Loading...
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
48static 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
64static inline int
65xfs_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
78static inline int
79xfs_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 */
98static inline void
99xfs_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 */
118static 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
130static inline void
131xfs_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 */
147void
148xfs_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
180static int
181xfs_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 */
203static void
204xfs_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
213struct 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 INIT_LIST_HEAD(&bp->b_li_list);
240 sema_init(&bp->b_sema, 0); /* held, no waiters */
241 spin_lock_init(&bp->b_lock);
242 XB_SET_OWNER(bp);
243 bp->b_target = target;
244 bp->b_flags = flags;
245
246 /*
247 * Set length and io_length to the same value initially.
248 * I/O routines should use io_length, which will be the same in
249 * most cases but may be reset (e.g. XFS recovery).
250 */
251 error = xfs_buf_get_maps(bp, nmaps);
252 if (error) {
253 kmem_zone_free(xfs_buf_zone, bp);
254 return NULL;
255 }
256
257 bp->b_bn = map[0].bm_bn;
258 bp->b_length = 0;
259 for (i = 0; i < nmaps; i++) {
260 bp->b_maps[i].bm_bn = map[i].bm_bn;
261 bp->b_maps[i].bm_len = map[i].bm_len;
262 bp->b_length += map[i].bm_len;
263 }
264 bp->b_io_length = bp->b_length;
265
266 atomic_set(&bp->b_pin_count, 0);
267 init_waitqueue_head(&bp->b_waiters);
268
269 XFS_STATS_INC(target->bt_mount, xb_create);
270 trace_xfs_buf_init(bp, _RET_IP_);
271
272 return bp;
273}
274
275/*
276 * Allocate a page array capable of holding a specified number
277 * of pages, and point the page buf at it.
278 */
279STATIC int
280_xfs_buf_get_pages(
281 xfs_buf_t *bp,
282 int page_count)
283{
284 /* Make sure that we have a page list */
285 if (bp->b_pages == NULL) {
286 bp->b_page_count = page_count;
287 if (page_count <= XB_PAGES) {
288 bp->b_pages = bp->b_page_array;
289 } else {
290 bp->b_pages = kmem_alloc(sizeof(struct page *) *
291 page_count, KM_NOFS);
292 if (bp->b_pages == NULL)
293 return -ENOMEM;
294 }
295 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
296 }
297 return 0;
298}
299
300/*
301 * Frees b_pages if it was allocated.
302 */
303STATIC void
304_xfs_buf_free_pages(
305 xfs_buf_t *bp)
306{
307 if (bp->b_pages != bp->b_page_array) {
308 kmem_free(bp->b_pages);
309 bp->b_pages = NULL;
310 }
311}
312
313/*
314 * Releases the specified buffer.
315 *
316 * The modification state of any associated pages is left unchanged.
317 * The buffer must not be on any hash - use xfs_buf_rele instead for
318 * hashed and refcounted buffers
319 */
320void
321xfs_buf_free(
322 xfs_buf_t *bp)
323{
324 trace_xfs_buf_free(bp, _RET_IP_);
325
326 ASSERT(list_empty(&bp->b_lru));
327
328 if (bp->b_flags & _XBF_PAGES) {
329 uint i;
330
331 if (xfs_buf_is_vmapped(bp))
332 vm_unmap_ram(bp->b_addr - bp->b_offset,
333 bp->b_page_count);
334
335 for (i = 0; i < bp->b_page_count; i++) {
336 struct page *page = bp->b_pages[i];
337
338 __free_page(page);
339 }
340 } else if (bp->b_flags & _XBF_KMEM)
341 kmem_free(bp->b_addr);
342 _xfs_buf_free_pages(bp);
343 xfs_buf_free_maps(bp);
344 kmem_zone_free(xfs_buf_zone, bp);
345}
346
347/*
348 * Allocates all the pages for buffer in question and builds it's page list.
349 */
350STATIC int
351xfs_buf_allocate_memory(
352 xfs_buf_t *bp,
353 uint flags)
354{
355 size_t size;
356 size_t nbytes, offset;
357 gfp_t gfp_mask = xb_to_gfp(flags);
358 unsigned short page_count, i;
359 xfs_off_t start, end;
360 int error;
361
362 /*
363 * for buffers that are contained within a single page, just allocate
364 * the memory from the heap - there's no need for the complexity of
365 * page arrays to keep allocation down to order 0.
366 */
367 size = BBTOB(bp->b_length);
368 if (size < PAGE_SIZE) {
369 bp->b_addr = kmem_alloc(size, KM_NOFS);
370 if (!bp->b_addr) {
371 /* low memory - use alloc_page loop instead */
372 goto use_alloc_page;
373 }
374
375 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
376 ((unsigned long)bp->b_addr & PAGE_MASK)) {
377 /* b_addr spans two pages - use alloc_page instead */
378 kmem_free(bp->b_addr);
379 bp->b_addr = NULL;
380 goto use_alloc_page;
381 }
382 bp->b_offset = offset_in_page(bp->b_addr);
383 bp->b_pages = bp->b_page_array;
384 bp->b_pages[0] = virt_to_page(bp->b_addr);
385 bp->b_page_count = 1;
386 bp->b_flags |= _XBF_KMEM;
387 return 0;
388 }
389
390use_alloc_page:
391 start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT;
392 end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1)
393 >> PAGE_SHIFT;
394 page_count = end - start;
395 error = _xfs_buf_get_pages(bp, page_count);
396 if (unlikely(error))
397 return error;
398
399 offset = bp->b_offset;
400 bp->b_flags |= _XBF_PAGES;
401
402 for (i = 0; i < bp->b_page_count; i++) {
403 struct page *page;
404 uint retries = 0;
405retry:
406 page = alloc_page(gfp_mask);
407 if (unlikely(page == NULL)) {
408 if (flags & XBF_READ_AHEAD) {
409 bp->b_page_count = i;
410 error = -ENOMEM;
411 goto out_free_pages;
412 }
413
414 /*
415 * This could deadlock.
416 *
417 * But until all the XFS lowlevel code is revamped to
418 * handle buffer allocation failures we can't do much.
419 */
420 if (!(++retries % 100))
421 xfs_err(NULL,
422 "%s(%u) possible memory allocation deadlock in %s (mode:0x%x)",
423 current->comm, current->pid,
424 __func__, gfp_mask);
425
426 XFS_STATS_INC(bp->b_target->bt_mount, xb_page_retries);
427 congestion_wait(BLK_RW_ASYNC, HZ/50);
428 goto retry;
429 }
430
431 XFS_STATS_INC(bp->b_target->bt_mount, xb_page_found);
432
433 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
434 size -= nbytes;
435 bp->b_pages[i] = page;
436 offset = 0;
437 }
438 return 0;
439
440out_free_pages:
441 for (i = 0; i < bp->b_page_count; i++)
442 __free_page(bp->b_pages[i]);
443 bp->b_flags &= ~_XBF_PAGES;
444 return error;
445}
446
447/*
448 * Map buffer into kernel address-space if necessary.
449 */
450STATIC int
451_xfs_buf_map_pages(
452 xfs_buf_t *bp,
453 uint flags)
454{
455 ASSERT(bp->b_flags & _XBF_PAGES);
456 if (bp->b_page_count == 1) {
457 /* A single page buffer is always mappable */
458 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
459 } else if (flags & XBF_UNMAPPED) {
460 bp->b_addr = NULL;
461 } else {
462 int retried = 0;
463 unsigned nofs_flag;
464
465 /*
466 * vm_map_ram() will allocate auxillary structures (e.g.
467 * pagetables) with GFP_KERNEL, yet we are likely to be under
468 * GFP_NOFS context here. Hence we need to tell memory reclaim
469 * that we are in such a context via PF_MEMALLOC_NOFS to prevent
470 * memory reclaim re-entering the filesystem here and
471 * potentially deadlocking.
472 */
473 nofs_flag = memalloc_nofs_save();
474 do {
475 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
476 -1, PAGE_KERNEL);
477 if (bp->b_addr)
478 break;
479 vm_unmap_aliases();
480 } while (retried++ <= 1);
481 memalloc_nofs_restore(nofs_flag);
482
483 if (!bp->b_addr)
484 return -ENOMEM;
485 bp->b_addr += bp->b_offset;
486 }
487
488 return 0;
489}
490
491/*
492 * Finding and Reading Buffers
493 */
494static int
495_xfs_buf_obj_cmp(
496 struct rhashtable_compare_arg *arg,
497 const void *obj)
498{
499 const struct xfs_buf_map *map = arg->key;
500 const struct xfs_buf *bp = obj;
501
502 /*
503 * The key hashing in the lookup path depends on the key being the
504 * first element of the compare_arg, make sure to assert this.
505 */
506 BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0);
507
508 if (bp->b_bn != map->bm_bn)
509 return 1;
510
511 if (unlikely(bp->b_length != map->bm_len)) {
512 /*
513 * found a block number match. If the range doesn't
514 * match, the only way this is allowed is if the buffer
515 * in the cache is stale and the transaction that made
516 * it stale has not yet committed. i.e. we are
517 * reallocating a busy extent. Skip this buffer and
518 * continue searching for an exact match.
519 */
520 ASSERT(bp->b_flags & XBF_STALE);
521 return 1;
522 }
523 return 0;
524}
525
526static const struct rhashtable_params xfs_buf_hash_params = {
527 .min_size = 32, /* empty AGs have minimal footprint */
528 .nelem_hint = 16,
529 .key_len = sizeof(xfs_daddr_t),
530 .key_offset = offsetof(struct xfs_buf, b_bn),
531 .head_offset = offsetof(struct xfs_buf, b_rhash_head),
532 .automatic_shrinking = true,
533 .obj_cmpfn = _xfs_buf_obj_cmp,
534};
535
536int
537xfs_buf_hash_init(
538 struct xfs_perag *pag)
539{
540 spin_lock_init(&pag->pag_buf_lock);
541 return rhashtable_init(&pag->pag_buf_hash, &xfs_buf_hash_params);
542}
543
544void
545xfs_buf_hash_destroy(
546 struct xfs_perag *pag)
547{
548 rhashtable_destroy(&pag->pag_buf_hash);
549}
550
551/*
552 * Look up, and creates if absent, a lockable buffer for
553 * a given range of an inode. The buffer is returned
554 * locked. No I/O is implied by this call.
555 */
556xfs_buf_t *
557_xfs_buf_find(
558 struct xfs_buftarg *btp,
559 struct xfs_buf_map *map,
560 int nmaps,
561 xfs_buf_flags_t flags,
562 xfs_buf_t *new_bp)
563{
564 struct xfs_perag *pag;
565 xfs_buf_t *bp;
566 struct xfs_buf_map cmap = { .bm_bn = map[0].bm_bn };
567 xfs_daddr_t eofs;
568 int i;
569
570 for (i = 0; i < nmaps; i++)
571 cmap.bm_len += map[i].bm_len;
572
573 /* Check for IOs smaller than the sector size / not sector aligned */
574 ASSERT(!(BBTOB(cmap.bm_len) < btp->bt_meta_sectorsize));
575 ASSERT(!(BBTOB(cmap.bm_bn) & (xfs_off_t)btp->bt_meta_sectormask));
576
577 /*
578 * Corrupted block numbers can get through to here, unfortunately, so we
579 * have to check that the buffer falls within the filesystem bounds.
580 */
581 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
582 if (cmap.bm_bn < 0 || cmap.bm_bn >= eofs) {
583 /*
584 * XXX (dgc): we should really be returning -EFSCORRUPTED here,
585 * but none of the higher level infrastructure supports
586 * returning a specific error on buffer lookup failures.
587 */
588 xfs_alert(btp->bt_mount,
589 "%s: daddr 0x%llx out of range, EOFS 0x%llx",
590 __func__, cmap.bm_bn, eofs);
591 WARN_ON(1);
592 return NULL;
593 }
594
595 pag = xfs_perag_get(btp->bt_mount,
596 xfs_daddr_to_agno(btp->bt_mount, cmap.bm_bn));
597
598 spin_lock(&pag->pag_buf_lock);
599 bp = rhashtable_lookup_fast(&pag->pag_buf_hash, &cmap,
600 xfs_buf_hash_params);
601 if (bp) {
602 atomic_inc(&bp->b_hold);
603 goto found;
604 }
605
606 /* No match found */
607 if (new_bp) {
608 /* the buffer keeps the perag reference until it is freed */
609 new_bp->b_pag = pag;
610 rhashtable_insert_fast(&pag->pag_buf_hash,
611 &new_bp->b_rhash_head,
612 xfs_buf_hash_params);
613 spin_unlock(&pag->pag_buf_lock);
614 } else {
615 XFS_STATS_INC(btp->bt_mount, xb_miss_locked);
616 spin_unlock(&pag->pag_buf_lock);
617 xfs_perag_put(pag);
618 }
619 return new_bp;
620
621found:
622 spin_unlock(&pag->pag_buf_lock);
623 xfs_perag_put(pag);
624
625 if (!xfs_buf_trylock(bp)) {
626 if (flags & XBF_TRYLOCK) {
627 xfs_buf_rele(bp);
628 XFS_STATS_INC(btp->bt_mount, xb_busy_locked);
629 return NULL;
630 }
631 xfs_buf_lock(bp);
632 XFS_STATS_INC(btp->bt_mount, xb_get_locked_waited);
633 }
634
635 /*
636 * if the buffer is stale, clear all the external state associated with
637 * it. We need to keep flags such as how we allocated the buffer memory
638 * intact here.
639 */
640 if (bp->b_flags & XBF_STALE) {
641 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
642 ASSERT(bp->b_iodone == NULL);
643 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
644 bp->b_ops = NULL;
645 }
646
647 trace_xfs_buf_find(bp, flags, _RET_IP_);
648 XFS_STATS_INC(btp->bt_mount, xb_get_locked);
649 return bp;
650}
651
652/*
653 * Assembles a buffer covering the specified range. The code is optimised for
654 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
655 * more hits than misses.
656 */
657struct xfs_buf *
658xfs_buf_get_map(
659 struct xfs_buftarg *target,
660 struct xfs_buf_map *map,
661 int nmaps,
662 xfs_buf_flags_t flags)
663{
664 struct xfs_buf *bp;
665 struct xfs_buf *new_bp;
666 int error = 0;
667
668 bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
669 if (likely(bp))
670 goto found;
671
672 new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
673 if (unlikely(!new_bp))
674 return NULL;
675
676 error = xfs_buf_allocate_memory(new_bp, flags);
677 if (error) {
678 xfs_buf_free(new_bp);
679 return NULL;
680 }
681
682 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
683 if (!bp) {
684 xfs_buf_free(new_bp);
685 return NULL;
686 }
687
688 if (bp != new_bp)
689 xfs_buf_free(new_bp);
690
691found:
692 if (!bp->b_addr) {
693 error = _xfs_buf_map_pages(bp, flags);
694 if (unlikely(error)) {
695 xfs_warn(target->bt_mount,
696 "%s: failed to map pagesn", __func__);
697 xfs_buf_relse(bp);
698 return NULL;
699 }
700 }
701
702 /*
703 * Clear b_error if this is a lookup from a caller that doesn't expect
704 * valid data to be found in the buffer.
705 */
706 if (!(flags & XBF_READ))
707 xfs_buf_ioerror(bp, 0);
708
709 XFS_STATS_INC(target->bt_mount, xb_get);
710 trace_xfs_buf_get(bp, flags, _RET_IP_);
711 return bp;
712}
713
714STATIC int
715_xfs_buf_read(
716 xfs_buf_t *bp,
717 xfs_buf_flags_t flags)
718{
719 ASSERT(!(flags & XBF_WRITE));
720 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
721
722 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
723 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
724
725 if (flags & XBF_ASYNC) {
726 xfs_buf_submit(bp);
727 return 0;
728 }
729 return xfs_buf_submit_wait(bp);
730}
731
732xfs_buf_t *
733xfs_buf_read_map(
734 struct xfs_buftarg *target,
735 struct xfs_buf_map *map,
736 int nmaps,
737 xfs_buf_flags_t flags,
738 const struct xfs_buf_ops *ops)
739{
740 struct xfs_buf *bp;
741
742 flags |= XBF_READ;
743
744 bp = xfs_buf_get_map(target, map, nmaps, flags);
745 if (bp) {
746 trace_xfs_buf_read(bp, flags, _RET_IP_);
747
748 if (!(bp->b_flags & XBF_DONE)) {
749 XFS_STATS_INC(target->bt_mount, xb_get_read);
750 bp->b_ops = ops;
751 _xfs_buf_read(bp, flags);
752 } else if (flags & XBF_ASYNC) {
753 /*
754 * Read ahead call which is already satisfied,
755 * drop the buffer
756 */
757 xfs_buf_relse(bp);
758 return NULL;
759 } else {
760 /* We do not want read in the flags */
761 bp->b_flags &= ~XBF_READ;
762 }
763 }
764
765 return bp;
766}
767
768/*
769 * If we are not low on memory then do the readahead in a deadlock
770 * safe manner.
771 */
772void
773xfs_buf_readahead_map(
774 struct xfs_buftarg *target,
775 struct xfs_buf_map *map,
776 int nmaps,
777 const struct xfs_buf_ops *ops)
778{
779 if (bdi_read_congested(target->bt_bdev->bd_bdi))
780 return;
781
782 xfs_buf_read_map(target, map, nmaps,
783 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops);
784}
785
786/*
787 * Read an uncached buffer from disk. Allocates and returns a locked
788 * buffer containing the disk contents or nothing.
789 */
790int
791xfs_buf_read_uncached(
792 struct xfs_buftarg *target,
793 xfs_daddr_t daddr,
794 size_t numblks,
795 int flags,
796 struct xfs_buf **bpp,
797 const struct xfs_buf_ops *ops)
798{
799 struct xfs_buf *bp;
800
801 *bpp = NULL;
802
803 bp = xfs_buf_get_uncached(target, numblks, flags);
804 if (!bp)
805 return -ENOMEM;
806
807 /* set up the buffer for a read IO */
808 ASSERT(bp->b_map_count == 1);
809 bp->b_bn = XFS_BUF_DADDR_NULL; /* always null for uncached buffers */
810 bp->b_maps[0].bm_bn = daddr;
811 bp->b_flags |= XBF_READ;
812 bp->b_ops = ops;
813
814 xfs_buf_submit_wait(bp);
815 if (bp->b_error) {
816 int error = bp->b_error;
817 xfs_buf_relse(bp);
818 return error;
819 }
820
821 *bpp = bp;
822 return 0;
823}
824
825/*
826 * Return a buffer allocated as an empty buffer and associated to external
827 * memory via xfs_buf_associate_memory() back to it's empty state.
828 */
829void
830xfs_buf_set_empty(
831 struct xfs_buf *bp,
832 size_t numblks)
833{
834 if (bp->b_pages)
835 _xfs_buf_free_pages(bp);
836
837 bp->b_pages = NULL;
838 bp->b_page_count = 0;
839 bp->b_addr = NULL;
840 bp->b_length = numblks;
841 bp->b_io_length = numblks;
842
843 ASSERT(bp->b_map_count == 1);
844 bp->b_bn = XFS_BUF_DADDR_NULL;
845 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
846 bp->b_maps[0].bm_len = bp->b_length;
847}
848
849static inline struct page *
850mem_to_page(
851 void *addr)
852{
853 if ((!is_vmalloc_addr(addr))) {
854 return virt_to_page(addr);
855 } else {
856 return vmalloc_to_page(addr);
857 }
858}
859
860int
861xfs_buf_associate_memory(
862 xfs_buf_t *bp,
863 void *mem,
864 size_t len)
865{
866 int rval;
867 int i = 0;
868 unsigned long pageaddr;
869 unsigned long offset;
870 size_t buflen;
871 int page_count;
872
873 pageaddr = (unsigned long)mem & PAGE_MASK;
874 offset = (unsigned long)mem - pageaddr;
875 buflen = PAGE_ALIGN(len + offset);
876 page_count = buflen >> PAGE_SHIFT;
877
878 /* Free any previous set of page pointers */
879 if (bp->b_pages)
880 _xfs_buf_free_pages(bp);
881
882 bp->b_pages = NULL;
883 bp->b_addr = mem;
884
885 rval = _xfs_buf_get_pages(bp, page_count);
886 if (rval)
887 return rval;
888
889 bp->b_offset = offset;
890
891 for (i = 0; i < bp->b_page_count; i++) {
892 bp->b_pages[i] = mem_to_page((void *)pageaddr);
893 pageaddr += PAGE_SIZE;
894 }
895
896 bp->b_io_length = BTOBB(len);
897 bp->b_length = BTOBB(buflen);
898
899 return 0;
900}
901
902xfs_buf_t *
903xfs_buf_get_uncached(
904 struct xfs_buftarg *target,
905 size_t numblks,
906 int flags)
907{
908 unsigned long page_count;
909 int error, i;
910 struct xfs_buf *bp;
911 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
912
913 /* flags might contain irrelevant bits, pass only what we care about */
914 bp = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT);
915 if (unlikely(bp == NULL))
916 goto fail;
917
918 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
919 error = _xfs_buf_get_pages(bp, page_count);
920 if (error)
921 goto fail_free_buf;
922
923 for (i = 0; i < page_count; i++) {
924 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
925 if (!bp->b_pages[i])
926 goto fail_free_mem;
927 }
928 bp->b_flags |= _XBF_PAGES;
929
930 error = _xfs_buf_map_pages(bp, 0);
931 if (unlikely(error)) {
932 xfs_warn(target->bt_mount,
933 "%s: failed to map pages", __func__);
934 goto fail_free_mem;
935 }
936
937 trace_xfs_buf_get_uncached(bp, _RET_IP_);
938 return bp;
939
940 fail_free_mem:
941 while (--i >= 0)
942 __free_page(bp->b_pages[i]);
943 _xfs_buf_free_pages(bp);
944 fail_free_buf:
945 xfs_buf_free_maps(bp);
946 kmem_zone_free(xfs_buf_zone, bp);
947 fail:
948 return NULL;
949}
950
951/*
952 * Increment reference count on buffer, to hold the buffer concurrently
953 * with another thread which may release (free) the buffer asynchronously.
954 * Must hold the buffer already to call this function.
955 */
956void
957xfs_buf_hold(
958 xfs_buf_t *bp)
959{
960 trace_xfs_buf_hold(bp, _RET_IP_);
961 atomic_inc(&bp->b_hold);
962}
963
964/*
965 * Release a hold on the specified buffer. If the hold count is 1, the buffer is
966 * placed on LRU or freed (depending on b_lru_ref).
967 */
968void
969xfs_buf_rele(
970 xfs_buf_t *bp)
971{
972 struct xfs_perag *pag = bp->b_pag;
973 bool release;
974 bool freebuf = false;
975
976 trace_xfs_buf_rele(bp, _RET_IP_);
977
978 if (!pag) {
979 ASSERT(list_empty(&bp->b_lru));
980 if (atomic_dec_and_test(&bp->b_hold)) {
981 xfs_buf_ioacct_dec(bp);
982 xfs_buf_free(bp);
983 }
984 return;
985 }
986
987 ASSERT(atomic_read(&bp->b_hold) > 0);
988
989 release = atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock);
990 spin_lock(&bp->b_lock);
991 if (!release) {
992 /*
993 * Drop the in-flight state if the buffer is already on the LRU
994 * and it holds the only reference. This is racy because we
995 * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT
996 * ensures the decrement occurs only once per-buf.
997 */
998 if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru))
999 __xfs_buf_ioacct_dec(bp);
1000 goto out_unlock;
1001 }
1002
1003 /* the last reference has been dropped ... */
1004 __xfs_buf_ioacct_dec(bp);
1005 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
1006 /*
1007 * If the buffer is added to the LRU take a new reference to the
1008 * buffer for the LRU and clear the (now stale) dispose list
1009 * state flag
1010 */
1011 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) {
1012 bp->b_state &= ~XFS_BSTATE_DISPOSE;
1013 atomic_inc(&bp->b_hold);
1014 }
1015 spin_unlock(&pag->pag_buf_lock);
1016 } else {
1017 /*
1018 * most of the time buffers will already be removed from the
1019 * LRU, so optimise that case by checking for the
1020 * XFS_BSTATE_DISPOSE flag indicating the last list the buffer
1021 * was on was the disposal list
1022 */
1023 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
1024 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru);
1025 } else {
1026 ASSERT(list_empty(&bp->b_lru));
1027 }
1028
1029 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1030 rhashtable_remove_fast(&pag->pag_buf_hash, &bp->b_rhash_head,
1031 xfs_buf_hash_params);
1032 spin_unlock(&pag->pag_buf_lock);
1033 xfs_perag_put(pag);
1034 freebuf = true;
1035 }
1036
1037out_unlock:
1038 spin_unlock(&bp->b_lock);
1039
1040 if (freebuf)
1041 xfs_buf_free(bp);
1042}
1043
1044
1045/*
1046 * Lock a buffer object, if it is not already locked.
1047 *
1048 * If we come across a stale, pinned, locked buffer, we know that we are
1049 * being asked to lock a buffer that has been reallocated. Because it is
1050 * pinned, we know that the log has not been pushed to disk and hence it
1051 * will still be locked. Rather than continuing to have trylock attempts
1052 * fail until someone else pushes the log, push it ourselves before
1053 * returning. This means that the xfsaild will not get stuck trying
1054 * to push on stale inode buffers.
1055 */
1056int
1057xfs_buf_trylock(
1058 struct xfs_buf *bp)
1059{
1060 int locked;
1061
1062 locked = down_trylock(&bp->b_sema) == 0;
1063 if (locked) {
1064 XB_SET_OWNER(bp);
1065 trace_xfs_buf_trylock(bp, _RET_IP_);
1066 } else {
1067 trace_xfs_buf_trylock_fail(bp, _RET_IP_);
1068 }
1069 return locked;
1070}
1071
1072/*
1073 * Lock a buffer object.
1074 *
1075 * If we come across a stale, pinned, locked buffer, we know that we
1076 * are being asked to lock a buffer that has been reallocated. Because
1077 * it is pinned, we know that the log has not been pushed to disk and
1078 * hence it will still be locked. Rather than sleeping until someone
1079 * else pushes the log, push it ourselves before trying to get the lock.
1080 */
1081void
1082xfs_buf_lock(
1083 struct xfs_buf *bp)
1084{
1085 trace_xfs_buf_lock(bp, _RET_IP_);
1086
1087 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
1088 xfs_log_force(bp->b_target->bt_mount, 0);
1089 down(&bp->b_sema);
1090 XB_SET_OWNER(bp);
1091
1092 trace_xfs_buf_lock_done(bp, _RET_IP_);
1093}
1094
1095void
1096xfs_buf_unlock(
1097 struct xfs_buf *bp)
1098{
1099 ASSERT(xfs_buf_islocked(bp));
1100
1101 XB_CLEAR_OWNER(bp);
1102 up(&bp->b_sema);
1103
1104 trace_xfs_buf_unlock(bp, _RET_IP_);
1105}
1106
1107STATIC void
1108xfs_buf_wait_unpin(
1109 xfs_buf_t *bp)
1110{
1111 DECLARE_WAITQUEUE (wait, current);
1112
1113 if (atomic_read(&bp->b_pin_count) == 0)
1114 return;
1115
1116 add_wait_queue(&bp->b_waiters, &wait);
1117 for (;;) {
1118 set_current_state(TASK_UNINTERRUPTIBLE);
1119 if (atomic_read(&bp->b_pin_count) == 0)
1120 break;
1121 io_schedule();
1122 }
1123 remove_wait_queue(&bp->b_waiters, &wait);
1124 set_current_state(TASK_RUNNING);
1125}
1126
1127/*
1128 * Buffer Utility Routines
1129 */
1130
1131void
1132xfs_buf_ioend(
1133 struct xfs_buf *bp)
1134{
1135 bool read = bp->b_flags & XBF_READ;
1136
1137 trace_xfs_buf_iodone(bp, _RET_IP_);
1138
1139 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1140
1141 /*
1142 * Pull in IO completion errors now. We are guaranteed to be running
1143 * single threaded, so we don't need the lock to read b_io_error.
1144 */
1145 if (!bp->b_error && bp->b_io_error)
1146 xfs_buf_ioerror(bp, bp->b_io_error);
1147
1148 /* Only validate buffers that were read without errors */
1149 if (read && !bp->b_error && bp->b_ops) {
1150 ASSERT(!bp->b_iodone);
1151 bp->b_ops->verify_read(bp);
1152 }
1153
1154 if (!bp->b_error)
1155 bp->b_flags |= XBF_DONE;
1156
1157 if (bp->b_iodone)
1158 (*(bp->b_iodone))(bp);
1159 else if (bp->b_flags & XBF_ASYNC)
1160 xfs_buf_relse(bp);
1161 else
1162 complete(&bp->b_iowait);
1163}
1164
1165static void
1166xfs_buf_ioend_work(
1167 struct work_struct *work)
1168{
1169 struct xfs_buf *bp =
1170 container_of(work, xfs_buf_t, b_ioend_work);
1171
1172 xfs_buf_ioend(bp);
1173}
1174
1175static void
1176xfs_buf_ioend_async(
1177 struct xfs_buf *bp)
1178{
1179 INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
1180 queue_work(bp->b_ioend_wq, &bp->b_ioend_work);
1181}
1182
1183void
1184__xfs_buf_ioerror(
1185 xfs_buf_t *bp,
1186 int error,
1187 xfs_failaddr_t failaddr)
1188{
1189 ASSERT(error <= 0 && error >= -1000);
1190 bp->b_error = error;
1191 trace_xfs_buf_ioerror(bp, error, failaddr);
1192}
1193
1194void
1195xfs_buf_ioerror_alert(
1196 struct xfs_buf *bp,
1197 const char *func)
1198{
1199 xfs_alert(bp->b_target->bt_mount,
1200"metadata I/O error in \"%s\" at daddr 0x%llx len %d error %d",
1201 func, (uint64_t)XFS_BUF_ADDR(bp), bp->b_length,
1202 -bp->b_error);
1203}
1204
1205int
1206xfs_bwrite(
1207 struct xfs_buf *bp)
1208{
1209 int error;
1210
1211 ASSERT(xfs_buf_islocked(bp));
1212
1213 bp->b_flags |= XBF_WRITE;
1214 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1215 XBF_WRITE_FAIL | XBF_DONE);
1216
1217 error = xfs_buf_submit_wait(bp);
1218 if (error) {
1219 xfs_force_shutdown(bp->b_target->bt_mount,
1220 SHUTDOWN_META_IO_ERROR);
1221 }
1222 return error;
1223}
1224
1225static void
1226xfs_buf_bio_end_io(
1227 struct bio *bio)
1228{
1229 struct xfs_buf *bp = (struct xfs_buf *)bio->bi_private;
1230
1231 /*
1232 * don't overwrite existing errors - otherwise we can lose errors on
1233 * buffers that require multiple bios to complete.
1234 */
1235 if (bio->bi_status) {
1236 int error = blk_status_to_errno(bio->bi_status);
1237
1238 cmpxchg(&bp->b_io_error, 0, error);
1239 }
1240
1241 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1242 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1243
1244 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1245 xfs_buf_ioend_async(bp);
1246 bio_put(bio);
1247}
1248
1249static void
1250xfs_buf_ioapply_map(
1251 struct xfs_buf *bp,
1252 int map,
1253 int *buf_offset,
1254 int *count,
1255 int op,
1256 int op_flags)
1257{
1258 int page_index;
1259 int total_nr_pages = bp->b_page_count;
1260 int nr_pages;
1261 struct bio *bio;
1262 sector_t sector = bp->b_maps[map].bm_bn;
1263 int size;
1264 int offset;
1265
1266 /* skip the pages in the buffer before the start offset */
1267 page_index = 0;
1268 offset = *buf_offset;
1269 while (offset >= PAGE_SIZE) {
1270 page_index++;
1271 offset -= PAGE_SIZE;
1272 }
1273
1274 /*
1275 * Limit the IO size to the length of the current vector, and update the
1276 * remaining IO count for the next time around.
1277 */
1278 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1279 *count -= size;
1280 *buf_offset += size;
1281
1282next_chunk:
1283 atomic_inc(&bp->b_io_remaining);
1284 nr_pages = min(total_nr_pages, BIO_MAX_PAGES);
1285
1286 bio = bio_alloc(GFP_NOIO, nr_pages);
1287 bio_set_dev(bio, bp->b_target->bt_bdev);
1288 bio->bi_iter.bi_sector = sector;
1289 bio->bi_end_io = xfs_buf_bio_end_io;
1290 bio->bi_private = bp;
1291 bio_set_op_attrs(bio, op, op_flags);
1292
1293 for (; size && nr_pages; nr_pages--, page_index++) {
1294 int rbytes, nbytes = PAGE_SIZE - offset;
1295
1296 if (nbytes > size)
1297 nbytes = size;
1298
1299 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1300 offset);
1301 if (rbytes < nbytes)
1302 break;
1303
1304 offset = 0;
1305 sector += BTOBB(nbytes);
1306 size -= nbytes;
1307 total_nr_pages--;
1308 }
1309
1310 if (likely(bio->bi_iter.bi_size)) {
1311 if (xfs_buf_is_vmapped(bp)) {
1312 flush_kernel_vmap_range(bp->b_addr,
1313 xfs_buf_vmap_len(bp));
1314 }
1315 submit_bio(bio);
1316 if (size)
1317 goto next_chunk;
1318 } else {
1319 /*
1320 * This is guaranteed not to be the last io reference count
1321 * because the caller (xfs_buf_submit) holds a count itself.
1322 */
1323 atomic_dec(&bp->b_io_remaining);
1324 xfs_buf_ioerror(bp, -EIO);
1325 bio_put(bio);
1326 }
1327
1328}
1329
1330STATIC void
1331_xfs_buf_ioapply(
1332 struct xfs_buf *bp)
1333{
1334 struct blk_plug plug;
1335 int op;
1336 int op_flags = 0;
1337 int offset;
1338 int size;
1339 int i;
1340
1341 /*
1342 * Make sure we capture only current IO errors rather than stale errors
1343 * left over from previous use of the buffer (e.g. failed readahead).
1344 */
1345 bp->b_error = 0;
1346
1347 /*
1348 * Initialize the I/O completion workqueue if we haven't yet or the
1349 * submitter has not opted to specify a custom one.
1350 */
1351 if (!bp->b_ioend_wq)
1352 bp->b_ioend_wq = bp->b_target->bt_mount->m_buf_workqueue;
1353
1354 if (bp->b_flags & XBF_WRITE) {
1355 op = REQ_OP_WRITE;
1356 if (bp->b_flags & XBF_SYNCIO)
1357 op_flags = REQ_SYNC;
1358 if (bp->b_flags & XBF_FUA)
1359 op_flags |= REQ_FUA;
1360 if (bp->b_flags & XBF_FLUSH)
1361 op_flags |= REQ_PREFLUSH;
1362
1363 /*
1364 * Run the write verifier callback function if it exists. If
1365 * this function fails it will mark the buffer with an error and
1366 * the IO should not be dispatched.
1367 */
1368 if (bp->b_ops) {
1369 bp->b_ops->verify_write(bp);
1370 if (bp->b_error) {
1371 xfs_force_shutdown(bp->b_target->bt_mount,
1372 SHUTDOWN_CORRUPT_INCORE);
1373 return;
1374 }
1375 } else if (bp->b_bn != XFS_BUF_DADDR_NULL) {
1376 struct xfs_mount *mp = bp->b_target->bt_mount;
1377
1378 /*
1379 * non-crc filesystems don't attach verifiers during
1380 * log recovery, so don't warn for such filesystems.
1381 */
1382 if (xfs_sb_version_hascrc(&mp->m_sb)) {
1383 xfs_warn(mp,
1384 "%s: no buf ops on daddr 0x%llx len %d",
1385 __func__, bp->b_bn, bp->b_length);
1386 xfs_hex_dump(bp->b_addr,
1387 XFS_CORRUPTION_DUMP_LEN);
1388 dump_stack();
1389 }
1390 }
1391 } else if (bp->b_flags & XBF_READ_AHEAD) {
1392 op = REQ_OP_READ;
1393 op_flags = REQ_RAHEAD;
1394 } else {
1395 op = REQ_OP_READ;
1396 }
1397
1398 /* we only use the buffer cache for meta-data */
1399 op_flags |= REQ_META;
1400
1401 /*
1402 * Walk all the vectors issuing IO on them. Set up the initial offset
1403 * into the buffer and the desired IO size before we start -
1404 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1405 * subsequent call.
1406 */
1407 offset = bp->b_offset;
1408 size = BBTOB(bp->b_io_length);
1409 blk_start_plug(&plug);
1410 for (i = 0; i < bp->b_map_count; i++) {
1411 xfs_buf_ioapply_map(bp, i, &offset, &size, op, op_flags);
1412 if (bp->b_error)
1413 break;
1414 if (size <= 0)
1415 break; /* all done */
1416 }
1417 blk_finish_plug(&plug);
1418}
1419
1420/*
1421 * Asynchronous IO submission path. This transfers the buffer lock ownership and
1422 * the current reference to the IO. It is not safe to reference the buffer after
1423 * a call to this function unless the caller holds an additional reference
1424 * itself.
1425 */
1426void
1427xfs_buf_submit(
1428 struct xfs_buf *bp)
1429{
1430 trace_xfs_buf_submit(bp, _RET_IP_);
1431
1432 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1433 ASSERT(bp->b_flags & XBF_ASYNC);
1434
1435 /* on shutdown we stale and complete the buffer immediately */
1436 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1437 xfs_buf_ioerror(bp, -EIO);
1438 bp->b_flags &= ~XBF_DONE;
1439 xfs_buf_stale(bp);
1440 xfs_buf_ioend(bp);
1441 return;
1442 }
1443
1444 if (bp->b_flags & XBF_WRITE)
1445 xfs_buf_wait_unpin(bp);
1446
1447 /* clear the internal error state to avoid spurious errors */
1448 bp->b_io_error = 0;
1449
1450 /*
1451 * The caller's reference is released during I/O completion.
1452 * This occurs some time after the last b_io_remaining reference is
1453 * released, so after we drop our Io reference we have to have some
1454 * other reference to ensure the buffer doesn't go away from underneath
1455 * us. Take a direct reference to ensure we have safe access to the
1456 * buffer until we are finished with it.
1457 */
1458 xfs_buf_hold(bp);
1459
1460 /*
1461 * Set the count to 1 initially, this will stop an I/O completion
1462 * callout which happens before we have started all the I/O from calling
1463 * xfs_buf_ioend too early.
1464 */
1465 atomic_set(&bp->b_io_remaining, 1);
1466 xfs_buf_ioacct_inc(bp);
1467 _xfs_buf_ioapply(bp);
1468
1469 /*
1470 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1471 * reference we took above. If we drop it to zero, run completion so
1472 * that we don't return to the caller with completion still pending.
1473 */
1474 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1475 if (bp->b_error)
1476 xfs_buf_ioend(bp);
1477 else
1478 xfs_buf_ioend_async(bp);
1479 }
1480
1481 xfs_buf_rele(bp);
1482 /* Note: it is not safe to reference bp now we've dropped our ref */
1483}
1484
1485/*
1486 * Synchronous buffer IO submission path, read or write.
1487 */
1488int
1489xfs_buf_submit_wait(
1490 struct xfs_buf *bp)
1491{
1492 int error;
1493
1494 trace_xfs_buf_submit_wait(bp, _RET_IP_);
1495
1496 ASSERT(!(bp->b_flags & (_XBF_DELWRI_Q | XBF_ASYNC)));
1497
1498 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1499 xfs_buf_ioerror(bp, -EIO);
1500 xfs_buf_stale(bp);
1501 bp->b_flags &= ~XBF_DONE;
1502 return -EIO;
1503 }
1504
1505 if (bp->b_flags & XBF_WRITE)
1506 xfs_buf_wait_unpin(bp);
1507
1508 /* clear the internal error state to avoid spurious errors */
1509 bp->b_io_error = 0;
1510
1511 /*
1512 * For synchronous IO, the IO does not inherit the submitters reference
1513 * count, nor the buffer lock. Hence we cannot release the reference we
1514 * are about to take until we've waited for all IO completion to occur,
1515 * including any xfs_buf_ioend_async() work that may be pending.
1516 */
1517 xfs_buf_hold(bp);
1518
1519 /*
1520 * Set the count to 1 initially, this will stop an I/O completion
1521 * callout which happens before we have started all the I/O from calling
1522 * xfs_buf_ioend too early.
1523 */
1524 atomic_set(&bp->b_io_remaining, 1);
1525 _xfs_buf_ioapply(bp);
1526
1527 /*
1528 * make sure we run completion synchronously if it raced with us and is
1529 * already complete.
1530 */
1531 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1532 xfs_buf_ioend(bp);
1533
1534 /* wait for completion before gathering the error from the buffer */
1535 trace_xfs_buf_iowait(bp, _RET_IP_);
1536 wait_for_completion(&bp->b_iowait);
1537 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1538 error = bp->b_error;
1539
1540 /*
1541 * all done now, we can release the hold that keeps the buffer
1542 * referenced for the entire IO.
1543 */
1544 xfs_buf_rele(bp);
1545 return error;
1546}
1547
1548void *
1549xfs_buf_offset(
1550 struct xfs_buf *bp,
1551 size_t offset)
1552{
1553 struct page *page;
1554
1555 if (bp->b_addr)
1556 return bp->b_addr + offset;
1557
1558 offset += bp->b_offset;
1559 page = bp->b_pages[offset >> PAGE_SHIFT];
1560 return page_address(page) + (offset & (PAGE_SIZE-1));
1561}
1562
1563/*
1564 * Move data into or out of a buffer.
1565 */
1566void
1567xfs_buf_iomove(
1568 xfs_buf_t *bp, /* buffer to process */
1569 size_t boff, /* starting buffer offset */
1570 size_t bsize, /* length to copy */
1571 void *data, /* data address */
1572 xfs_buf_rw_t mode) /* read/write/zero flag */
1573{
1574 size_t bend;
1575
1576 bend = boff + bsize;
1577 while (boff < bend) {
1578 struct page *page;
1579 int page_index, page_offset, csize;
1580
1581 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1582 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1583 page = bp->b_pages[page_index];
1584 csize = min_t(size_t, PAGE_SIZE - page_offset,
1585 BBTOB(bp->b_io_length) - boff);
1586
1587 ASSERT((csize + page_offset) <= PAGE_SIZE);
1588
1589 switch (mode) {
1590 case XBRW_ZERO:
1591 memset(page_address(page) + page_offset, 0, csize);
1592 break;
1593 case XBRW_READ:
1594 memcpy(data, page_address(page) + page_offset, csize);
1595 break;
1596 case XBRW_WRITE:
1597 memcpy(page_address(page) + page_offset, data, csize);
1598 }
1599
1600 boff += csize;
1601 data += csize;
1602 }
1603}
1604
1605/*
1606 * Handling of buffer targets (buftargs).
1607 */
1608
1609/*
1610 * Wait for any bufs with callbacks that have been submitted but have not yet
1611 * returned. These buffers will have an elevated hold count, so wait on those
1612 * while freeing all the buffers only held by the LRU.
1613 */
1614static enum lru_status
1615xfs_buftarg_wait_rele(
1616 struct list_head *item,
1617 struct list_lru_one *lru,
1618 spinlock_t *lru_lock,
1619 void *arg)
1620
1621{
1622 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1623 struct list_head *dispose = arg;
1624
1625 if (atomic_read(&bp->b_hold) > 1) {
1626 /* need to wait, so skip it this pass */
1627 trace_xfs_buf_wait_buftarg(bp, _RET_IP_);
1628 return LRU_SKIP;
1629 }
1630 if (!spin_trylock(&bp->b_lock))
1631 return LRU_SKIP;
1632
1633 /*
1634 * clear the LRU reference count so the buffer doesn't get
1635 * ignored in xfs_buf_rele().
1636 */
1637 atomic_set(&bp->b_lru_ref, 0);
1638 bp->b_state |= XFS_BSTATE_DISPOSE;
1639 list_lru_isolate_move(lru, item, dispose);
1640 spin_unlock(&bp->b_lock);
1641 return LRU_REMOVED;
1642}
1643
1644void
1645xfs_wait_buftarg(
1646 struct xfs_buftarg *btp)
1647{
1648 LIST_HEAD(dispose);
1649 int loop = 0;
1650
1651 /*
1652 * First wait on the buftarg I/O count for all in-flight buffers to be
1653 * released. This is critical as new buffers do not make the LRU until
1654 * they are released.
1655 *
1656 * Next, flush the buffer workqueue to ensure all completion processing
1657 * has finished. Just waiting on buffer locks is not sufficient for
1658 * async IO as the reference count held over IO is not released until
1659 * after the buffer lock is dropped. Hence we need to ensure here that
1660 * all reference counts have been dropped before we start walking the
1661 * LRU list.
1662 */
1663 while (percpu_counter_sum(&btp->bt_io_count))
1664 delay(100);
1665 flush_workqueue(btp->bt_mount->m_buf_workqueue);
1666
1667 /* loop until there is nothing left on the lru list. */
1668 while (list_lru_count(&btp->bt_lru)) {
1669 list_lru_walk(&btp->bt_lru, xfs_buftarg_wait_rele,
1670 &dispose, LONG_MAX);
1671
1672 while (!list_empty(&dispose)) {
1673 struct xfs_buf *bp;
1674 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1675 list_del_init(&bp->b_lru);
1676 if (bp->b_flags & XBF_WRITE_FAIL) {
1677 xfs_alert(btp->bt_mount,
1678"Corruption Alert: Buffer at daddr 0x%llx had permanent write failures!",
1679 (long long)bp->b_bn);
1680 xfs_alert(btp->bt_mount,
1681"Please run xfs_repair to determine the extent of the problem.");
1682 }
1683 xfs_buf_rele(bp);
1684 }
1685 if (loop++ != 0)
1686 delay(100);
1687 }
1688}
1689
1690static enum lru_status
1691xfs_buftarg_isolate(
1692 struct list_head *item,
1693 struct list_lru_one *lru,
1694 spinlock_t *lru_lock,
1695 void *arg)
1696{
1697 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1698 struct list_head *dispose = arg;
1699
1700 /*
1701 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1702 * If we fail to get the lock, just skip it.
1703 */
1704 if (!spin_trylock(&bp->b_lock))
1705 return LRU_SKIP;
1706 /*
1707 * Decrement the b_lru_ref count unless the value is already
1708 * zero. If the value is already zero, we need to reclaim the
1709 * buffer, otherwise it gets another trip through the LRU.
1710 */
1711 if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1712 spin_unlock(&bp->b_lock);
1713 return LRU_ROTATE;
1714 }
1715
1716 bp->b_state |= XFS_BSTATE_DISPOSE;
1717 list_lru_isolate_move(lru, item, dispose);
1718 spin_unlock(&bp->b_lock);
1719 return LRU_REMOVED;
1720}
1721
1722static unsigned long
1723xfs_buftarg_shrink_scan(
1724 struct shrinker *shrink,
1725 struct shrink_control *sc)
1726{
1727 struct xfs_buftarg *btp = container_of(shrink,
1728 struct xfs_buftarg, bt_shrinker);
1729 LIST_HEAD(dispose);
1730 unsigned long freed;
1731
1732 freed = list_lru_shrink_walk(&btp->bt_lru, sc,
1733 xfs_buftarg_isolate, &dispose);
1734
1735 while (!list_empty(&dispose)) {
1736 struct xfs_buf *bp;
1737 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1738 list_del_init(&bp->b_lru);
1739 xfs_buf_rele(bp);
1740 }
1741
1742 return freed;
1743}
1744
1745static unsigned long
1746xfs_buftarg_shrink_count(
1747 struct shrinker *shrink,
1748 struct shrink_control *sc)
1749{
1750 struct xfs_buftarg *btp = container_of(shrink,
1751 struct xfs_buftarg, bt_shrinker);
1752 return list_lru_shrink_count(&btp->bt_lru, sc);
1753}
1754
1755void
1756xfs_free_buftarg(
1757 struct xfs_buftarg *btp)
1758{
1759 unregister_shrinker(&btp->bt_shrinker);
1760 ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
1761 percpu_counter_destroy(&btp->bt_io_count);
1762 list_lru_destroy(&btp->bt_lru);
1763
1764 xfs_blkdev_issue_flush(btp);
1765
1766 kmem_free(btp);
1767}
1768
1769int
1770xfs_setsize_buftarg(
1771 xfs_buftarg_t *btp,
1772 unsigned int sectorsize)
1773{
1774 /* Set up metadata sector size info */
1775 btp->bt_meta_sectorsize = sectorsize;
1776 btp->bt_meta_sectormask = sectorsize - 1;
1777
1778 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1779 xfs_warn(btp->bt_mount,
1780 "Cannot set_blocksize to %u on device %pg",
1781 sectorsize, btp->bt_bdev);
1782 return -EINVAL;
1783 }
1784
1785 /* Set up device logical sector size mask */
1786 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1787 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1788
1789 return 0;
1790}
1791
1792/*
1793 * When allocating the initial buffer target we have not yet
1794 * read in the superblock, so don't know what sized sectors
1795 * are being used at this early stage. Play safe.
1796 */
1797STATIC int
1798xfs_setsize_buftarg_early(
1799 xfs_buftarg_t *btp,
1800 struct block_device *bdev)
1801{
1802 return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev));
1803}
1804
1805xfs_buftarg_t *
1806xfs_alloc_buftarg(
1807 struct xfs_mount *mp,
1808 struct block_device *bdev,
1809 struct dax_device *dax_dev)
1810{
1811 xfs_buftarg_t *btp;
1812
1813 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS);
1814
1815 btp->bt_mount = mp;
1816 btp->bt_dev = bdev->bd_dev;
1817 btp->bt_bdev = bdev;
1818 btp->bt_daxdev = dax_dev;
1819
1820 if (xfs_setsize_buftarg_early(btp, bdev))
1821 goto error_free;
1822
1823 if (list_lru_init(&btp->bt_lru))
1824 goto error_free;
1825
1826 if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL))
1827 goto error_lru;
1828
1829 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
1830 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
1831 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1832 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
1833 if (register_shrinker(&btp->bt_shrinker))
1834 goto error_pcpu;
1835 return btp;
1836
1837error_pcpu:
1838 percpu_counter_destroy(&btp->bt_io_count);
1839error_lru:
1840 list_lru_destroy(&btp->bt_lru);
1841error_free:
1842 kmem_free(btp);
1843 return NULL;
1844}
1845
1846/*
1847 * Cancel a delayed write list.
1848 *
1849 * Remove each buffer from the list, clear the delwri queue flag and drop the
1850 * associated buffer reference.
1851 */
1852void
1853xfs_buf_delwri_cancel(
1854 struct list_head *list)
1855{
1856 struct xfs_buf *bp;
1857
1858 while (!list_empty(list)) {
1859 bp = list_first_entry(list, struct xfs_buf, b_list);
1860
1861 xfs_buf_lock(bp);
1862 bp->b_flags &= ~_XBF_DELWRI_Q;
1863 list_del_init(&bp->b_list);
1864 xfs_buf_relse(bp);
1865 }
1866}
1867
1868/*
1869 * Add a buffer to the delayed write list.
1870 *
1871 * This queues a buffer for writeout if it hasn't already been. Note that
1872 * neither this routine nor the buffer list submission functions perform
1873 * any internal synchronization. It is expected that the lists are thread-local
1874 * to the callers.
1875 *
1876 * Returns true if we queued up the buffer, or false if it already had
1877 * been on the buffer list.
1878 */
1879bool
1880xfs_buf_delwri_queue(
1881 struct xfs_buf *bp,
1882 struct list_head *list)
1883{
1884 ASSERT(xfs_buf_islocked(bp));
1885 ASSERT(!(bp->b_flags & XBF_READ));
1886
1887 /*
1888 * If the buffer is already marked delwri it already is queued up
1889 * by someone else for imediate writeout. Just ignore it in that
1890 * case.
1891 */
1892 if (bp->b_flags & _XBF_DELWRI_Q) {
1893 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1894 return false;
1895 }
1896
1897 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1898
1899 /*
1900 * If a buffer gets written out synchronously or marked stale while it
1901 * is on a delwri list we lazily remove it. To do this, the other party
1902 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1903 * It remains referenced and on the list. In a rare corner case it
1904 * might get readded to a delwri list after the synchronous writeout, in
1905 * which case we need just need to re-add the flag here.
1906 */
1907 bp->b_flags |= _XBF_DELWRI_Q;
1908 if (list_empty(&bp->b_list)) {
1909 atomic_inc(&bp->b_hold);
1910 list_add_tail(&bp->b_list, list);
1911 }
1912
1913 return true;
1914}
1915
1916/*
1917 * Compare function is more complex than it needs to be because
1918 * the return value is only 32 bits and we are doing comparisons
1919 * on 64 bit values
1920 */
1921static int
1922xfs_buf_cmp(
1923 void *priv,
1924 struct list_head *a,
1925 struct list_head *b)
1926{
1927 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1928 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1929 xfs_daddr_t diff;
1930
1931 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
1932 if (diff < 0)
1933 return -1;
1934 if (diff > 0)
1935 return 1;
1936 return 0;
1937}
1938
1939/*
1940 * submit buffers for write.
1941 *
1942 * When we have a large buffer list, we do not want to hold all the buffers
1943 * locked while we block on the request queue waiting for IO dispatch. To avoid
1944 * this problem, we lock and submit buffers in groups of 50, thereby minimising
1945 * the lock hold times for lists which may contain thousands of objects.
1946 *
1947 * To do this, we sort the buffer list before we walk the list to lock and
1948 * submit buffers, and we plug and unplug around each group of buffers we
1949 * submit.
1950 */
1951static int
1952xfs_buf_delwri_submit_buffers(
1953 struct list_head *buffer_list,
1954 struct list_head *wait_list)
1955{
1956 struct xfs_buf *bp, *n;
1957 LIST_HEAD (submit_list);
1958 int pinned = 0;
1959 struct blk_plug plug;
1960
1961 list_sort(NULL, buffer_list, xfs_buf_cmp);
1962
1963 blk_start_plug(&plug);
1964 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1965 if (!wait_list) {
1966 if (xfs_buf_ispinned(bp)) {
1967 pinned++;
1968 continue;
1969 }
1970 if (!xfs_buf_trylock(bp))
1971 continue;
1972 } else {
1973 xfs_buf_lock(bp);
1974 }
1975
1976 /*
1977 * Someone else might have written the buffer synchronously or
1978 * marked it stale in the meantime. In that case only the
1979 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1980 * reference and remove it from the list here.
1981 */
1982 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1983 list_del_init(&bp->b_list);
1984 xfs_buf_relse(bp);
1985 continue;
1986 }
1987
1988 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1989
1990 /*
1991 * We do all IO submission async. This means if we need
1992 * to wait for IO completion we need to take an extra
1993 * reference so the buffer is still valid on the other
1994 * side. We need to move the buffer onto the io_list
1995 * at this point so the caller can still access it.
1996 */
1997 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_WRITE_FAIL);
1998 bp->b_flags |= XBF_WRITE | XBF_ASYNC;
1999 if (wait_list) {
2000 xfs_buf_hold(bp);
2001 list_move_tail(&bp->b_list, wait_list);
2002 } else
2003 list_del_init(&bp->b_list);
2004
2005 xfs_buf_submit(bp);
2006 }
2007 blk_finish_plug(&plug);
2008
2009 return pinned;
2010}
2011
2012/*
2013 * Write out a buffer list asynchronously.
2014 *
2015 * This will take the @buffer_list, write all non-locked and non-pinned buffers
2016 * out and not wait for I/O completion on any of the buffers. This interface
2017 * is only safely useable for callers that can track I/O completion by higher
2018 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
2019 * function.
2020 */
2021int
2022xfs_buf_delwri_submit_nowait(
2023 struct list_head *buffer_list)
2024{
2025 return xfs_buf_delwri_submit_buffers(buffer_list, NULL);
2026}
2027
2028/*
2029 * Write out a buffer list synchronously.
2030 *
2031 * This will take the @buffer_list, write all buffers out and wait for I/O
2032 * completion on all of the buffers. @buffer_list is consumed by the function,
2033 * so callers must have some other way of tracking buffers if they require such
2034 * functionality.
2035 */
2036int
2037xfs_buf_delwri_submit(
2038 struct list_head *buffer_list)
2039{
2040 LIST_HEAD (wait_list);
2041 int error = 0, error2;
2042 struct xfs_buf *bp;
2043
2044 xfs_buf_delwri_submit_buffers(buffer_list, &wait_list);
2045
2046 /* Wait for IO to complete. */
2047 while (!list_empty(&wait_list)) {
2048 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
2049
2050 list_del_init(&bp->b_list);
2051
2052 /* locking the buffer will wait for async IO completion. */
2053 xfs_buf_lock(bp);
2054 error2 = bp->b_error;
2055 xfs_buf_relse(bp);
2056 if (!error)
2057 error = error2;
2058 }
2059
2060 return error;
2061}
2062
2063/*
2064 * Push a single buffer on a delwri queue.
2065 *
2066 * The purpose of this function is to submit a single buffer of a delwri queue
2067 * and return with the buffer still on the original queue. The waiting delwri
2068 * buffer submission infrastructure guarantees transfer of the delwri queue
2069 * buffer reference to a temporary wait list. We reuse this infrastructure to
2070 * transfer the buffer back to the original queue.
2071 *
2072 * Note the buffer transitions from the queued state, to the submitted and wait
2073 * listed state and back to the queued state during this call. The buffer
2074 * locking and queue management logic between _delwri_pushbuf() and
2075 * _delwri_queue() guarantee that the buffer cannot be queued to another list
2076 * before returning.
2077 */
2078int
2079xfs_buf_delwri_pushbuf(
2080 struct xfs_buf *bp,
2081 struct list_head *buffer_list)
2082{
2083 LIST_HEAD (submit_list);
2084 int error;
2085
2086 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
2087
2088 trace_xfs_buf_delwri_pushbuf(bp, _RET_IP_);
2089
2090 /*
2091 * Isolate the buffer to a new local list so we can submit it for I/O
2092 * independently from the rest of the original list.
2093 */
2094 xfs_buf_lock(bp);
2095 list_move(&bp->b_list, &submit_list);
2096 xfs_buf_unlock(bp);
2097
2098 /*
2099 * Delwri submission clears the DELWRI_Q buffer flag and returns with
2100 * the buffer on the wait list with an associated reference. Rather than
2101 * bounce the buffer from a local wait list back to the original list
2102 * after I/O completion, reuse the original list as the wait list.
2103 */
2104 xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
2105
2106 /*
2107 * The buffer is now under I/O and wait listed as during typical delwri
2108 * submission. Lock the buffer to wait for I/O completion. Rather than
2109 * remove the buffer from the wait list and release the reference, we
2110 * want to return with the buffer queued to the original list. The
2111 * buffer already sits on the original list with a wait list reference,
2112 * however. If we let the queue inherit that wait list reference, all we
2113 * need to do is reset the DELWRI_Q flag.
2114 */
2115 xfs_buf_lock(bp);
2116 error = bp->b_error;
2117 bp->b_flags |= _XBF_DELWRI_Q;
2118 xfs_buf_unlock(bp);
2119
2120 return error;
2121}
2122
2123int __init
2124xfs_buf_init(void)
2125{
2126 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
2127 KM_ZONE_HWALIGN, NULL);
2128 if (!xfs_buf_zone)
2129 goto out;
2130
2131 return 0;
2132
2133 out:
2134 return -ENOMEM;
2135}
2136
2137void
2138xfs_buf_terminate(void)
2139{
2140 kmem_zone_destroy(xfs_buf_zone);
2141}
2142
2143void xfs_buf_set_ref(struct xfs_buf *bp, int lru_ref)
2144{
2145 /*
2146 * Set the lru reference count to 0 based on the error injection tag.
2147 * This allows userspace to disrupt buffer caching for debug/testing
2148 * purposes.
2149 */
2150 if (XFS_TEST_ERROR(false, bp->b_target->bt_mount,
2151 XFS_ERRTAG_BUF_LRU_REF))
2152 lru_ref = 0;
2153
2154 atomic_set(&bp->b_lru_ref, lru_ref);
2155}
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#include <linux/dax.h>
9
10#include "xfs_shared.h"
11#include "xfs_format.h"
12#include "xfs_log_format.h"
13#include "xfs_trans_resv.h"
14#include "xfs_mount.h"
15#include "xfs_trace.h"
16#include "xfs_log.h"
17#include "xfs_log_recover.h"
18#include "xfs_log_priv.h"
19#include "xfs_trans.h"
20#include "xfs_buf_item.h"
21#include "xfs_errortag.h"
22#include "xfs_error.h"
23#include "xfs_ag.h"
24#include "xfs_buf_mem.h"
25
26struct kmem_cache *xfs_buf_cache;
27
28/*
29 * Locking orders
30 *
31 * xfs_buf_ioacct_inc:
32 * xfs_buf_ioacct_dec:
33 * b_sema (caller holds)
34 * b_lock
35 *
36 * xfs_buf_stale:
37 * b_sema (caller holds)
38 * b_lock
39 * lru_lock
40 *
41 * xfs_buf_rele:
42 * b_lock
43 * pag_buf_lock
44 * lru_lock
45 *
46 * xfs_buftarg_drain_rele
47 * lru_lock
48 * b_lock (trylock due to inversion)
49 *
50 * xfs_buftarg_isolate
51 * lru_lock
52 * b_lock (trylock due to inversion)
53 */
54
55static int __xfs_buf_submit(struct xfs_buf *bp, bool wait);
56
57static inline int
58xfs_buf_submit(
59 struct xfs_buf *bp)
60{
61 return __xfs_buf_submit(bp, !(bp->b_flags & XBF_ASYNC));
62}
63
64static inline bool xfs_buf_is_uncached(struct xfs_buf *bp)
65{
66 return bp->b_rhash_key == XFS_BUF_DADDR_NULL;
67}
68
69static inline int
70xfs_buf_is_vmapped(
71 struct xfs_buf *bp)
72{
73 /*
74 * Return true if the buffer is vmapped.
75 *
76 * b_addr is null if the buffer is not mapped, but the code is clever
77 * enough to know it doesn't have to map a single page, so the check has
78 * to be both for b_addr and bp->b_page_count > 1.
79 */
80 return bp->b_addr && bp->b_page_count > 1;
81}
82
83static inline int
84xfs_buf_vmap_len(
85 struct xfs_buf *bp)
86{
87 return (bp->b_page_count * PAGE_SIZE);
88}
89
90/*
91 * Bump the I/O in flight count on the buftarg if we haven't yet done so for
92 * this buffer. The count is incremented once per buffer (per hold cycle)
93 * because the corresponding decrement is deferred to buffer release. Buffers
94 * can undergo I/O multiple times in a hold-release cycle and per buffer I/O
95 * tracking adds unnecessary overhead. This is used for sychronization purposes
96 * with unmount (see xfs_buftarg_drain()), so all we really need is a count of
97 * in-flight buffers.
98 *
99 * Buffers that are never released (e.g., superblock, iclog buffers) must set
100 * the XBF_NO_IOACCT flag before I/O submission. Otherwise, the buftarg count
101 * never reaches zero and unmount hangs indefinitely.
102 */
103static inline void
104xfs_buf_ioacct_inc(
105 struct xfs_buf *bp)
106{
107 if (bp->b_flags & XBF_NO_IOACCT)
108 return;
109
110 ASSERT(bp->b_flags & XBF_ASYNC);
111 spin_lock(&bp->b_lock);
112 if (!(bp->b_state & XFS_BSTATE_IN_FLIGHT)) {
113 bp->b_state |= XFS_BSTATE_IN_FLIGHT;
114 percpu_counter_inc(&bp->b_target->bt_io_count);
115 }
116 spin_unlock(&bp->b_lock);
117}
118
119/*
120 * Clear the in-flight state on a buffer about to be released to the LRU or
121 * freed and unaccount from the buftarg.
122 */
123static inline void
124__xfs_buf_ioacct_dec(
125 struct xfs_buf *bp)
126{
127 lockdep_assert_held(&bp->b_lock);
128
129 if (bp->b_state & XFS_BSTATE_IN_FLIGHT) {
130 bp->b_state &= ~XFS_BSTATE_IN_FLIGHT;
131 percpu_counter_dec(&bp->b_target->bt_io_count);
132 }
133}
134
135static inline void
136xfs_buf_ioacct_dec(
137 struct xfs_buf *bp)
138{
139 spin_lock(&bp->b_lock);
140 __xfs_buf_ioacct_dec(bp);
141 spin_unlock(&bp->b_lock);
142}
143
144/*
145 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
146 * b_lru_ref count so that the buffer is freed immediately when the buffer
147 * reference count falls to zero. If the buffer is already on the LRU, we need
148 * to remove the reference that LRU holds on the buffer.
149 *
150 * This prevents build-up of stale buffers on the LRU.
151 */
152void
153xfs_buf_stale(
154 struct xfs_buf *bp)
155{
156 ASSERT(xfs_buf_islocked(bp));
157
158 bp->b_flags |= XBF_STALE;
159
160 /*
161 * Clear the delwri status so that a delwri queue walker will not
162 * flush this buffer to disk now that it is stale. The delwri queue has
163 * a reference to the buffer, so this is safe to do.
164 */
165 bp->b_flags &= ~_XBF_DELWRI_Q;
166
167 /*
168 * Once the buffer is marked stale and unlocked, a subsequent lookup
169 * could reset b_flags. There is no guarantee that the buffer is
170 * unaccounted (released to LRU) before that occurs. Drop in-flight
171 * status now to preserve accounting consistency.
172 */
173 spin_lock(&bp->b_lock);
174 __xfs_buf_ioacct_dec(bp);
175
176 atomic_set(&bp->b_lru_ref, 0);
177 if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
178 (list_lru_del_obj(&bp->b_target->bt_lru, &bp->b_lru)))
179 atomic_dec(&bp->b_hold);
180
181 ASSERT(atomic_read(&bp->b_hold) >= 1);
182 spin_unlock(&bp->b_lock);
183}
184
185static int
186xfs_buf_get_maps(
187 struct xfs_buf *bp,
188 int map_count)
189{
190 ASSERT(bp->b_maps == NULL);
191 bp->b_map_count = map_count;
192
193 if (map_count == 1) {
194 bp->b_maps = &bp->__b_map;
195 return 0;
196 }
197
198 bp->b_maps = kzalloc(map_count * sizeof(struct xfs_buf_map),
199 GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL);
200 if (!bp->b_maps)
201 return -ENOMEM;
202 return 0;
203}
204
205/*
206 * Frees b_pages if it was allocated.
207 */
208static void
209xfs_buf_free_maps(
210 struct xfs_buf *bp)
211{
212 if (bp->b_maps != &bp->__b_map) {
213 kfree(bp->b_maps);
214 bp->b_maps = NULL;
215 }
216}
217
218static int
219_xfs_buf_alloc(
220 struct xfs_buftarg *target,
221 struct xfs_buf_map *map,
222 int nmaps,
223 xfs_buf_flags_t flags,
224 struct xfs_buf **bpp)
225{
226 struct xfs_buf *bp;
227 int error;
228 int i;
229
230 *bpp = NULL;
231 bp = kmem_cache_zalloc(xfs_buf_cache,
232 GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL);
233
234 /*
235 * We don't want certain flags to appear in b_flags unless they are
236 * specifically set by later operations on the buffer.
237 */
238 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
239
240 atomic_set(&bp->b_hold, 1);
241 atomic_set(&bp->b_lru_ref, 1);
242 init_completion(&bp->b_iowait);
243 INIT_LIST_HEAD(&bp->b_lru);
244 INIT_LIST_HEAD(&bp->b_list);
245 INIT_LIST_HEAD(&bp->b_li_list);
246 sema_init(&bp->b_sema, 0); /* held, no waiters */
247 spin_lock_init(&bp->b_lock);
248 bp->b_target = target;
249 bp->b_mount = target->bt_mount;
250 bp->b_flags = flags;
251
252 /*
253 * Set length and io_length to the same value initially.
254 * I/O routines should use io_length, which will be the same in
255 * most cases but may be reset (e.g. XFS recovery).
256 */
257 error = xfs_buf_get_maps(bp, nmaps);
258 if (error) {
259 kmem_cache_free(xfs_buf_cache, bp);
260 return error;
261 }
262
263 bp->b_rhash_key = map[0].bm_bn;
264 bp->b_length = 0;
265 for (i = 0; i < nmaps; i++) {
266 bp->b_maps[i].bm_bn = map[i].bm_bn;
267 bp->b_maps[i].bm_len = map[i].bm_len;
268 bp->b_length += map[i].bm_len;
269 }
270
271 atomic_set(&bp->b_pin_count, 0);
272 init_waitqueue_head(&bp->b_waiters);
273
274 XFS_STATS_INC(bp->b_mount, xb_create);
275 trace_xfs_buf_init(bp, _RET_IP_);
276
277 *bpp = bp;
278 return 0;
279}
280
281static void
282xfs_buf_free_pages(
283 struct xfs_buf *bp)
284{
285 uint i;
286
287 ASSERT(bp->b_flags & _XBF_PAGES);
288
289 if (xfs_buf_is_vmapped(bp))
290 vm_unmap_ram(bp->b_addr, bp->b_page_count);
291
292 for (i = 0; i < bp->b_page_count; i++) {
293 if (bp->b_pages[i])
294 __free_page(bp->b_pages[i]);
295 }
296 mm_account_reclaimed_pages(bp->b_page_count);
297
298 if (bp->b_pages != bp->b_page_array)
299 kfree(bp->b_pages);
300 bp->b_pages = NULL;
301 bp->b_flags &= ~_XBF_PAGES;
302}
303
304static void
305xfs_buf_free_callback(
306 struct callback_head *cb)
307{
308 struct xfs_buf *bp = container_of(cb, struct xfs_buf, b_rcu);
309
310 xfs_buf_free_maps(bp);
311 kmem_cache_free(xfs_buf_cache, bp);
312}
313
314static void
315xfs_buf_free(
316 struct xfs_buf *bp)
317{
318 trace_xfs_buf_free(bp, _RET_IP_);
319
320 ASSERT(list_empty(&bp->b_lru));
321
322 if (xfs_buftarg_is_mem(bp->b_target))
323 xmbuf_unmap_page(bp);
324 else if (bp->b_flags & _XBF_PAGES)
325 xfs_buf_free_pages(bp);
326 else if (bp->b_flags & _XBF_KMEM)
327 kfree(bp->b_addr);
328
329 call_rcu(&bp->b_rcu, xfs_buf_free_callback);
330}
331
332static int
333xfs_buf_alloc_kmem(
334 struct xfs_buf *bp,
335 xfs_buf_flags_t flags)
336{
337 gfp_t gfp_mask = GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL;
338 size_t size = BBTOB(bp->b_length);
339
340 /* Assure zeroed buffer for non-read cases. */
341 if (!(flags & XBF_READ))
342 gfp_mask |= __GFP_ZERO;
343
344 bp->b_addr = kmalloc(size, gfp_mask);
345 if (!bp->b_addr)
346 return -ENOMEM;
347
348 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
349 ((unsigned long)bp->b_addr & PAGE_MASK)) {
350 /* b_addr spans two pages - use alloc_page instead */
351 kfree(bp->b_addr);
352 bp->b_addr = NULL;
353 return -ENOMEM;
354 }
355 bp->b_offset = offset_in_page(bp->b_addr);
356 bp->b_pages = bp->b_page_array;
357 bp->b_pages[0] = kmem_to_page(bp->b_addr);
358 bp->b_page_count = 1;
359 bp->b_flags |= _XBF_KMEM;
360 return 0;
361}
362
363static int
364xfs_buf_alloc_pages(
365 struct xfs_buf *bp,
366 xfs_buf_flags_t flags)
367{
368 gfp_t gfp_mask = GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOWARN;
369 long filled = 0;
370
371 if (flags & XBF_READ_AHEAD)
372 gfp_mask |= __GFP_NORETRY;
373
374 /* Make sure that we have a page list */
375 bp->b_page_count = DIV_ROUND_UP(BBTOB(bp->b_length), PAGE_SIZE);
376 if (bp->b_page_count <= XB_PAGES) {
377 bp->b_pages = bp->b_page_array;
378 } else {
379 bp->b_pages = kzalloc(sizeof(struct page *) * bp->b_page_count,
380 gfp_mask);
381 if (!bp->b_pages)
382 return -ENOMEM;
383 }
384 bp->b_flags |= _XBF_PAGES;
385
386 /* Assure zeroed buffer for non-read cases. */
387 if (!(flags & XBF_READ))
388 gfp_mask |= __GFP_ZERO;
389
390 /*
391 * Bulk filling of pages can take multiple calls. Not filling the entire
392 * array is not an allocation failure, so don't back off if we get at
393 * least one extra page.
394 */
395 for (;;) {
396 long last = filled;
397
398 filled = alloc_pages_bulk_array(gfp_mask, bp->b_page_count,
399 bp->b_pages);
400 if (filled == bp->b_page_count) {
401 XFS_STATS_INC(bp->b_mount, xb_page_found);
402 break;
403 }
404
405 if (filled != last)
406 continue;
407
408 if (flags & XBF_READ_AHEAD) {
409 xfs_buf_free_pages(bp);
410 return -ENOMEM;
411 }
412
413 XFS_STATS_INC(bp->b_mount, xb_page_retries);
414 memalloc_retry_wait(gfp_mask);
415 }
416 return 0;
417}
418
419/*
420 * Map buffer into kernel address-space if necessary.
421 */
422STATIC int
423_xfs_buf_map_pages(
424 struct xfs_buf *bp,
425 xfs_buf_flags_t flags)
426{
427 ASSERT(bp->b_flags & _XBF_PAGES);
428 if (bp->b_page_count == 1) {
429 /* A single page buffer is always mappable */
430 bp->b_addr = page_address(bp->b_pages[0]);
431 } else if (flags & XBF_UNMAPPED) {
432 bp->b_addr = NULL;
433 } else {
434 int retried = 0;
435 unsigned nofs_flag;
436
437 /*
438 * vm_map_ram() will allocate auxiliary structures (e.g.
439 * pagetables) with GFP_KERNEL, yet we often under a scoped nofs
440 * context here. Mixing GFP_KERNEL with GFP_NOFS allocations
441 * from the same call site that can be run from both above and
442 * below memory reclaim causes lockdep false positives. Hence we
443 * always need to force this allocation to nofs context because
444 * we can't pass __GFP_NOLOCKDEP down to auxillary structures to
445 * prevent false positive lockdep reports.
446 *
447 * XXX(dgc): I think dquot reclaim is the only place we can get
448 * to this function from memory reclaim context now. If we fix
449 * that like we've fixed inode reclaim to avoid writeback from
450 * reclaim, this nofs wrapping can go away.
451 */
452 nofs_flag = memalloc_nofs_save();
453 do {
454 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
455 -1);
456 if (bp->b_addr)
457 break;
458 vm_unmap_aliases();
459 } while (retried++ <= 1);
460 memalloc_nofs_restore(nofs_flag);
461
462 if (!bp->b_addr)
463 return -ENOMEM;
464 }
465
466 return 0;
467}
468
469/*
470 * Finding and Reading Buffers
471 */
472static int
473_xfs_buf_obj_cmp(
474 struct rhashtable_compare_arg *arg,
475 const void *obj)
476{
477 const struct xfs_buf_map *map = arg->key;
478 const struct xfs_buf *bp = obj;
479
480 /*
481 * The key hashing in the lookup path depends on the key being the
482 * first element of the compare_arg, make sure to assert this.
483 */
484 BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0);
485
486 if (bp->b_rhash_key != map->bm_bn)
487 return 1;
488
489 if (unlikely(bp->b_length != map->bm_len)) {
490 /*
491 * found a block number match. If the range doesn't
492 * match, the only way this is allowed is if the buffer
493 * in the cache is stale and the transaction that made
494 * it stale has not yet committed. i.e. we are
495 * reallocating a busy extent. Skip this buffer and
496 * continue searching for an exact match.
497 *
498 * Note: If we're scanning for incore buffers to stale, don't
499 * complain if we find non-stale buffers.
500 */
501 if (!(map->bm_flags & XBM_LIVESCAN))
502 ASSERT(bp->b_flags & XBF_STALE);
503 return 1;
504 }
505 return 0;
506}
507
508static const struct rhashtable_params xfs_buf_hash_params = {
509 .min_size = 32, /* empty AGs have minimal footprint */
510 .nelem_hint = 16,
511 .key_len = sizeof(xfs_daddr_t),
512 .key_offset = offsetof(struct xfs_buf, b_rhash_key),
513 .head_offset = offsetof(struct xfs_buf, b_rhash_head),
514 .automatic_shrinking = true,
515 .obj_cmpfn = _xfs_buf_obj_cmp,
516};
517
518int
519xfs_buf_cache_init(
520 struct xfs_buf_cache *bch)
521{
522 spin_lock_init(&bch->bc_lock);
523 return rhashtable_init(&bch->bc_hash, &xfs_buf_hash_params);
524}
525
526void
527xfs_buf_cache_destroy(
528 struct xfs_buf_cache *bch)
529{
530 rhashtable_destroy(&bch->bc_hash);
531}
532
533static int
534xfs_buf_map_verify(
535 struct xfs_buftarg *btp,
536 struct xfs_buf_map *map)
537{
538 xfs_daddr_t eofs;
539
540 /* Check for IOs smaller than the sector size / not sector aligned */
541 ASSERT(!(BBTOB(map->bm_len) < btp->bt_meta_sectorsize));
542 ASSERT(!(BBTOB(map->bm_bn) & (xfs_off_t)btp->bt_meta_sectormask));
543
544 /*
545 * Corrupted block numbers can get through to here, unfortunately, so we
546 * have to check that the buffer falls within the filesystem bounds.
547 */
548 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
549 if (map->bm_bn < 0 || map->bm_bn >= eofs) {
550 xfs_alert(btp->bt_mount,
551 "%s: daddr 0x%llx out of range, EOFS 0x%llx",
552 __func__, map->bm_bn, eofs);
553 WARN_ON(1);
554 return -EFSCORRUPTED;
555 }
556 return 0;
557}
558
559static int
560xfs_buf_find_lock(
561 struct xfs_buf *bp,
562 xfs_buf_flags_t flags)
563{
564 if (flags & XBF_TRYLOCK) {
565 if (!xfs_buf_trylock(bp)) {
566 XFS_STATS_INC(bp->b_mount, xb_busy_locked);
567 return -EAGAIN;
568 }
569 } else {
570 xfs_buf_lock(bp);
571 XFS_STATS_INC(bp->b_mount, xb_get_locked_waited);
572 }
573
574 /*
575 * if the buffer is stale, clear all the external state associated with
576 * it. We need to keep flags such as how we allocated the buffer memory
577 * intact here.
578 */
579 if (bp->b_flags & XBF_STALE) {
580 if (flags & XBF_LIVESCAN) {
581 xfs_buf_unlock(bp);
582 return -ENOENT;
583 }
584 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
585 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
586 bp->b_ops = NULL;
587 }
588 return 0;
589}
590
591static inline int
592xfs_buf_lookup(
593 struct xfs_buf_cache *bch,
594 struct xfs_buf_map *map,
595 xfs_buf_flags_t flags,
596 struct xfs_buf **bpp)
597{
598 struct xfs_buf *bp;
599 int error;
600
601 rcu_read_lock();
602 bp = rhashtable_lookup(&bch->bc_hash, map, xfs_buf_hash_params);
603 if (!bp || !atomic_inc_not_zero(&bp->b_hold)) {
604 rcu_read_unlock();
605 return -ENOENT;
606 }
607 rcu_read_unlock();
608
609 error = xfs_buf_find_lock(bp, flags);
610 if (error) {
611 xfs_buf_rele(bp);
612 return error;
613 }
614
615 trace_xfs_buf_find(bp, flags, _RET_IP_);
616 *bpp = bp;
617 return 0;
618}
619
620/*
621 * Insert the new_bp into the hash table. This consumes the perag reference
622 * taken for the lookup regardless of the result of the insert.
623 */
624static int
625xfs_buf_find_insert(
626 struct xfs_buftarg *btp,
627 struct xfs_buf_cache *bch,
628 struct xfs_perag *pag,
629 struct xfs_buf_map *cmap,
630 struct xfs_buf_map *map,
631 int nmaps,
632 xfs_buf_flags_t flags,
633 struct xfs_buf **bpp)
634{
635 struct xfs_buf *new_bp;
636 struct xfs_buf *bp;
637 int error;
638
639 error = _xfs_buf_alloc(btp, map, nmaps, flags, &new_bp);
640 if (error)
641 goto out_drop_pag;
642
643 if (xfs_buftarg_is_mem(new_bp->b_target)) {
644 error = xmbuf_map_page(new_bp);
645 } else if (BBTOB(new_bp->b_length) >= PAGE_SIZE ||
646 xfs_buf_alloc_kmem(new_bp, flags) < 0) {
647 /*
648 * For buffers that fit entirely within a single page, first
649 * attempt to allocate the memory from the heap to minimise
650 * memory usage. If we can't get heap memory for these small
651 * buffers, we fall back to using the page allocator.
652 */
653 error = xfs_buf_alloc_pages(new_bp, flags);
654 }
655 if (error)
656 goto out_free_buf;
657
658 spin_lock(&bch->bc_lock);
659 bp = rhashtable_lookup_get_insert_fast(&bch->bc_hash,
660 &new_bp->b_rhash_head, xfs_buf_hash_params);
661 if (IS_ERR(bp)) {
662 error = PTR_ERR(bp);
663 spin_unlock(&bch->bc_lock);
664 goto out_free_buf;
665 }
666 if (bp && atomic_inc_not_zero(&bp->b_hold)) {
667 /* found an existing buffer */
668 spin_unlock(&bch->bc_lock);
669 error = xfs_buf_find_lock(bp, flags);
670 if (error)
671 xfs_buf_rele(bp);
672 else
673 *bpp = bp;
674 goto out_free_buf;
675 }
676
677 /* The new buffer keeps the perag reference until it is freed. */
678 new_bp->b_pag = pag;
679 spin_unlock(&bch->bc_lock);
680 *bpp = new_bp;
681 return 0;
682
683out_free_buf:
684 xfs_buf_free(new_bp);
685out_drop_pag:
686 if (pag)
687 xfs_perag_put(pag);
688 return error;
689}
690
691static inline struct xfs_perag *
692xfs_buftarg_get_pag(
693 struct xfs_buftarg *btp,
694 const struct xfs_buf_map *map)
695{
696 struct xfs_mount *mp = btp->bt_mount;
697
698 if (xfs_buftarg_is_mem(btp))
699 return NULL;
700 return xfs_perag_get(mp, xfs_daddr_to_agno(mp, map->bm_bn));
701}
702
703static inline struct xfs_buf_cache *
704xfs_buftarg_buf_cache(
705 struct xfs_buftarg *btp,
706 struct xfs_perag *pag)
707{
708 if (pag)
709 return &pag->pag_bcache;
710 return btp->bt_cache;
711}
712
713/*
714 * Assembles a buffer covering the specified range. The code is optimised for
715 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
716 * more hits than misses.
717 */
718int
719xfs_buf_get_map(
720 struct xfs_buftarg *btp,
721 struct xfs_buf_map *map,
722 int nmaps,
723 xfs_buf_flags_t flags,
724 struct xfs_buf **bpp)
725{
726 struct xfs_buf_cache *bch;
727 struct xfs_perag *pag;
728 struct xfs_buf *bp = NULL;
729 struct xfs_buf_map cmap = { .bm_bn = map[0].bm_bn };
730 int error;
731 int i;
732
733 if (flags & XBF_LIVESCAN)
734 cmap.bm_flags |= XBM_LIVESCAN;
735 for (i = 0; i < nmaps; i++)
736 cmap.bm_len += map[i].bm_len;
737
738 error = xfs_buf_map_verify(btp, &cmap);
739 if (error)
740 return error;
741
742 pag = xfs_buftarg_get_pag(btp, &cmap);
743 bch = xfs_buftarg_buf_cache(btp, pag);
744
745 error = xfs_buf_lookup(bch, &cmap, flags, &bp);
746 if (error && error != -ENOENT)
747 goto out_put_perag;
748
749 /* cache hits always outnumber misses by at least 10:1 */
750 if (unlikely(!bp)) {
751 XFS_STATS_INC(btp->bt_mount, xb_miss_locked);
752
753 if (flags & XBF_INCORE)
754 goto out_put_perag;
755
756 /* xfs_buf_find_insert() consumes the perag reference. */
757 error = xfs_buf_find_insert(btp, bch, pag, &cmap, map, nmaps,
758 flags, &bp);
759 if (error)
760 return error;
761 } else {
762 XFS_STATS_INC(btp->bt_mount, xb_get_locked);
763 if (pag)
764 xfs_perag_put(pag);
765 }
766
767 /* We do not hold a perag reference anymore. */
768 if (!bp->b_addr) {
769 error = _xfs_buf_map_pages(bp, flags);
770 if (unlikely(error)) {
771 xfs_warn_ratelimited(btp->bt_mount,
772 "%s: failed to map %u pages", __func__,
773 bp->b_page_count);
774 xfs_buf_relse(bp);
775 return error;
776 }
777 }
778
779 /*
780 * Clear b_error if this is a lookup from a caller that doesn't expect
781 * valid data to be found in the buffer.
782 */
783 if (!(flags & XBF_READ))
784 xfs_buf_ioerror(bp, 0);
785
786 XFS_STATS_INC(btp->bt_mount, xb_get);
787 trace_xfs_buf_get(bp, flags, _RET_IP_);
788 *bpp = bp;
789 return 0;
790
791out_put_perag:
792 if (pag)
793 xfs_perag_put(pag);
794 return error;
795}
796
797int
798_xfs_buf_read(
799 struct xfs_buf *bp,
800 xfs_buf_flags_t flags)
801{
802 ASSERT(!(flags & XBF_WRITE));
803 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
804
805 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD | XBF_DONE);
806 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
807
808 return xfs_buf_submit(bp);
809}
810
811/*
812 * Reverify a buffer found in cache without an attached ->b_ops.
813 *
814 * If the caller passed an ops structure and the buffer doesn't have ops
815 * assigned, set the ops and use it to verify the contents. If verification
816 * fails, clear XBF_DONE. We assume the buffer has no recorded errors and is
817 * already in XBF_DONE state on entry.
818 *
819 * Under normal operations, every in-core buffer is verified on read I/O
820 * completion. There are two scenarios that can lead to in-core buffers without
821 * an assigned ->b_ops. The first is during log recovery of buffers on a V4
822 * filesystem, though these buffers are purged at the end of recovery. The
823 * other is online repair, which intentionally reads with a NULL buffer ops to
824 * run several verifiers across an in-core buffer in order to establish buffer
825 * type. If repair can't establish that, the buffer will be left in memory
826 * with NULL buffer ops.
827 */
828int
829xfs_buf_reverify(
830 struct xfs_buf *bp,
831 const struct xfs_buf_ops *ops)
832{
833 ASSERT(bp->b_flags & XBF_DONE);
834 ASSERT(bp->b_error == 0);
835
836 if (!ops || bp->b_ops)
837 return 0;
838
839 bp->b_ops = ops;
840 bp->b_ops->verify_read(bp);
841 if (bp->b_error)
842 bp->b_flags &= ~XBF_DONE;
843 return bp->b_error;
844}
845
846int
847xfs_buf_read_map(
848 struct xfs_buftarg *target,
849 struct xfs_buf_map *map,
850 int nmaps,
851 xfs_buf_flags_t flags,
852 struct xfs_buf **bpp,
853 const struct xfs_buf_ops *ops,
854 xfs_failaddr_t fa)
855{
856 struct xfs_buf *bp;
857 int error;
858
859 flags |= XBF_READ;
860 *bpp = NULL;
861
862 error = xfs_buf_get_map(target, map, nmaps, flags, &bp);
863 if (error)
864 return error;
865
866 trace_xfs_buf_read(bp, flags, _RET_IP_);
867
868 if (!(bp->b_flags & XBF_DONE)) {
869 /* Initiate the buffer read and wait. */
870 XFS_STATS_INC(target->bt_mount, xb_get_read);
871 bp->b_ops = ops;
872 error = _xfs_buf_read(bp, flags);
873
874 /* Readahead iodone already dropped the buffer, so exit. */
875 if (flags & XBF_ASYNC)
876 return 0;
877 } else {
878 /* Buffer already read; all we need to do is check it. */
879 error = xfs_buf_reverify(bp, ops);
880
881 /* Readahead already finished; drop the buffer and exit. */
882 if (flags & XBF_ASYNC) {
883 xfs_buf_relse(bp);
884 return 0;
885 }
886
887 /* We do not want read in the flags */
888 bp->b_flags &= ~XBF_READ;
889 ASSERT(bp->b_ops != NULL || ops == NULL);
890 }
891
892 /*
893 * If we've had a read error, then the contents of the buffer are
894 * invalid and should not be used. To ensure that a followup read tries
895 * to pull the buffer from disk again, we clear the XBF_DONE flag and
896 * mark the buffer stale. This ensures that anyone who has a current
897 * reference to the buffer will interpret it's contents correctly and
898 * future cache lookups will also treat it as an empty, uninitialised
899 * buffer.
900 */
901 if (error) {
902 /*
903 * Check against log shutdown for error reporting because
904 * metadata writeback may require a read first and we need to
905 * report errors in metadata writeback until the log is shut
906 * down. High level transaction read functions already check
907 * against mount shutdown, anyway, so we only need to be
908 * concerned about low level IO interactions here.
909 */
910 if (!xlog_is_shutdown(target->bt_mount->m_log))
911 xfs_buf_ioerror_alert(bp, fa);
912
913 bp->b_flags &= ~XBF_DONE;
914 xfs_buf_stale(bp);
915 xfs_buf_relse(bp);
916
917 /* bad CRC means corrupted metadata */
918 if (error == -EFSBADCRC)
919 error = -EFSCORRUPTED;
920 return error;
921 }
922
923 *bpp = bp;
924 return 0;
925}
926
927/*
928 * If we are not low on memory then do the readahead in a deadlock
929 * safe manner.
930 */
931void
932xfs_buf_readahead_map(
933 struct xfs_buftarg *target,
934 struct xfs_buf_map *map,
935 int nmaps,
936 const struct xfs_buf_ops *ops)
937{
938 struct xfs_buf *bp;
939
940 /*
941 * Currently we don't have a good means or justification for performing
942 * xmbuf_map_page asynchronously, so we don't do readahead.
943 */
944 if (xfs_buftarg_is_mem(target))
945 return;
946
947 xfs_buf_read_map(target, map, nmaps,
948 XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD, &bp, ops,
949 __this_address);
950}
951
952/*
953 * Read an uncached buffer from disk. Allocates and returns a locked
954 * buffer containing the disk contents or nothing. Uncached buffers always have
955 * a cache index of XFS_BUF_DADDR_NULL so we can easily determine if the buffer
956 * is cached or uncached during fault diagnosis.
957 */
958int
959xfs_buf_read_uncached(
960 struct xfs_buftarg *target,
961 xfs_daddr_t daddr,
962 size_t numblks,
963 xfs_buf_flags_t flags,
964 struct xfs_buf **bpp,
965 const struct xfs_buf_ops *ops)
966{
967 struct xfs_buf *bp;
968 int error;
969
970 *bpp = NULL;
971
972 error = xfs_buf_get_uncached(target, numblks, flags, &bp);
973 if (error)
974 return error;
975
976 /* set up the buffer for a read IO */
977 ASSERT(bp->b_map_count == 1);
978 bp->b_rhash_key = XFS_BUF_DADDR_NULL;
979 bp->b_maps[0].bm_bn = daddr;
980 bp->b_flags |= XBF_READ;
981 bp->b_ops = ops;
982
983 xfs_buf_submit(bp);
984 if (bp->b_error) {
985 error = bp->b_error;
986 xfs_buf_relse(bp);
987 return error;
988 }
989
990 *bpp = bp;
991 return 0;
992}
993
994int
995xfs_buf_get_uncached(
996 struct xfs_buftarg *target,
997 size_t numblks,
998 xfs_buf_flags_t flags,
999 struct xfs_buf **bpp)
1000{
1001 int error;
1002 struct xfs_buf *bp;
1003 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
1004
1005 *bpp = NULL;
1006
1007 /* flags might contain irrelevant bits, pass only what we care about */
1008 error = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT, &bp);
1009 if (error)
1010 return error;
1011
1012 if (xfs_buftarg_is_mem(bp->b_target))
1013 error = xmbuf_map_page(bp);
1014 else
1015 error = xfs_buf_alloc_pages(bp, flags);
1016 if (error)
1017 goto fail_free_buf;
1018
1019 error = _xfs_buf_map_pages(bp, 0);
1020 if (unlikely(error)) {
1021 xfs_warn(target->bt_mount,
1022 "%s: failed to map pages", __func__);
1023 goto fail_free_buf;
1024 }
1025
1026 trace_xfs_buf_get_uncached(bp, _RET_IP_);
1027 *bpp = bp;
1028 return 0;
1029
1030fail_free_buf:
1031 xfs_buf_free(bp);
1032 return error;
1033}
1034
1035/*
1036 * Increment reference count on buffer, to hold the buffer concurrently
1037 * with another thread which may release (free) the buffer asynchronously.
1038 * Must hold the buffer already to call this function.
1039 */
1040void
1041xfs_buf_hold(
1042 struct xfs_buf *bp)
1043{
1044 trace_xfs_buf_hold(bp, _RET_IP_);
1045 atomic_inc(&bp->b_hold);
1046}
1047
1048static void
1049xfs_buf_rele_uncached(
1050 struct xfs_buf *bp)
1051{
1052 ASSERT(list_empty(&bp->b_lru));
1053 if (atomic_dec_and_test(&bp->b_hold)) {
1054 xfs_buf_ioacct_dec(bp);
1055 xfs_buf_free(bp);
1056 }
1057}
1058
1059static void
1060xfs_buf_rele_cached(
1061 struct xfs_buf *bp)
1062{
1063 struct xfs_buftarg *btp = bp->b_target;
1064 struct xfs_perag *pag = bp->b_pag;
1065 struct xfs_buf_cache *bch = xfs_buftarg_buf_cache(btp, pag);
1066 bool release;
1067 bool freebuf = false;
1068
1069 trace_xfs_buf_rele(bp, _RET_IP_);
1070
1071 ASSERT(atomic_read(&bp->b_hold) > 0);
1072
1073 /*
1074 * We grab the b_lock here first to serialise racing xfs_buf_rele()
1075 * calls. The pag_buf_lock being taken on the last reference only
1076 * serialises against racing lookups in xfs_buf_find(). IOWs, the second
1077 * to last reference we drop here is not serialised against the last
1078 * reference until we take bp->b_lock. Hence if we don't grab b_lock
1079 * first, the last "release" reference can win the race to the lock and
1080 * free the buffer before the second-to-last reference is processed,
1081 * leading to a use-after-free scenario.
1082 */
1083 spin_lock(&bp->b_lock);
1084 release = atomic_dec_and_lock(&bp->b_hold, &bch->bc_lock);
1085 if (!release) {
1086 /*
1087 * Drop the in-flight state if the buffer is already on the LRU
1088 * and it holds the only reference. This is racy because we
1089 * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT
1090 * ensures the decrement occurs only once per-buf.
1091 */
1092 if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru))
1093 __xfs_buf_ioacct_dec(bp);
1094 goto out_unlock;
1095 }
1096
1097 /* the last reference has been dropped ... */
1098 __xfs_buf_ioacct_dec(bp);
1099 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
1100 /*
1101 * If the buffer is added to the LRU take a new reference to the
1102 * buffer for the LRU and clear the (now stale) dispose list
1103 * state flag
1104 */
1105 if (list_lru_add_obj(&btp->bt_lru, &bp->b_lru)) {
1106 bp->b_state &= ~XFS_BSTATE_DISPOSE;
1107 atomic_inc(&bp->b_hold);
1108 }
1109 spin_unlock(&bch->bc_lock);
1110 } else {
1111 /*
1112 * most of the time buffers will already be removed from the
1113 * LRU, so optimise that case by checking for the
1114 * XFS_BSTATE_DISPOSE flag indicating the last list the buffer
1115 * was on was the disposal list
1116 */
1117 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
1118 list_lru_del_obj(&btp->bt_lru, &bp->b_lru);
1119 } else {
1120 ASSERT(list_empty(&bp->b_lru));
1121 }
1122
1123 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1124 rhashtable_remove_fast(&bch->bc_hash, &bp->b_rhash_head,
1125 xfs_buf_hash_params);
1126 spin_unlock(&bch->bc_lock);
1127 if (pag)
1128 xfs_perag_put(pag);
1129 freebuf = true;
1130 }
1131
1132out_unlock:
1133 spin_unlock(&bp->b_lock);
1134
1135 if (freebuf)
1136 xfs_buf_free(bp);
1137}
1138
1139/*
1140 * Release a hold on the specified buffer.
1141 */
1142void
1143xfs_buf_rele(
1144 struct xfs_buf *bp)
1145{
1146 trace_xfs_buf_rele(bp, _RET_IP_);
1147 if (xfs_buf_is_uncached(bp))
1148 xfs_buf_rele_uncached(bp);
1149 else
1150 xfs_buf_rele_cached(bp);
1151}
1152
1153/*
1154 * Lock a buffer object, if it is not already locked.
1155 *
1156 * If we come across a stale, pinned, locked buffer, we know that we are
1157 * being asked to lock a buffer that has been reallocated. Because it is
1158 * pinned, we know that the log has not been pushed to disk and hence it
1159 * will still be locked. Rather than continuing to have trylock attempts
1160 * fail until someone else pushes the log, push it ourselves before
1161 * returning. This means that the xfsaild will not get stuck trying
1162 * to push on stale inode buffers.
1163 */
1164int
1165xfs_buf_trylock(
1166 struct xfs_buf *bp)
1167{
1168 int locked;
1169
1170 locked = down_trylock(&bp->b_sema) == 0;
1171 if (locked)
1172 trace_xfs_buf_trylock(bp, _RET_IP_);
1173 else
1174 trace_xfs_buf_trylock_fail(bp, _RET_IP_);
1175 return locked;
1176}
1177
1178/*
1179 * Lock a buffer object.
1180 *
1181 * If we come across a stale, pinned, locked buffer, we know that we
1182 * are being asked to lock a buffer that has been reallocated. Because
1183 * it is pinned, we know that the log has not been pushed to disk and
1184 * hence it will still be locked. Rather than sleeping until someone
1185 * else pushes the log, push it ourselves before trying to get the lock.
1186 */
1187void
1188xfs_buf_lock(
1189 struct xfs_buf *bp)
1190{
1191 trace_xfs_buf_lock(bp, _RET_IP_);
1192
1193 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
1194 xfs_log_force(bp->b_mount, 0);
1195 down(&bp->b_sema);
1196
1197 trace_xfs_buf_lock_done(bp, _RET_IP_);
1198}
1199
1200void
1201xfs_buf_unlock(
1202 struct xfs_buf *bp)
1203{
1204 ASSERT(xfs_buf_islocked(bp));
1205
1206 up(&bp->b_sema);
1207 trace_xfs_buf_unlock(bp, _RET_IP_);
1208}
1209
1210STATIC void
1211xfs_buf_wait_unpin(
1212 struct xfs_buf *bp)
1213{
1214 DECLARE_WAITQUEUE (wait, current);
1215
1216 if (atomic_read(&bp->b_pin_count) == 0)
1217 return;
1218
1219 add_wait_queue(&bp->b_waiters, &wait);
1220 for (;;) {
1221 set_current_state(TASK_UNINTERRUPTIBLE);
1222 if (atomic_read(&bp->b_pin_count) == 0)
1223 break;
1224 io_schedule();
1225 }
1226 remove_wait_queue(&bp->b_waiters, &wait);
1227 set_current_state(TASK_RUNNING);
1228}
1229
1230static void
1231xfs_buf_ioerror_alert_ratelimited(
1232 struct xfs_buf *bp)
1233{
1234 static unsigned long lasttime;
1235 static struct xfs_buftarg *lasttarg;
1236
1237 if (bp->b_target != lasttarg ||
1238 time_after(jiffies, (lasttime + 5*HZ))) {
1239 lasttime = jiffies;
1240 xfs_buf_ioerror_alert(bp, __this_address);
1241 }
1242 lasttarg = bp->b_target;
1243}
1244
1245/*
1246 * Account for this latest trip around the retry handler, and decide if
1247 * we've failed enough times to constitute a permanent failure.
1248 */
1249static bool
1250xfs_buf_ioerror_permanent(
1251 struct xfs_buf *bp,
1252 struct xfs_error_cfg *cfg)
1253{
1254 struct xfs_mount *mp = bp->b_mount;
1255
1256 if (cfg->max_retries != XFS_ERR_RETRY_FOREVER &&
1257 ++bp->b_retries > cfg->max_retries)
1258 return true;
1259 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1260 time_after(jiffies, cfg->retry_timeout + bp->b_first_retry_time))
1261 return true;
1262
1263 /* At unmount we may treat errors differently */
1264 if (xfs_is_unmounting(mp) && mp->m_fail_unmount)
1265 return true;
1266
1267 return false;
1268}
1269
1270/*
1271 * On a sync write or shutdown we just want to stale the buffer and let the
1272 * caller handle the error in bp->b_error appropriately.
1273 *
1274 * If the write was asynchronous then no one will be looking for the error. If
1275 * this is the first failure of this type, clear the error state and write the
1276 * buffer out again. This means we always retry an async write failure at least
1277 * once, but we also need to set the buffer up to behave correctly now for
1278 * repeated failures.
1279 *
1280 * If we get repeated async write failures, then we take action according to the
1281 * error configuration we have been set up to use.
1282 *
1283 * Returns true if this function took care of error handling and the caller must
1284 * not touch the buffer again. Return false if the caller should proceed with
1285 * normal I/O completion handling.
1286 */
1287static bool
1288xfs_buf_ioend_handle_error(
1289 struct xfs_buf *bp)
1290{
1291 struct xfs_mount *mp = bp->b_mount;
1292 struct xfs_error_cfg *cfg;
1293
1294 /*
1295 * If we've already shutdown the journal because of I/O errors, there's
1296 * no point in giving this a retry.
1297 */
1298 if (xlog_is_shutdown(mp->m_log))
1299 goto out_stale;
1300
1301 xfs_buf_ioerror_alert_ratelimited(bp);
1302
1303 /*
1304 * We're not going to bother about retrying this during recovery.
1305 * One strike!
1306 */
1307 if (bp->b_flags & _XBF_LOGRECOVERY) {
1308 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1309 return false;
1310 }
1311
1312 /*
1313 * Synchronous writes will have callers process the error.
1314 */
1315 if (!(bp->b_flags & XBF_ASYNC))
1316 goto out_stale;
1317
1318 trace_xfs_buf_iodone_async(bp, _RET_IP_);
1319
1320 cfg = xfs_error_get_cfg(mp, XFS_ERR_METADATA, bp->b_error);
1321 if (bp->b_last_error != bp->b_error ||
1322 !(bp->b_flags & (XBF_STALE | XBF_WRITE_FAIL))) {
1323 bp->b_last_error = bp->b_error;
1324 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1325 !bp->b_first_retry_time)
1326 bp->b_first_retry_time = jiffies;
1327 goto resubmit;
1328 }
1329
1330 /*
1331 * Permanent error - we need to trigger a shutdown if we haven't already
1332 * to indicate that inconsistency will result from this action.
1333 */
1334 if (xfs_buf_ioerror_permanent(bp, cfg)) {
1335 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1336 goto out_stale;
1337 }
1338
1339 /* Still considered a transient error. Caller will schedule retries. */
1340 if (bp->b_flags & _XBF_INODES)
1341 xfs_buf_inode_io_fail(bp);
1342 else if (bp->b_flags & _XBF_DQUOTS)
1343 xfs_buf_dquot_io_fail(bp);
1344 else
1345 ASSERT(list_empty(&bp->b_li_list));
1346 xfs_buf_ioerror(bp, 0);
1347 xfs_buf_relse(bp);
1348 return true;
1349
1350resubmit:
1351 xfs_buf_ioerror(bp, 0);
1352 bp->b_flags |= (XBF_DONE | XBF_WRITE_FAIL);
1353 xfs_buf_submit(bp);
1354 return true;
1355out_stale:
1356 xfs_buf_stale(bp);
1357 bp->b_flags |= XBF_DONE;
1358 bp->b_flags &= ~XBF_WRITE;
1359 trace_xfs_buf_error_relse(bp, _RET_IP_);
1360 return false;
1361}
1362
1363static void
1364xfs_buf_ioend(
1365 struct xfs_buf *bp)
1366{
1367 trace_xfs_buf_iodone(bp, _RET_IP_);
1368
1369 /*
1370 * Pull in IO completion errors now. We are guaranteed to be running
1371 * single threaded, so we don't need the lock to read b_io_error.
1372 */
1373 if (!bp->b_error && bp->b_io_error)
1374 xfs_buf_ioerror(bp, bp->b_io_error);
1375
1376 if (bp->b_flags & XBF_READ) {
1377 if (!bp->b_error && bp->b_ops)
1378 bp->b_ops->verify_read(bp);
1379 if (!bp->b_error)
1380 bp->b_flags |= XBF_DONE;
1381 } else {
1382 if (!bp->b_error) {
1383 bp->b_flags &= ~XBF_WRITE_FAIL;
1384 bp->b_flags |= XBF_DONE;
1385 }
1386
1387 if (unlikely(bp->b_error) && xfs_buf_ioend_handle_error(bp))
1388 return;
1389
1390 /* clear the retry state */
1391 bp->b_last_error = 0;
1392 bp->b_retries = 0;
1393 bp->b_first_retry_time = 0;
1394
1395 /*
1396 * Note that for things like remote attribute buffers, there may
1397 * not be a buffer log item here, so processing the buffer log
1398 * item must remain optional.
1399 */
1400 if (bp->b_log_item)
1401 xfs_buf_item_done(bp);
1402
1403 if (bp->b_flags & _XBF_INODES)
1404 xfs_buf_inode_iodone(bp);
1405 else if (bp->b_flags & _XBF_DQUOTS)
1406 xfs_buf_dquot_iodone(bp);
1407
1408 }
1409
1410 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD |
1411 _XBF_LOGRECOVERY);
1412
1413 if (bp->b_flags & XBF_ASYNC)
1414 xfs_buf_relse(bp);
1415 else
1416 complete(&bp->b_iowait);
1417}
1418
1419static void
1420xfs_buf_ioend_work(
1421 struct work_struct *work)
1422{
1423 struct xfs_buf *bp =
1424 container_of(work, struct xfs_buf, b_ioend_work);
1425
1426 xfs_buf_ioend(bp);
1427}
1428
1429static void
1430xfs_buf_ioend_async(
1431 struct xfs_buf *bp)
1432{
1433 INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
1434 queue_work(bp->b_mount->m_buf_workqueue, &bp->b_ioend_work);
1435}
1436
1437void
1438__xfs_buf_ioerror(
1439 struct xfs_buf *bp,
1440 int error,
1441 xfs_failaddr_t failaddr)
1442{
1443 ASSERT(error <= 0 && error >= -1000);
1444 bp->b_error = error;
1445 trace_xfs_buf_ioerror(bp, error, failaddr);
1446}
1447
1448void
1449xfs_buf_ioerror_alert(
1450 struct xfs_buf *bp,
1451 xfs_failaddr_t func)
1452{
1453 xfs_buf_alert_ratelimited(bp, "XFS: metadata IO error",
1454 "metadata I/O error in \"%pS\" at daddr 0x%llx len %d error %d",
1455 func, (uint64_t)xfs_buf_daddr(bp),
1456 bp->b_length, -bp->b_error);
1457}
1458
1459/*
1460 * To simulate an I/O failure, the buffer must be locked and held with at least
1461 * three references. The LRU reference is dropped by the stale call. The buf
1462 * item reference is dropped via ioend processing. The third reference is owned
1463 * by the caller and is dropped on I/O completion if the buffer is XBF_ASYNC.
1464 */
1465void
1466xfs_buf_ioend_fail(
1467 struct xfs_buf *bp)
1468{
1469 bp->b_flags &= ~XBF_DONE;
1470 xfs_buf_stale(bp);
1471 xfs_buf_ioerror(bp, -EIO);
1472 xfs_buf_ioend(bp);
1473}
1474
1475int
1476xfs_bwrite(
1477 struct xfs_buf *bp)
1478{
1479 int error;
1480
1481 ASSERT(xfs_buf_islocked(bp));
1482
1483 bp->b_flags |= XBF_WRITE;
1484 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1485 XBF_DONE);
1486
1487 error = xfs_buf_submit(bp);
1488 if (error)
1489 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
1490 return error;
1491}
1492
1493static void
1494xfs_buf_bio_end_io(
1495 struct bio *bio)
1496{
1497 struct xfs_buf *bp = (struct xfs_buf *)bio->bi_private;
1498
1499 if (!bio->bi_status &&
1500 (bp->b_flags & XBF_WRITE) && (bp->b_flags & XBF_ASYNC) &&
1501 XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_IOERROR))
1502 bio->bi_status = BLK_STS_IOERR;
1503
1504 /*
1505 * don't overwrite existing errors - otherwise we can lose errors on
1506 * buffers that require multiple bios to complete.
1507 */
1508 if (bio->bi_status) {
1509 int error = blk_status_to_errno(bio->bi_status);
1510
1511 cmpxchg(&bp->b_io_error, 0, error);
1512 }
1513
1514 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1515 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1516
1517 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1518 xfs_buf_ioend_async(bp);
1519 bio_put(bio);
1520}
1521
1522static void
1523xfs_buf_ioapply_map(
1524 struct xfs_buf *bp,
1525 int map,
1526 int *buf_offset,
1527 int *count,
1528 blk_opf_t op)
1529{
1530 int page_index;
1531 unsigned int total_nr_pages = bp->b_page_count;
1532 int nr_pages;
1533 struct bio *bio;
1534 sector_t sector = bp->b_maps[map].bm_bn;
1535 int size;
1536 int offset;
1537
1538 /* skip the pages in the buffer before the start offset */
1539 page_index = 0;
1540 offset = *buf_offset;
1541 while (offset >= PAGE_SIZE) {
1542 page_index++;
1543 offset -= PAGE_SIZE;
1544 }
1545
1546 /*
1547 * Limit the IO size to the length of the current vector, and update the
1548 * remaining IO count for the next time around.
1549 */
1550 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1551 *count -= size;
1552 *buf_offset += size;
1553
1554next_chunk:
1555 atomic_inc(&bp->b_io_remaining);
1556 nr_pages = bio_max_segs(total_nr_pages);
1557
1558 bio = bio_alloc(bp->b_target->bt_bdev, nr_pages, op, GFP_NOIO);
1559 bio->bi_iter.bi_sector = sector;
1560 bio->bi_end_io = xfs_buf_bio_end_io;
1561 bio->bi_private = bp;
1562
1563 for (; size && nr_pages; nr_pages--, page_index++) {
1564 int rbytes, nbytes = PAGE_SIZE - offset;
1565
1566 if (nbytes > size)
1567 nbytes = size;
1568
1569 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1570 offset);
1571 if (rbytes < nbytes)
1572 break;
1573
1574 offset = 0;
1575 sector += BTOBB(nbytes);
1576 size -= nbytes;
1577 total_nr_pages--;
1578 }
1579
1580 if (likely(bio->bi_iter.bi_size)) {
1581 if (xfs_buf_is_vmapped(bp)) {
1582 flush_kernel_vmap_range(bp->b_addr,
1583 xfs_buf_vmap_len(bp));
1584 }
1585 submit_bio(bio);
1586 if (size)
1587 goto next_chunk;
1588 } else {
1589 /*
1590 * This is guaranteed not to be the last io reference count
1591 * because the caller (xfs_buf_submit) holds a count itself.
1592 */
1593 atomic_dec(&bp->b_io_remaining);
1594 xfs_buf_ioerror(bp, -EIO);
1595 bio_put(bio);
1596 }
1597
1598}
1599
1600STATIC void
1601_xfs_buf_ioapply(
1602 struct xfs_buf *bp)
1603{
1604 struct blk_plug plug;
1605 blk_opf_t op;
1606 int offset;
1607 int size;
1608 int i;
1609
1610 /*
1611 * Make sure we capture only current IO errors rather than stale errors
1612 * left over from previous use of the buffer (e.g. failed readahead).
1613 */
1614 bp->b_error = 0;
1615
1616 if (bp->b_flags & XBF_WRITE) {
1617 op = REQ_OP_WRITE;
1618
1619 /*
1620 * Run the write verifier callback function if it exists. If
1621 * this function fails it will mark the buffer with an error and
1622 * the IO should not be dispatched.
1623 */
1624 if (bp->b_ops) {
1625 bp->b_ops->verify_write(bp);
1626 if (bp->b_error) {
1627 xfs_force_shutdown(bp->b_mount,
1628 SHUTDOWN_CORRUPT_INCORE);
1629 return;
1630 }
1631 } else if (bp->b_rhash_key != XFS_BUF_DADDR_NULL) {
1632 struct xfs_mount *mp = bp->b_mount;
1633
1634 /*
1635 * non-crc filesystems don't attach verifiers during
1636 * log recovery, so don't warn for such filesystems.
1637 */
1638 if (xfs_has_crc(mp)) {
1639 xfs_warn(mp,
1640 "%s: no buf ops on daddr 0x%llx len %d",
1641 __func__, xfs_buf_daddr(bp),
1642 bp->b_length);
1643 xfs_hex_dump(bp->b_addr,
1644 XFS_CORRUPTION_DUMP_LEN);
1645 dump_stack();
1646 }
1647 }
1648 } else {
1649 op = REQ_OP_READ;
1650 if (bp->b_flags & XBF_READ_AHEAD)
1651 op |= REQ_RAHEAD;
1652 }
1653
1654 /* we only use the buffer cache for meta-data */
1655 op |= REQ_META;
1656
1657 /* in-memory targets are directly mapped, no IO required. */
1658 if (xfs_buftarg_is_mem(bp->b_target)) {
1659 xfs_buf_ioend(bp);
1660 return;
1661 }
1662
1663 /*
1664 * Walk all the vectors issuing IO on them. Set up the initial offset
1665 * into the buffer and the desired IO size before we start -
1666 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1667 * subsequent call.
1668 */
1669 offset = bp->b_offset;
1670 size = BBTOB(bp->b_length);
1671 blk_start_plug(&plug);
1672 for (i = 0; i < bp->b_map_count; i++) {
1673 xfs_buf_ioapply_map(bp, i, &offset, &size, op);
1674 if (bp->b_error)
1675 break;
1676 if (size <= 0)
1677 break; /* all done */
1678 }
1679 blk_finish_plug(&plug);
1680}
1681
1682/*
1683 * Wait for I/O completion of a sync buffer and return the I/O error code.
1684 */
1685static int
1686xfs_buf_iowait(
1687 struct xfs_buf *bp)
1688{
1689 ASSERT(!(bp->b_flags & XBF_ASYNC));
1690
1691 trace_xfs_buf_iowait(bp, _RET_IP_);
1692 wait_for_completion(&bp->b_iowait);
1693 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1694
1695 return bp->b_error;
1696}
1697
1698/*
1699 * Buffer I/O submission path, read or write. Asynchronous submission transfers
1700 * the buffer lock ownership and the current reference to the IO. It is not
1701 * safe to reference the buffer after a call to this function unless the caller
1702 * holds an additional reference itself.
1703 */
1704static int
1705__xfs_buf_submit(
1706 struct xfs_buf *bp,
1707 bool wait)
1708{
1709 int error = 0;
1710
1711 trace_xfs_buf_submit(bp, _RET_IP_);
1712
1713 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1714
1715 /*
1716 * On log shutdown we stale and complete the buffer immediately. We can
1717 * be called to read the superblock before the log has been set up, so
1718 * be careful checking the log state.
1719 *
1720 * Checking the mount shutdown state here can result in the log tail
1721 * moving inappropriately on disk as the log may not yet be shut down.
1722 * i.e. failing this buffer on mount shutdown can remove it from the AIL
1723 * and move the tail of the log forwards without having written this
1724 * buffer to disk. This corrupts the log tail state in memory, and
1725 * because the log may not be shut down yet, it can then be propagated
1726 * to disk before the log is shutdown. Hence we check log shutdown
1727 * state here rather than mount state to avoid corrupting the log tail
1728 * on shutdown.
1729 */
1730 if (bp->b_mount->m_log &&
1731 xlog_is_shutdown(bp->b_mount->m_log)) {
1732 xfs_buf_ioend_fail(bp);
1733 return -EIO;
1734 }
1735
1736 /*
1737 * Grab a reference so the buffer does not go away underneath us. For
1738 * async buffers, I/O completion drops the callers reference, which
1739 * could occur before submission returns.
1740 */
1741 xfs_buf_hold(bp);
1742
1743 if (bp->b_flags & XBF_WRITE)
1744 xfs_buf_wait_unpin(bp);
1745
1746 /* clear the internal error state to avoid spurious errors */
1747 bp->b_io_error = 0;
1748
1749 /*
1750 * Set the count to 1 initially, this will stop an I/O completion
1751 * callout which happens before we have started all the I/O from calling
1752 * xfs_buf_ioend too early.
1753 */
1754 atomic_set(&bp->b_io_remaining, 1);
1755 if (bp->b_flags & XBF_ASYNC)
1756 xfs_buf_ioacct_inc(bp);
1757 _xfs_buf_ioapply(bp);
1758
1759 /*
1760 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1761 * reference we took above. If we drop it to zero, run completion so
1762 * that we don't return to the caller with completion still pending.
1763 */
1764 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1765 if (bp->b_error || !(bp->b_flags & XBF_ASYNC))
1766 xfs_buf_ioend(bp);
1767 else
1768 xfs_buf_ioend_async(bp);
1769 }
1770
1771 if (wait)
1772 error = xfs_buf_iowait(bp);
1773
1774 /*
1775 * Release the hold that keeps the buffer referenced for the entire
1776 * I/O. Note that if the buffer is async, it is not safe to reference
1777 * after this release.
1778 */
1779 xfs_buf_rele(bp);
1780 return error;
1781}
1782
1783void *
1784xfs_buf_offset(
1785 struct xfs_buf *bp,
1786 size_t offset)
1787{
1788 struct page *page;
1789
1790 if (bp->b_addr)
1791 return bp->b_addr + offset;
1792
1793 page = bp->b_pages[offset >> PAGE_SHIFT];
1794 return page_address(page) + (offset & (PAGE_SIZE-1));
1795}
1796
1797void
1798xfs_buf_zero(
1799 struct xfs_buf *bp,
1800 size_t boff,
1801 size_t bsize)
1802{
1803 size_t bend;
1804
1805 bend = boff + bsize;
1806 while (boff < bend) {
1807 struct page *page;
1808 int page_index, page_offset, csize;
1809
1810 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1811 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1812 page = bp->b_pages[page_index];
1813 csize = min_t(size_t, PAGE_SIZE - page_offset,
1814 BBTOB(bp->b_length) - boff);
1815
1816 ASSERT((csize + page_offset) <= PAGE_SIZE);
1817
1818 memset(page_address(page) + page_offset, 0, csize);
1819
1820 boff += csize;
1821 }
1822}
1823
1824/*
1825 * Log a message about and stale a buffer that a caller has decided is corrupt.
1826 *
1827 * This function should be called for the kinds of metadata corruption that
1828 * cannot be detect from a verifier, such as incorrect inter-block relationship
1829 * data. Do /not/ call this function from a verifier function.
1830 *
1831 * The buffer must be XBF_DONE prior to the call. Afterwards, the buffer will
1832 * be marked stale, but b_error will not be set. The caller is responsible for
1833 * releasing the buffer or fixing it.
1834 */
1835void
1836__xfs_buf_mark_corrupt(
1837 struct xfs_buf *bp,
1838 xfs_failaddr_t fa)
1839{
1840 ASSERT(bp->b_flags & XBF_DONE);
1841
1842 xfs_buf_corruption_error(bp, fa);
1843 xfs_buf_stale(bp);
1844}
1845
1846/*
1847 * Handling of buffer targets (buftargs).
1848 */
1849
1850/*
1851 * Wait for any bufs with callbacks that have been submitted but have not yet
1852 * returned. These buffers will have an elevated hold count, so wait on those
1853 * while freeing all the buffers only held by the LRU.
1854 */
1855static enum lru_status
1856xfs_buftarg_drain_rele(
1857 struct list_head *item,
1858 struct list_lru_one *lru,
1859 void *arg)
1860
1861{
1862 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1863 struct list_head *dispose = arg;
1864
1865 if (atomic_read(&bp->b_hold) > 1) {
1866 /* need to wait, so skip it this pass */
1867 trace_xfs_buf_drain_buftarg(bp, _RET_IP_);
1868 return LRU_SKIP;
1869 }
1870 if (!spin_trylock(&bp->b_lock))
1871 return LRU_SKIP;
1872
1873 /*
1874 * clear the LRU reference count so the buffer doesn't get
1875 * ignored in xfs_buf_rele().
1876 */
1877 atomic_set(&bp->b_lru_ref, 0);
1878 bp->b_state |= XFS_BSTATE_DISPOSE;
1879 list_lru_isolate_move(lru, item, dispose);
1880 spin_unlock(&bp->b_lock);
1881 return LRU_REMOVED;
1882}
1883
1884/*
1885 * Wait for outstanding I/O on the buftarg to complete.
1886 */
1887void
1888xfs_buftarg_wait(
1889 struct xfs_buftarg *btp)
1890{
1891 /*
1892 * First wait on the buftarg I/O count for all in-flight buffers to be
1893 * released. This is critical as new buffers do not make the LRU until
1894 * they are released.
1895 *
1896 * Next, flush the buffer workqueue to ensure all completion processing
1897 * has finished. Just waiting on buffer locks is not sufficient for
1898 * async IO as the reference count held over IO is not released until
1899 * after the buffer lock is dropped. Hence we need to ensure here that
1900 * all reference counts have been dropped before we start walking the
1901 * LRU list.
1902 */
1903 while (percpu_counter_sum(&btp->bt_io_count))
1904 delay(100);
1905 flush_workqueue(btp->bt_mount->m_buf_workqueue);
1906}
1907
1908void
1909xfs_buftarg_drain(
1910 struct xfs_buftarg *btp)
1911{
1912 LIST_HEAD(dispose);
1913 int loop = 0;
1914 bool write_fail = false;
1915
1916 xfs_buftarg_wait(btp);
1917
1918 /* loop until there is nothing left on the lru list. */
1919 while (list_lru_count(&btp->bt_lru)) {
1920 list_lru_walk(&btp->bt_lru, xfs_buftarg_drain_rele,
1921 &dispose, LONG_MAX);
1922
1923 while (!list_empty(&dispose)) {
1924 struct xfs_buf *bp;
1925 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1926 list_del_init(&bp->b_lru);
1927 if (bp->b_flags & XBF_WRITE_FAIL) {
1928 write_fail = true;
1929 xfs_buf_alert_ratelimited(bp,
1930 "XFS: Corruption Alert",
1931"Corruption Alert: Buffer at daddr 0x%llx had permanent write failures!",
1932 (long long)xfs_buf_daddr(bp));
1933 }
1934 xfs_buf_rele(bp);
1935 }
1936 if (loop++ != 0)
1937 delay(100);
1938 }
1939
1940 /*
1941 * If one or more failed buffers were freed, that means dirty metadata
1942 * was thrown away. This should only ever happen after I/O completion
1943 * handling has elevated I/O error(s) to permanent failures and shuts
1944 * down the journal.
1945 */
1946 if (write_fail) {
1947 ASSERT(xlog_is_shutdown(btp->bt_mount->m_log));
1948 xfs_alert(btp->bt_mount,
1949 "Please run xfs_repair to determine the extent of the problem.");
1950 }
1951}
1952
1953static enum lru_status
1954xfs_buftarg_isolate(
1955 struct list_head *item,
1956 struct list_lru_one *lru,
1957 void *arg)
1958{
1959 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1960 struct list_head *dispose = arg;
1961
1962 /*
1963 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1964 * If we fail to get the lock, just skip it.
1965 */
1966 if (!spin_trylock(&bp->b_lock))
1967 return LRU_SKIP;
1968 /*
1969 * Decrement the b_lru_ref count unless the value is already
1970 * zero. If the value is already zero, we need to reclaim the
1971 * buffer, otherwise it gets another trip through the LRU.
1972 */
1973 if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1974 spin_unlock(&bp->b_lock);
1975 return LRU_ROTATE;
1976 }
1977
1978 bp->b_state |= XFS_BSTATE_DISPOSE;
1979 list_lru_isolate_move(lru, item, dispose);
1980 spin_unlock(&bp->b_lock);
1981 return LRU_REMOVED;
1982}
1983
1984static unsigned long
1985xfs_buftarg_shrink_scan(
1986 struct shrinker *shrink,
1987 struct shrink_control *sc)
1988{
1989 struct xfs_buftarg *btp = shrink->private_data;
1990 LIST_HEAD(dispose);
1991 unsigned long freed;
1992
1993 freed = list_lru_shrink_walk(&btp->bt_lru, sc,
1994 xfs_buftarg_isolate, &dispose);
1995
1996 while (!list_empty(&dispose)) {
1997 struct xfs_buf *bp;
1998 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1999 list_del_init(&bp->b_lru);
2000 xfs_buf_rele(bp);
2001 }
2002
2003 return freed;
2004}
2005
2006static unsigned long
2007xfs_buftarg_shrink_count(
2008 struct shrinker *shrink,
2009 struct shrink_control *sc)
2010{
2011 struct xfs_buftarg *btp = shrink->private_data;
2012 return list_lru_shrink_count(&btp->bt_lru, sc);
2013}
2014
2015void
2016xfs_destroy_buftarg(
2017 struct xfs_buftarg *btp)
2018{
2019 shrinker_free(btp->bt_shrinker);
2020 ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
2021 percpu_counter_destroy(&btp->bt_io_count);
2022 list_lru_destroy(&btp->bt_lru);
2023}
2024
2025void
2026xfs_free_buftarg(
2027 struct xfs_buftarg *btp)
2028{
2029 xfs_destroy_buftarg(btp);
2030 fs_put_dax(btp->bt_daxdev, btp->bt_mount);
2031 /* the main block device is closed by kill_block_super */
2032 if (btp->bt_bdev != btp->bt_mount->m_super->s_bdev)
2033 bdev_fput(btp->bt_bdev_file);
2034 kfree(btp);
2035}
2036
2037int
2038xfs_setsize_buftarg(
2039 struct xfs_buftarg *btp,
2040 unsigned int sectorsize)
2041{
2042 /* Set up metadata sector size info */
2043 btp->bt_meta_sectorsize = sectorsize;
2044 btp->bt_meta_sectormask = sectorsize - 1;
2045
2046 if (set_blocksize(btp->bt_bdev_file, sectorsize)) {
2047 xfs_warn(btp->bt_mount,
2048 "Cannot set_blocksize to %u on device %pg",
2049 sectorsize, btp->bt_bdev);
2050 return -EINVAL;
2051 }
2052
2053 return 0;
2054}
2055
2056int
2057xfs_init_buftarg(
2058 struct xfs_buftarg *btp,
2059 size_t logical_sectorsize,
2060 const char *descr)
2061{
2062 /* Set up device logical sector size mask */
2063 btp->bt_logical_sectorsize = logical_sectorsize;
2064 btp->bt_logical_sectormask = logical_sectorsize - 1;
2065
2066 /*
2067 * Buffer IO error rate limiting. Limit it to no more than 10 messages
2068 * per 30 seconds so as to not spam logs too much on repeated errors.
2069 */
2070 ratelimit_state_init(&btp->bt_ioerror_rl, 30 * HZ,
2071 DEFAULT_RATELIMIT_BURST);
2072
2073 if (list_lru_init(&btp->bt_lru))
2074 return -ENOMEM;
2075 if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL))
2076 goto out_destroy_lru;
2077
2078 btp->bt_shrinker =
2079 shrinker_alloc(SHRINKER_NUMA_AWARE, "xfs-buf:%s", descr);
2080 if (!btp->bt_shrinker)
2081 goto out_destroy_io_count;
2082 btp->bt_shrinker->count_objects = xfs_buftarg_shrink_count;
2083 btp->bt_shrinker->scan_objects = xfs_buftarg_shrink_scan;
2084 btp->bt_shrinker->private_data = btp;
2085 shrinker_register(btp->bt_shrinker);
2086 return 0;
2087
2088out_destroy_io_count:
2089 percpu_counter_destroy(&btp->bt_io_count);
2090out_destroy_lru:
2091 list_lru_destroy(&btp->bt_lru);
2092 return -ENOMEM;
2093}
2094
2095struct xfs_buftarg *
2096xfs_alloc_buftarg(
2097 struct xfs_mount *mp,
2098 struct file *bdev_file)
2099{
2100 struct xfs_buftarg *btp;
2101 const struct dax_holder_operations *ops = NULL;
2102
2103#if defined(CONFIG_FS_DAX) && defined(CONFIG_MEMORY_FAILURE)
2104 ops = &xfs_dax_holder_operations;
2105#endif
2106 btp = kzalloc(sizeof(*btp), GFP_KERNEL | __GFP_NOFAIL);
2107
2108 btp->bt_mount = mp;
2109 btp->bt_bdev_file = bdev_file;
2110 btp->bt_bdev = file_bdev(bdev_file);
2111 btp->bt_dev = btp->bt_bdev->bd_dev;
2112 btp->bt_daxdev = fs_dax_get_by_bdev(btp->bt_bdev, &btp->bt_dax_part_off,
2113 mp, ops);
2114
2115 if (bdev_can_atomic_write(btp->bt_bdev)) {
2116 btp->bt_bdev_awu_min = bdev_atomic_write_unit_min_bytes(
2117 btp->bt_bdev);
2118 btp->bt_bdev_awu_max = bdev_atomic_write_unit_max_bytes(
2119 btp->bt_bdev);
2120 }
2121
2122 /*
2123 * When allocating the buftargs we have not yet read the super block and
2124 * thus don't know the file system sector size yet.
2125 */
2126 if (xfs_setsize_buftarg(btp, bdev_logical_block_size(btp->bt_bdev)))
2127 goto error_free;
2128 if (xfs_init_buftarg(btp, bdev_logical_block_size(btp->bt_bdev),
2129 mp->m_super->s_id))
2130 goto error_free;
2131
2132 return btp;
2133
2134error_free:
2135 kfree(btp);
2136 return NULL;
2137}
2138
2139static inline void
2140xfs_buf_list_del(
2141 struct xfs_buf *bp)
2142{
2143 list_del_init(&bp->b_list);
2144 wake_up_var(&bp->b_list);
2145}
2146
2147/*
2148 * Cancel a delayed write list.
2149 *
2150 * Remove each buffer from the list, clear the delwri queue flag and drop the
2151 * associated buffer reference.
2152 */
2153void
2154xfs_buf_delwri_cancel(
2155 struct list_head *list)
2156{
2157 struct xfs_buf *bp;
2158
2159 while (!list_empty(list)) {
2160 bp = list_first_entry(list, struct xfs_buf, b_list);
2161
2162 xfs_buf_lock(bp);
2163 bp->b_flags &= ~_XBF_DELWRI_Q;
2164 xfs_buf_list_del(bp);
2165 xfs_buf_relse(bp);
2166 }
2167}
2168
2169/*
2170 * Add a buffer to the delayed write list.
2171 *
2172 * This queues a buffer for writeout if it hasn't already been. Note that
2173 * neither this routine nor the buffer list submission functions perform
2174 * any internal synchronization. It is expected that the lists are thread-local
2175 * to the callers.
2176 *
2177 * Returns true if we queued up the buffer, or false if it already had
2178 * been on the buffer list.
2179 */
2180bool
2181xfs_buf_delwri_queue(
2182 struct xfs_buf *bp,
2183 struct list_head *list)
2184{
2185 ASSERT(xfs_buf_islocked(bp));
2186 ASSERT(!(bp->b_flags & XBF_READ));
2187
2188 /*
2189 * If the buffer is already marked delwri it already is queued up
2190 * by someone else for imediate writeout. Just ignore it in that
2191 * case.
2192 */
2193 if (bp->b_flags & _XBF_DELWRI_Q) {
2194 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
2195 return false;
2196 }
2197
2198 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
2199
2200 /*
2201 * If a buffer gets written out synchronously or marked stale while it
2202 * is on a delwri list we lazily remove it. To do this, the other party
2203 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
2204 * It remains referenced and on the list. In a rare corner case it
2205 * might get readded to a delwri list after the synchronous writeout, in
2206 * which case we need just need to re-add the flag here.
2207 */
2208 bp->b_flags |= _XBF_DELWRI_Q;
2209 if (list_empty(&bp->b_list)) {
2210 atomic_inc(&bp->b_hold);
2211 list_add_tail(&bp->b_list, list);
2212 }
2213
2214 return true;
2215}
2216
2217/*
2218 * Queue a buffer to this delwri list as part of a data integrity operation.
2219 * If the buffer is on any other delwri list, we'll wait for that to clear
2220 * so that the caller can submit the buffer for IO and wait for the result.
2221 * Callers must ensure the buffer is not already on the list.
2222 */
2223void
2224xfs_buf_delwri_queue_here(
2225 struct xfs_buf *bp,
2226 struct list_head *buffer_list)
2227{
2228 /*
2229 * We need this buffer to end up on the /caller's/ delwri list, not any
2230 * old list. This can happen if the buffer is marked stale (which
2231 * clears DELWRI_Q) after the AIL queues the buffer to its list but
2232 * before the AIL has a chance to submit the list.
2233 */
2234 while (!list_empty(&bp->b_list)) {
2235 xfs_buf_unlock(bp);
2236 wait_var_event(&bp->b_list, list_empty(&bp->b_list));
2237 xfs_buf_lock(bp);
2238 }
2239
2240 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
2241
2242 xfs_buf_delwri_queue(bp, buffer_list);
2243}
2244
2245/*
2246 * Compare function is more complex than it needs to be because
2247 * the return value is only 32 bits and we are doing comparisons
2248 * on 64 bit values
2249 */
2250static int
2251xfs_buf_cmp(
2252 void *priv,
2253 const struct list_head *a,
2254 const struct list_head *b)
2255{
2256 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
2257 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
2258 xfs_daddr_t diff;
2259
2260 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
2261 if (diff < 0)
2262 return -1;
2263 if (diff > 0)
2264 return 1;
2265 return 0;
2266}
2267
2268/*
2269 * Submit buffers for write. If wait_list is specified, the buffers are
2270 * submitted using sync I/O and placed on the wait list such that the caller can
2271 * iowait each buffer. Otherwise async I/O is used and the buffers are released
2272 * at I/O completion time. In either case, buffers remain locked until I/O
2273 * completes and the buffer is released from the queue.
2274 */
2275static int
2276xfs_buf_delwri_submit_buffers(
2277 struct list_head *buffer_list,
2278 struct list_head *wait_list)
2279{
2280 struct xfs_buf *bp, *n;
2281 int pinned = 0;
2282 struct blk_plug plug;
2283
2284 list_sort(NULL, buffer_list, xfs_buf_cmp);
2285
2286 blk_start_plug(&plug);
2287 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
2288 if (!wait_list) {
2289 if (!xfs_buf_trylock(bp))
2290 continue;
2291 if (xfs_buf_ispinned(bp)) {
2292 xfs_buf_unlock(bp);
2293 pinned++;
2294 continue;
2295 }
2296 } else {
2297 xfs_buf_lock(bp);
2298 }
2299
2300 /*
2301 * Someone else might have written the buffer synchronously or
2302 * marked it stale in the meantime. In that case only the
2303 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
2304 * reference and remove it from the list here.
2305 */
2306 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
2307 xfs_buf_list_del(bp);
2308 xfs_buf_relse(bp);
2309 continue;
2310 }
2311
2312 trace_xfs_buf_delwri_split(bp, _RET_IP_);
2313
2314 /*
2315 * If we have a wait list, each buffer (and associated delwri
2316 * queue reference) transfers to it and is submitted
2317 * synchronously. Otherwise, drop the buffer from the delwri
2318 * queue and submit async.
2319 */
2320 bp->b_flags &= ~_XBF_DELWRI_Q;
2321 bp->b_flags |= XBF_WRITE;
2322 if (wait_list) {
2323 bp->b_flags &= ~XBF_ASYNC;
2324 list_move_tail(&bp->b_list, wait_list);
2325 } else {
2326 bp->b_flags |= XBF_ASYNC;
2327 xfs_buf_list_del(bp);
2328 }
2329 __xfs_buf_submit(bp, false);
2330 }
2331 blk_finish_plug(&plug);
2332
2333 return pinned;
2334}
2335
2336/*
2337 * Write out a buffer list asynchronously.
2338 *
2339 * This will take the @buffer_list, write all non-locked and non-pinned buffers
2340 * out and not wait for I/O completion on any of the buffers. This interface
2341 * is only safely useable for callers that can track I/O completion by higher
2342 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
2343 * function.
2344 *
2345 * Note: this function will skip buffers it would block on, and in doing so
2346 * leaves them on @buffer_list so they can be retried on a later pass. As such,
2347 * it is up to the caller to ensure that the buffer list is fully submitted or
2348 * cancelled appropriately when they are finished with the list. Failure to
2349 * cancel or resubmit the list until it is empty will result in leaked buffers
2350 * at unmount time.
2351 */
2352int
2353xfs_buf_delwri_submit_nowait(
2354 struct list_head *buffer_list)
2355{
2356 return xfs_buf_delwri_submit_buffers(buffer_list, NULL);
2357}
2358
2359/*
2360 * Write out a buffer list synchronously.
2361 *
2362 * This will take the @buffer_list, write all buffers out and wait for I/O
2363 * completion on all of the buffers. @buffer_list is consumed by the function,
2364 * so callers must have some other way of tracking buffers if they require such
2365 * functionality.
2366 */
2367int
2368xfs_buf_delwri_submit(
2369 struct list_head *buffer_list)
2370{
2371 LIST_HEAD (wait_list);
2372 int error = 0, error2;
2373 struct xfs_buf *bp;
2374
2375 xfs_buf_delwri_submit_buffers(buffer_list, &wait_list);
2376
2377 /* Wait for IO to complete. */
2378 while (!list_empty(&wait_list)) {
2379 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
2380
2381 xfs_buf_list_del(bp);
2382
2383 /*
2384 * Wait on the locked buffer, check for errors and unlock and
2385 * release the delwri queue reference.
2386 */
2387 error2 = xfs_buf_iowait(bp);
2388 xfs_buf_relse(bp);
2389 if (!error)
2390 error = error2;
2391 }
2392
2393 return error;
2394}
2395
2396/*
2397 * Push a single buffer on a delwri queue.
2398 *
2399 * The purpose of this function is to submit a single buffer of a delwri queue
2400 * and return with the buffer still on the original queue. The waiting delwri
2401 * buffer submission infrastructure guarantees transfer of the delwri queue
2402 * buffer reference to a temporary wait list. We reuse this infrastructure to
2403 * transfer the buffer back to the original queue.
2404 *
2405 * Note the buffer transitions from the queued state, to the submitted and wait
2406 * listed state and back to the queued state during this call. The buffer
2407 * locking and queue management logic between _delwri_pushbuf() and
2408 * _delwri_queue() guarantee that the buffer cannot be queued to another list
2409 * before returning.
2410 */
2411int
2412xfs_buf_delwri_pushbuf(
2413 struct xfs_buf *bp,
2414 struct list_head *buffer_list)
2415{
2416 LIST_HEAD (submit_list);
2417 int error;
2418
2419 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
2420
2421 trace_xfs_buf_delwri_pushbuf(bp, _RET_IP_);
2422
2423 /*
2424 * Isolate the buffer to a new local list so we can submit it for I/O
2425 * independently from the rest of the original list.
2426 */
2427 xfs_buf_lock(bp);
2428 list_move(&bp->b_list, &submit_list);
2429 xfs_buf_unlock(bp);
2430
2431 /*
2432 * Delwri submission clears the DELWRI_Q buffer flag and returns with
2433 * the buffer on the wait list with the original reference. Rather than
2434 * bounce the buffer from a local wait list back to the original list
2435 * after I/O completion, reuse the original list as the wait list.
2436 */
2437 xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
2438
2439 /*
2440 * The buffer is now locked, under I/O and wait listed on the original
2441 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
2442 * return with the buffer unlocked and on the original queue.
2443 */
2444 error = xfs_buf_iowait(bp);
2445 bp->b_flags |= _XBF_DELWRI_Q;
2446 xfs_buf_unlock(bp);
2447
2448 return error;
2449}
2450
2451void xfs_buf_set_ref(struct xfs_buf *bp, int lru_ref)
2452{
2453 /*
2454 * Set the lru reference count to 0 based on the error injection tag.
2455 * This allows userspace to disrupt buffer caching for debug/testing
2456 * purposes.
2457 */
2458 if (XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_LRU_REF))
2459 lru_ref = 0;
2460
2461 atomic_set(&bp->b_lru_ref, lru_ref);
2462}
2463
2464/*
2465 * Verify an on-disk magic value against the magic value specified in the
2466 * verifier structure. The verifier magic is in disk byte order so the caller is
2467 * expected to pass the value directly from disk.
2468 */
2469bool
2470xfs_verify_magic(
2471 struct xfs_buf *bp,
2472 __be32 dmagic)
2473{
2474 struct xfs_mount *mp = bp->b_mount;
2475 int idx;
2476
2477 idx = xfs_has_crc(mp);
2478 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic[idx]))
2479 return false;
2480 return dmagic == bp->b_ops->magic[idx];
2481}
2482/*
2483 * Verify an on-disk magic value against the magic value specified in the
2484 * verifier structure. The verifier magic is in disk byte order so the caller is
2485 * expected to pass the value directly from disk.
2486 */
2487bool
2488xfs_verify_magic16(
2489 struct xfs_buf *bp,
2490 __be16 dmagic)
2491{
2492 struct xfs_mount *mp = bp->b_mount;
2493 int idx;
2494
2495 idx = xfs_has_crc(mp);
2496 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic16[idx]))
2497 return false;
2498 return dmagic == bp->b_ops->magic16[idx];
2499}