1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6 #include "xfs.h"
7 #include <linux/backing-dev.h>
8
9 #include "xfs_shared.h"
10 #include "xfs_format.h"
11 #include "xfs_log_format.h"
12 #include "xfs_trans_resv.h"
13 #include "xfs_mount.h"
14 #include "xfs_trace.h"
15 #include "xfs_log.h"
16 #include "xfs_log_recover.h"
17 #include "xfs_log_priv.h"
18 #include "xfs_trans.h"
19 #include "xfs_buf_item.h"
20 #include "xfs_errortag.h"
21 #include "xfs_error.h"
22 #include "xfs_ag.h"
23
24 static struct kmem_cache *xfs_buf_cache;
25
26 /*
27 * Locking orders
28 *
29 * xfs_buf_ioacct_inc:
30 * xfs_buf_ioacct_dec:
31 * b_sema (caller holds)
32 * b_lock
33 *
34 * xfs_buf_stale:
35 * b_sema (caller holds)
36 * b_lock
37 * lru_lock
38 *
39 * xfs_buf_rele:
40 * b_lock
41 * pag_buf_lock
42 * lru_lock
43 *
44 * xfs_buftarg_drain_rele
45 * lru_lock
46 * b_lock (trylock due to inversion)
47 *
48 * xfs_buftarg_isolate
49 * lru_lock
50 * b_lock (trylock due to inversion)
51 */
52
53 static int __xfs_buf_submit(struct xfs_buf *bp, bool wait);
54
55 static inline int
xfs_buf_submit(struct xfs_buf * bp)56 xfs_buf_submit(
57 struct xfs_buf *bp)
58 {
59 return __xfs_buf_submit(bp, !(bp->b_flags & XBF_ASYNC));
60 }
61
62 static inline int
xfs_buf_is_vmapped(struct xfs_buf * bp)63 xfs_buf_is_vmapped(
64 struct xfs_buf *bp)
65 {
66 /*
67 * Return true if the buffer is vmapped.
68 *
69 * b_addr is null if the buffer is not mapped, but the code is clever
70 * enough to know it doesn't have to map a single page, so the check has
71 * to be both for b_addr and bp->b_page_count > 1.
72 */
73 return bp->b_addr && bp->b_page_count > 1;
74 }
75
76 static inline int
xfs_buf_vmap_len(struct xfs_buf * bp)77 xfs_buf_vmap_len(
78 struct xfs_buf *bp)
79 {
80 return (bp->b_page_count * PAGE_SIZE);
81 }
82
83 /*
84 * Bump the I/O in flight count on the buftarg if we haven't yet done so for
85 * this buffer. The count is incremented once per buffer (per hold cycle)
86 * because the corresponding decrement is deferred to buffer release. Buffers
87 * can undergo I/O multiple times in a hold-release cycle and per buffer I/O
88 * tracking adds unnecessary overhead. This is used for sychronization purposes
89 * with unmount (see xfs_buftarg_drain()), so all we really need is a count of
90 * in-flight buffers.
91 *
92 * Buffers that are never released (e.g., superblock, iclog buffers) must set
93 * the XBF_NO_IOACCT flag before I/O submission. Otherwise, the buftarg count
94 * never reaches zero and unmount hangs indefinitely.
95 */
96 static inline void
xfs_buf_ioacct_inc(struct xfs_buf * bp)97 xfs_buf_ioacct_inc(
98 struct xfs_buf *bp)
99 {
100 if (bp->b_flags & XBF_NO_IOACCT)
101 return;
102
103 ASSERT(bp->b_flags & XBF_ASYNC);
104 spin_lock(&bp->b_lock);
105 if (!(bp->b_state & XFS_BSTATE_IN_FLIGHT)) {
106 bp->b_state |= XFS_BSTATE_IN_FLIGHT;
107 percpu_counter_inc(&bp->b_target->bt_io_count);
108 }
109 spin_unlock(&bp->b_lock);
110 }
111
112 /*
113 * Clear the in-flight state on a buffer about to be released to the LRU or
114 * freed and unaccount from the buftarg.
115 */
116 static inline void
__xfs_buf_ioacct_dec(struct xfs_buf * bp)117 __xfs_buf_ioacct_dec(
118 struct xfs_buf *bp)
119 {
120 lockdep_assert_held(&bp->b_lock);
121
122 if (bp->b_state & XFS_BSTATE_IN_FLIGHT) {
123 bp->b_state &= ~XFS_BSTATE_IN_FLIGHT;
124 percpu_counter_dec(&bp->b_target->bt_io_count);
125 }
126 }
127
128 static inline void
xfs_buf_ioacct_dec(struct xfs_buf * bp)129 xfs_buf_ioacct_dec(
130 struct xfs_buf *bp)
131 {
132 spin_lock(&bp->b_lock);
133 __xfs_buf_ioacct_dec(bp);
134 spin_unlock(&bp->b_lock);
135 }
136
137 /*
138 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
139 * b_lru_ref count so that the buffer is freed immediately when the buffer
140 * reference count falls to zero. If the buffer is already on the LRU, we need
141 * to remove the reference that LRU holds on the buffer.
142 *
143 * This prevents build-up of stale buffers on the LRU.
144 */
145 void
xfs_buf_stale(struct xfs_buf * bp)146 xfs_buf_stale(
147 struct xfs_buf *bp)
148 {
149 ASSERT(xfs_buf_islocked(bp));
150
151 bp->b_flags |= XBF_STALE;
152
153 /*
154 * Clear the delwri status so that a delwri queue walker will not
155 * flush this buffer to disk now that it is stale. The delwri queue has
156 * a reference to the buffer, so this is safe to do.
157 */
158 bp->b_flags &= ~_XBF_DELWRI_Q;
159
160 /*
161 * Once the buffer is marked stale and unlocked, a subsequent lookup
162 * could reset b_flags. There is no guarantee that the buffer is
163 * unaccounted (released to LRU) before that occurs. Drop in-flight
164 * status now to preserve accounting consistency.
165 */
166 spin_lock(&bp->b_lock);
167 __xfs_buf_ioacct_dec(bp);
168
169 atomic_set(&bp->b_lru_ref, 0);
170 if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
171 (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru)))
172 atomic_dec(&bp->b_hold);
173
174 ASSERT(atomic_read(&bp->b_hold) >= 1);
175 spin_unlock(&bp->b_lock);
176 }
177
178 static int
xfs_buf_get_maps(struct xfs_buf * bp,int map_count)179 xfs_buf_get_maps(
180 struct xfs_buf *bp,
181 int map_count)
182 {
183 ASSERT(bp->b_maps == NULL);
184 bp->b_map_count = map_count;
185
186 if (map_count == 1) {
187 bp->b_maps = &bp->__b_map;
188 return 0;
189 }
190
191 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
192 KM_NOFS);
193 if (!bp->b_maps)
194 return -ENOMEM;
195 return 0;
196 }
197
198 /*
199 * Frees b_pages if it was allocated.
200 */
201 static void
xfs_buf_free_maps(struct xfs_buf * bp)202 xfs_buf_free_maps(
203 struct xfs_buf *bp)
204 {
205 if (bp->b_maps != &bp->__b_map) {
206 kmem_free(bp->b_maps);
207 bp->b_maps = NULL;
208 }
209 }
210
211 static int
_xfs_buf_alloc(struct xfs_buftarg * target,struct xfs_buf_map * map,int nmaps,xfs_buf_flags_t flags,struct xfs_buf ** bpp)212 _xfs_buf_alloc(
213 struct xfs_buftarg *target,
214 struct xfs_buf_map *map,
215 int nmaps,
216 xfs_buf_flags_t flags,
217 struct xfs_buf **bpp)
218 {
219 struct xfs_buf *bp;
220 int error;
221 int i;
222
223 *bpp = NULL;
224 bp = kmem_cache_zalloc(xfs_buf_cache, GFP_NOFS | __GFP_NOFAIL);
225
226 /*
227 * We don't want certain flags to appear in b_flags unless they are
228 * specifically set by later operations on the buffer.
229 */
230 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
231
232 atomic_set(&bp->b_hold, 1);
233 atomic_set(&bp->b_lru_ref, 1);
234 init_completion(&bp->b_iowait);
235 INIT_LIST_HEAD(&bp->b_lru);
236 INIT_LIST_HEAD(&bp->b_list);
237 INIT_LIST_HEAD(&bp->b_li_list);
238 sema_init(&bp->b_sema, 0); /* held, no waiters */
239 spin_lock_init(&bp->b_lock);
240 bp->b_target = target;
241 bp->b_mount = target->bt_mount;
242 bp->b_flags = flags;
243
244 /*
245 * Set length and io_length to the same value initially.
246 * I/O routines should use io_length, which will be the same in
247 * most cases but may be reset (e.g. XFS recovery).
248 */
249 error = xfs_buf_get_maps(bp, nmaps);
250 if (error) {
251 kmem_cache_free(xfs_buf_cache, bp);
252 return error;
253 }
254
255 bp->b_rhash_key = map[0].bm_bn;
256 bp->b_length = 0;
257 for (i = 0; i < nmaps; i++) {
258 bp->b_maps[i].bm_bn = map[i].bm_bn;
259 bp->b_maps[i].bm_len = map[i].bm_len;
260 bp->b_length += map[i].bm_len;
261 }
262
263 atomic_set(&bp->b_pin_count, 0);
264 init_waitqueue_head(&bp->b_waiters);
265
266 XFS_STATS_INC(bp->b_mount, xb_create);
267 trace_xfs_buf_init(bp, _RET_IP_);
268
269 *bpp = bp;
270 return 0;
271 }
272
273 static void
xfs_buf_free_pages(struct xfs_buf * bp)274 xfs_buf_free_pages(
275 struct xfs_buf *bp)
276 {
277 uint i;
278
279 ASSERT(bp->b_flags & _XBF_PAGES);
280
281 if (xfs_buf_is_vmapped(bp))
282 vm_unmap_ram(bp->b_addr, bp->b_page_count);
283
284 for (i = 0; i < bp->b_page_count; i++) {
285 if (bp->b_pages[i])
286 __free_page(bp->b_pages[i]);
287 }
288 if (current->reclaim_state)
289 current->reclaim_state->reclaimed_slab += bp->b_page_count;
290
291 if (bp->b_pages != bp->b_page_array)
292 kmem_free(bp->b_pages);
293 bp->b_pages = NULL;
294 bp->b_flags &= ~_XBF_PAGES;
295 }
296
297 static void
xfs_buf_free(struct xfs_buf * bp)298 xfs_buf_free(
299 struct xfs_buf *bp)
300 {
301 trace_xfs_buf_free(bp, _RET_IP_);
302
303 ASSERT(list_empty(&bp->b_lru));
304
305 if (bp->b_flags & _XBF_PAGES)
306 xfs_buf_free_pages(bp);
307 else if (bp->b_flags & _XBF_KMEM)
308 kmem_free(bp->b_addr);
309
310 xfs_buf_free_maps(bp);
311 kmem_cache_free(xfs_buf_cache, bp);
312 }
313
314 static int
xfs_buf_alloc_kmem(struct xfs_buf * bp,xfs_buf_flags_t flags)315 xfs_buf_alloc_kmem(
316 struct xfs_buf *bp,
317 xfs_buf_flags_t flags)
318 {
319 xfs_km_flags_t kmflag_mask = KM_NOFS;
320 size_t size = BBTOB(bp->b_length);
321
322 /* Assure zeroed buffer for non-read cases. */
323 if (!(flags & XBF_READ))
324 kmflag_mask |= KM_ZERO;
325
326 bp->b_addr = kmem_alloc(size, kmflag_mask);
327 if (!bp->b_addr)
328 return -ENOMEM;
329
330 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
331 ((unsigned long)bp->b_addr & PAGE_MASK)) {
332 /* b_addr spans two pages - use alloc_page instead */
333 kmem_free(bp->b_addr);
334 bp->b_addr = NULL;
335 return -ENOMEM;
336 }
337 bp->b_offset = offset_in_page(bp->b_addr);
338 bp->b_pages = bp->b_page_array;
339 bp->b_pages[0] = kmem_to_page(bp->b_addr);
340 bp->b_page_count = 1;
341 bp->b_flags |= _XBF_KMEM;
342 return 0;
343 }
344
345 static int
xfs_buf_alloc_pages(struct xfs_buf * bp,xfs_buf_flags_t flags)346 xfs_buf_alloc_pages(
347 struct xfs_buf *bp,
348 xfs_buf_flags_t flags)
349 {
350 gfp_t gfp_mask = __GFP_NOWARN;
351 long filled = 0;
352
353 if (flags & XBF_READ_AHEAD)
354 gfp_mask |= __GFP_NORETRY;
355 else
356 gfp_mask |= GFP_NOFS;
357
358 /* Make sure that we have a page list */
359 bp->b_page_count = DIV_ROUND_UP(BBTOB(bp->b_length), PAGE_SIZE);
360 if (bp->b_page_count <= XB_PAGES) {
361 bp->b_pages = bp->b_page_array;
362 } else {
363 bp->b_pages = kzalloc(sizeof(struct page *) * bp->b_page_count,
364 gfp_mask);
365 if (!bp->b_pages)
366 return -ENOMEM;
367 }
368 bp->b_flags |= _XBF_PAGES;
369
370 /* Assure zeroed buffer for non-read cases. */
371 if (!(flags & XBF_READ))
372 gfp_mask |= __GFP_ZERO;
373
374 /*
375 * Bulk filling of pages can take multiple calls. Not filling the entire
376 * array is not an allocation failure, so don't back off if we get at
377 * least one extra page.
378 */
379 for (;;) {
380 long last = filled;
381
382 filled = alloc_pages_bulk_array(gfp_mask, bp->b_page_count,
383 bp->b_pages);
384 if (filled == bp->b_page_count) {
385 XFS_STATS_INC(bp->b_mount, xb_page_found);
386 break;
387 }
388
389 if (filled != last)
390 continue;
391
392 if (flags & XBF_READ_AHEAD) {
393 xfs_buf_free_pages(bp);
394 return -ENOMEM;
395 }
396
397 XFS_STATS_INC(bp->b_mount, xb_page_retries);
398 memalloc_retry_wait(gfp_mask);
399 }
400 return 0;
401 }
402
403 /*
404 * Map buffer into kernel address-space if necessary.
405 */
406 STATIC int
_xfs_buf_map_pages(struct xfs_buf * bp,xfs_buf_flags_t flags)407 _xfs_buf_map_pages(
408 struct xfs_buf *bp,
409 xfs_buf_flags_t flags)
410 {
411 ASSERT(bp->b_flags & _XBF_PAGES);
412 if (bp->b_page_count == 1) {
413 /* A single page buffer is always mappable */
414 bp->b_addr = page_address(bp->b_pages[0]);
415 } else if (flags & XBF_UNMAPPED) {
416 bp->b_addr = NULL;
417 } else {
418 int retried = 0;
419 unsigned nofs_flag;
420
421 /*
422 * vm_map_ram() will allocate auxiliary structures (e.g.
423 * pagetables) with GFP_KERNEL, yet we are likely to be under
424 * GFP_NOFS context here. Hence we need to tell memory reclaim
425 * that we are in such a context via PF_MEMALLOC_NOFS to prevent
426 * memory reclaim re-entering the filesystem here and
427 * potentially deadlocking.
428 */
429 nofs_flag = memalloc_nofs_save();
430 do {
431 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
432 -1);
433 if (bp->b_addr)
434 break;
435 vm_unmap_aliases();
436 } while (retried++ <= 1);
437 memalloc_nofs_restore(nofs_flag);
438
439 if (!bp->b_addr)
440 return -ENOMEM;
441 }
442
443 return 0;
444 }
445
446 /*
447 * Finding and Reading Buffers
448 */
449 static int
_xfs_buf_obj_cmp(struct rhashtable_compare_arg * arg,const void * obj)450 _xfs_buf_obj_cmp(
451 struct rhashtable_compare_arg *arg,
452 const void *obj)
453 {
454 const struct xfs_buf_map *map = arg->key;
455 const struct xfs_buf *bp = obj;
456
457 /*
458 * The key hashing in the lookup path depends on the key being the
459 * first element of the compare_arg, make sure to assert this.
460 */
461 BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0);
462
463 if (bp->b_rhash_key != map->bm_bn)
464 return 1;
465
466 if (unlikely(bp->b_length != map->bm_len)) {
467 /*
468 * found a block number match. If the range doesn't
469 * match, the only way this is allowed is if the buffer
470 * in the cache is stale and the transaction that made
471 * it stale has not yet committed. i.e. we are
472 * reallocating a busy extent. Skip this buffer and
473 * continue searching for an exact match.
474 */
475 ASSERT(bp->b_flags & XBF_STALE);
476 return 1;
477 }
478 return 0;
479 }
480
481 static const struct rhashtable_params xfs_buf_hash_params = {
482 .min_size = 32, /* empty AGs have minimal footprint */
483 .nelem_hint = 16,
484 .key_len = sizeof(xfs_daddr_t),
485 .key_offset = offsetof(struct xfs_buf, b_rhash_key),
486 .head_offset = offsetof(struct xfs_buf, b_rhash_head),
487 .automatic_shrinking = true,
488 .obj_cmpfn = _xfs_buf_obj_cmp,
489 };
490
491 int
xfs_buf_hash_init(struct xfs_perag * pag)492 xfs_buf_hash_init(
493 struct xfs_perag *pag)
494 {
495 spin_lock_init(&pag->pag_buf_lock);
496 return rhashtable_init(&pag->pag_buf_hash, &xfs_buf_hash_params);
497 }
498
499 void
xfs_buf_hash_destroy(struct xfs_perag * pag)500 xfs_buf_hash_destroy(
501 struct xfs_perag *pag)
502 {
503 rhashtable_destroy(&pag->pag_buf_hash);
504 }
505
506 /*
507 * Look up a buffer in the buffer cache and return it referenced and locked
508 * in @found_bp.
509 *
510 * If @new_bp is supplied and we have a lookup miss, insert @new_bp into the
511 * cache.
512 *
513 * If XBF_TRYLOCK is set in @flags, only try to lock the buffer and return
514 * -EAGAIN if we fail to lock it.
515 *
516 * Return values are:
517 * -EFSCORRUPTED if have been supplied with an invalid address
518 * -EAGAIN on trylock failure
519 * -ENOENT if we fail to find a match and @new_bp was NULL
520 * 0, with @found_bp:
521 * - @new_bp if we inserted it into the cache
522 * - the buffer we found and locked.
523 */
524 static int
xfs_buf_find(struct xfs_buftarg * btp,struct xfs_buf_map * map,int nmaps,xfs_buf_flags_t flags,struct xfs_buf * new_bp,struct xfs_buf ** found_bp)525 xfs_buf_find(
526 struct xfs_buftarg *btp,
527 struct xfs_buf_map *map,
528 int nmaps,
529 xfs_buf_flags_t flags,
530 struct xfs_buf *new_bp,
531 struct xfs_buf **found_bp)
532 {
533 struct xfs_perag *pag;
534 struct xfs_buf *bp;
535 struct xfs_buf_map cmap = { .bm_bn = map[0].bm_bn };
536 xfs_daddr_t eofs;
537 int i;
538
539 *found_bp = NULL;
540
541 for (i = 0; i < nmaps; i++)
542 cmap.bm_len += map[i].bm_len;
543
544 /* Check for IOs smaller than the sector size / not sector aligned */
545 ASSERT(!(BBTOB(cmap.bm_len) < btp->bt_meta_sectorsize));
546 ASSERT(!(BBTOB(cmap.bm_bn) & (xfs_off_t)btp->bt_meta_sectormask));
547
548 /*
549 * Corrupted block numbers can get through to here, unfortunately, so we
550 * have to check that the buffer falls within the filesystem bounds.
551 */
552 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
553 if (cmap.bm_bn < 0 || cmap.bm_bn >= eofs) {
554 xfs_alert(btp->bt_mount,
555 "%s: daddr 0x%llx out of range, EOFS 0x%llx",
556 __func__, cmap.bm_bn, eofs);
557 WARN_ON(1);
558 return -EFSCORRUPTED;
559 }
560
561 pag = xfs_perag_get(btp->bt_mount,
562 xfs_daddr_to_agno(btp->bt_mount, cmap.bm_bn));
563
564 spin_lock(&pag->pag_buf_lock);
565 bp = rhashtable_lookup_fast(&pag->pag_buf_hash, &cmap,
566 xfs_buf_hash_params);
567 if (bp) {
568 atomic_inc(&bp->b_hold);
569 goto found;
570 }
571
572 /* No match found */
573 if (!new_bp) {
574 XFS_STATS_INC(btp->bt_mount, xb_miss_locked);
575 spin_unlock(&pag->pag_buf_lock);
576 xfs_perag_put(pag);
577 return -ENOENT;
578 }
579
580 /* the buffer keeps the perag reference until it is freed */
581 new_bp->b_pag = pag;
582 rhashtable_insert_fast(&pag->pag_buf_hash, &new_bp->b_rhash_head,
583 xfs_buf_hash_params);
584 spin_unlock(&pag->pag_buf_lock);
585 *found_bp = new_bp;
586 return 0;
587
588 found:
589 spin_unlock(&pag->pag_buf_lock);
590 xfs_perag_put(pag);
591
592 if (!xfs_buf_trylock(bp)) {
593 if (flags & XBF_TRYLOCK) {
594 xfs_buf_rele(bp);
595 XFS_STATS_INC(btp->bt_mount, xb_busy_locked);
596 return -EAGAIN;
597 }
598 xfs_buf_lock(bp);
599 XFS_STATS_INC(btp->bt_mount, xb_get_locked_waited);
600 }
601
602 /*
603 * if the buffer is stale, clear all the external state associated with
604 * it. We need to keep flags such as how we allocated the buffer memory
605 * intact here.
606 */
607 if (bp->b_flags & XBF_STALE) {
608 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
609 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
610 bp->b_ops = NULL;
611 }
612
613 trace_xfs_buf_find(bp, flags, _RET_IP_);
614 XFS_STATS_INC(btp->bt_mount, xb_get_locked);
615 *found_bp = bp;
616 return 0;
617 }
618
619 struct xfs_buf *
xfs_buf_incore(struct xfs_buftarg * target,xfs_daddr_t blkno,size_t numblks,xfs_buf_flags_t flags)620 xfs_buf_incore(
621 struct xfs_buftarg *target,
622 xfs_daddr_t blkno,
623 size_t numblks,
624 xfs_buf_flags_t flags)
625 {
626 struct xfs_buf *bp;
627 int error;
628 DEFINE_SINGLE_BUF_MAP(map, blkno, numblks);
629
630 error = xfs_buf_find(target, &map, 1, flags, NULL, &bp);
631 if (error)
632 return NULL;
633 return bp;
634 }
635
636 /*
637 * Assembles a buffer covering the specified range. The code is optimised for
638 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
639 * more hits than misses.
640 */
641 int
xfs_buf_get_map(struct xfs_buftarg * target,struct xfs_buf_map * map,int nmaps,xfs_buf_flags_t flags,struct xfs_buf ** bpp)642 xfs_buf_get_map(
643 struct xfs_buftarg *target,
644 struct xfs_buf_map *map,
645 int nmaps,
646 xfs_buf_flags_t flags,
647 struct xfs_buf **bpp)
648 {
649 struct xfs_buf *bp;
650 struct xfs_buf *new_bp;
651 int error;
652
653 *bpp = NULL;
654 error = xfs_buf_find(target, map, nmaps, flags, NULL, &bp);
655 if (!error)
656 goto found;
657 if (error != -ENOENT)
658 return error;
659
660 error = _xfs_buf_alloc(target, map, nmaps, flags, &new_bp);
661 if (error)
662 return error;
663
664 /*
665 * For buffers that fit entirely within a single page, first attempt to
666 * allocate the memory from the heap to minimise memory usage. If we
667 * can't get heap memory for these small buffers, we fall back to using
668 * the page allocator.
669 */
670 if (BBTOB(new_bp->b_length) >= PAGE_SIZE ||
671 xfs_buf_alloc_kmem(new_bp, flags) < 0) {
672 error = xfs_buf_alloc_pages(new_bp, flags);
673 if (error)
674 goto out_free_buf;
675 }
676
677 error = xfs_buf_find(target, map, nmaps, flags, new_bp, &bp);
678 if (error)
679 goto out_free_buf;
680
681 if (bp != new_bp)
682 xfs_buf_free(new_bp);
683
684 found:
685 if (!bp->b_addr) {
686 error = _xfs_buf_map_pages(bp, flags);
687 if (unlikely(error)) {
688 xfs_warn_ratelimited(target->bt_mount,
689 "%s: failed to map %u pages", __func__,
690 bp->b_page_count);
691 xfs_buf_relse(bp);
692 return error;
693 }
694 }
695
696 /*
697 * Clear b_error if this is a lookup from a caller that doesn't expect
698 * valid data to be found in the buffer.
699 */
700 if (!(flags & XBF_READ))
701 xfs_buf_ioerror(bp, 0);
702
703 XFS_STATS_INC(target->bt_mount, xb_get);
704 trace_xfs_buf_get(bp, flags, _RET_IP_);
705 *bpp = bp;
706 return 0;
707 out_free_buf:
708 xfs_buf_free(new_bp);
709 return error;
710 }
711
712 int
_xfs_buf_read(struct xfs_buf * bp,xfs_buf_flags_t flags)713 _xfs_buf_read(
714 struct xfs_buf *bp,
715 xfs_buf_flags_t flags)
716 {
717 ASSERT(!(flags & XBF_WRITE));
718 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
719
720 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD | XBF_DONE);
721 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
722
723 return xfs_buf_submit(bp);
724 }
725
726 /*
727 * Reverify a buffer found in cache without an attached ->b_ops.
728 *
729 * If the caller passed an ops structure and the buffer doesn't have ops
730 * assigned, set the ops and use it to verify the contents. If verification
731 * fails, clear XBF_DONE. We assume the buffer has no recorded errors and is
732 * already in XBF_DONE state on entry.
733 *
734 * Under normal operations, every in-core buffer is verified on read I/O
735 * completion. There are two scenarios that can lead to in-core buffers without
736 * an assigned ->b_ops. The first is during log recovery of buffers on a V4
737 * filesystem, though these buffers are purged at the end of recovery. The
738 * other is online repair, which intentionally reads with a NULL buffer ops to
739 * run several verifiers across an in-core buffer in order to establish buffer
740 * type. If repair can't establish that, the buffer will be left in memory
741 * with NULL buffer ops.
742 */
743 int
xfs_buf_reverify(struct xfs_buf * bp,const struct xfs_buf_ops * ops)744 xfs_buf_reverify(
745 struct xfs_buf *bp,
746 const struct xfs_buf_ops *ops)
747 {
748 ASSERT(bp->b_flags & XBF_DONE);
749 ASSERT(bp->b_error == 0);
750
751 if (!ops || bp->b_ops)
752 return 0;
753
754 bp->b_ops = ops;
755 bp->b_ops->verify_read(bp);
756 if (bp->b_error)
757 bp->b_flags &= ~XBF_DONE;
758 return bp->b_error;
759 }
760
761 int
xfs_buf_read_map(struct xfs_buftarg * target,struct xfs_buf_map * map,int nmaps,xfs_buf_flags_t flags,struct xfs_buf ** bpp,const struct xfs_buf_ops * ops,xfs_failaddr_t fa)762 xfs_buf_read_map(
763 struct xfs_buftarg *target,
764 struct xfs_buf_map *map,
765 int nmaps,
766 xfs_buf_flags_t flags,
767 struct xfs_buf **bpp,
768 const struct xfs_buf_ops *ops,
769 xfs_failaddr_t fa)
770 {
771 struct xfs_buf *bp;
772 int error;
773
774 flags |= XBF_READ;
775 *bpp = NULL;
776
777 error = xfs_buf_get_map(target, map, nmaps, flags, &bp);
778 if (error)
779 return error;
780
781 trace_xfs_buf_read(bp, flags, _RET_IP_);
782
783 if (!(bp->b_flags & XBF_DONE)) {
784 /* Initiate the buffer read and wait. */
785 XFS_STATS_INC(target->bt_mount, xb_get_read);
786 bp->b_ops = ops;
787 error = _xfs_buf_read(bp, flags);
788
789 /* Readahead iodone already dropped the buffer, so exit. */
790 if (flags & XBF_ASYNC)
791 return 0;
792 } else {
793 /* Buffer already read; all we need to do is check it. */
794 error = xfs_buf_reverify(bp, ops);
795
796 /* Readahead already finished; drop the buffer and exit. */
797 if (flags & XBF_ASYNC) {
798 xfs_buf_relse(bp);
799 return 0;
800 }
801
802 /* We do not want read in the flags */
803 bp->b_flags &= ~XBF_READ;
804 ASSERT(bp->b_ops != NULL || ops == NULL);
805 }
806
807 /*
808 * If we've had a read error, then the contents of the buffer are
809 * invalid and should not be used. To ensure that a followup read tries
810 * to pull the buffer from disk again, we clear the XBF_DONE flag and
811 * mark the buffer stale. This ensures that anyone who has a current
812 * reference to the buffer will interpret it's contents correctly and
813 * future cache lookups will also treat it as an empty, uninitialised
814 * buffer.
815 */
816 if (error) {
817 /*
818 * Check against log shutdown for error reporting because
819 * metadata writeback may require a read first and we need to
820 * report errors in metadata writeback until the log is shut
821 * down. High level transaction read functions already check
822 * against mount shutdown, anyway, so we only need to be
823 * concerned about low level IO interactions here.
824 */
825 if (!xlog_is_shutdown(target->bt_mount->m_log))
826 xfs_buf_ioerror_alert(bp, fa);
827
828 bp->b_flags &= ~XBF_DONE;
829 xfs_buf_stale(bp);
830 xfs_buf_relse(bp);
831
832 /* bad CRC means corrupted metadata */
833 if (error == -EFSBADCRC)
834 error = -EFSCORRUPTED;
835 return error;
836 }
837
838 *bpp = bp;
839 return 0;
840 }
841
842 /*
843 * If we are not low on memory then do the readahead in a deadlock
844 * safe manner.
845 */
846 void
xfs_buf_readahead_map(struct xfs_buftarg * target,struct xfs_buf_map * map,int nmaps,const struct xfs_buf_ops * ops)847 xfs_buf_readahead_map(
848 struct xfs_buftarg *target,
849 struct xfs_buf_map *map,
850 int nmaps,
851 const struct xfs_buf_ops *ops)
852 {
853 struct xfs_buf *bp;
854
855 xfs_buf_read_map(target, map, nmaps,
856 XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD, &bp, ops,
857 __this_address);
858 }
859
860 /*
861 * Read an uncached buffer from disk. Allocates and returns a locked
862 * buffer containing the disk contents or nothing. Uncached buffers always have
863 * a cache index of XFS_BUF_DADDR_NULL so we can easily determine if the buffer
864 * is cached or uncached during fault diagnosis.
865 */
866 int
xfs_buf_read_uncached(struct xfs_buftarg * target,xfs_daddr_t daddr,size_t numblks,xfs_buf_flags_t flags,struct xfs_buf ** bpp,const struct xfs_buf_ops * ops)867 xfs_buf_read_uncached(
868 struct xfs_buftarg *target,
869 xfs_daddr_t daddr,
870 size_t numblks,
871 xfs_buf_flags_t flags,
872 struct xfs_buf **bpp,
873 const struct xfs_buf_ops *ops)
874 {
875 struct xfs_buf *bp;
876 int error;
877
878 *bpp = NULL;
879
880 error = xfs_buf_get_uncached(target, numblks, flags, &bp);
881 if (error)
882 return error;
883
884 /* set up the buffer for a read IO */
885 ASSERT(bp->b_map_count == 1);
886 bp->b_rhash_key = XFS_BUF_DADDR_NULL;
887 bp->b_maps[0].bm_bn = daddr;
888 bp->b_flags |= XBF_READ;
889 bp->b_ops = ops;
890
891 xfs_buf_submit(bp);
892 if (bp->b_error) {
893 error = bp->b_error;
894 xfs_buf_relse(bp);
895 return error;
896 }
897
898 *bpp = bp;
899 return 0;
900 }
901
902 int
xfs_buf_get_uncached(struct xfs_buftarg * target,size_t numblks,xfs_buf_flags_t flags,struct xfs_buf ** bpp)903 xfs_buf_get_uncached(
904 struct xfs_buftarg *target,
905 size_t numblks,
906 xfs_buf_flags_t flags,
907 struct xfs_buf **bpp)
908 {
909 int error;
910 struct xfs_buf *bp;
911 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
912
913 *bpp = NULL;
914
915 /* flags might contain irrelevant bits, pass only what we care about */
916 error = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT, &bp);
917 if (error)
918 return error;
919
920 error = xfs_buf_alloc_pages(bp, flags);
921 if (error)
922 goto fail_free_buf;
923
924 error = _xfs_buf_map_pages(bp, 0);
925 if (unlikely(error)) {
926 xfs_warn(target->bt_mount,
927 "%s: failed to map pages", __func__);
928 goto fail_free_buf;
929 }
930
931 trace_xfs_buf_get_uncached(bp, _RET_IP_);
932 *bpp = bp;
933 return 0;
934
935 fail_free_buf:
936 xfs_buf_free(bp);
937 return error;
938 }
939
940 /*
941 * Increment reference count on buffer, to hold the buffer concurrently
942 * with another thread which may release (free) the buffer asynchronously.
943 * Must hold the buffer already to call this function.
944 */
945 void
xfs_buf_hold(struct xfs_buf * bp)946 xfs_buf_hold(
947 struct xfs_buf *bp)
948 {
949 trace_xfs_buf_hold(bp, _RET_IP_);
950 atomic_inc(&bp->b_hold);
951 }
952
953 /*
954 * Release a hold on the specified buffer. If the hold count is 1, the buffer is
955 * placed on LRU or freed (depending on b_lru_ref).
956 */
957 void
xfs_buf_rele(struct xfs_buf * bp)958 xfs_buf_rele(
959 struct xfs_buf *bp)
960 {
961 struct xfs_perag *pag = bp->b_pag;
962 bool release;
963 bool freebuf = false;
964
965 trace_xfs_buf_rele(bp, _RET_IP_);
966
967 if (!pag) {
968 ASSERT(list_empty(&bp->b_lru));
969 if (atomic_dec_and_test(&bp->b_hold)) {
970 xfs_buf_ioacct_dec(bp);
971 xfs_buf_free(bp);
972 }
973 return;
974 }
975
976 ASSERT(atomic_read(&bp->b_hold) > 0);
977
978 /*
979 * We grab the b_lock here first to serialise racing xfs_buf_rele()
980 * calls. The pag_buf_lock being taken on the last reference only
981 * serialises against racing lookups in xfs_buf_find(). IOWs, the second
982 * to last reference we drop here is not serialised against the last
983 * reference until we take bp->b_lock. Hence if we don't grab b_lock
984 * first, the last "release" reference can win the race to the lock and
985 * free the buffer before the second-to-last reference is processed,
986 * leading to a use-after-free scenario.
987 */
988 spin_lock(&bp->b_lock);
989 release = atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_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
xfs_buf_trylock(struct xfs_buf * bp)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 trace_xfs_buf_trylock(bp, _RET_IP_);
1064 else
1065 trace_xfs_buf_trylock_fail(bp, _RET_IP_);
1066 return locked;
1067 }
1068
1069 /*
1070 * Lock a buffer object.
1071 *
1072 * If we come across a stale, pinned, locked buffer, we know that we
1073 * are being asked to lock a buffer that has been reallocated. Because
1074 * it is pinned, we know that the log has not been pushed to disk and
1075 * hence it will still be locked. Rather than sleeping until someone
1076 * else pushes the log, push it ourselves before trying to get the lock.
1077 */
1078 void
xfs_buf_lock(struct xfs_buf * bp)1079 xfs_buf_lock(
1080 struct xfs_buf *bp)
1081 {
1082 trace_xfs_buf_lock(bp, _RET_IP_);
1083
1084 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
1085 xfs_log_force(bp->b_mount, 0);
1086 down(&bp->b_sema);
1087
1088 trace_xfs_buf_lock_done(bp, _RET_IP_);
1089 }
1090
1091 void
xfs_buf_unlock(struct xfs_buf * bp)1092 xfs_buf_unlock(
1093 struct xfs_buf *bp)
1094 {
1095 ASSERT(xfs_buf_islocked(bp));
1096
1097 up(&bp->b_sema);
1098 trace_xfs_buf_unlock(bp, _RET_IP_);
1099 }
1100
1101 STATIC void
xfs_buf_wait_unpin(struct xfs_buf * bp)1102 xfs_buf_wait_unpin(
1103 struct xfs_buf *bp)
1104 {
1105 DECLARE_WAITQUEUE (wait, current);
1106
1107 if (atomic_read(&bp->b_pin_count) == 0)
1108 return;
1109
1110 add_wait_queue(&bp->b_waiters, &wait);
1111 for (;;) {
1112 set_current_state(TASK_UNINTERRUPTIBLE);
1113 if (atomic_read(&bp->b_pin_count) == 0)
1114 break;
1115 io_schedule();
1116 }
1117 remove_wait_queue(&bp->b_waiters, &wait);
1118 set_current_state(TASK_RUNNING);
1119 }
1120
1121 static void
xfs_buf_ioerror_alert_ratelimited(struct xfs_buf * bp)1122 xfs_buf_ioerror_alert_ratelimited(
1123 struct xfs_buf *bp)
1124 {
1125 static unsigned long lasttime;
1126 static struct xfs_buftarg *lasttarg;
1127
1128 if (bp->b_target != lasttarg ||
1129 time_after(jiffies, (lasttime + 5*HZ))) {
1130 lasttime = jiffies;
1131 xfs_buf_ioerror_alert(bp, __this_address);
1132 }
1133 lasttarg = bp->b_target;
1134 }
1135
1136 /*
1137 * Account for this latest trip around the retry handler, and decide if
1138 * we've failed enough times to constitute a permanent failure.
1139 */
1140 static bool
xfs_buf_ioerror_permanent(struct xfs_buf * bp,struct xfs_error_cfg * cfg)1141 xfs_buf_ioerror_permanent(
1142 struct xfs_buf *bp,
1143 struct xfs_error_cfg *cfg)
1144 {
1145 struct xfs_mount *mp = bp->b_mount;
1146
1147 if (cfg->max_retries != XFS_ERR_RETRY_FOREVER &&
1148 ++bp->b_retries > cfg->max_retries)
1149 return true;
1150 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1151 time_after(jiffies, cfg->retry_timeout + bp->b_first_retry_time))
1152 return true;
1153
1154 /* At unmount we may treat errors differently */
1155 if (xfs_is_unmounting(mp) && mp->m_fail_unmount)
1156 return true;
1157
1158 return false;
1159 }
1160
1161 /*
1162 * On a sync write or shutdown we just want to stale the buffer and let the
1163 * caller handle the error in bp->b_error appropriately.
1164 *
1165 * If the write was asynchronous then no one will be looking for the error. If
1166 * this is the first failure of this type, clear the error state and write the
1167 * buffer out again. This means we always retry an async write failure at least
1168 * once, but we also need to set the buffer up to behave correctly now for
1169 * repeated failures.
1170 *
1171 * If we get repeated async write failures, then we take action according to the
1172 * error configuration we have been set up to use.
1173 *
1174 * Returns true if this function took care of error handling and the caller must
1175 * not touch the buffer again. Return false if the caller should proceed with
1176 * normal I/O completion handling.
1177 */
1178 static bool
xfs_buf_ioend_handle_error(struct xfs_buf * bp)1179 xfs_buf_ioend_handle_error(
1180 struct xfs_buf *bp)
1181 {
1182 struct xfs_mount *mp = bp->b_mount;
1183 struct xfs_error_cfg *cfg;
1184
1185 /*
1186 * If we've already shutdown the journal because of I/O errors, there's
1187 * no point in giving this a retry.
1188 */
1189 if (xlog_is_shutdown(mp->m_log))
1190 goto out_stale;
1191
1192 xfs_buf_ioerror_alert_ratelimited(bp);
1193
1194 /*
1195 * We're not going to bother about retrying this during recovery.
1196 * One strike!
1197 */
1198 if (bp->b_flags & _XBF_LOGRECOVERY) {
1199 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1200 return false;
1201 }
1202
1203 /*
1204 * Synchronous writes will have callers process the error.
1205 */
1206 if (!(bp->b_flags & XBF_ASYNC))
1207 goto out_stale;
1208
1209 trace_xfs_buf_iodone_async(bp, _RET_IP_);
1210
1211 cfg = xfs_error_get_cfg(mp, XFS_ERR_METADATA, bp->b_error);
1212 if (bp->b_last_error != bp->b_error ||
1213 !(bp->b_flags & (XBF_STALE | XBF_WRITE_FAIL))) {
1214 bp->b_last_error = bp->b_error;
1215 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1216 !bp->b_first_retry_time)
1217 bp->b_first_retry_time = jiffies;
1218 goto resubmit;
1219 }
1220
1221 /*
1222 * Permanent error - we need to trigger a shutdown if we haven't already
1223 * to indicate that inconsistency will result from this action.
1224 */
1225 if (xfs_buf_ioerror_permanent(bp, cfg)) {
1226 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1227 goto out_stale;
1228 }
1229
1230 /* Still considered a transient error. Caller will schedule retries. */
1231 if (bp->b_flags & _XBF_INODES)
1232 xfs_buf_inode_io_fail(bp);
1233 else if (bp->b_flags & _XBF_DQUOTS)
1234 xfs_buf_dquot_io_fail(bp);
1235 else
1236 ASSERT(list_empty(&bp->b_li_list));
1237 xfs_buf_ioerror(bp, 0);
1238 xfs_buf_relse(bp);
1239 return true;
1240
1241 resubmit:
1242 xfs_buf_ioerror(bp, 0);
1243 bp->b_flags |= (XBF_DONE | XBF_WRITE_FAIL);
1244 xfs_buf_submit(bp);
1245 return true;
1246 out_stale:
1247 xfs_buf_stale(bp);
1248 bp->b_flags |= XBF_DONE;
1249 bp->b_flags &= ~XBF_WRITE;
1250 trace_xfs_buf_error_relse(bp, _RET_IP_);
1251 return false;
1252 }
1253
1254 static void
xfs_buf_ioend(struct xfs_buf * bp)1255 xfs_buf_ioend(
1256 struct xfs_buf *bp)
1257 {
1258 trace_xfs_buf_iodone(bp, _RET_IP_);
1259
1260 /*
1261 * Pull in IO completion errors now. We are guaranteed to be running
1262 * single threaded, so we don't need the lock to read b_io_error.
1263 */
1264 if (!bp->b_error && bp->b_io_error)
1265 xfs_buf_ioerror(bp, bp->b_io_error);
1266
1267 if (bp->b_flags & XBF_READ) {
1268 if (!bp->b_error && bp->b_ops)
1269 bp->b_ops->verify_read(bp);
1270 if (!bp->b_error)
1271 bp->b_flags |= XBF_DONE;
1272 } else {
1273 if (!bp->b_error) {
1274 bp->b_flags &= ~XBF_WRITE_FAIL;
1275 bp->b_flags |= XBF_DONE;
1276 }
1277
1278 if (unlikely(bp->b_error) && xfs_buf_ioend_handle_error(bp))
1279 return;
1280
1281 /* clear the retry state */
1282 bp->b_last_error = 0;
1283 bp->b_retries = 0;
1284 bp->b_first_retry_time = 0;
1285
1286 /*
1287 * Note that for things like remote attribute buffers, there may
1288 * not be a buffer log item here, so processing the buffer log
1289 * item must remain optional.
1290 */
1291 if (bp->b_log_item)
1292 xfs_buf_item_done(bp);
1293
1294 if (bp->b_flags & _XBF_INODES)
1295 xfs_buf_inode_iodone(bp);
1296 else if (bp->b_flags & _XBF_DQUOTS)
1297 xfs_buf_dquot_iodone(bp);
1298
1299 }
1300
1301 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD |
1302 _XBF_LOGRECOVERY);
1303
1304 if (bp->b_flags & XBF_ASYNC)
1305 xfs_buf_relse(bp);
1306 else
1307 complete(&bp->b_iowait);
1308 }
1309
1310 static void
xfs_buf_ioend_work(struct work_struct * work)1311 xfs_buf_ioend_work(
1312 struct work_struct *work)
1313 {
1314 struct xfs_buf *bp =
1315 container_of(work, struct xfs_buf, b_ioend_work);
1316
1317 xfs_buf_ioend(bp);
1318 }
1319
1320 static void
xfs_buf_ioend_async(struct xfs_buf * bp)1321 xfs_buf_ioend_async(
1322 struct xfs_buf *bp)
1323 {
1324 INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
1325 queue_work(bp->b_mount->m_buf_workqueue, &bp->b_ioend_work);
1326 }
1327
1328 void
__xfs_buf_ioerror(struct xfs_buf * bp,int error,xfs_failaddr_t failaddr)1329 __xfs_buf_ioerror(
1330 struct xfs_buf *bp,
1331 int error,
1332 xfs_failaddr_t failaddr)
1333 {
1334 ASSERT(error <= 0 && error >= -1000);
1335 bp->b_error = error;
1336 trace_xfs_buf_ioerror(bp, error, failaddr);
1337 }
1338
1339 void
xfs_buf_ioerror_alert(struct xfs_buf * bp,xfs_failaddr_t func)1340 xfs_buf_ioerror_alert(
1341 struct xfs_buf *bp,
1342 xfs_failaddr_t func)
1343 {
1344 xfs_buf_alert_ratelimited(bp, "XFS: metadata IO error",
1345 "metadata I/O error in \"%pS\" at daddr 0x%llx len %d error %d",
1346 func, (uint64_t)xfs_buf_daddr(bp),
1347 bp->b_length, -bp->b_error);
1348 }
1349
1350 /*
1351 * To simulate an I/O failure, the buffer must be locked and held with at least
1352 * three references. The LRU reference is dropped by the stale call. The buf
1353 * item reference is dropped via ioend processing. The third reference is owned
1354 * by the caller and is dropped on I/O completion if the buffer is XBF_ASYNC.
1355 */
1356 void
xfs_buf_ioend_fail(struct xfs_buf * bp)1357 xfs_buf_ioend_fail(
1358 struct xfs_buf *bp)
1359 {
1360 bp->b_flags &= ~XBF_DONE;
1361 xfs_buf_stale(bp);
1362 xfs_buf_ioerror(bp, -EIO);
1363 xfs_buf_ioend(bp);
1364 }
1365
1366 int
xfs_bwrite(struct xfs_buf * bp)1367 xfs_bwrite(
1368 struct xfs_buf *bp)
1369 {
1370 int error;
1371
1372 ASSERT(xfs_buf_islocked(bp));
1373
1374 bp->b_flags |= XBF_WRITE;
1375 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1376 XBF_DONE);
1377
1378 error = xfs_buf_submit(bp);
1379 if (error)
1380 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
1381 return error;
1382 }
1383
1384 static void
xfs_buf_bio_end_io(struct bio * bio)1385 xfs_buf_bio_end_io(
1386 struct bio *bio)
1387 {
1388 struct xfs_buf *bp = (struct xfs_buf *)bio->bi_private;
1389
1390 if (!bio->bi_status &&
1391 (bp->b_flags & XBF_WRITE) && (bp->b_flags & XBF_ASYNC) &&
1392 XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_IOERROR))
1393 bio->bi_status = BLK_STS_IOERR;
1394
1395 /*
1396 * don't overwrite existing errors - otherwise we can lose errors on
1397 * buffers that require multiple bios to complete.
1398 */
1399 if (bio->bi_status) {
1400 int error = blk_status_to_errno(bio->bi_status);
1401
1402 cmpxchg(&bp->b_io_error, 0, error);
1403 }
1404
1405 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1406 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1407
1408 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1409 xfs_buf_ioend_async(bp);
1410 bio_put(bio);
1411 }
1412
1413 static void
xfs_buf_ioapply_map(struct xfs_buf * bp,int map,int * buf_offset,int * count,int op)1414 xfs_buf_ioapply_map(
1415 struct xfs_buf *bp,
1416 int map,
1417 int *buf_offset,
1418 int *count,
1419 int op)
1420 {
1421 int page_index;
1422 unsigned int total_nr_pages = bp->b_page_count;
1423 int nr_pages;
1424 struct bio *bio;
1425 sector_t sector = bp->b_maps[map].bm_bn;
1426 int size;
1427 int offset;
1428
1429 /* skip the pages in the buffer before the start offset */
1430 page_index = 0;
1431 offset = *buf_offset;
1432 while (offset >= PAGE_SIZE) {
1433 page_index++;
1434 offset -= PAGE_SIZE;
1435 }
1436
1437 /*
1438 * Limit the IO size to the length of the current vector, and update the
1439 * remaining IO count for the next time around.
1440 */
1441 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1442 *count -= size;
1443 *buf_offset += size;
1444
1445 next_chunk:
1446 atomic_inc(&bp->b_io_remaining);
1447 nr_pages = bio_max_segs(total_nr_pages);
1448
1449 bio = bio_alloc(bp->b_target->bt_bdev, nr_pages, op, GFP_NOIO);
1450 bio->bi_iter.bi_sector = sector;
1451 bio->bi_end_io = xfs_buf_bio_end_io;
1452 bio->bi_private = bp;
1453
1454 for (; size && nr_pages; nr_pages--, page_index++) {
1455 int rbytes, nbytes = PAGE_SIZE - offset;
1456
1457 if (nbytes > size)
1458 nbytes = size;
1459
1460 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1461 offset);
1462 if (rbytes < nbytes)
1463 break;
1464
1465 offset = 0;
1466 sector += BTOBB(nbytes);
1467 size -= nbytes;
1468 total_nr_pages--;
1469 }
1470
1471 if (likely(bio->bi_iter.bi_size)) {
1472 if (xfs_buf_is_vmapped(bp)) {
1473 flush_kernel_vmap_range(bp->b_addr,
1474 xfs_buf_vmap_len(bp));
1475 }
1476 submit_bio(bio);
1477 if (size)
1478 goto next_chunk;
1479 } else {
1480 /*
1481 * This is guaranteed not to be the last io reference count
1482 * because the caller (xfs_buf_submit) holds a count itself.
1483 */
1484 atomic_dec(&bp->b_io_remaining);
1485 xfs_buf_ioerror(bp, -EIO);
1486 bio_put(bio);
1487 }
1488
1489 }
1490
1491 STATIC void
_xfs_buf_ioapply(struct xfs_buf * bp)1492 _xfs_buf_ioapply(
1493 struct xfs_buf *bp)
1494 {
1495 struct blk_plug plug;
1496 int op;
1497 int offset;
1498 int size;
1499 int i;
1500
1501 /*
1502 * Make sure we capture only current IO errors rather than stale errors
1503 * left over from previous use of the buffer (e.g. failed readahead).
1504 */
1505 bp->b_error = 0;
1506
1507 if (bp->b_flags & XBF_WRITE) {
1508 op = REQ_OP_WRITE;
1509
1510 /*
1511 * Run the write verifier callback function if it exists. If
1512 * this function fails it will mark the buffer with an error and
1513 * the IO should not be dispatched.
1514 */
1515 if (bp->b_ops) {
1516 bp->b_ops->verify_write(bp);
1517 if (bp->b_error) {
1518 xfs_force_shutdown(bp->b_mount,
1519 SHUTDOWN_CORRUPT_INCORE);
1520 return;
1521 }
1522 } else if (bp->b_rhash_key != XFS_BUF_DADDR_NULL) {
1523 struct xfs_mount *mp = bp->b_mount;
1524
1525 /*
1526 * non-crc filesystems don't attach verifiers during
1527 * log recovery, so don't warn for such filesystems.
1528 */
1529 if (xfs_has_crc(mp)) {
1530 xfs_warn(mp,
1531 "%s: no buf ops on daddr 0x%llx len %d",
1532 __func__, xfs_buf_daddr(bp),
1533 bp->b_length);
1534 xfs_hex_dump(bp->b_addr,
1535 XFS_CORRUPTION_DUMP_LEN);
1536 dump_stack();
1537 }
1538 }
1539 } else {
1540 op = REQ_OP_READ;
1541 if (bp->b_flags & XBF_READ_AHEAD)
1542 op |= REQ_RAHEAD;
1543 }
1544
1545 /* we only use the buffer cache for meta-data */
1546 op |= REQ_META;
1547
1548 /*
1549 * Walk all the vectors issuing IO on them. Set up the initial offset
1550 * into the buffer and the desired IO size before we start -
1551 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1552 * subsequent call.
1553 */
1554 offset = bp->b_offset;
1555 size = BBTOB(bp->b_length);
1556 blk_start_plug(&plug);
1557 for (i = 0; i < bp->b_map_count; i++) {
1558 xfs_buf_ioapply_map(bp, i, &offset, &size, op);
1559 if (bp->b_error)
1560 break;
1561 if (size <= 0)
1562 break; /* all done */
1563 }
1564 blk_finish_plug(&plug);
1565 }
1566
1567 /*
1568 * Wait for I/O completion of a sync buffer and return the I/O error code.
1569 */
1570 static int
xfs_buf_iowait(struct xfs_buf * bp)1571 xfs_buf_iowait(
1572 struct xfs_buf *bp)
1573 {
1574 ASSERT(!(bp->b_flags & XBF_ASYNC));
1575
1576 trace_xfs_buf_iowait(bp, _RET_IP_);
1577 wait_for_completion(&bp->b_iowait);
1578 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1579
1580 return bp->b_error;
1581 }
1582
1583 /*
1584 * Buffer I/O submission path, read or write. Asynchronous submission transfers
1585 * the buffer lock ownership and the current reference to the IO. It is not
1586 * safe to reference the buffer after a call to this function unless the caller
1587 * holds an additional reference itself.
1588 */
1589 static int
__xfs_buf_submit(struct xfs_buf * bp,bool wait)1590 __xfs_buf_submit(
1591 struct xfs_buf *bp,
1592 bool wait)
1593 {
1594 int error = 0;
1595
1596 trace_xfs_buf_submit(bp, _RET_IP_);
1597
1598 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1599
1600 /*
1601 * On log shutdown we stale and complete the buffer immediately. We can
1602 * be called to read the superblock before the log has been set up, so
1603 * be careful checking the log state.
1604 *
1605 * Checking the mount shutdown state here can result in the log tail
1606 * moving inappropriately on disk as the log may not yet be shut down.
1607 * i.e. failing this buffer on mount shutdown can remove it from the AIL
1608 * and move the tail of the log forwards without having written this
1609 * buffer to disk. This corrupts the log tail state in memory, and
1610 * because the log may not be shut down yet, it can then be propagated
1611 * to disk before the log is shutdown. Hence we check log shutdown
1612 * state here rather than mount state to avoid corrupting the log tail
1613 * on shutdown.
1614 */
1615 if (bp->b_mount->m_log &&
1616 xlog_is_shutdown(bp->b_mount->m_log)) {
1617 xfs_buf_ioend_fail(bp);
1618 return -EIO;
1619 }
1620
1621 /*
1622 * Grab a reference so the buffer does not go away underneath us. For
1623 * async buffers, I/O completion drops the callers reference, which
1624 * could occur before submission returns.
1625 */
1626 xfs_buf_hold(bp);
1627
1628 if (bp->b_flags & XBF_WRITE)
1629 xfs_buf_wait_unpin(bp);
1630
1631 /* clear the internal error state to avoid spurious errors */
1632 bp->b_io_error = 0;
1633
1634 /*
1635 * Set the count to 1 initially, this will stop an I/O completion
1636 * callout which happens before we have started all the I/O from calling
1637 * xfs_buf_ioend too early.
1638 */
1639 atomic_set(&bp->b_io_remaining, 1);
1640 if (bp->b_flags & XBF_ASYNC)
1641 xfs_buf_ioacct_inc(bp);
1642 _xfs_buf_ioapply(bp);
1643
1644 /*
1645 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1646 * reference we took above. If we drop it to zero, run completion so
1647 * that we don't return to the caller with completion still pending.
1648 */
1649 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1650 if (bp->b_error || !(bp->b_flags & XBF_ASYNC))
1651 xfs_buf_ioend(bp);
1652 else
1653 xfs_buf_ioend_async(bp);
1654 }
1655
1656 if (wait)
1657 error = xfs_buf_iowait(bp);
1658
1659 /*
1660 * Release the hold that keeps the buffer referenced for the entire
1661 * I/O. Note that if the buffer is async, it is not safe to reference
1662 * after this release.
1663 */
1664 xfs_buf_rele(bp);
1665 return error;
1666 }
1667
1668 void *
xfs_buf_offset(struct xfs_buf * bp,size_t offset)1669 xfs_buf_offset(
1670 struct xfs_buf *bp,
1671 size_t offset)
1672 {
1673 struct page *page;
1674
1675 if (bp->b_addr)
1676 return bp->b_addr + offset;
1677
1678 page = bp->b_pages[offset >> PAGE_SHIFT];
1679 return page_address(page) + (offset & (PAGE_SIZE-1));
1680 }
1681
1682 void
xfs_buf_zero(struct xfs_buf * bp,size_t boff,size_t bsize)1683 xfs_buf_zero(
1684 struct xfs_buf *bp,
1685 size_t boff,
1686 size_t bsize)
1687 {
1688 size_t bend;
1689
1690 bend = boff + bsize;
1691 while (boff < bend) {
1692 struct page *page;
1693 int page_index, page_offset, csize;
1694
1695 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1696 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1697 page = bp->b_pages[page_index];
1698 csize = min_t(size_t, PAGE_SIZE - page_offset,
1699 BBTOB(bp->b_length) - boff);
1700
1701 ASSERT((csize + page_offset) <= PAGE_SIZE);
1702
1703 memset(page_address(page) + page_offset, 0, csize);
1704
1705 boff += csize;
1706 }
1707 }
1708
1709 /*
1710 * Log a message about and stale a buffer that a caller has decided is corrupt.
1711 *
1712 * This function should be called for the kinds of metadata corruption that
1713 * cannot be detect from a verifier, such as incorrect inter-block relationship
1714 * data. Do /not/ call this function from a verifier function.
1715 *
1716 * The buffer must be XBF_DONE prior to the call. Afterwards, the buffer will
1717 * be marked stale, but b_error will not be set. The caller is responsible for
1718 * releasing the buffer or fixing it.
1719 */
1720 void
__xfs_buf_mark_corrupt(struct xfs_buf * bp,xfs_failaddr_t fa)1721 __xfs_buf_mark_corrupt(
1722 struct xfs_buf *bp,
1723 xfs_failaddr_t fa)
1724 {
1725 ASSERT(bp->b_flags & XBF_DONE);
1726
1727 xfs_buf_corruption_error(bp, fa);
1728 xfs_buf_stale(bp);
1729 }
1730
1731 /*
1732 * Handling of buffer targets (buftargs).
1733 */
1734
1735 /*
1736 * Wait for any bufs with callbacks that have been submitted but have not yet
1737 * returned. These buffers will have an elevated hold count, so wait on those
1738 * while freeing all the buffers only held by the LRU.
1739 */
1740 static enum lru_status
xfs_buftarg_drain_rele(struct list_head * item,struct list_lru_one * lru,spinlock_t * lru_lock,void * arg)1741 xfs_buftarg_drain_rele(
1742 struct list_head *item,
1743 struct list_lru_one *lru,
1744 spinlock_t *lru_lock,
1745 void *arg)
1746
1747 {
1748 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1749 struct list_head *dispose = arg;
1750
1751 if (atomic_read(&bp->b_hold) > 1) {
1752 /* need to wait, so skip it this pass */
1753 trace_xfs_buf_drain_buftarg(bp, _RET_IP_);
1754 return LRU_SKIP;
1755 }
1756 if (!spin_trylock(&bp->b_lock))
1757 return LRU_SKIP;
1758
1759 /*
1760 * clear the LRU reference count so the buffer doesn't get
1761 * ignored in xfs_buf_rele().
1762 */
1763 atomic_set(&bp->b_lru_ref, 0);
1764 bp->b_state |= XFS_BSTATE_DISPOSE;
1765 list_lru_isolate_move(lru, item, dispose);
1766 spin_unlock(&bp->b_lock);
1767 return LRU_REMOVED;
1768 }
1769
1770 /*
1771 * Wait for outstanding I/O on the buftarg to complete.
1772 */
1773 void
xfs_buftarg_wait(struct xfs_buftarg * btp)1774 xfs_buftarg_wait(
1775 struct xfs_buftarg *btp)
1776 {
1777 /*
1778 * First wait on the buftarg I/O count for all in-flight buffers to be
1779 * released. This is critical as new buffers do not make the LRU until
1780 * they are released.
1781 *
1782 * Next, flush the buffer workqueue to ensure all completion processing
1783 * has finished. Just waiting on buffer locks is not sufficient for
1784 * async IO as the reference count held over IO is not released until
1785 * after the buffer lock is dropped. Hence we need to ensure here that
1786 * all reference counts have been dropped before we start walking the
1787 * LRU list.
1788 */
1789 while (percpu_counter_sum(&btp->bt_io_count))
1790 delay(100);
1791 flush_workqueue(btp->bt_mount->m_buf_workqueue);
1792 }
1793
1794 void
xfs_buftarg_drain(struct xfs_buftarg * btp)1795 xfs_buftarg_drain(
1796 struct xfs_buftarg *btp)
1797 {
1798 LIST_HEAD(dispose);
1799 int loop = 0;
1800 bool write_fail = false;
1801
1802 xfs_buftarg_wait(btp);
1803
1804 /* loop until there is nothing left on the lru list. */
1805 while (list_lru_count(&btp->bt_lru)) {
1806 list_lru_walk(&btp->bt_lru, xfs_buftarg_drain_rele,
1807 &dispose, LONG_MAX);
1808
1809 while (!list_empty(&dispose)) {
1810 struct xfs_buf *bp;
1811 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1812 list_del_init(&bp->b_lru);
1813 if (bp->b_flags & XBF_WRITE_FAIL) {
1814 write_fail = true;
1815 xfs_buf_alert_ratelimited(bp,
1816 "XFS: Corruption Alert",
1817 "Corruption Alert: Buffer at daddr 0x%llx had permanent write failures!",
1818 (long long)xfs_buf_daddr(bp));
1819 }
1820 xfs_buf_rele(bp);
1821 }
1822 if (loop++ != 0)
1823 delay(100);
1824 }
1825
1826 /*
1827 * If one or more failed buffers were freed, that means dirty metadata
1828 * was thrown away. This should only ever happen after I/O completion
1829 * handling has elevated I/O error(s) to permanent failures and shuts
1830 * down the journal.
1831 */
1832 if (write_fail) {
1833 ASSERT(xlog_is_shutdown(btp->bt_mount->m_log));
1834 xfs_alert(btp->bt_mount,
1835 "Please run xfs_repair to determine the extent of the problem.");
1836 }
1837 }
1838
1839 static enum lru_status
xfs_buftarg_isolate(struct list_head * item,struct list_lru_one * lru,spinlock_t * lru_lock,void * arg)1840 xfs_buftarg_isolate(
1841 struct list_head *item,
1842 struct list_lru_one *lru,
1843 spinlock_t *lru_lock,
1844 void *arg)
1845 {
1846 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1847 struct list_head *dispose = arg;
1848
1849 /*
1850 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1851 * If we fail to get the lock, just skip it.
1852 */
1853 if (!spin_trylock(&bp->b_lock))
1854 return LRU_SKIP;
1855 /*
1856 * Decrement the b_lru_ref count unless the value is already
1857 * zero. If the value is already zero, we need to reclaim the
1858 * buffer, otherwise it gets another trip through the LRU.
1859 */
1860 if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1861 spin_unlock(&bp->b_lock);
1862 return LRU_ROTATE;
1863 }
1864
1865 bp->b_state |= XFS_BSTATE_DISPOSE;
1866 list_lru_isolate_move(lru, item, dispose);
1867 spin_unlock(&bp->b_lock);
1868 return LRU_REMOVED;
1869 }
1870
1871 static unsigned long
xfs_buftarg_shrink_scan(struct shrinker * shrink,struct shrink_control * sc)1872 xfs_buftarg_shrink_scan(
1873 struct shrinker *shrink,
1874 struct shrink_control *sc)
1875 {
1876 struct xfs_buftarg *btp = container_of(shrink,
1877 struct xfs_buftarg, bt_shrinker);
1878 LIST_HEAD(dispose);
1879 unsigned long freed;
1880
1881 freed = list_lru_shrink_walk(&btp->bt_lru, sc,
1882 xfs_buftarg_isolate, &dispose);
1883
1884 while (!list_empty(&dispose)) {
1885 struct xfs_buf *bp;
1886 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1887 list_del_init(&bp->b_lru);
1888 xfs_buf_rele(bp);
1889 }
1890
1891 return freed;
1892 }
1893
1894 static unsigned long
xfs_buftarg_shrink_count(struct shrinker * shrink,struct shrink_control * sc)1895 xfs_buftarg_shrink_count(
1896 struct shrinker *shrink,
1897 struct shrink_control *sc)
1898 {
1899 struct xfs_buftarg *btp = container_of(shrink,
1900 struct xfs_buftarg, bt_shrinker);
1901 return list_lru_shrink_count(&btp->bt_lru, sc);
1902 }
1903
1904 void
xfs_free_buftarg(struct xfs_buftarg * btp)1905 xfs_free_buftarg(
1906 struct xfs_buftarg *btp)
1907 {
1908 unregister_shrinker(&btp->bt_shrinker);
1909 ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
1910 percpu_counter_destroy(&btp->bt_io_count);
1911 list_lru_destroy(&btp->bt_lru);
1912
1913 blkdev_issue_flush(btp->bt_bdev);
1914 fs_put_dax(btp->bt_daxdev);
1915
1916 kmem_free(btp);
1917 }
1918
1919 int
xfs_setsize_buftarg(xfs_buftarg_t * btp,unsigned int sectorsize)1920 xfs_setsize_buftarg(
1921 xfs_buftarg_t *btp,
1922 unsigned int sectorsize)
1923 {
1924 /* Set up metadata sector size info */
1925 btp->bt_meta_sectorsize = sectorsize;
1926 btp->bt_meta_sectormask = sectorsize - 1;
1927
1928 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1929 xfs_warn(btp->bt_mount,
1930 "Cannot set_blocksize to %u on device %pg",
1931 sectorsize, btp->bt_bdev);
1932 return -EINVAL;
1933 }
1934
1935 /* Set up device logical sector size mask */
1936 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1937 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1938
1939 return 0;
1940 }
1941
1942 /*
1943 * When allocating the initial buffer target we have not yet
1944 * read in the superblock, so don't know what sized sectors
1945 * are being used at this early stage. Play safe.
1946 */
1947 STATIC int
xfs_setsize_buftarg_early(xfs_buftarg_t * btp,struct block_device * bdev)1948 xfs_setsize_buftarg_early(
1949 xfs_buftarg_t *btp,
1950 struct block_device *bdev)
1951 {
1952 return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev));
1953 }
1954
1955 struct xfs_buftarg *
xfs_alloc_buftarg(struct xfs_mount * mp,struct block_device * bdev)1956 xfs_alloc_buftarg(
1957 struct xfs_mount *mp,
1958 struct block_device *bdev)
1959 {
1960 xfs_buftarg_t *btp;
1961
1962 btp = kmem_zalloc(sizeof(*btp), KM_NOFS);
1963
1964 btp->bt_mount = mp;
1965 btp->bt_dev = bdev->bd_dev;
1966 btp->bt_bdev = bdev;
1967 btp->bt_daxdev = fs_dax_get_by_bdev(bdev, &btp->bt_dax_part_off);
1968
1969 /*
1970 * Buffer IO error rate limiting. Limit it to no more than 10 messages
1971 * per 30 seconds so as to not spam logs too much on repeated errors.
1972 */
1973 ratelimit_state_init(&btp->bt_ioerror_rl, 30 * HZ,
1974 DEFAULT_RATELIMIT_BURST);
1975
1976 if (xfs_setsize_buftarg_early(btp, bdev))
1977 goto error_free;
1978
1979 if (list_lru_init(&btp->bt_lru))
1980 goto error_free;
1981
1982 if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL))
1983 goto error_lru;
1984
1985 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
1986 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
1987 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1988 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
1989 if (register_shrinker(&btp->bt_shrinker))
1990 goto error_pcpu;
1991 return btp;
1992
1993 error_pcpu:
1994 percpu_counter_destroy(&btp->bt_io_count);
1995 error_lru:
1996 list_lru_destroy(&btp->bt_lru);
1997 error_free:
1998 kmem_free(btp);
1999 return NULL;
2000 }
2001
2002 /*
2003 * Cancel a delayed write list.
2004 *
2005 * Remove each buffer from the list, clear the delwri queue flag and drop the
2006 * associated buffer reference.
2007 */
2008 void
xfs_buf_delwri_cancel(struct list_head * list)2009 xfs_buf_delwri_cancel(
2010 struct list_head *list)
2011 {
2012 struct xfs_buf *bp;
2013
2014 while (!list_empty(list)) {
2015 bp = list_first_entry(list, struct xfs_buf, b_list);
2016
2017 xfs_buf_lock(bp);
2018 bp->b_flags &= ~_XBF_DELWRI_Q;
2019 list_del_init(&bp->b_list);
2020 xfs_buf_relse(bp);
2021 }
2022 }
2023
2024 /*
2025 * Add a buffer to the delayed write list.
2026 *
2027 * This queues a buffer for writeout if it hasn't already been. Note that
2028 * neither this routine nor the buffer list submission functions perform
2029 * any internal synchronization. It is expected that the lists are thread-local
2030 * to the callers.
2031 *
2032 * Returns true if we queued up the buffer, or false if it already had
2033 * been on the buffer list.
2034 */
2035 bool
xfs_buf_delwri_queue(struct xfs_buf * bp,struct list_head * list)2036 xfs_buf_delwri_queue(
2037 struct xfs_buf *bp,
2038 struct list_head *list)
2039 {
2040 ASSERT(xfs_buf_islocked(bp));
2041 ASSERT(!(bp->b_flags & XBF_READ));
2042
2043 /*
2044 * If the buffer is already marked delwri it already is queued up
2045 * by someone else for imediate writeout. Just ignore it in that
2046 * case.
2047 */
2048 if (bp->b_flags & _XBF_DELWRI_Q) {
2049 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
2050 return false;
2051 }
2052
2053 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
2054
2055 /*
2056 * If a buffer gets written out synchronously or marked stale while it
2057 * is on a delwri list we lazily remove it. To do this, the other party
2058 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
2059 * It remains referenced and on the list. In a rare corner case it
2060 * might get readded to a delwri list after the synchronous writeout, in
2061 * which case we need just need to re-add the flag here.
2062 */
2063 bp->b_flags |= _XBF_DELWRI_Q;
2064 if (list_empty(&bp->b_list)) {
2065 atomic_inc(&bp->b_hold);
2066 list_add_tail(&bp->b_list, list);
2067 }
2068
2069 return true;
2070 }
2071
2072 /*
2073 * Compare function is more complex than it needs to be because
2074 * the return value is only 32 bits and we are doing comparisons
2075 * on 64 bit values
2076 */
2077 static int
xfs_buf_cmp(void * priv,const struct list_head * a,const struct list_head * b)2078 xfs_buf_cmp(
2079 void *priv,
2080 const struct list_head *a,
2081 const struct list_head *b)
2082 {
2083 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
2084 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
2085 xfs_daddr_t diff;
2086
2087 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
2088 if (diff < 0)
2089 return -1;
2090 if (diff > 0)
2091 return 1;
2092 return 0;
2093 }
2094
2095 /*
2096 * Submit buffers for write. If wait_list is specified, the buffers are
2097 * submitted using sync I/O and placed on the wait list such that the caller can
2098 * iowait each buffer. Otherwise async I/O is used and the buffers are released
2099 * at I/O completion time. In either case, buffers remain locked until I/O
2100 * completes and the buffer is released from the queue.
2101 */
2102 static int
xfs_buf_delwri_submit_buffers(struct list_head * buffer_list,struct list_head * wait_list)2103 xfs_buf_delwri_submit_buffers(
2104 struct list_head *buffer_list,
2105 struct list_head *wait_list)
2106 {
2107 struct xfs_buf *bp, *n;
2108 int pinned = 0;
2109 struct blk_plug plug;
2110
2111 list_sort(NULL, buffer_list, xfs_buf_cmp);
2112
2113 blk_start_plug(&plug);
2114 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
2115 if (!wait_list) {
2116 if (!xfs_buf_trylock(bp))
2117 continue;
2118 if (xfs_buf_ispinned(bp)) {
2119 xfs_buf_unlock(bp);
2120 pinned++;
2121 continue;
2122 }
2123 } else {
2124 xfs_buf_lock(bp);
2125 }
2126
2127 /*
2128 * Someone else might have written the buffer synchronously or
2129 * marked it stale in the meantime. In that case only the
2130 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
2131 * reference and remove it from the list here.
2132 */
2133 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
2134 list_del_init(&bp->b_list);
2135 xfs_buf_relse(bp);
2136 continue;
2137 }
2138
2139 trace_xfs_buf_delwri_split(bp, _RET_IP_);
2140
2141 /*
2142 * If we have a wait list, each buffer (and associated delwri
2143 * queue reference) transfers to it and is submitted
2144 * synchronously. Otherwise, drop the buffer from the delwri
2145 * queue and submit async.
2146 */
2147 bp->b_flags &= ~_XBF_DELWRI_Q;
2148 bp->b_flags |= XBF_WRITE;
2149 if (wait_list) {
2150 bp->b_flags &= ~XBF_ASYNC;
2151 list_move_tail(&bp->b_list, wait_list);
2152 } else {
2153 bp->b_flags |= XBF_ASYNC;
2154 list_del_init(&bp->b_list);
2155 }
2156 __xfs_buf_submit(bp, false);
2157 }
2158 blk_finish_plug(&plug);
2159
2160 return pinned;
2161 }
2162
2163 /*
2164 * Write out a buffer list asynchronously.
2165 *
2166 * This will take the @buffer_list, write all non-locked and non-pinned buffers
2167 * out and not wait for I/O completion on any of the buffers. This interface
2168 * is only safely useable for callers that can track I/O completion by higher
2169 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
2170 * function.
2171 *
2172 * Note: this function will skip buffers it would block on, and in doing so
2173 * leaves them on @buffer_list so they can be retried on a later pass. As such,
2174 * it is up to the caller to ensure that the buffer list is fully submitted or
2175 * cancelled appropriately when they are finished with the list. Failure to
2176 * cancel or resubmit the list until it is empty will result in leaked buffers
2177 * at unmount time.
2178 */
2179 int
xfs_buf_delwri_submit_nowait(struct list_head * buffer_list)2180 xfs_buf_delwri_submit_nowait(
2181 struct list_head *buffer_list)
2182 {
2183 return xfs_buf_delwri_submit_buffers(buffer_list, NULL);
2184 }
2185
2186 /*
2187 * Write out a buffer list synchronously.
2188 *
2189 * This will take the @buffer_list, write all buffers out and wait for I/O
2190 * completion on all of the buffers. @buffer_list is consumed by the function,
2191 * so callers must have some other way of tracking buffers if they require such
2192 * functionality.
2193 */
2194 int
xfs_buf_delwri_submit(struct list_head * buffer_list)2195 xfs_buf_delwri_submit(
2196 struct list_head *buffer_list)
2197 {
2198 LIST_HEAD (wait_list);
2199 int error = 0, error2;
2200 struct xfs_buf *bp;
2201
2202 xfs_buf_delwri_submit_buffers(buffer_list, &wait_list);
2203
2204 /* Wait for IO to complete. */
2205 while (!list_empty(&wait_list)) {
2206 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
2207
2208 list_del_init(&bp->b_list);
2209
2210 /*
2211 * Wait on the locked buffer, check for errors and unlock and
2212 * release the delwri queue reference.
2213 */
2214 error2 = xfs_buf_iowait(bp);
2215 xfs_buf_relse(bp);
2216 if (!error)
2217 error = error2;
2218 }
2219
2220 return error;
2221 }
2222
2223 /*
2224 * Push a single buffer on a delwri queue.
2225 *
2226 * The purpose of this function is to submit a single buffer of a delwri queue
2227 * and return with the buffer still on the original queue. The waiting delwri
2228 * buffer submission infrastructure guarantees transfer of the delwri queue
2229 * buffer reference to a temporary wait list. We reuse this infrastructure to
2230 * transfer the buffer back to the original queue.
2231 *
2232 * Note the buffer transitions from the queued state, to the submitted and wait
2233 * listed state and back to the queued state during this call. The buffer
2234 * locking and queue management logic between _delwri_pushbuf() and
2235 * _delwri_queue() guarantee that the buffer cannot be queued to another list
2236 * before returning.
2237 */
2238 int
xfs_buf_delwri_pushbuf(struct xfs_buf * bp,struct list_head * buffer_list)2239 xfs_buf_delwri_pushbuf(
2240 struct xfs_buf *bp,
2241 struct list_head *buffer_list)
2242 {
2243 LIST_HEAD (submit_list);
2244 int error;
2245
2246 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
2247
2248 trace_xfs_buf_delwri_pushbuf(bp, _RET_IP_);
2249
2250 /*
2251 * Isolate the buffer to a new local list so we can submit it for I/O
2252 * independently from the rest of the original list.
2253 */
2254 xfs_buf_lock(bp);
2255 list_move(&bp->b_list, &submit_list);
2256 xfs_buf_unlock(bp);
2257
2258 /*
2259 * Delwri submission clears the DELWRI_Q buffer flag and returns with
2260 * the buffer on the wait list with the original reference. Rather than
2261 * bounce the buffer from a local wait list back to the original list
2262 * after I/O completion, reuse the original list as the wait list.
2263 */
2264 xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
2265
2266 /*
2267 * The buffer is now locked, under I/O and wait listed on the original
2268 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
2269 * return with the buffer unlocked and on the original queue.
2270 */
2271 error = xfs_buf_iowait(bp);
2272 bp->b_flags |= _XBF_DELWRI_Q;
2273 xfs_buf_unlock(bp);
2274
2275 return error;
2276 }
2277
2278 int __init
xfs_buf_init(void)2279 xfs_buf_init(void)
2280 {
2281 xfs_buf_cache = kmem_cache_create("xfs_buf", sizeof(struct xfs_buf), 0,
2282 SLAB_HWCACHE_ALIGN |
2283 SLAB_RECLAIM_ACCOUNT |
2284 SLAB_MEM_SPREAD,
2285 NULL);
2286 if (!xfs_buf_cache)
2287 goto out;
2288
2289 return 0;
2290
2291 out:
2292 return -ENOMEM;
2293 }
2294
2295 void
xfs_buf_terminate(void)2296 xfs_buf_terminate(void)
2297 {
2298 kmem_cache_destroy(xfs_buf_cache);
2299 }
2300
xfs_buf_set_ref(struct xfs_buf * bp,int lru_ref)2301 void xfs_buf_set_ref(struct xfs_buf *bp, int lru_ref)
2302 {
2303 /*
2304 * Set the lru reference count to 0 based on the error injection tag.
2305 * This allows userspace to disrupt buffer caching for debug/testing
2306 * purposes.
2307 */
2308 if (XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_LRU_REF))
2309 lru_ref = 0;
2310
2311 atomic_set(&bp->b_lru_ref, lru_ref);
2312 }
2313
2314 /*
2315 * Verify an on-disk magic value against the magic value specified in the
2316 * verifier structure. The verifier magic is in disk byte order so the caller is
2317 * expected to pass the value directly from disk.
2318 */
2319 bool
xfs_verify_magic(struct xfs_buf * bp,__be32 dmagic)2320 xfs_verify_magic(
2321 struct xfs_buf *bp,
2322 __be32 dmagic)
2323 {
2324 struct xfs_mount *mp = bp->b_mount;
2325 int idx;
2326
2327 idx = xfs_has_crc(mp);
2328 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic[idx]))
2329 return false;
2330 return dmagic == bp->b_ops->magic[idx];
2331 }
2332 /*
2333 * Verify an on-disk magic value against the magic value specified in the
2334 * verifier structure. The verifier magic is in disk byte order so the caller is
2335 * expected to pass the value directly from disk.
2336 */
2337 bool
xfs_verify_magic16(struct xfs_buf * bp,__be16 dmagic)2338 xfs_verify_magic16(
2339 struct xfs_buf *bp,
2340 __be16 dmagic)
2341 {
2342 struct xfs_mount *mp = bp->b_mount;
2343 int idx;
2344
2345 idx = xfs_has_crc(mp);
2346 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic16[idx]))
2347 return false;
2348 return dmagic == bp->b_ops->magic16[idx];
2349 }
2350