just skip it. | |
1701 | */ | |
1702 | if (!spin_trylock(&bp->b_lock)) | |
1703 | return LRU_SKIP; | |
1704 | /* | |
1705 | * Decrement the b_lru_ref count unless the value is already | |
1706 | * zero. If the value is already zero, we need to reclaim the | |
1707 | * buffer, otherwise it gets another trip through the LRU. | |
1708 | */ | |
1709 | if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) { | |
1710 | spin_unlock(&bp->b_lock); | |
1711 | return LRU_ROTATE; | |
1712 | } | |
1713 | ||
1714 | bp->b_state |= XFS_BSTATE_DISPOSE; | |
1715 | list_lru_isolate_move(lru, item, dispose); | |
1716 | spin_unlock(&bp->b_lock); | |
1717 | return LRU_REMOVED; | |
1718 | } | |
1719 | ||
1720 | static unsigned long | |
1721 | xfs_buftarg_shrink_scan( | |
1722 | struct shrinker *shrink, | |
1723 | struct shrink_control *sc) | |
1724 | { | |
1725 | struct xfs_buftarg *btp = container_of(shrink, | |
1726 | struct xfs_buftarg, bt_shrinker); | |
1727 | LIST_HEAD(dispose); | |
1728 | unsigned long freed; | |
1729 | ||
1730 | freed = list_lru_shrink_walk(&btp->bt_lru, sc, | |
1731 | xfs_buftarg_isolate, &dispose); | |
1732 | ||
1733 | while (!list_empty(&dispose)) { | |
1734 | struct xfs_buf *bp; | |
1735 | bp = list_first_entry(&dispose, struct xfs_buf, b_lru); | |
1736 | list_del_init(&bp->b_lru); | |
1737 | xfs_buf_rele(bp); | |
1738 | } | |
1739 | ||
1740 | return freed; | |
1741 | } | |
1742 | ||
1743 | static unsigned long | |
1744 | xfs_buftarg_shrink_count( | |
1745 | struct shrinker *shrink, | |
1746 | struct shrink_control *sc) | |
1747 | { | |
1748 | struct xfs_buftarg *btp = container_of(shrink, | |
1749 | struct xfs_buftarg, bt_shrinker); | |
1750 | return list_lru_shrink_count(&btp->bt_lru, sc); | |
1751 | } | |
1752 | ||
1753 | void | |
1754 | xfs_free_buftarg( | |
1755 | struct xfs_mount *mp, | |
1756 | struct xfs_buftarg *btp) | |
1757 | { | |
1758 | unregister_shrinker(&btp->bt_shrinker); | |
1759 | ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0); | |
1760 | percpu_counter_destroy(&btp->bt_io_count); | |
1761 | list_lru_destroy(&btp->bt_lru); | |
1762 | ||
1763 | xfs_blkdev_issue_flush(btp); | |
1764 | ||
1765 | kmem_free(btp); | |
1766 | } | |
1767 | ||
1768 | int | |
1769 | xfs_setsize_buftarg( | |
1770 | xfs_buftarg_t *btp, | |
1771 | unsigned int sectorsize) | |
1772 | { | |
1773 | /* Set up metadata sector size info */ | |
1774 | btp->bt_meta_sectorsize = sectorsize; | |
1775 | btp->bt_meta_sectormask = sectorsize - 1; | |
1776 | ||
1777 | if (set_blocksize(btp->bt_bdev, sectorsize)) { | |
1778 | xfs_warn(btp->bt_mount, | |
1779 | "Cannot set_blocksize to %u on device %pg", | |
1780 | sectorsize, btp->bt_bdev); | |
1781 | return -EINVAL; | |
1782 | } | |
1783 | ||
1784 | /* Set up device logical sector size mask */ | |
1785 | btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev); | |
1786 | btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1; | |
1787 | ||
1788 | return 0; | |
1789 | } | |
1790 | ||
1791 | /* | |
1792 | * When allocating the initial buffer target we have not yet | |
1793 | * read in the superblock, so don't know what sized sectors | |
1794 | * are being used at this early stage. Play safe. | |
1795 | */ | |
1796 | STATIC int | |
1797 | xfs_setsize_buftarg_early( | |
1798 | xfs_buftarg_t *btp, | |
1799 | struct block_device *bdev) | |
1800 | { | |
1801 | return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev)); | |
1802 | } | |
1803 | ||
1804 | xfs_buftarg_t * | |
1805 | xfs_alloc_buftarg( | |
1806 | struct xfs_mount *mp, | |
1807 | struct block_device *bdev, | |
1808 | struct dax_device *dax_dev) | |
1809 | { | |
1810 | xfs_buftarg_t *btp; | |
1811 | ||
1812 | btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS); | |
1813 | ||
1814 | btp->bt_mount = mp; | |
1815 | btp->bt_dev = bdev->bd_dev; | |
1816 | btp->bt_bdev = bdev; | |
1817 | btp->bt_daxdev = dax_dev; | |
1818 | ||
1819 | if (xfs_setsize_buftarg_early(btp, bdev)) | |
1820 | goto error_free; | |
1821 | ||
1822 | if (list_lru_init(&btp->bt_lru)) | |
1823 | goto error_free; | |
1824 | ||
1825 | if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL)) | |
1826 | goto error_lru; | |
1827 | ||
1828 | btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count; | |
1829 | btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan; | |
1830 | btp->bt_shrinker.seeks = DEFAULT_SEEKS; | |
1831 | btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE; | |
1832 | if (register_shrinker(&btp->bt_shrinker)) | |
1833 | goto error_pcpu; | |
1834 | return btp; | |
1835 | ||
1836 | error_pcpu: | |
1837 | percpu_counter_destroy(&btp->bt_io_count); | |
1838 | error_lru: | |
1839 | list_lru_destroy(&btp->bt_lru); | |
1840 | error_free: | |
1841 | kmem_free(btp); | |
1842 | return NULL; | |
1843 | } | |
1844 | ||
1845 | /* | |
1846 | * Cancel a delayed write list. | |
1847 | * | |
1848 | * Remove each buffer from the list, clear the delwri queue flag and drop the | |
1849 | * associated buffer reference. | |
1850 | */ | |
1851 | void | |
1852 | xfs_buf_delwri_cancel( | |
1853 | struct list_head *list) | |
1854 | { | |
1855 | struct xfs_buf *bp; | |
1856 | ||
1857 | while (!list_empty(list)) { | |
1858 | bp = list_first_entry(list, struct xfs_buf, b_list); | |
1859 | ||
1860 | xfs_buf_lock(bp); | |
1861 | bp->b_flags &= ~_XBF_DELWRI_Q; | |
1862 | list_del_init(&bp->b_list); | |
1863 | xfs_buf_relse(bp); | |
1864 | } | |
1865 | } | |
1866 | ||
1867 | /* | |
1868 | * Add a buffer to the delayed write list. | |
1869 | * | |
1870 | * This queues a buffer for writeout if it hasn't already been. Note that | |
1871 | * neither this routine nor the buffer list submission functions perform | |
1872 | * any internal synchronization. It is expected that the lists are thread-local | |
1873 | * to the callers. | |
1874 | * | |
1875 | * Returns true if we queued up the buffer, or false if it already had | |
1876 | * been on the buffer list. | |
1877 | */ | |
1878 | bool | |
1879 | xfs_buf_delwri_queue( | |
1880 | struct xfs_buf *bp, | |
1881 | struct list_head *list) | |
1882 | { | |
1883 | ASSERT(xfs_buf_islocked(bp)); | |
1884 | ASSERT(!(bp->b_flags & XBF_READ)); | |
1885 | ||
1886 | /* | |
1887 | * If the buffer is already marked delwri it already is queued up | |
1888 | * by someone else for imediate writeout. Just ignore it in that | |
1889 | * case. | |
1890 | */ | |
1891 | if (bp->b_flags & _XBF_DELWRI_Q) { | |
1892 | trace_xfs_buf_delwri_queued(bp, _RET_IP_); | |
1893 | return false; | |
1894 | } | |
1895 | ||
1896 | trace_xfs_buf_delwri_queue(bp, _RET_IP_); | |
1897 | ||
1898 | /* | |
1899 | * If a buffer gets written out synchronously or marked stale while it | |
1900 | * is on a delwri list we lazily remove it. To do this, the other party | |
1901 | * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone. | |
1902 | * It remains referenced and on the list. In a rare corner case it | |
1903 | * might get readded to a delwri list after the synchronous writeout, in | |
1904 | * which case we need just need to re-add the flag here. | |
1905 | */ | |
1906 | bp->b_flags |= _XBF_DELWRI_Q; | |
1907 | if (list_empty(&bp->b_list)) { | |
1908 | atomic_inc(&bp->b_hold); | |
1909 | list_add_tail(&bp->b_list, list); | |
1910 | } | |
1911 | ||
1912 | return true; | |
1913 | } | |
1914 | ||
1915 | /* | |
1916 | * Compare function is more complex than it needs to be because | |
1917 | * the return value is only 32 bits and we are doing comparisons | |
1918 | * on 64 bit values | |
1919 | */ | |
1920 | static int | |
1921 | xfs_buf_cmp( | |
1922 | void *priv, | |
1923 | struct list_head *a, | |
1924 | struct list_head *b) | |
1925 | { | |
1926 | struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list); | |
1927 | struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list); | |
1928 | xfs_daddr_t diff; | |
1929 | ||
1930 | diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn; | |
1931 | if (diff < 0) | |
1932 | return -1; | |
1933 | if (diff > 0) | |
1934 | return 1; | |
1935 | return 0; | |
1936 | } | |
1937 | ||
1938 | /* | |
1939 | * submit buffers for write. | |
1940 | * | |
1941 | * When we have a large buffer list, we do not want to hold all the buffers | |
1942 | * locked while we block on the request queue waiting for IO dispatch. To avoid | |
1943 | * this problem, we lock and submit buffers in groups of 50, thereby minimising | |
1944 | * the lock hold times for lists which may contain thousands of objects. | |
1945 | * | |
1946 | * To do this, we sort the buffer list before we walk the list to lock and | |
1947 | * submit buffers, and we plug and unplug around each group of buffers we | |
1948 | * submit. | |
1949 | */ | |
1950 | static int | |
1951 | xfs_buf_delwri_submit_buffers( | |
1952 | struct list_head *buffer_list, | |
1953 | struct list_head *wait_list) | |
1954 | { | |
1955 | struct xfs_buf *bp, *n; | |
1956 | LIST_HEAD (submit_list); | |
1957 | int pinned = 0; | |
1958 | struct blk_plug plug; | |
1959 | ||
1960 | list_sort(NULL, buffer_list, xfs_buf_cmp); | |
1961 | ||
1962 | blk_start_plug(&plug); | |
1963 | list_for_each_entry_safe(bp, n, buffer_list, b_list) { | |
1964 | if (!wait_list) { | |
1965 | if (xfs_buf_ispinned(bp)) { | |
1966 | pinned++; | |
1967 | continue; | |
1968 | } | |
1969 | if (!xfs_buf_trylock(bp)) | |
1970 | continue; | |
1971 | } else { | |
1972 | xfs_buf_lock(bp); | |
1973 | } | |
1974 | ||
1975 | /* | |
1976 | * Someone else might have written the buffer synchronously or | |
1977 | * marked it stale in the meantime. In that case only the | |
1978 | * _XBF_DELWRI_Q flag got cleared, and we have to drop the | |
1979 | * reference and remove it from the list here. | |
1980 | */ | |
1981 | if (!(bp->b_flags & _XBF_DELWRI_Q)) { | |
1982 | list_del_init(&bp->b_list); | |
1983 | xfs_buf_relse(bp); | |
1984 | continue; | |
1985 | } | |
1986 | ||
1987 | trace_xfs_buf_delwri_split(bp, _RET_IP_); | |
1988 | ||
1989 | /* | |
1990 | * We do all IO submission async. This means if we need | |
1991 | * to wait for IO completion we need to take an extra | |
1992 | * reference so the buffer is still valid on the other | |
1993 | * side. We need to move the buffer onto the io_list | |
1994 | * at this point so the caller can still access it. | |
1995 | */ | |
1996 | bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_WRITE_FAIL); | |
1997 | bp->b_flags |= XBF_WRITE | XBF_ASYNC; | |
1998 | if (wait_list) { | |
1999 | xfs_buf_hold(bp); | |
2000 | list_move_tail(&bp->b_list, wait_list); | |
2001 | } else | |
2002 | list_del_init(&bp->b_list); | |
2003 | ||
2004 | xfs_buf_submit(bp); | |
2005 | } | |
2006 | blk_finish_plug(&plug); | |
2007 | ||
2008 | return pinned; | |
2009 | } | |
2010 | ||
2011 | /* | |
2012 | * Write out a buffer list asynchronously. | |
2013 | * | |
2014 | * This will take the @buffer_list, write all non-locked and non-pinned buffers | |
2015 | * out and not wait for I/O completion on any of the buffers. This interface | |
2016 | * is only safely useable for callers that can track I/O completion by higher | |
2017 | * level means, e.g. AIL pushing as the @buffer_list is consumed in this | |
2018 | * function. | |
2019 | */ | |
2020 | int | |
2021 | xfs_buf_delwri_submit_nowait( | |
2022 | struct list_head *buffer_list) | |
2023 | { | |
2024 | return xfs_buf_delwri_submit_buffers(buffer_list, NULL); | |
2025 | } | |
2026 | ||
2027 | /* | |
2028 | * Write out a buffer list synchronously. | |
2029 | * | |
2030 | * This will take the @buffer_list, write all buffers out and wait for I/O | |
2031 | * completion on all of the buffers. @buffer_list is consumed by the function, | |
2032 | * so callers must have some other way of tracking buffers if they require such | |
2033 | * functionality. | |
2034 | */ | |
2035 | int | |
2036 | xfs_buf_delwri_submit( | |
2037 | struct list_head *buffer_list) | |
2038 | { | |
2039 | LIST_HEAD (wait_list); | |
2040 | int error = 0, error2; | |
2041 | struct xfs_buf *bp; | |
2042 | ||
2043 | xfs_buf_delwri_submit_buffers(buffer_list, &wait_list); | |
2044 | ||
2045 | /* Wait for IO to complete. */ | |
2046 | while (!list_empty(&wait_list)) { | |
2047 | bp = list_first_entry(&wait_list, struct xfs_buf, b_list); | |
2048 | ||
2049 | list_del_init(&bp->b_list); | |
2050 | ||
2051 | /* locking the buffer will wait for async IO completion. */ | |
2052 | xfs_buf_lock(bp); | |
2053 | error2 = bp->b_error; | |
2054 | xfs_buf_relse(bp); | |
2055 | if (!error) | |
2056 | error = error2; | |
2057 | } | |
2058 | ||
2059 | return error; | |
2060 | } | |
2061 | ||
2062 | /* | |
2063 | * Push a single buffer on a delwri queue. | |
2064 | * | |
2065 | * The purpose of this function is to submit a single buffer of a delwri queue | |
2066 | * and return with the buffer still on the original queue. The waiting delwri | |
2067 | * buffer submission infrastructure guarantees transfer of the delwri queue | |
2068 | * buffer reference to a temporary wait list. We reuse this infrastructure to | |
2069 | * transfer the buffer back to the original queue. | |
2070 | * | |
2071 | * Note the buffer transitions from the queued state, to the submitted and wait | |
2072 | * listed state and back to the queued state during this call. The buffer | |
2073 | * locking and queue management logic between _delwri_pushbuf() and | |
2074 | * _delwri_queue() guarantee that the buffer cannot be queued to another list | |
2075 | * before returning. | |
2076 | */ | |
2077 | int | |
2078 | xfs_buf_delwri_pushbuf( | |
2079 | struct xfs_buf *bp, | |
2080 | struct list_head *buffer_list) | |
2081 | { | |
2082 | LIST_HEAD (submit_list); | |
2083 | int error; | |
2084 | ||
2085 | ASSERT(bp->b_flags & _XBF_DELWRI_Q); | |
2086 | ||
2087 | trace_xfs_buf_delwri_pushbuf(bp, _RET_IP_); | |
2088 | ||
2089 | /* | |
2090 | * Isolate the buffer to a new local list so we can submit it for I/O | |
2091 | * independently from the rest of the original list. | |
2092 | */ | |
2093 | xfs_buf_lock(bp); | |
2094 | list_move(&bp->b_list, &submit_list); | |
2095 | xfs_buf_unlock(bp); | |
2096 | ||
2097 | /* | |
2098 | * Delwri submission clears the DELWRI_Q buffer flag and returns with | |
2099 | * the buffer on the wait list with an associated reference. Rather than | |
2100 | * bounce the buffer from a local wait list back to the original list | |
2101 | * after I/O completion, reuse the original list as the wait list. | |
2102 | */ | |
2103 | xfs_buf_delwri_submit_buffers(&submit_list, buffer_list); | |
2104 | ||
2105 | /* | |
2106 | * The buffer is now under I/O and wait listed as during typical delwri | |
2107 | * submission. Lock the buffer to wait for I/O completion. Rather than | |
2108 | * remove the buffer from the wait list and release the reference, we | |
2109 | * want to return with the buffer queued to the original list. The | |
2110 | * buffer already sits on the original list with a wait list reference, | |
2111 | * however. If we let the queue inherit that wait list reference, all we | |
2112 | * need to do is reset the DELWRI_Q flag. | |
2113 | */ | |
2114 | xfs_buf_lock(bp); | |
2115 | error = bp->b_error; | |
2116 | bp->b_flags |= _XBF_DELWRI_Q; | |
2117 | xfs_buf_unlock(bp); | |
2118 | ||
2119 | return error; | |
2120 | } | |
2121 | ||
2122 | int __init | |
2123 | xfs_buf_init(void) | |
2124 | { | |
2125 | xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf", | |
2126 | KM_ZONE_HWALIGN, NULL); | |
2127 | if (!xfs_buf_zone) | |
2128 | goto out; | |
2129 | ||
2130 | return 0; | |
2131 | ||
2132 | out: | |
2133 | return -ENOMEM; | |
2134 | } | |
2135 | ||
2136 | void | |
2137 | xfs_buf_terminate(void) | |
2138 | { | |
2139 | kmem_zone_destroy(xfs_buf_zone); | |
2140 | } | |
2141 | ||
2142 | void xfs_buf_set_ref(struct xfs_buf *bp, int lru_ref) | |
2143 | { | |
2144 | /* | |
2145 | * Set the lru reference count to 0 based on the error injection tag. | |
2146 | * This allows userspace to disrupt buffer caching for debug/testing | |
2147 | * purposes. | |
2148 | */ | |
2149 | if (XFS_TEST_ERROR(false, bp->b_target->bt_mount, | |
2150 | XFS_ERRTAG_BUF_LRU_REF)) | |
2151 | lru_ref = 0; | |
2152 | ||
2153 | atomic_set(&bp->b_lru_ref, lru_ref); | |
2154 | } |