1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
7 #include <linux/backing-dev.h>
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"
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"
25 struct kmem_cache *xfs_buf_cache;
32 * b_sema (caller holds)
36 * b_sema (caller holds)
45 * xfs_buftarg_drain_rele
47 * b_lock (trylock due to inversion)
51 * b_lock (trylock due to inversion)
54 static int __xfs_buf_submit(struct xfs_buf *bp, bool wait);
60 return __xfs_buf_submit(bp, !(bp->b_flags & XBF_ASYNC));
68 * Return true if the buffer is vmapped.
70 * b_addr is null if the buffer is not mapped, but the code is clever
71 * enough to know it doesn't have to map a single page, so the check has
72 * to be both for b_addr and bp->b_page_count > 1.
74 return bp->b_addr && bp->b_page_count > 1;
81 return (bp->b_page_count * PAGE_SIZE);
85 * Bump the I/O in flight count on the buftarg if we haven't yet done so for
86 * this buffer. The count is incremented once per buffer (per hold cycle)
87 * because the corresponding decrement is deferred to buffer release. Buffers
88 * can undergo I/O multiple times in a hold-release cycle and per buffer I/O
89 * tracking adds unnecessary overhead. This is used for sychronization purposes
90 * with unmount (see xfs_buftarg_drain()), so all we really need is a count of
93 * Buffers that are never released (e.g., superblock, iclog buffers) must set
94 * the XBF_NO_IOACCT flag before I/O submission. Otherwise, the buftarg count
95 * never reaches zero and unmount hangs indefinitely.
101 if (bp->b_flags & XBF_NO_IOACCT)
104 ASSERT(bp->b_flags & XBF_ASYNC);
105 spin_lock(&bp->b_lock);
106 if (!(bp->b_state & XFS_BSTATE_IN_FLIGHT)) {
107 bp->b_state |= XFS_BSTATE_IN_FLIGHT;
108 percpu_counter_inc(&bp->b_target->bt_io_count);
110 spin_unlock(&bp->b_lock);
114 * Clear the in-flight state on a buffer about to be released to the LRU or
115 * freed and unaccount from the buftarg.
118 __xfs_buf_ioacct_dec(
121 lockdep_assert_held(&bp->b_lock);
123 if (bp->b_state & XFS_BSTATE_IN_FLIGHT) {
124 bp->b_state &= ~XFS_BSTATE_IN_FLIGHT;
125 percpu_counter_dec(&bp->b_target->bt_io_count);
133 spin_lock(&bp->b_lock);
134 __xfs_buf_ioacct_dec(bp);
135 spin_unlock(&bp->b_lock);
139 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
140 * b_lru_ref count so that the buffer is freed immediately when the buffer
141 * reference count falls to zero. If the buffer is already on the LRU, we need
142 * to remove the reference that LRU holds on the buffer.
144 * This prevents build-up of stale buffers on the LRU.
150 ASSERT(xfs_buf_islocked(bp));
152 bp->b_flags |= XBF_STALE;
155 * Clear the delwri status so that a delwri queue walker will not
156 * flush this buffer to disk now that it is stale. The delwri queue has
157 * a reference to the buffer, so this is safe to do.
159 bp->b_flags &= ~_XBF_DELWRI_Q;
162 * Once the buffer is marked stale and unlocked, a subsequent lookup
163 * could reset b_flags. There is no guarantee that the buffer is
164 * unaccounted (released to LRU) before that occurs. Drop in-flight
165 * status now to preserve accounting consistency.
167 spin_lock(&bp->b_lock);
168 __xfs_buf_ioacct_dec(bp);
170 atomic_set(&bp->b_lru_ref, 0);
171 if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
172 (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru)))
173 atomic_dec(&bp->b_hold);
175 ASSERT(atomic_read(&bp->b_hold) >= 1);
176 spin_unlock(&bp->b_lock);
184 ASSERT(bp->b_maps == NULL);
185 bp->b_map_count = map_count;
187 if (map_count == 1) {
188 bp->b_maps = &bp->__b_map;
192 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
200 * Frees b_pages if it was allocated.
206 if (bp->b_maps != &bp->__b_map) {
207 kmem_free(bp->b_maps);
214 struct xfs_buftarg *target,
215 struct xfs_buf_map *map,
217 xfs_buf_flags_t flags,
218 struct xfs_buf **bpp)
225 bp = kmem_cache_zalloc(xfs_buf_cache, GFP_NOFS | __GFP_NOFAIL);
228 * We don't want certain flags to appear in b_flags unless they are
229 * specifically set by later operations on the buffer.
231 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
233 atomic_set(&bp->b_hold, 1);
234 atomic_set(&bp->b_lru_ref, 1);
235 init_completion(&bp->b_iowait);
236 INIT_LIST_HEAD(&bp->b_lru);
237 INIT_LIST_HEAD(&bp->b_list);
238 INIT_LIST_HEAD(&bp->b_li_list);
239 sema_init(&bp->b_sema, 0); /* held, no waiters */
240 spin_lock_init(&bp->b_lock);
241 bp->b_target = target;
242 bp->b_mount = target->bt_mount;
246 * Set length and io_length to the same value initially.
247 * I/O routines should use io_length, which will be the same in
248 * most cases but may be reset (e.g. XFS recovery).
250 error = xfs_buf_get_maps(bp, nmaps);
252 kmem_cache_free(xfs_buf_cache, bp);
256 bp->b_rhash_key = map[0].bm_bn;
258 for (i = 0; i < nmaps; i++) {
259 bp->b_maps[i].bm_bn = map[i].bm_bn;
260 bp->b_maps[i].bm_len = map[i].bm_len;
261 bp->b_length += map[i].bm_len;
264 atomic_set(&bp->b_pin_count, 0);
265 init_waitqueue_head(&bp->b_waiters);
267 XFS_STATS_INC(bp->b_mount, xb_create);
268 trace_xfs_buf_init(bp, _RET_IP_);
280 ASSERT(bp->b_flags & _XBF_PAGES);
282 if (xfs_buf_is_vmapped(bp))
283 vm_unmap_ram(bp->b_addr, bp->b_page_count);
285 for (i = 0; i < bp->b_page_count; i++) {
287 __free_page(bp->b_pages[i]);
289 mm_account_reclaimed_pages(bp->b_page_count);
291 if (bp->b_pages != bp->b_page_array)
292 kmem_free(bp->b_pages);
294 bp->b_flags &= ~_XBF_PAGES;
298 xfs_buf_free_callback(
299 struct callback_head *cb)
301 struct xfs_buf *bp = container_of(cb, struct xfs_buf, b_rcu);
303 xfs_buf_free_maps(bp);
304 kmem_cache_free(xfs_buf_cache, bp);
311 trace_xfs_buf_free(bp, _RET_IP_);
313 ASSERT(list_empty(&bp->b_lru));
315 if (bp->b_flags & _XBF_PAGES)
316 xfs_buf_free_pages(bp);
317 else if (bp->b_flags & _XBF_KMEM)
318 kmem_free(bp->b_addr);
320 call_rcu(&bp->b_rcu, xfs_buf_free_callback);
326 xfs_buf_flags_t flags)
328 xfs_km_flags_t kmflag_mask = KM_NOFS;
329 size_t size = BBTOB(bp->b_length);
331 /* Assure zeroed buffer for non-read cases. */
332 if (!(flags & XBF_READ))
333 kmflag_mask |= KM_ZERO;
335 bp->b_addr = kmem_alloc(size, kmflag_mask);
339 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
340 ((unsigned long)bp->b_addr & PAGE_MASK)) {
341 /* b_addr spans two pages - use alloc_page instead */
342 kmem_free(bp->b_addr);
346 bp->b_offset = offset_in_page(bp->b_addr);
347 bp->b_pages = bp->b_page_array;
348 bp->b_pages[0] = kmem_to_page(bp->b_addr);
349 bp->b_page_count = 1;
350 bp->b_flags |= _XBF_KMEM;
357 xfs_buf_flags_t flags)
359 gfp_t gfp_mask = __GFP_NOWARN;
362 if (flags & XBF_READ_AHEAD)
363 gfp_mask |= __GFP_NORETRY;
365 gfp_mask |= GFP_NOFS;
367 /* Make sure that we have a page list */
368 bp->b_page_count = DIV_ROUND_UP(BBTOB(bp->b_length), PAGE_SIZE);
369 if (bp->b_page_count <= XB_PAGES) {
370 bp->b_pages = bp->b_page_array;
372 bp->b_pages = kzalloc(sizeof(struct page *) * bp->b_page_count,
377 bp->b_flags |= _XBF_PAGES;
379 /* Assure zeroed buffer for non-read cases. */
380 if (!(flags & XBF_READ))
381 gfp_mask |= __GFP_ZERO;
384 * Bulk filling of pages can take multiple calls. Not filling the entire
385 * array is not an allocation failure, so don't back off if we get at
386 * least one extra page.
391 filled = alloc_pages_bulk_array(gfp_mask, bp->b_page_count,
393 if (filled == bp->b_page_count) {
394 XFS_STATS_INC(bp->b_mount, xb_page_found);
401 if (flags & XBF_READ_AHEAD) {
402 xfs_buf_free_pages(bp);
406 XFS_STATS_INC(bp->b_mount, xb_page_retries);
407 memalloc_retry_wait(gfp_mask);
413 * Map buffer into kernel address-space if necessary.
418 xfs_buf_flags_t flags)
420 ASSERT(bp->b_flags & _XBF_PAGES);
421 if (bp->b_page_count == 1) {
422 /* A single page buffer is always mappable */
423 bp->b_addr = page_address(bp->b_pages[0]);
424 } else if (flags & XBF_UNMAPPED) {
431 * vm_map_ram() will allocate auxiliary structures (e.g.
432 * pagetables) with GFP_KERNEL, yet we are likely to be under
433 * GFP_NOFS context here. Hence we need to tell memory reclaim
434 * that we are in such a context via PF_MEMALLOC_NOFS to prevent
435 * memory reclaim re-entering the filesystem here and
436 * potentially deadlocking.
438 nofs_flag = memalloc_nofs_save();
440 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
445 } while (retried++ <= 1);
446 memalloc_nofs_restore(nofs_flag);
456 * Finding and Reading Buffers
460 struct rhashtable_compare_arg *arg,
463 const struct xfs_buf_map *map = arg->key;
464 const struct xfs_buf *bp = obj;
467 * The key hashing in the lookup path depends on the key being the
468 * first element of the compare_arg, make sure to assert this.
470 BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0);
472 if (bp->b_rhash_key != map->bm_bn)
475 if (unlikely(bp->b_length != map->bm_len)) {
477 * found a block number match. If the range doesn't
478 * match, the only way this is allowed is if the buffer
479 * in the cache is stale and the transaction that made
480 * it stale has not yet committed. i.e. we are
481 * reallocating a busy extent. Skip this buffer and
482 * continue searching for an exact match.
484 ASSERT(bp->b_flags & XBF_STALE);
490 static const struct rhashtable_params xfs_buf_hash_params = {
491 .min_size = 32, /* empty AGs have minimal footprint */
493 .key_len = sizeof(xfs_daddr_t),
494 .key_offset = offsetof(struct xfs_buf, b_rhash_key),
495 .head_offset = offsetof(struct xfs_buf, b_rhash_head),
496 .automatic_shrinking = true,
497 .obj_cmpfn = _xfs_buf_obj_cmp,
502 struct xfs_perag *pag)
504 spin_lock_init(&pag->pag_buf_lock);
505 return rhashtable_init(&pag->pag_buf_hash, &xfs_buf_hash_params);
509 xfs_buf_hash_destroy(
510 struct xfs_perag *pag)
512 rhashtable_destroy(&pag->pag_buf_hash);
517 struct xfs_buftarg *btp,
518 struct xfs_buf_map *map)
522 /* Check for IOs smaller than the sector size / not sector aligned */
523 ASSERT(!(BBTOB(map->bm_len) < btp->bt_meta_sectorsize));
524 ASSERT(!(BBTOB(map->bm_bn) & (xfs_off_t)btp->bt_meta_sectormask));
527 * Corrupted block numbers can get through to here, unfortunately, so we
528 * have to check that the buffer falls within the filesystem bounds.
530 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
531 if (map->bm_bn < 0 || map->bm_bn >= eofs) {
532 xfs_alert(btp->bt_mount,
533 "%s: daddr 0x%llx out of range, EOFS 0x%llx",
534 __func__, map->bm_bn, eofs);
536 return -EFSCORRUPTED;
544 xfs_buf_flags_t flags)
546 if (flags & XBF_TRYLOCK) {
547 if (!xfs_buf_trylock(bp)) {
548 XFS_STATS_INC(bp->b_mount, xb_busy_locked);
553 XFS_STATS_INC(bp->b_mount, xb_get_locked_waited);
557 * if the buffer is stale, clear all the external state associated with
558 * it. We need to keep flags such as how we allocated the buffer memory
561 if (bp->b_flags & XBF_STALE) {
562 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
563 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
571 struct xfs_perag *pag,
572 struct xfs_buf_map *map,
573 xfs_buf_flags_t flags,
574 struct xfs_buf **bpp)
580 bp = rhashtable_lookup(&pag->pag_buf_hash, map, xfs_buf_hash_params);
581 if (!bp || !atomic_inc_not_zero(&bp->b_hold)) {
587 error = xfs_buf_find_lock(bp, flags);
593 trace_xfs_buf_find(bp, flags, _RET_IP_);
599 * Insert the new_bp into the hash table. This consumes the perag reference
600 * taken for the lookup regardless of the result of the insert.
604 struct xfs_buftarg *btp,
605 struct xfs_perag *pag,
606 struct xfs_buf_map *cmap,
607 struct xfs_buf_map *map,
609 xfs_buf_flags_t flags,
610 struct xfs_buf **bpp)
612 struct xfs_buf *new_bp;
616 error = _xfs_buf_alloc(btp, map, nmaps, flags, &new_bp);
621 * For buffers that fit entirely within a single page, first attempt to
622 * allocate the memory from the heap to minimise memory usage. If we
623 * can't get heap memory for these small buffers, we fall back to using
624 * the page allocator.
626 if (BBTOB(new_bp->b_length) >= PAGE_SIZE ||
627 xfs_buf_alloc_kmem(new_bp, flags) < 0) {
628 error = xfs_buf_alloc_pages(new_bp, flags);
633 spin_lock(&pag->pag_buf_lock);
634 bp = rhashtable_lookup_get_insert_fast(&pag->pag_buf_hash,
635 &new_bp->b_rhash_head, xfs_buf_hash_params);
638 spin_unlock(&pag->pag_buf_lock);
642 /* found an existing buffer */
643 atomic_inc(&bp->b_hold);
644 spin_unlock(&pag->pag_buf_lock);
645 error = xfs_buf_find_lock(bp, flags);
653 /* The new buffer keeps the perag reference until it is freed. */
655 spin_unlock(&pag->pag_buf_lock);
660 xfs_buf_free(new_bp);
667 * Assembles a buffer covering the specified range. The code is optimised for
668 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
669 * more hits than misses.
673 struct xfs_buftarg *btp,
674 struct xfs_buf_map *map,
676 xfs_buf_flags_t flags,
677 struct xfs_buf **bpp)
679 struct xfs_perag *pag;
680 struct xfs_buf *bp = NULL;
681 struct xfs_buf_map cmap = { .bm_bn = map[0].bm_bn };
685 for (i = 0; i < nmaps; i++)
686 cmap.bm_len += map[i].bm_len;
688 error = xfs_buf_map_verify(btp, &cmap);
692 pag = xfs_perag_get(btp->bt_mount,
693 xfs_daddr_to_agno(btp->bt_mount, cmap.bm_bn));
695 error = xfs_buf_lookup(pag, &cmap, flags, &bp);
696 if (error && error != -ENOENT)
699 /* cache hits always outnumber misses by at least 10:1 */
701 XFS_STATS_INC(btp->bt_mount, xb_miss_locked);
703 if (flags & XBF_INCORE)
706 /* xfs_buf_find_insert() consumes the perag reference. */
707 error = xfs_buf_find_insert(btp, pag, &cmap, map, nmaps,
712 XFS_STATS_INC(btp->bt_mount, xb_get_locked);
716 /* We do not hold a perag reference anymore. */
718 error = _xfs_buf_map_pages(bp, flags);
719 if (unlikely(error)) {
720 xfs_warn_ratelimited(btp->bt_mount,
721 "%s: failed to map %u pages", __func__,
729 * Clear b_error if this is a lookup from a caller that doesn't expect
730 * valid data to be found in the buffer.
732 if (!(flags & XBF_READ))
733 xfs_buf_ioerror(bp, 0);
735 XFS_STATS_INC(btp->bt_mount, xb_get);
736 trace_xfs_buf_get(bp, flags, _RET_IP_);
748 xfs_buf_flags_t flags)
750 ASSERT(!(flags & XBF_WRITE));
751 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
753 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD | XBF_DONE);
754 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
756 return xfs_buf_submit(bp);
760 * Reverify a buffer found in cache without an attached ->b_ops.
762 * If the caller passed an ops structure and the buffer doesn't have ops
763 * assigned, set the ops and use it to verify the contents. If verification
764 * fails, clear XBF_DONE. We assume the buffer has no recorded errors and is
765 * already in XBF_DONE state on entry.
767 * Under normal operations, every in-core buffer is verified on read I/O
768 * completion. There are two scenarios that can lead to in-core buffers without
769 * an assigned ->b_ops. The first is during log recovery of buffers on a V4
770 * filesystem, though these buffers are purged at the end of recovery. The
771 * other is online repair, which intentionally reads with a NULL buffer ops to
772 * run several verifiers across an in-core buffer in order to establish buffer
773 * type. If repair can't establish that, the buffer will be left in memory
774 * with NULL buffer ops.
779 const struct xfs_buf_ops *ops)
781 ASSERT(bp->b_flags & XBF_DONE);
782 ASSERT(bp->b_error == 0);
784 if (!ops || bp->b_ops)
788 bp->b_ops->verify_read(bp);
790 bp->b_flags &= ~XBF_DONE;
796 struct xfs_buftarg *target,
797 struct xfs_buf_map *map,
799 xfs_buf_flags_t flags,
800 struct xfs_buf **bpp,
801 const struct xfs_buf_ops *ops,
810 error = xfs_buf_get_map(target, map, nmaps, flags, &bp);
814 trace_xfs_buf_read(bp, flags, _RET_IP_);
816 if (!(bp->b_flags & XBF_DONE)) {
817 /* Initiate the buffer read and wait. */
818 XFS_STATS_INC(target->bt_mount, xb_get_read);
820 error = _xfs_buf_read(bp, flags);
822 /* Readahead iodone already dropped the buffer, so exit. */
823 if (flags & XBF_ASYNC)
826 /* Buffer already read; all we need to do is check it. */
827 error = xfs_buf_reverify(bp, ops);
829 /* Readahead already finished; drop the buffer and exit. */
830 if (flags & XBF_ASYNC) {
835 /* We do not want read in the flags */
836 bp->b_flags &= ~XBF_READ;
837 ASSERT(bp->b_ops != NULL || ops == NULL);
841 * If we've had a read error, then the contents of the buffer are
842 * invalid and should not be used. To ensure that a followup read tries
843 * to pull the buffer from disk again, we clear the XBF_DONE flag and
844 * mark the buffer stale. This ensures that anyone who has a current
845 * reference to the buffer will interpret it's contents correctly and
846 * future cache lookups will also treat it as an empty, uninitialised
851 * Check against log shutdown for error reporting because
852 * metadata writeback may require a read first and we need to
853 * report errors in metadata writeback until the log is shut
854 * down. High level transaction read functions already check
855 * against mount shutdown, anyway, so we only need to be
856 * concerned about low level IO interactions here.
858 if (!xlog_is_shutdown(target->bt_mount->m_log))
859 xfs_buf_ioerror_alert(bp, fa);
861 bp->b_flags &= ~XBF_DONE;
865 /* bad CRC means corrupted metadata */
866 if (error == -EFSBADCRC)
867 error = -EFSCORRUPTED;
876 * If we are not low on memory then do the readahead in a deadlock
880 xfs_buf_readahead_map(
881 struct xfs_buftarg *target,
882 struct xfs_buf_map *map,
884 const struct xfs_buf_ops *ops)
888 xfs_buf_read_map(target, map, nmaps,
889 XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD, &bp, ops,
894 * Read an uncached buffer from disk. Allocates and returns a locked
895 * buffer containing the disk contents or nothing. Uncached buffers always have
896 * a cache index of XFS_BUF_DADDR_NULL so we can easily determine if the buffer
897 * is cached or uncached during fault diagnosis.
900 xfs_buf_read_uncached(
901 struct xfs_buftarg *target,
904 xfs_buf_flags_t flags,
905 struct xfs_buf **bpp,
906 const struct xfs_buf_ops *ops)
913 error = xfs_buf_get_uncached(target, numblks, flags, &bp);
917 /* set up the buffer for a read IO */
918 ASSERT(bp->b_map_count == 1);
919 bp->b_rhash_key = XFS_BUF_DADDR_NULL;
920 bp->b_maps[0].bm_bn = daddr;
921 bp->b_flags |= XBF_READ;
936 xfs_buf_get_uncached(
937 struct xfs_buftarg *target,
939 xfs_buf_flags_t flags,
940 struct xfs_buf **bpp)
944 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
948 /* flags might contain irrelevant bits, pass only what we care about */
949 error = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT, &bp);
953 error = xfs_buf_alloc_pages(bp, flags);
957 error = _xfs_buf_map_pages(bp, 0);
958 if (unlikely(error)) {
959 xfs_warn(target->bt_mount,
960 "%s: failed to map pages", __func__);
964 trace_xfs_buf_get_uncached(bp, _RET_IP_);
974 * Increment reference count on buffer, to hold the buffer concurrently
975 * with another thread which may release (free) the buffer asynchronously.
976 * Must hold the buffer already to call this function.
982 trace_xfs_buf_hold(bp, _RET_IP_);
983 atomic_inc(&bp->b_hold);
987 * Release a hold on the specified buffer. If the hold count is 1, the buffer is
988 * placed on LRU or freed (depending on b_lru_ref).
994 struct xfs_perag *pag = bp->b_pag;
996 bool freebuf = false;
998 trace_xfs_buf_rele(bp, _RET_IP_);
1001 ASSERT(list_empty(&bp->b_lru));
1002 if (atomic_dec_and_test(&bp->b_hold)) {
1003 xfs_buf_ioacct_dec(bp);
1009 ASSERT(atomic_read(&bp->b_hold) > 0);
1012 * We grab the b_lock here first to serialise racing xfs_buf_rele()
1013 * calls. The pag_buf_lock being taken on the last reference only
1014 * serialises against racing lookups in xfs_buf_find(). IOWs, the second
1015 * to last reference we drop here is not serialised against the last
1016 * reference until we take bp->b_lock. Hence if we don't grab b_lock
1017 * first, the last "release" reference can win the race to the lock and
1018 * free the buffer before the second-to-last reference is processed,
1019 * leading to a use-after-free scenario.
1021 spin_lock(&bp->b_lock);
1022 release = atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock);
1025 * Drop the in-flight state if the buffer is already on the LRU
1026 * and it holds the only reference. This is racy because we
1027 * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT
1028 * ensures the decrement occurs only once per-buf.
1030 if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru))
1031 __xfs_buf_ioacct_dec(bp);
1035 /* the last reference has been dropped ... */
1036 __xfs_buf_ioacct_dec(bp);
1037 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
1039 * If the buffer is added to the LRU take a new reference to the
1040 * buffer for the LRU and clear the (now stale) dispose list
1043 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) {
1044 bp->b_state &= ~XFS_BSTATE_DISPOSE;
1045 atomic_inc(&bp->b_hold);
1047 spin_unlock(&pag->pag_buf_lock);
1050 * most of the time buffers will already be removed from the
1051 * LRU, so optimise that case by checking for the
1052 * XFS_BSTATE_DISPOSE flag indicating the last list the buffer
1053 * was on was the disposal list
1055 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
1056 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru);
1058 ASSERT(list_empty(&bp->b_lru));
1061 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1062 rhashtable_remove_fast(&pag->pag_buf_hash, &bp->b_rhash_head,
1063 xfs_buf_hash_params);
1064 spin_unlock(&pag->pag_buf_lock);
1070 spin_unlock(&bp->b_lock);
1078 * Lock a buffer object, if it is not already locked.
1080 * If we come across a stale, pinned, locked buffer, we know that we are
1081 * being asked to lock a buffer that has been reallocated. Because it is
1082 * pinned, we know that the log has not been pushed to disk and hence it
1083 * will still be locked. Rather than continuing to have trylock attempts
1084 * fail until someone else pushes the log, push it ourselves before
1085 * returning. This means that the xfsaild will not get stuck trying
1086 * to push on stale inode buffers.
1094 locked = down_trylock(&bp->b_sema) == 0;
1096 trace_xfs_buf_trylock(bp, _RET_IP_);
1098 trace_xfs_buf_trylock_fail(bp, _RET_IP_);
1103 * Lock a buffer object.
1105 * If we come across a stale, pinned, locked buffer, we know that we
1106 * are being asked to lock a buffer that has been reallocated. Because
1107 * it is pinned, we know that the log has not been pushed to disk and
1108 * hence it will still be locked. Rather than sleeping until someone
1109 * else pushes the log, push it ourselves before trying to get the lock.
1115 trace_xfs_buf_lock(bp, _RET_IP_);
1117 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
1118 xfs_log_force(bp->b_mount, 0);
1121 trace_xfs_buf_lock_done(bp, _RET_IP_);
1128 ASSERT(xfs_buf_islocked(bp));
1131 trace_xfs_buf_unlock(bp, _RET_IP_);
1138 DECLARE_WAITQUEUE (wait, current);
1140 if (atomic_read(&bp->b_pin_count) == 0)
1143 add_wait_queue(&bp->b_waiters, &wait);
1145 set_current_state(TASK_UNINTERRUPTIBLE);
1146 if (atomic_read(&bp->b_pin_count) == 0)
1150 remove_wait_queue(&bp->b_waiters, &wait);
1151 set_current_state(TASK_RUNNING);
1155 xfs_buf_ioerror_alert_ratelimited(
1158 static unsigned long lasttime;
1159 static struct xfs_buftarg *lasttarg;
1161 if (bp->b_target != lasttarg ||
1162 time_after(jiffies, (lasttime + 5*HZ))) {
1164 xfs_buf_ioerror_alert(bp, __this_address);
1166 lasttarg = bp->b_target;
1170 * Account for this latest trip around the retry handler, and decide if
1171 * we've failed enough times to constitute a permanent failure.
1174 xfs_buf_ioerror_permanent(
1176 struct xfs_error_cfg *cfg)
1178 struct xfs_mount *mp = bp->b_mount;
1180 if (cfg->max_retries != XFS_ERR_RETRY_FOREVER &&
1181 ++bp->b_retries > cfg->max_retries)
1183 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1184 time_after(jiffies, cfg->retry_timeout + bp->b_first_retry_time))
1187 /* At unmount we may treat errors differently */
1188 if (xfs_is_unmounting(mp) && mp->m_fail_unmount)
1195 * On a sync write or shutdown we just want to stale the buffer and let the
1196 * caller handle the error in bp->b_error appropriately.
1198 * If the write was asynchronous then no one will be looking for the error. If
1199 * this is the first failure of this type, clear the error state and write the
1200 * buffer out again. This means we always retry an async write failure at least
1201 * once, but we also need to set the buffer up to behave correctly now for
1202 * repeated failures.
1204 * If we get repeated async write failures, then we take action according to the
1205 * error configuration we have been set up to use.
1207 * Returns true if this function took care of error handling and the caller must
1208 * not touch the buffer again. Return false if the caller should proceed with
1209 * normal I/O completion handling.
1212 xfs_buf_ioend_handle_error(
1215 struct xfs_mount *mp = bp->b_mount;
1216 struct xfs_error_cfg *cfg;
1219 * If we've already shutdown the journal because of I/O errors, there's
1220 * no point in giving this a retry.
1222 if (xlog_is_shutdown(mp->m_log))
1225 xfs_buf_ioerror_alert_ratelimited(bp);
1228 * We're not going to bother about retrying this during recovery.
1231 if (bp->b_flags & _XBF_LOGRECOVERY) {
1232 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1237 * Synchronous writes will have callers process the error.
1239 if (!(bp->b_flags & XBF_ASYNC))
1242 trace_xfs_buf_iodone_async(bp, _RET_IP_);
1244 cfg = xfs_error_get_cfg(mp, XFS_ERR_METADATA, bp->b_error);
1245 if (bp->b_last_error != bp->b_error ||
1246 !(bp->b_flags & (XBF_STALE | XBF_WRITE_FAIL))) {
1247 bp->b_last_error = bp->b_error;
1248 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1249 !bp->b_first_retry_time)
1250 bp->b_first_retry_time = jiffies;
1255 * Permanent error - we need to trigger a shutdown if we haven't already
1256 * to indicate that inconsistency will result from this action.
1258 if (xfs_buf_ioerror_permanent(bp, cfg)) {
1259 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1263 /* Still considered a transient error. Caller will schedule retries. */
1264 if (bp->b_flags & _XBF_INODES)
1265 xfs_buf_inode_io_fail(bp);
1266 else if (bp->b_flags & _XBF_DQUOTS)
1267 xfs_buf_dquot_io_fail(bp);
1269 ASSERT(list_empty(&bp->b_li_list));
1270 xfs_buf_ioerror(bp, 0);
1275 xfs_buf_ioerror(bp, 0);
1276 bp->b_flags |= (XBF_DONE | XBF_WRITE_FAIL);
1281 bp->b_flags |= XBF_DONE;
1282 bp->b_flags &= ~XBF_WRITE;
1283 trace_xfs_buf_error_relse(bp, _RET_IP_);
1291 trace_xfs_buf_iodone(bp, _RET_IP_);
1294 * Pull in IO completion errors now. We are guaranteed to be running
1295 * single threaded, so we don't need the lock to read b_io_error.
1297 if (!bp->b_error && bp->b_io_error)
1298 xfs_buf_ioerror(bp, bp->b_io_error);
1300 if (bp->b_flags & XBF_READ) {
1301 if (!bp->b_error && bp->b_ops)
1302 bp->b_ops->verify_read(bp);
1304 bp->b_flags |= XBF_DONE;
1307 bp->b_flags &= ~XBF_WRITE_FAIL;
1308 bp->b_flags |= XBF_DONE;
1311 if (unlikely(bp->b_error) && xfs_buf_ioend_handle_error(bp))
1314 /* clear the retry state */
1315 bp->b_last_error = 0;
1317 bp->b_first_retry_time = 0;
1320 * Note that for things like remote attribute buffers, there may
1321 * not be a buffer log item here, so processing the buffer log
1322 * item must remain optional.
1325 xfs_buf_item_done(bp);
1327 if (bp->b_flags & _XBF_INODES)
1328 xfs_buf_inode_iodone(bp);
1329 else if (bp->b_flags & _XBF_DQUOTS)
1330 xfs_buf_dquot_iodone(bp);
1334 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD |
1337 if (bp->b_flags & XBF_ASYNC)
1340 complete(&bp->b_iowait);
1345 struct work_struct *work)
1347 struct xfs_buf *bp =
1348 container_of(work, struct xfs_buf, b_ioend_work);
1354 xfs_buf_ioend_async(
1357 INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
1358 queue_work(bp->b_mount->m_buf_workqueue, &bp->b_ioend_work);
1365 xfs_failaddr_t failaddr)
1367 ASSERT(error <= 0 && error >= -1000);
1368 bp->b_error = error;
1369 trace_xfs_buf_ioerror(bp, error, failaddr);
1373 xfs_buf_ioerror_alert(
1375 xfs_failaddr_t func)
1377 xfs_buf_alert_ratelimited(bp, "XFS: metadata IO error",
1378 "metadata I/O error in \"%pS\" at daddr 0x%llx len %d error %d",
1379 func, (uint64_t)xfs_buf_daddr(bp),
1380 bp->b_length, -bp->b_error);
1384 * To simulate an I/O failure, the buffer must be locked and held with at least
1385 * three references. The LRU reference is dropped by the stale call. The buf
1386 * item reference is dropped via ioend processing. The third reference is owned
1387 * by the caller and is dropped on I/O completion if the buffer is XBF_ASYNC.
1393 bp->b_flags &= ~XBF_DONE;
1395 xfs_buf_ioerror(bp, -EIO);
1405 ASSERT(xfs_buf_islocked(bp));
1407 bp->b_flags |= XBF_WRITE;
1408 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1411 error = xfs_buf_submit(bp);
1413 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
1421 struct xfs_buf *bp = (struct xfs_buf *)bio->bi_private;
1423 if (!bio->bi_status &&
1424 (bp->b_flags & XBF_WRITE) && (bp->b_flags & XBF_ASYNC) &&
1425 XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_IOERROR))
1426 bio->bi_status = BLK_STS_IOERR;
1429 * don't overwrite existing errors - otherwise we can lose errors on
1430 * buffers that require multiple bios to complete.
1432 if (bio->bi_status) {
1433 int error = blk_status_to_errno(bio->bi_status);
1435 cmpxchg(&bp->b_io_error, 0, error);
1438 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1439 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1441 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1442 xfs_buf_ioend_async(bp);
1447 xfs_buf_ioapply_map(
1455 unsigned int total_nr_pages = bp->b_page_count;
1458 sector_t sector = bp->b_maps[map].bm_bn;
1462 /* skip the pages in the buffer before the start offset */
1464 offset = *buf_offset;
1465 while (offset >= PAGE_SIZE) {
1467 offset -= PAGE_SIZE;
1471 * Limit the IO size to the length of the current vector, and update the
1472 * remaining IO count for the next time around.
1474 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1476 *buf_offset += size;
1479 atomic_inc(&bp->b_io_remaining);
1480 nr_pages = bio_max_segs(total_nr_pages);
1482 bio = bio_alloc(bp->b_target->bt_bdev, nr_pages, op, GFP_NOIO);
1483 bio->bi_iter.bi_sector = sector;
1484 bio->bi_end_io = xfs_buf_bio_end_io;
1485 bio->bi_private = bp;
1487 for (; size && nr_pages; nr_pages--, page_index++) {
1488 int rbytes, nbytes = PAGE_SIZE - offset;
1493 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1495 if (rbytes < nbytes)
1499 sector += BTOBB(nbytes);
1504 if (likely(bio->bi_iter.bi_size)) {
1505 if (xfs_buf_is_vmapped(bp)) {
1506 flush_kernel_vmap_range(bp->b_addr,
1507 xfs_buf_vmap_len(bp));
1514 * This is guaranteed not to be the last io reference count
1515 * because the caller (xfs_buf_submit) holds a count itself.
1517 atomic_dec(&bp->b_io_remaining);
1518 xfs_buf_ioerror(bp, -EIO);
1528 struct blk_plug plug;
1535 * Make sure we capture only current IO errors rather than stale errors
1536 * left over from previous use of the buffer (e.g. failed readahead).
1540 if (bp->b_flags & XBF_WRITE) {
1544 * Run the write verifier callback function if it exists. If
1545 * this function fails it will mark the buffer with an error and
1546 * the IO should not be dispatched.
1549 bp->b_ops->verify_write(bp);
1551 xfs_force_shutdown(bp->b_mount,
1552 SHUTDOWN_CORRUPT_INCORE);
1555 } else if (bp->b_rhash_key != XFS_BUF_DADDR_NULL) {
1556 struct xfs_mount *mp = bp->b_mount;
1559 * non-crc filesystems don't attach verifiers during
1560 * log recovery, so don't warn for such filesystems.
1562 if (xfs_has_crc(mp)) {
1564 "%s: no buf ops on daddr 0x%llx len %d",
1565 __func__, xfs_buf_daddr(bp),
1567 xfs_hex_dump(bp->b_addr,
1568 XFS_CORRUPTION_DUMP_LEN);
1574 if (bp->b_flags & XBF_READ_AHEAD)
1578 /* we only use the buffer cache for meta-data */
1582 * Walk all the vectors issuing IO on them. Set up the initial offset
1583 * into the buffer and the desired IO size before we start -
1584 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1587 offset = bp->b_offset;
1588 size = BBTOB(bp->b_length);
1589 blk_start_plug(&plug);
1590 for (i = 0; i < bp->b_map_count; i++) {
1591 xfs_buf_ioapply_map(bp, i, &offset, &size, op);
1595 break; /* all done */
1597 blk_finish_plug(&plug);
1601 * Wait for I/O completion of a sync buffer and return the I/O error code.
1607 ASSERT(!(bp->b_flags & XBF_ASYNC));
1609 trace_xfs_buf_iowait(bp, _RET_IP_);
1610 wait_for_completion(&bp->b_iowait);
1611 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1617 * Buffer I/O submission path, read or write. Asynchronous submission transfers
1618 * the buffer lock ownership and the current reference to the IO. It is not
1619 * safe to reference the buffer after a call to this function unless the caller
1620 * holds an additional reference itself.
1629 trace_xfs_buf_submit(bp, _RET_IP_);
1631 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1634 * On log shutdown we stale and complete the buffer immediately. We can
1635 * be called to read the superblock before the log has been set up, so
1636 * be careful checking the log state.
1638 * Checking the mount shutdown state here can result in the log tail
1639 * moving inappropriately on disk as the log may not yet be shut down.
1640 * i.e. failing this buffer on mount shutdown can remove it from the AIL
1641 * and move the tail of the log forwards without having written this
1642 * buffer to disk. This corrupts the log tail state in memory, and
1643 * because the log may not be shut down yet, it can then be propagated
1644 * to disk before the log is shutdown. Hence we check log shutdown
1645 * state here rather than mount state to avoid corrupting the log tail
1648 if (bp->b_mount->m_log &&
1649 xlog_is_shutdown(bp->b_mount->m_log)) {
1650 xfs_buf_ioend_fail(bp);
1655 * Grab a reference so the buffer does not go away underneath us. For
1656 * async buffers, I/O completion drops the callers reference, which
1657 * could occur before submission returns.
1661 if (bp->b_flags & XBF_WRITE)
1662 xfs_buf_wait_unpin(bp);
1664 /* clear the internal error state to avoid spurious errors */
1668 * Set the count to 1 initially, this will stop an I/O completion
1669 * callout which happens before we have started all the I/O from calling
1670 * xfs_buf_ioend too early.
1672 atomic_set(&bp->b_io_remaining, 1);
1673 if (bp->b_flags & XBF_ASYNC)
1674 xfs_buf_ioacct_inc(bp);
1675 _xfs_buf_ioapply(bp);
1678 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1679 * reference we took above. If we drop it to zero, run completion so
1680 * that we don't return to the caller with completion still pending.
1682 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1683 if (bp->b_error || !(bp->b_flags & XBF_ASYNC))
1686 xfs_buf_ioend_async(bp);
1690 error = xfs_buf_iowait(bp);
1693 * Release the hold that keeps the buffer referenced for the entire
1694 * I/O. Note that if the buffer is async, it is not safe to reference
1695 * after this release.
1709 return bp->b_addr + offset;
1711 page = bp->b_pages[offset >> PAGE_SHIFT];
1712 return page_address(page) + (offset & (PAGE_SIZE-1));
1723 bend = boff + bsize;
1724 while (boff < bend) {
1726 int page_index, page_offset, csize;
1728 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1729 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1730 page = bp->b_pages[page_index];
1731 csize = min_t(size_t, PAGE_SIZE - page_offset,
1732 BBTOB(bp->b_length) - boff);
1734 ASSERT((csize + page_offset) <= PAGE_SIZE);
1736 memset(page_address(page) + page_offset, 0, csize);
1743 * Log a message about and stale a buffer that a caller has decided is corrupt.
1745 * This function should be called for the kinds of metadata corruption that
1746 * cannot be detect from a verifier, such as incorrect inter-block relationship
1747 * data. Do /not/ call this function from a verifier function.
1749 * The buffer must be XBF_DONE prior to the call. Afterwards, the buffer will
1750 * be marked stale, but b_error will not be set. The caller is responsible for
1751 * releasing the buffer or fixing it.
1754 __xfs_buf_mark_corrupt(
1758 ASSERT(bp->b_flags & XBF_DONE);
1760 xfs_buf_corruption_error(bp, fa);
1765 * Handling of buffer targets (buftargs).
1769 * Wait for any bufs with callbacks that have been submitted but have not yet
1770 * returned. These buffers will have an elevated hold count, so wait on those
1771 * while freeing all the buffers only held by the LRU.
1773 static enum lru_status
1774 xfs_buftarg_drain_rele(
1775 struct list_head *item,
1776 struct list_lru_one *lru,
1777 spinlock_t *lru_lock,
1781 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1782 struct list_head *dispose = arg;
1784 if (atomic_read(&bp->b_hold) > 1) {
1785 /* need to wait, so skip it this pass */
1786 trace_xfs_buf_drain_buftarg(bp, _RET_IP_);
1789 if (!spin_trylock(&bp->b_lock))
1793 * clear the LRU reference count so the buffer doesn't get
1794 * ignored in xfs_buf_rele().
1796 atomic_set(&bp->b_lru_ref, 0);
1797 bp->b_state |= XFS_BSTATE_DISPOSE;
1798 list_lru_isolate_move(lru, item, dispose);
1799 spin_unlock(&bp->b_lock);
1804 * Wait for outstanding I/O on the buftarg to complete.
1808 struct xfs_buftarg *btp)
1811 * First wait on the buftarg I/O count for all in-flight buffers to be
1812 * released. This is critical as new buffers do not make the LRU until
1813 * they are released.
1815 * Next, flush the buffer workqueue to ensure all completion processing
1816 * has finished. Just waiting on buffer locks is not sufficient for
1817 * async IO as the reference count held over IO is not released until
1818 * after the buffer lock is dropped. Hence we need to ensure here that
1819 * all reference counts have been dropped before we start walking the
1822 while (percpu_counter_sum(&btp->bt_io_count))
1824 flush_workqueue(btp->bt_mount->m_buf_workqueue);
1829 struct xfs_buftarg *btp)
1833 bool write_fail = false;
1835 xfs_buftarg_wait(btp);
1837 /* loop until there is nothing left on the lru list. */
1838 while (list_lru_count(&btp->bt_lru)) {
1839 list_lru_walk(&btp->bt_lru, xfs_buftarg_drain_rele,
1840 &dispose, LONG_MAX);
1842 while (!list_empty(&dispose)) {
1844 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1845 list_del_init(&bp->b_lru);
1846 if (bp->b_flags & XBF_WRITE_FAIL) {
1848 xfs_buf_alert_ratelimited(bp,
1849 "XFS: Corruption Alert",
1850 "Corruption Alert: Buffer at daddr 0x%llx had permanent write failures!",
1851 (long long)xfs_buf_daddr(bp));
1860 * If one or more failed buffers were freed, that means dirty metadata
1861 * was thrown away. This should only ever happen after I/O completion
1862 * handling has elevated I/O error(s) to permanent failures and shuts
1866 ASSERT(xlog_is_shutdown(btp->bt_mount->m_log));
1867 xfs_alert(btp->bt_mount,
1868 "Please run xfs_repair to determine the extent of the problem.");
1872 static enum lru_status
1873 xfs_buftarg_isolate(
1874 struct list_head *item,
1875 struct list_lru_one *lru,
1876 spinlock_t *lru_lock,
1879 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1880 struct list_head *dispose = arg;
1883 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1884 * If we fail to get the lock, just skip it.
1886 if (!spin_trylock(&bp->b_lock))
1889 * Decrement the b_lru_ref count unless the value is already
1890 * zero. If the value is already zero, we need to reclaim the
1891 * buffer, otherwise it gets another trip through the LRU.
1893 if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1894 spin_unlock(&bp->b_lock);
1898 bp->b_state |= XFS_BSTATE_DISPOSE;
1899 list_lru_isolate_move(lru, item, dispose);
1900 spin_unlock(&bp->b_lock);
1904 static unsigned long
1905 xfs_buftarg_shrink_scan(
1906 struct shrinker *shrink,
1907 struct shrink_control *sc)
1909 struct xfs_buftarg *btp = container_of(shrink,
1910 struct xfs_buftarg, bt_shrinker);
1912 unsigned long freed;
1914 freed = list_lru_shrink_walk(&btp->bt_lru, sc,
1915 xfs_buftarg_isolate, &dispose);
1917 while (!list_empty(&dispose)) {
1919 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1920 list_del_init(&bp->b_lru);
1927 static unsigned long
1928 xfs_buftarg_shrink_count(
1929 struct shrinker *shrink,
1930 struct shrink_control *sc)
1932 struct xfs_buftarg *btp = container_of(shrink,
1933 struct xfs_buftarg, bt_shrinker);
1934 return list_lru_shrink_count(&btp->bt_lru, sc);
1939 struct xfs_buftarg *btp)
1941 struct block_device *bdev = btp->bt_bdev;
1943 unregister_shrinker(&btp->bt_shrinker);
1944 ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
1945 percpu_counter_destroy(&btp->bt_io_count);
1946 list_lru_destroy(&btp->bt_lru);
1948 fs_put_dax(btp->bt_daxdev, btp->bt_mount);
1949 /* the main block device is closed by kill_block_super */
1950 if (bdev != btp->bt_mount->m_super->s_bdev)
1951 blkdev_put(bdev, btp->bt_mount->m_super);
1957 xfs_setsize_buftarg(
1959 unsigned int sectorsize)
1961 /* Set up metadata sector size info */
1962 btp->bt_meta_sectorsize = sectorsize;
1963 btp->bt_meta_sectormask = sectorsize - 1;
1965 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1966 xfs_warn(btp->bt_mount,
1967 "Cannot set_blocksize to %u on device %pg",
1968 sectorsize, btp->bt_bdev);
1972 /* Set up device logical sector size mask */
1973 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1974 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1980 * When allocating the initial buffer target we have not yet
1981 * read in the superblock, so don't know what sized sectors
1982 * are being used at this early stage. Play safe.
1985 xfs_setsize_buftarg_early(
1987 struct block_device *bdev)
1989 return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev));
1992 struct xfs_buftarg *
1994 struct xfs_mount *mp,
1995 struct block_device *bdev)
1998 const struct dax_holder_operations *ops = NULL;
2000 #if defined(CONFIG_FS_DAX) && defined(CONFIG_MEMORY_FAILURE)
2001 ops = &xfs_dax_holder_operations;
2003 btp = kmem_zalloc(sizeof(*btp), KM_NOFS);
2006 btp->bt_dev = bdev->bd_dev;
2007 btp->bt_bdev = bdev;
2008 btp->bt_daxdev = fs_dax_get_by_bdev(bdev, &btp->bt_dax_part_off,
2012 * Buffer IO error rate limiting. Limit it to no more than 10 messages
2013 * per 30 seconds so as to not spam logs too much on repeated errors.
2015 ratelimit_state_init(&btp->bt_ioerror_rl, 30 * HZ,
2016 DEFAULT_RATELIMIT_BURST);
2018 if (xfs_setsize_buftarg_early(btp, bdev))
2021 if (list_lru_init(&btp->bt_lru))
2024 if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL))
2027 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
2028 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
2029 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
2030 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
2031 if (register_shrinker(&btp->bt_shrinker, "xfs-buf:%s",
2037 percpu_counter_destroy(&btp->bt_io_count);
2039 list_lru_destroy(&btp->bt_lru);
2046 * Cancel a delayed write list.
2048 * Remove each buffer from the list, clear the delwri queue flag and drop the
2049 * associated buffer reference.
2052 xfs_buf_delwri_cancel(
2053 struct list_head *list)
2057 while (!list_empty(list)) {
2058 bp = list_first_entry(list, struct xfs_buf, b_list);
2061 bp->b_flags &= ~_XBF_DELWRI_Q;
2062 list_del_init(&bp->b_list);
2068 * Add a buffer to the delayed write list.
2070 * This queues a buffer for writeout if it hasn't already been. Note that
2071 * neither this routine nor the buffer list submission functions perform
2072 * any internal synchronization. It is expected that the lists are thread-local
2075 * Returns true if we queued up the buffer, or false if it already had
2076 * been on the buffer list.
2079 xfs_buf_delwri_queue(
2081 struct list_head *list)
2083 ASSERT(xfs_buf_islocked(bp));
2084 ASSERT(!(bp->b_flags & XBF_READ));
2087 * If the buffer is already marked delwri it already is queued up
2088 * by someone else for imediate writeout. Just ignore it in that
2091 if (bp->b_flags & _XBF_DELWRI_Q) {
2092 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
2096 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
2099 * If a buffer gets written out synchronously or marked stale while it
2100 * is on a delwri list we lazily remove it. To do this, the other party
2101 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
2102 * It remains referenced and on the list. In a rare corner case it
2103 * might get readded to a delwri list after the synchronous writeout, in
2104 * which case we need just need to re-add the flag here.
2106 bp->b_flags |= _XBF_DELWRI_Q;
2107 if (list_empty(&bp->b_list)) {
2108 atomic_inc(&bp->b_hold);
2109 list_add_tail(&bp->b_list, list);
2116 * Compare function is more complex than it needs to be because
2117 * the return value is only 32 bits and we are doing comparisons
2123 const struct list_head *a,
2124 const struct list_head *b)
2126 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
2127 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
2130 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
2139 * Submit buffers for write. If wait_list is specified, the buffers are
2140 * submitted using sync I/O and placed on the wait list such that the caller can
2141 * iowait each buffer. Otherwise async I/O is used and the buffers are released
2142 * at I/O completion time. In either case, buffers remain locked until I/O
2143 * completes and the buffer is released from the queue.
2146 xfs_buf_delwri_submit_buffers(
2147 struct list_head *buffer_list,
2148 struct list_head *wait_list)
2150 struct xfs_buf *bp, *n;
2152 struct blk_plug plug;
2154 list_sort(NULL, buffer_list, xfs_buf_cmp);
2156 blk_start_plug(&plug);
2157 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
2159 if (!xfs_buf_trylock(bp))
2161 if (xfs_buf_ispinned(bp)) {
2171 * Someone else might have written the buffer synchronously or
2172 * marked it stale in the meantime. In that case only the
2173 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
2174 * reference and remove it from the list here.
2176 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
2177 list_del_init(&bp->b_list);
2182 trace_xfs_buf_delwri_split(bp, _RET_IP_);
2185 * If we have a wait list, each buffer (and associated delwri
2186 * queue reference) transfers to it and is submitted
2187 * synchronously. Otherwise, drop the buffer from the delwri
2188 * queue and submit async.
2190 bp->b_flags &= ~_XBF_DELWRI_Q;
2191 bp->b_flags |= XBF_WRITE;
2193 bp->b_flags &= ~XBF_ASYNC;
2194 list_move_tail(&bp->b_list, wait_list);
2196 bp->b_flags |= XBF_ASYNC;
2197 list_del_init(&bp->b_list);
2199 __xfs_buf_submit(bp, false);
2201 blk_finish_plug(&plug);
2207 * Write out a buffer list asynchronously.
2209 * This will take the @buffer_list, write all non-locked and non-pinned buffers
2210 * out and not wait for I/O completion on any of the buffers. This interface
2211 * is only safely useable for callers that can track I/O completion by higher
2212 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
2215 * Note: this function will skip buffers it would block on, and in doing so
2216 * leaves them on @buffer_list so they can be retried on a later pass. As such,
2217 * it is up to the caller to ensure that the buffer list is fully submitted or
2218 * cancelled appropriately when they are finished with the list. Failure to
2219 * cancel or resubmit the list until it is empty will result in leaked buffers
2223 xfs_buf_delwri_submit_nowait(
2224 struct list_head *buffer_list)
2226 return xfs_buf_delwri_submit_buffers(buffer_list, NULL);
2230 * Write out a buffer list synchronously.
2232 * This will take the @buffer_list, write all buffers out and wait for I/O
2233 * completion on all of the buffers. @buffer_list is consumed by the function,
2234 * so callers must have some other way of tracking buffers if they require such
2238 xfs_buf_delwri_submit(
2239 struct list_head *buffer_list)
2241 LIST_HEAD (wait_list);
2242 int error = 0, error2;
2245 xfs_buf_delwri_submit_buffers(buffer_list, &wait_list);
2247 /* Wait for IO to complete. */
2248 while (!list_empty(&wait_list)) {
2249 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
2251 list_del_init(&bp->b_list);
2254 * Wait on the locked buffer, check for errors and unlock and
2255 * release the delwri queue reference.
2257 error2 = xfs_buf_iowait(bp);
2267 * Push a single buffer on a delwri queue.
2269 * The purpose of this function is to submit a single buffer of a delwri queue
2270 * and return with the buffer still on the original queue. The waiting delwri
2271 * buffer submission infrastructure guarantees transfer of the delwri queue
2272 * buffer reference to a temporary wait list. We reuse this infrastructure to
2273 * transfer the buffer back to the original queue.
2275 * Note the buffer transitions from the queued state, to the submitted and wait
2276 * listed state and back to the queued state during this call. The buffer
2277 * locking and queue management logic between _delwri_pushbuf() and
2278 * _delwri_queue() guarantee that the buffer cannot be queued to another list
2282 xfs_buf_delwri_pushbuf(
2284 struct list_head *buffer_list)
2286 LIST_HEAD (submit_list);
2289 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
2291 trace_xfs_buf_delwri_pushbuf(bp, _RET_IP_);
2294 * Isolate the buffer to a new local list so we can submit it for I/O
2295 * independently from the rest of the original list.
2298 list_move(&bp->b_list, &submit_list);
2302 * Delwri submission clears the DELWRI_Q buffer flag and returns with
2303 * the buffer on the wait list with the original reference. Rather than
2304 * bounce the buffer from a local wait list back to the original list
2305 * after I/O completion, reuse the original list as the wait list.
2307 xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
2310 * The buffer is now locked, under I/O and wait listed on the original
2311 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
2312 * return with the buffer unlocked and on the original queue.
2314 error = xfs_buf_iowait(bp);
2315 bp->b_flags |= _XBF_DELWRI_Q;
2321 void xfs_buf_set_ref(struct xfs_buf *bp, int lru_ref)
2324 * Set the lru reference count to 0 based on the error injection tag.
2325 * This allows userspace to disrupt buffer caching for debug/testing
2328 if (XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_LRU_REF))
2331 atomic_set(&bp->b_lru_ref, lru_ref);
2335 * Verify an on-disk magic value against the magic value specified in the
2336 * verifier structure. The verifier magic is in disk byte order so the caller is
2337 * expected to pass the value directly from disk.
2344 struct xfs_mount *mp = bp->b_mount;
2347 idx = xfs_has_crc(mp);
2348 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic[idx]))
2350 return dmagic == bp->b_ops->magic[idx];
2353 * Verify an on-disk magic value against the magic value specified in the
2354 * verifier structure. The verifier magic is in disk byte order so the caller is
2355 * expected to pass the value directly from disk.
2362 struct xfs_mount *mp = bp->b_mount;
2365 idx = xfs_has_crc(mp);
2366 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic16[idx]))
2368 return dmagic == bp->b_ops->magic16[idx];