1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
7 #include <linux/backing-dev.h>
9 #include "xfs_shared.h"
10 #include "xfs_format.h"
11 #include "xfs_log_format.h"
12 #include "xfs_trans_resv.h"
13 #include "xfs_mount.h"
14 #include "xfs_trace.h"
16 #include "xfs_log_recover.h"
17 #include "xfs_log_priv.h"
18 #include "xfs_trans.h"
19 #include "xfs_buf_item.h"
20 #include "xfs_errortag.h"
21 #include "xfs_error.h"
24 static struct kmem_cache *xfs_buf_cache;
31 * b_sema (caller holds)
35 * b_sema (caller holds)
44 * xfs_buftarg_drain_rele
46 * b_lock (trylock due to inversion)
50 * b_lock (trylock due to inversion)
53 static int __xfs_buf_submit(struct xfs_buf *bp, bool wait);
59 return __xfs_buf_submit(bp, !(bp->b_flags & XBF_ASYNC));
67 * Return true if the buffer is vmapped.
69 * b_addr is null if the buffer is not mapped, but the code is clever
70 * enough to know it doesn't have to map a single page, so the check has
71 * to be both for b_addr and bp->b_page_count > 1.
73 return bp->b_addr && bp->b_page_count > 1;
80 return (bp->b_page_count * PAGE_SIZE);
84 * Bump the I/O in flight count on the buftarg if we haven't yet done so for
85 * this buffer. The count is incremented once per buffer (per hold cycle)
86 * because the corresponding decrement is deferred to buffer release. Buffers
87 * can undergo I/O multiple times in a hold-release cycle and per buffer I/O
88 * tracking adds unnecessary overhead. This is used for sychronization purposes
89 * with unmount (see xfs_buftarg_drain()), so all we really need is a count of
92 * Buffers that are never released (e.g., superblock, iclog buffers) must set
93 * the XBF_NO_IOACCT flag before I/O submission. Otherwise, the buftarg count
94 * never reaches zero and unmount hangs indefinitely.
100 if (bp->b_flags & XBF_NO_IOACCT)
103 ASSERT(bp->b_flags & XBF_ASYNC);
104 spin_lock(&bp->b_lock);
105 if (!(bp->b_state & XFS_BSTATE_IN_FLIGHT)) {
106 bp->b_state |= XFS_BSTATE_IN_FLIGHT;
107 percpu_counter_inc(&bp->b_target->bt_io_count);
109 spin_unlock(&bp->b_lock);
113 * Clear the in-flight state on a buffer about to be released to the LRU or
114 * freed and unaccount from the buftarg.
117 __xfs_buf_ioacct_dec(
120 lockdep_assert_held(&bp->b_lock);
122 if (bp->b_state & XFS_BSTATE_IN_FLIGHT) {
123 bp->b_state &= ~XFS_BSTATE_IN_FLIGHT;
124 percpu_counter_dec(&bp->b_target->bt_io_count);
132 spin_lock(&bp->b_lock);
133 __xfs_buf_ioacct_dec(bp);
134 spin_unlock(&bp->b_lock);
138 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
139 * b_lru_ref count so that the buffer is freed immediately when the buffer
140 * reference count falls to zero. If the buffer is already on the LRU, we need
141 * to remove the reference that LRU holds on the buffer.
143 * This prevents build-up of stale buffers on the LRU.
149 ASSERT(xfs_buf_islocked(bp));
151 bp->b_flags |= XBF_STALE;
154 * Clear the delwri status so that a delwri queue walker will not
155 * flush this buffer to disk now that it is stale. The delwri queue has
156 * a reference to the buffer, so this is safe to do.
158 bp->b_flags &= ~_XBF_DELWRI_Q;
161 * Once the buffer is marked stale and unlocked, a subsequent lookup
162 * could reset b_flags. There is no guarantee that the buffer is
163 * unaccounted (released to LRU) before that occurs. Drop in-flight
164 * status now to preserve accounting consistency.
166 spin_lock(&bp->b_lock);
167 __xfs_buf_ioacct_dec(bp);
169 atomic_set(&bp->b_lru_ref, 0);
170 if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
171 (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru)))
172 atomic_dec(&bp->b_hold);
174 ASSERT(atomic_read(&bp->b_hold) >= 1);
175 spin_unlock(&bp->b_lock);
183 ASSERT(bp->b_maps == NULL);
184 bp->b_map_count = map_count;
186 if (map_count == 1) {
187 bp->b_maps = &bp->__b_map;
191 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
199 * Frees b_pages if it was allocated.
205 if (bp->b_maps != &bp->__b_map) {
206 kmem_free(bp->b_maps);
213 struct xfs_buftarg *target,
214 struct xfs_buf_map *map,
216 xfs_buf_flags_t flags,
217 struct xfs_buf **bpp)
224 bp = kmem_cache_zalloc(xfs_buf_cache, GFP_NOFS | __GFP_NOFAIL);
227 * We don't want certain flags to appear in b_flags unless they are
228 * specifically set by later operations on the buffer.
230 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
232 atomic_set(&bp->b_hold, 1);
233 atomic_set(&bp->b_lru_ref, 1);
234 init_completion(&bp->b_iowait);
235 INIT_LIST_HEAD(&bp->b_lru);
236 INIT_LIST_HEAD(&bp->b_list);
237 INIT_LIST_HEAD(&bp->b_li_list);
238 sema_init(&bp->b_sema, 0); /* held, no waiters */
239 spin_lock_init(&bp->b_lock);
240 bp->b_target = target;
241 bp->b_mount = target->bt_mount;
245 * Set length and io_length to the same value initially.
246 * I/O routines should use io_length, which will be the same in
247 * most cases but may be reset (e.g. XFS recovery).
249 error = xfs_buf_get_maps(bp, nmaps);
251 kmem_cache_free(xfs_buf_cache, bp);
255 bp->b_rhash_key = map[0].bm_bn;
257 for (i = 0; i < nmaps; i++) {
258 bp->b_maps[i].bm_bn = map[i].bm_bn;
259 bp->b_maps[i].bm_len = map[i].bm_len;
260 bp->b_length += map[i].bm_len;
263 atomic_set(&bp->b_pin_count, 0);
264 init_waitqueue_head(&bp->b_waiters);
266 XFS_STATS_INC(bp->b_mount, xb_create);
267 trace_xfs_buf_init(bp, _RET_IP_);
279 ASSERT(bp->b_flags & _XBF_PAGES);
281 if (xfs_buf_is_vmapped(bp))
282 vm_unmap_ram(bp->b_addr, bp->b_page_count);
284 for (i = 0; i < bp->b_page_count; i++) {
286 __free_page(bp->b_pages[i]);
288 if (current->reclaim_state)
289 current->reclaim_state->reclaimed_slab += 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;
301 trace_xfs_buf_free(bp, _RET_IP_);
303 ASSERT(list_empty(&bp->b_lru));
305 if (bp->b_flags & _XBF_PAGES)
306 xfs_buf_free_pages(bp);
307 else if (bp->b_flags & _XBF_KMEM)
308 kmem_free(bp->b_addr);
310 xfs_buf_free_maps(bp);
311 kmem_cache_free(xfs_buf_cache, bp);
317 xfs_buf_flags_t flags)
319 xfs_km_flags_t kmflag_mask = KM_NOFS;
320 size_t size = BBTOB(bp->b_length);
322 /* Assure zeroed buffer for non-read cases. */
323 if (!(flags & XBF_READ))
324 kmflag_mask |= KM_ZERO;
326 bp->b_addr = kmem_alloc(size, kmflag_mask);
330 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
331 ((unsigned long)bp->b_addr & PAGE_MASK)) {
332 /* b_addr spans two pages - use alloc_page instead */
333 kmem_free(bp->b_addr);
337 bp->b_offset = offset_in_page(bp->b_addr);
338 bp->b_pages = bp->b_page_array;
339 bp->b_pages[0] = kmem_to_page(bp->b_addr);
340 bp->b_page_count = 1;
341 bp->b_flags |= _XBF_KMEM;
348 xfs_buf_flags_t flags)
350 gfp_t gfp_mask = __GFP_NOWARN;
353 if (flags & XBF_READ_AHEAD)
354 gfp_mask |= __GFP_NORETRY;
356 gfp_mask |= GFP_NOFS;
358 /* Make sure that we have a page list */
359 bp->b_page_count = DIV_ROUND_UP(BBTOB(bp->b_length), PAGE_SIZE);
360 if (bp->b_page_count <= XB_PAGES) {
361 bp->b_pages = bp->b_page_array;
363 bp->b_pages = kzalloc(sizeof(struct page *) * bp->b_page_count,
368 bp->b_flags |= _XBF_PAGES;
370 /* Assure zeroed buffer for non-read cases. */
371 if (!(flags & XBF_READ))
372 gfp_mask |= __GFP_ZERO;
375 * Bulk filling of pages can take multiple calls. Not filling the entire
376 * array is not an allocation failure, so don't back off if we get at
377 * least one extra page.
382 filled = alloc_pages_bulk_array(gfp_mask, bp->b_page_count,
384 if (filled == bp->b_page_count) {
385 XFS_STATS_INC(bp->b_mount, xb_page_found);
392 if (flags & XBF_READ_AHEAD) {
393 xfs_buf_free_pages(bp);
397 XFS_STATS_INC(bp->b_mount, xb_page_retries);
398 memalloc_retry_wait(gfp_mask);
404 * Map buffer into kernel address-space if necessary.
411 ASSERT(bp->b_flags & _XBF_PAGES);
412 if (bp->b_page_count == 1) {
413 /* A single page buffer is always mappable */
414 bp->b_addr = page_address(bp->b_pages[0]);
415 } else if (flags & XBF_UNMAPPED) {
422 * vm_map_ram() will allocate auxiliary structures (e.g.
423 * pagetables) with GFP_KERNEL, yet we are likely to be under
424 * GFP_NOFS context here. Hence we need to tell memory reclaim
425 * that we are in such a context via PF_MEMALLOC_NOFS to prevent
426 * memory reclaim re-entering the filesystem here and
427 * potentially deadlocking.
429 nofs_flag = memalloc_nofs_save();
431 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
436 } while (retried++ <= 1);
437 memalloc_nofs_restore(nofs_flag);
447 * Finding and Reading Buffers
451 struct rhashtable_compare_arg *arg,
454 const struct xfs_buf_map *map = arg->key;
455 const struct xfs_buf *bp = obj;
458 * The key hashing in the lookup path depends on the key being the
459 * first element of the compare_arg, make sure to assert this.
461 BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0);
463 if (bp->b_rhash_key != map->bm_bn)
466 if (unlikely(bp->b_length != map->bm_len)) {
468 * found a block number match. If the range doesn't
469 * match, the only way this is allowed is if the buffer
470 * in the cache is stale and the transaction that made
471 * it stale has not yet committed. i.e. we are
472 * reallocating a busy extent. Skip this buffer and
473 * continue searching for an exact match.
475 ASSERT(bp->b_flags & XBF_STALE);
481 static const struct rhashtable_params xfs_buf_hash_params = {
482 .min_size = 32, /* empty AGs have minimal footprint */
484 .key_len = sizeof(xfs_daddr_t),
485 .key_offset = offsetof(struct xfs_buf, b_rhash_key),
486 .head_offset = offsetof(struct xfs_buf, b_rhash_head),
487 .automatic_shrinking = true,
488 .obj_cmpfn = _xfs_buf_obj_cmp,
493 struct xfs_perag *pag)
495 spin_lock_init(&pag->pag_buf_lock);
496 return rhashtable_init(&pag->pag_buf_hash, &xfs_buf_hash_params);
500 xfs_buf_hash_destroy(
501 struct xfs_perag *pag)
503 rhashtable_destroy(&pag->pag_buf_hash);
507 * Look up a buffer in the buffer cache and return it referenced and locked
510 * If @new_bp is supplied and we have a lookup miss, insert @new_bp into the
513 * If XBF_TRYLOCK is set in @flags, only try to lock the buffer and return
514 * -EAGAIN if we fail to lock it.
517 * -EFSCORRUPTED if have been supplied with an invalid address
518 * -EAGAIN on trylock failure
519 * -ENOENT if we fail to find a match and @new_bp was NULL
521 * - @new_bp if we inserted it into the cache
522 * - the buffer we found and locked.
526 struct xfs_buftarg *btp,
527 struct xfs_buf_map *map,
529 xfs_buf_flags_t flags,
530 struct xfs_buf *new_bp,
531 struct xfs_buf **found_bp)
533 struct xfs_perag *pag;
535 struct xfs_buf_map cmap = { .bm_bn = map[0].bm_bn };
541 for (i = 0; i < nmaps; i++)
542 cmap.bm_len += map[i].bm_len;
544 /* Check for IOs smaller than the sector size / not sector aligned */
545 ASSERT(!(BBTOB(cmap.bm_len) < btp->bt_meta_sectorsize));
546 ASSERT(!(BBTOB(cmap.bm_bn) & (xfs_off_t)btp->bt_meta_sectormask));
549 * Corrupted block numbers can get through to here, unfortunately, so we
550 * have to check that the buffer falls within the filesystem bounds.
552 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
553 if (cmap.bm_bn < 0 || cmap.bm_bn >= eofs) {
554 xfs_alert(btp->bt_mount,
555 "%s: daddr 0x%llx out of range, EOFS 0x%llx",
556 __func__, cmap.bm_bn, eofs);
558 return -EFSCORRUPTED;
561 pag = xfs_perag_get(btp->bt_mount,
562 xfs_daddr_to_agno(btp->bt_mount, cmap.bm_bn));
564 spin_lock(&pag->pag_buf_lock);
565 bp = rhashtable_lookup_fast(&pag->pag_buf_hash, &cmap,
566 xfs_buf_hash_params);
568 atomic_inc(&bp->b_hold);
574 XFS_STATS_INC(btp->bt_mount, xb_miss_locked);
575 spin_unlock(&pag->pag_buf_lock);
580 /* the buffer keeps the perag reference until it is freed */
582 rhashtable_insert_fast(&pag->pag_buf_hash, &new_bp->b_rhash_head,
583 xfs_buf_hash_params);
584 spin_unlock(&pag->pag_buf_lock);
589 spin_unlock(&pag->pag_buf_lock);
592 if (!xfs_buf_trylock(bp)) {
593 if (flags & XBF_TRYLOCK) {
595 XFS_STATS_INC(btp->bt_mount, xb_busy_locked);
599 XFS_STATS_INC(btp->bt_mount, xb_get_locked_waited);
603 * if the buffer is stale, clear all the external state associated with
604 * it. We need to keep flags such as how we allocated the buffer memory
607 if (bp->b_flags & XBF_STALE) {
608 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
609 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
613 trace_xfs_buf_find(bp, flags, _RET_IP_);
614 XFS_STATS_INC(btp->bt_mount, xb_get_locked);
621 struct xfs_buftarg *target,
624 xfs_buf_flags_t flags)
628 DEFINE_SINGLE_BUF_MAP(map, blkno, numblks);
630 error = xfs_buf_find(target, &map, 1, flags, NULL, &bp);
637 * Assembles a buffer covering the specified range. The code is optimised for
638 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
639 * more hits than misses.
643 struct xfs_buftarg *target,
644 struct xfs_buf_map *map,
646 xfs_buf_flags_t flags,
647 struct xfs_buf **bpp)
650 struct xfs_buf *new_bp;
654 error = xfs_buf_find(target, map, nmaps, flags, NULL, &bp);
657 if (error != -ENOENT)
660 error = _xfs_buf_alloc(target, map, nmaps, flags, &new_bp);
665 * For buffers that fit entirely within a single page, first attempt to
666 * allocate the memory from the heap to minimise memory usage. If we
667 * can't get heap memory for these small buffers, we fall back to using
668 * the page allocator.
670 if (BBTOB(new_bp->b_length) >= PAGE_SIZE ||
671 xfs_buf_alloc_kmem(new_bp, flags) < 0) {
672 error = xfs_buf_alloc_pages(new_bp, flags);
677 error = xfs_buf_find(target, map, nmaps, flags, new_bp, &bp);
682 xfs_buf_free(new_bp);
686 error = _xfs_buf_map_pages(bp, flags);
687 if (unlikely(error)) {
688 xfs_warn_ratelimited(target->bt_mount,
689 "%s: failed to map %u pages", __func__,
697 * Clear b_error if this is a lookup from a caller that doesn't expect
698 * valid data to be found in the buffer.
700 if (!(flags & XBF_READ))
701 xfs_buf_ioerror(bp, 0);
703 XFS_STATS_INC(target->bt_mount, xb_get);
704 trace_xfs_buf_get(bp, flags, _RET_IP_);
708 xfs_buf_free(new_bp);
715 xfs_buf_flags_t flags)
717 ASSERT(!(flags & XBF_WRITE));
718 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
720 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD | XBF_DONE);
721 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
723 return xfs_buf_submit(bp);
727 * Reverify a buffer found in cache without an attached ->b_ops.
729 * If the caller passed an ops structure and the buffer doesn't have ops
730 * assigned, set the ops and use it to verify the contents. If verification
731 * fails, clear XBF_DONE. We assume the buffer has no recorded errors and is
732 * already in XBF_DONE state on entry.
734 * Under normal operations, every in-core buffer is verified on read I/O
735 * completion. There are two scenarios that can lead to in-core buffers without
736 * an assigned ->b_ops. The first is during log recovery of buffers on a V4
737 * filesystem, though these buffers are purged at the end of recovery. The
738 * other is online repair, which intentionally reads with a NULL buffer ops to
739 * run several verifiers across an in-core buffer in order to establish buffer
740 * type. If repair can't establish that, the buffer will be left in memory
741 * with NULL buffer ops.
746 const struct xfs_buf_ops *ops)
748 ASSERT(bp->b_flags & XBF_DONE);
749 ASSERT(bp->b_error == 0);
751 if (!ops || bp->b_ops)
755 bp->b_ops->verify_read(bp);
757 bp->b_flags &= ~XBF_DONE;
763 struct xfs_buftarg *target,
764 struct xfs_buf_map *map,
766 xfs_buf_flags_t flags,
767 struct xfs_buf **bpp,
768 const struct xfs_buf_ops *ops,
777 error = xfs_buf_get_map(target, map, nmaps, flags, &bp);
781 trace_xfs_buf_read(bp, flags, _RET_IP_);
783 if (!(bp->b_flags & XBF_DONE)) {
784 /* Initiate the buffer read and wait. */
785 XFS_STATS_INC(target->bt_mount, xb_get_read);
787 error = _xfs_buf_read(bp, flags);
789 /* Readahead iodone already dropped the buffer, so exit. */
790 if (flags & XBF_ASYNC)
793 /* Buffer already read; all we need to do is check it. */
794 error = xfs_buf_reverify(bp, ops);
796 /* Readahead already finished; drop the buffer and exit. */
797 if (flags & XBF_ASYNC) {
802 /* We do not want read in the flags */
803 bp->b_flags &= ~XBF_READ;
804 ASSERT(bp->b_ops != NULL || ops == NULL);
808 * If we've had a read error, then the contents of the buffer are
809 * invalid and should not be used. To ensure that a followup read tries
810 * to pull the buffer from disk again, we clear the XBF_DONE flag and
811 * mark the buffer stale. This ensures that anyone who has a current
812 * reference to the buffer will interpret it's contents correctly and
813 * future cache lookups will also treat it as an empty, uninitialised
818 * Check against log shutdown for error reporting because
819 * metadata writeback may require a read first and we need to
820 * report errors in metadata writeback until the log is shut
821 * down. High level transaction read functions already check
822 * against mount shutdown, anyway, so we only need to be
823 * concerned about low level IO interactions here.
825 if (!xlog_is_shutdown(target->bt_mount->m_log))
826 xfs_buf_ioerror_alert(bp, fa);
828 bp->b_flags &= ~XBF_DONE;
832 /* bad CRC means corrupted metadata */
833 if (error == -EFSBADCRC)
834 error = -EFSCORRUPTED;
843 * If we are not low on memory then do the readahead in a deadlock
847 xfs_buf_readahead_map(
848 struct xfs_buftarg *target,
849 struct xfs_buf_map *map,
851 const struct xfs_buf_ops *ops)
855 if (bdi_read_congested(target->bt_bdev->bd_disk->bdi))
858 xfs_buf_read_map(target, map, nmaps,
859 XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD, &bp, ops,
864 * Read an uncached buffer from disk. Allocates and returns a locked
865 * buffer containing the disk contents or nothing. Uncached buffers always have
866 * a cache index of XFS_BUF_DADDR_NULL so we can easily determine if the buffer
867 * is cached or uncached during fault diagnosis.
870 xfs_buf_read_uncached(
871 struct xfs_buftarg *target,
875 struct xfs_buf **bpp,
876 const struct xfs_buf_ops *ops)
883 error = xfs_buf_get_uncached(target, numblks, flags, &bp);
887 /* set up the buffer for a read IO */
888 ASSERT(bp->b_map_count == 1);
889 bp->b_rhash_key = XFS_BUF_DADDR_NULL;
890 bp->b_maps[0].bm_bn = daddr;
891 bp->b_flags |= XBF_READ;
906 xfs_buf_get_uncached(
907 struct xfs_buftarg *target,
910 struct xfs_buf **bpp)
914 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
918 /* flags might contain irrelevant bits, pass only what we care about */
919 error = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT, &bp);
923 error = xfs_buf_alloc_pages(bp, flags);
927 error = _xfs_buf_map_pages(bp, 0);
928 if (unlikely(error)) {
929 xfs_warn(target->bt_mount,
930 "%s: failed to map pages", __func__);
934 trace_xfs_buf_get_uncached(bp, _RET_IP_);
944 * Increment reference count on buffer, to hold the buffer concurrently
945 * with another thread which may release (free) the buffer asynchronously.
946 * Must hold the buffer already to call this function.
952 trace_xfs_buf_hold(bp, _RET_IP_);
953 atomic_inc(&bp->b_hold);
957 * Release a hold on the specified buffer. If the hold count is 1, the buffer is
958 * placed on LRU or freed (depending on b_lru_ref).
964 struct xfs_perag *pag = bp->b_pag;
966 bool freebuf = false;
968 trace_xfs_buf_rele(bp, _RET_IP_);
971 ASSERT(list_empty(&bp->b_lru));
972 if (atomic_dec_and_test(&bp->b_hold)) {
973 xfs_buf_ioacct_dec(bp);
979 ASSERT(atomic_read(&bp->b_hold) > 0);
982 * We grab the b_lock here first to serialise racing xfs_buf_rele()
983 * calls. The pag_buf_lock being taken on the last reference only
984 * serialises against racing lookups in xfs_buf_find(). IOWs, the second
985 * to last reference we drop here is not serialised against the last
986 * reference until we take bp->b_lock. Hence if we don't grab b_lock
987 * first, the last "release" reference can win the race to the lock and
988 * free the buffer before the second-to-last reference is processed,
989 * leading to a use-after-free scenario.
991 spin_lock(&bp->b_lock);
992 release = atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock);
995 * Drop the in-flight state if the buffer is already on the LRU
996 * and it holds the only reference. This is racy because we
997 * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT
998 * ensures the decrement occurs only once per-buf.
1000 if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru))
1001 __xfs_buf_ioacct_dec(bp);
1005 /* the last reference has been dropped ... */
1006 __xfs_buf_ioacct_dec(bp);
1007 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
1009 * If the buffer is added to the LRU take a new reference to the
1010 * buffer for the LRU and clear the (now stale) dispose list
1013 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) {
1014 bp->b_state &= ~XFS_BSTATE_DISPOSE;
1015 atomic_inc(&bp->b_hold);
1017 spin_unlock(&pag->pag_buf_lock);
1020 * most of the time buffers will already be removed from the
1021 * LRU, so optimise that case by checking for the
1022 * XFS_BSTATE_DISPOSE flag indicating the last list the buffer
1023 * was on was the disposal list
1025 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
1026 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru);
1028 ASSERT(list_empty(&bp->b_lru));
1031 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1032 rhashtable_remove_fast(&pag->pag_buf_hash, &bp->b_rhash_head,
1033 xfs_buf_hash_params);
1034 spin_unlock(&pag->pag_buf_lock);
1040 spin_unlock(&bp->b_lock);
1048 * Lock a buffer object, if it is not already locked.
1050 * If we come across a stale, pinned, locked buffer, we know that we are
1051 * being asked to lock a buffer that has been reallocated. Because it is
1052 * pinned, we know that the log has not been pushed to disk and hence it
1053 * will still be locked. Rather than continuing to have trylock attempts
1054 * fail until someone else pushes the log, push it ourselves before
1055 * returning. This means that the xfsaild will not get stuck trying
1056 * to push on stale inode buffers.
1064 locked = down_trylock(&bp->b_sema) == 0;
1066 trace_xfs_buf_trylock(bp, _RET_IP_);
1068 trace_xfs_buf_trylock_fail(bp, _RET_IP_);
1073 * Lock a buffer object.
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.
1085 trace_xfs_buf_lock(bp, _RET_IP_);
1087 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
1088 xfs_log_force(bp->b_mount, 0);
1091 trace_xfs_buf_lock_done(bp, _RET_IP_);
1098 ASSERT(xfs_buf_islocked(bp));
1101 trace_xfs_buf_unlock(bp, _RET_IP_);
1108 DECLARE_WAITQUEUE (wait, current);
1110 if (atomic_read(&bp->b_pin_count) == 0)
1113 add_wait_queue(&bp->b_waiters, &wait);
1115 set_current_state(TASK_UNINTERRUPTIBLE);
1116 if (atomic_read(&bp->b_pin_count) == 0)
1120 remove_wait_queue(&bp->b_waiters, &wait);
1121 set_current_state(TASK_RUNNING);
1125 xfs_buf_ioerror_alert_ratelimited(
1128 static unsigned long lasttime;
1129 static struct xfs_buftarg *lasttarg;
1131 if (bp->b_target != lasttarg ||
1132 time_after(jiffies, (lasttime + 5*HZ))) {
1134 xfs_buf_ioerror_alert(bp, __this_address);
1136 lasttarg = bp->b_target;
1140 * Account for this latest trip around the retry handler, and decide if
1141 * we've failed enough times to constitute a permanent failure.
1144 xfs_buf_ioerror_permanent(
1146 struct xfs_error_cfg *cfg)
1148 struct xfs_mount *mp = bp->b_mount;
1150 if (cfg->max_retries != XFS_ERR_RETRY_FOREVER &&
1151 ++bp->b_retries > cfg->max_retries)
1153 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1154 time_after(jiffies, cfg->retry_timeout + bp->b_first_retry_time))
1157 /* At unmount we may treat errors differently */
1158 if (xfs_is_unmounting(mp) && mp->m_fail_unmount)
1165 * On a sync write or shutdown we just want to stale the buffer and let the
1166 * caller handle the error in bp->b_error appropriately.
1168 * If the write was asynchronous then no one will be looking for the error. If
1169 * this is the first failure of this type, clear the error state and write the
1170 * buffer out again. This means we always retry an async write failure at least
1171 * once, but we also need to set the buffer up to behave correctly now for
1172 * repeated failures.
1174 * If we get repeated async write failures, then we take action according to the
1175 * error configuration we have been set up to use.
1177 * Returns true if this function took care of error handling and the caller must
1178 * not touch the buffer again. Return false if the caller should proceed with
1179 * normal I/O completion handling.
1182 xfs_buf_ioend_handle_error(
1185 struct xfs_mount *mp = bp->b_mount;
1186 struct xfs_error_cfg *cfg;
1189 * If we've already shutdown the journal because of I/O errors, there's
1190 * no point in giving this a retry.
1192 if (xlog_is_shutdown(mp->m_log))
1195 xfs_buf_ioerror_alert_ratelimited(bp);
1198 * We're not going to bother about retrying this during recovery.
1201 if (bp->b_flags & _XBF_LOGRECOVERY) {
1202 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1207 * Synchronous writes will have callers process the error.
1209 if (!(bp->b_flags & XBF_ASYNC))
1212 trace_xfs_buf_iodone_async(bp, _RET_IP_);
1214 cfg = xfs_error_get_cfg(mp, XFS_ERR_METADATA, bp->b_error);
1215 if (bp->b_last_error != bp->b_error ||
1216 !(bp->b_flags & (XBF_STALE | XBF_WRITE_FAIL))) {
1217 bp->b_last_error = bp->b_error;
1218 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1219 !bp->b_first_retry_time)
1220 bp->b_first_retry_time = jiffies;
1225 * Permanent error - we need to trigger a shutdown if we haven't already
1226 * to indicate that inconsistency will result from this action.
1228 if (xfs_buf_ioerror_permanent(bp, cfg)) {
1229 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1233 /* Still considered a transient error. Caller will schedule retries. */
1234 if (bp->b_flags & _XBF_INODES)
1235 xfs_buf_inode_io_fail(bp);
1236 else if (bp->b_flags & _XBF_DQUOTS)
1237 xfs_buf_dquot_io_fail(bp);
1239 ASSERT(list_empty(&bp->b_li_list));
1240 xfs_buf_ioerror(bp, 0);
1245 xfs_buf_ioerror(bp, 0);
1246 bp->b_flags |= (XBF_DONE | XBF_WRITE_FAIL);
1251 bp->b_flags |= XBF_DONE;
1252 bp->b_flags &= ~XBF_WRITE;
1253 trace_xfs_buf_error_relse(bp, _RET_IP_);
1261 trace_xfs_buf_iodone(bp, _RET_IP_);
1264 * Pull in IO completion errors now. We are guaranteed to be running
1265 * single threaded, so we don't need the lock to read b_io_error.
1267 if (!bp->b_error && bp->b_io_error)
1268 xfs_buf_ioerror(bp, bp->b_io_error);
1270 if (bp->b_flags & XBF_READ) {
1271 if (!bp->b_error && bp->b_ops)
1272 bp->b_ops->verify_read(bp);
1274 bp->b_flags |= XBF_DONE;
1277 bp->b_flags &= ~XBF_WRITE_FAIL;
1278 bp->b_flags |= XBF_DONE;
1281 if (unlikely(bp->b_error) && xfs_buf_ioend_handle_error(bp))
1284 /* clear the retry state */
1285 bp->b_last_error = 0;
1287 bp->b_first_retry_time = 0;
1290 * Note that for things like remote attribute buffers, there may
1291 * not be a buffer log item here, so processing the buffer log
1292 * item must remain optional.
1295 xfs_buf_item_done(bp);
1297 if (bp->b_flags & _XBF_INODES)
1298 xfs_buf_inode_iodone(bp);
1299 else if (bp->b_flags & _XBF_DQUOTS)
1300 xfs_buf_dquot_iodone(bp);
1304 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD |
1307 if (bp->b_flags & XBF_ASYNC)
1310 complete(&bp->b_iowait);
1315 struct work_struct *work)
1317 struct xfs_buf *bp =
1318 container_of(work, struct xfs_buf, b_ioend_work);
1324 xfs_buf_ioend_async(
1327 INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
1328 queue_work(bp->b_mount->m_buf_workqueue, &bp->b_ioend_work);
1335 xfs_failaddr_t failaddr)
1337 ASSERT(error <= 0 && error >= -1000);
1338 bp->b_error = error;
1339 trace_xfs_buf_ioerror(bp, error, failaddr);
1343 xfs_buf_ioerror_alert(
1345 xfs_failaddr_t func)
1347 xfs_buf_alert_ratelimited(bp, "XFS: metadata IO error",
1348 "metadata I/O error in \"%pS\" at daddr 0x%llx len %d error %d",
1349 func, (uint64_t)xfs_buf_daddr(bp),
1350 bp->b_length, -bp->b_error);
1354 * To simulate an I/O failure, the buffer must be locked and held with at least
1355 * three references. The LRU reference is dropped by the stale call. The buf
1356 * item reference is dropped via ioend processing. The third reference is owned
1357 * by the caller and is dropped on I/O completion if the buffer is XBF_ASYNC.
1363 bp->b_flags &= ~XBF_DONE;
1365 xfs_buf_ioerror(bp, -EIO);
1375 ASSERT(xfs_buf_islocked(bp));
1377 bp->b_flags |= XBF_WRITE;
1378 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1381 error = xfs_buf_submit(bp);
1383 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
1391 struct xfs_buf *bp = (struct xfs_buf *)bio->bi_private;
1393 if (!bio->bi_status &&
1394 (bp->b_flags & XBF_WRITE) && (bp->b_flags & XBF_ASYNC) &&
1395 XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_IOERROR))
1396 bio->bi_status = BLK_STS_IOERR;
1399 * don't overwrite existing errors - otherwise we can lose errors on
1400 * buffers that require multiple bios to complete.
1402 if (bio->bi_status) {
1403 int error = blk_status_to_errno(bio->bi_status);
1405 cmpxchg(&bp->b_io_error, 0, error);
1408 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1409 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1411 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1412 xfs_buf_ioend_async(bp);
1417 xfs_buf_ioapply_map(
1425 unsigned int total_nr_pages = bp->b_page_count;
1428 sector_t sector = bp->b_maps[map].bm_bn;
1432 /* skip the pages in the buffer before the start offset */
1434 offset = *buf_offset;
1435 while (offset >= PAGE_SIZE) {
1437 offset -= PAGE_SIZE;
1441 * Limit the IO size to the length of the current vector, and update the
1442 * remaining IO count for the next time around.
1444 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1446 *buf_offset += size;
1449 atomic_inc(&bp->b_io_remaining);
1450 nr_pages = bio_max_segs(total_nr_pages);
1452 bio = bio_alloc(GFP_NOIO, nr_pages);
1453 bio_set_dev(bio, bp->b_target->bt_bdev);
1454 bio->bi_iter.bi_sector = sector;
1455 bio->bi_end_io = xfs_buf_bio_end_io;
1456 bio->bi_private = bp;
1459 for (; size && nr_pages; nr_pages--, page_index++) {
1460 int rbytes, nbytes = PAGE_SIZE - offset;
1465 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1467 if (rbytes < nbytes)
1471 sector += BTOBB(nbytes);
1476 if (likely(bio->bi_iter.bi_size)) {
1477 if (xfs_buf_is_vmapped(bp)) {
1478 flush_kernel_vmap_range(bp->b_addr,
1479 xfs_buf_vmap_len(bp));
1486 * This is guaranteed not to be the last io reference count
1487 * because the caller (xfs_buf_submit) holds a count itself.
1489 atomic_dec(&bp->b_io_remaining);
1490 xfs_buf_ioerror(bp, -EIO);
1500 struct blk_plug plug;
1507 * Make sure we capture only current IO errors rather than stale errors
1508 * left over from previous use of the buffer (e.g. failed readahead).
1512 if (bp->b_flags & XBF_WRITE) {
1516 * Run the write verifier callback function if it exists. If
1517 * this function fails it will mark the buffer with an error and
1518 * the IO should not be dispatched.
1521 bp->b_ops->verify_write(bp);
1523 xfs_force_shutdown(bp->b_mount,
1524 SHUTDOWN_CORRUPT_INCORE);
1527 } else if (bp->b_rhash_key != XFS_BUF_DADDR_NULL) {
1528 struct xfs_mount *mp = bp->b_mount;
1531 * non-crc filesystems don't attach verifiers during
1532 * log recovery, so don't warn for such filesystems.
1534 if (xfs_has_crc(mp)) {
1536 "%s: no buf ops on daddr 0x%llx len %d",
1537 __func__, xfs_buf_daddr(bp),
1539 xfs_hex_dump(bp->b_addr,
1540 XFS_CORRUPTION_DUMP_LEN);
1546 if (bp->b_flags & XBF_READ_AHEAD)
1550 /* we only use the buffer cache for meta-data */
1554 * Walk all the vectors issuing IO on them. Set up the initial offset
1555 * into the buffer and the desired IO size before we start -
1556 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1559 offset = bp->b_offset;
1560 size = BBTOB(bp->b_length);
1561 blk_start_plug(&plug);
1562 for (i = 0; i < bp->b_map_count; i++) {
1563 xfs_buf_ioapply_map(bp, i, &offset, &size, op);
1567 break; /* all done */
1569 blk_finish_plug(&plug);
1573 * Wait for I/O completion of a sync buffer and return the I/O error code.
1579 ASSERT(!(bp->b_flags & XBF_ASYNC));
1581 trace_xfs_buf_iowait(bp, _RET_IP_);
1582 wait_for_completion(&bp->b_iowait);
1583 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1589 * Buffer I/O submission path, read or write. Asynchronous submission transfers
1590 * the buffer lock ownership and the current reference to the IO. It is not
1591 * safe to reference the buffer after a call to this function unless the caller
1592 * holds an additional reference itself.
1601 trace_xfs_buf_submit(bp, _RET_IP_);
1603 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1606 * On log shutdown we stale and complete the buffer immediately. We can
1607 * be called to read the superblock before the log has been set up, so
1608 * be careful checking the log state.
1610 * Checking the mount shutdown state here can result in the log tail
1611 * moving inappropriately on disk as the log may not yet be shut down.
1612 * i.e. failing this buffer on mount shutdown can remove it from the AIL
1613 * and move the tail of the log forwards without having written this
1614 * buffer to disk. This corrupts the log tail state in memory, and
1615 * because the log may not be shut down yet, it can then be propagated
1616 * to disk before the log is shutdown. Hence we check log shutdown
1617 * state here rather than mount state to avoid corrupting the log tail
1620 if (bp->b_mount->m_log &&
1621 xlog_is_shutdown(bp->b_mount->m_log)) {
1622 xfs_buf_ioend_fail(bp);
1627 * Grab a reference so the buffer does not go away underneath us. For
1628 * async buffers, I/O completion drops the callers reference, which
1629 * could occur before submission returns.
1633 if (bp->b_flags & XBF_WRITE)
1634 xfs_buf_wait_unpin(bp);
1636 /* clear the internal error state to avoid spurious errors */
1640 * Set the count to 1 initially, this will stop an I/O completion
1641 * callout which happens before we have started all the I/O from calling
1642 * xfs_buf_ioend too early.
1644 atomic_set(&bp->b_io_remaining, 1);
1645 if (bp->b_flags & XBF_ASYNC)
1646 xfs_buf_ioacct_inc(bp);
1647 _xfs_buf_ioapply(bp);
1650 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1651 * reference we took above. If we drop it to zero, run completion so
1652 * that we don't return to the caller with completion still pending.
1654 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1655 if (bp->b_error || !(bp->b_flags & XBF_ASYNC))
1658 xfs_buf_ioend_async(bp);
1662 error = xfs_buf_iowait(bp);
1665 * Release the hold that keeps the buffer referenced for the entire
1666 * I/O. Note that if the buffer is async, it is not safe to reference
1667 * after this release.
1681 return bp->b_addr + offset;
1683 page = bp->b_pages[offset >> PAGE_SHIFT];
1684 return page_address(page) + (offset & (PAGE_SIZE-1));
1695 bend = boff + bsize;
1696 while (boff < bend) {
1698 int page_index, page_offset, csize;
1700 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1701 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1702 page = bp->b_pages[page_index];
1703 csize = min_t(size_t, PAGE_SIZE - page_offset,
1704 BBTOB(bp->b_length) - boff);
1706 ASSERT((csize + page_offset) <= PAGE_SIZE);
1708 memset(page_address(page) + page_offset, 0, csize);
1715 * Log a message about and stale a buffer that a caller has decided is corrupt.
1717 * This function should be called for the kinds of metadata corruption that
1718 * cannot be detect from a verifier, such as incorrect inter-block relationship
1719 * data. Do /not/ call this function from a verifier function.
1721 * The buffer must be XBF_DONE prior to the call. Afterwards, the buffer will
1722 * be marked stale, but b_error will not be set. The caller is responsible for
1723 * releasing the buffer or fixing it.
1726 __xfs_buf_mark_corrupt(
1730 ASSERT(bp->b_flags & XBF_DONE);
1732 xfs_buf_corruption_error(bp, fa);
1737 * Handling of buffer targets (buftargs).
1741 * Wait for any bufs with callbacks that have been submitted but have not yet
1742 * returned. These buffers will have an elevated hold count, so wait on those
1743 * while freeing all the buffers only held by the LRU.
1745 static enum lru_status
1746 xfs_buftarg_drain_rele(
1747 struct list_head *item,
1748 struct list_lru_one *lru,
1749 spinlock_t *lru_lock,
1753 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1754 struct list_head *dispose = arg;
1756 if (atomic_read(&bp->b_hold) > 1) {
1757 /* need to wait, so skip it this pass */
1758 trace_xfs_buf_drain_buftarg(bp, _RET_IP_);
1761 if (!spin_trylock(&bp->b_lock))
1765 * clear the LRU reference count so the buffer doesn't get
1766 * ignored in xfs_buf_rele().
1768 atomic_set(&bp->b_lru_ref, 0);
1769 bp->b_state |= XFS_BSTATE_DISPOSE;
1770 list_lru_isolate_move(lru, item, dispose);
1771 spin_unlock(&bp->b_lock);
1776 * Wait for outstanding I/O on the buftarg to complete.
1780 struct xfs_buftarg *btp)
1783 * First wait on the buftarg I/O count for all in-flight buffers to be
1784 * released. This is critical as new buffers do not make the LRU until
1785 * they are released.
1787 * Next, flush the buffer workqueue to ensure all completion processing
1788 * has finished. Just waiting on buffer locks is not sufficient for
1789 * async IO as the reference count held over IO is not released until
1790 * after the buffer lock is dropped. Hence we need to ensure here that
1791 * all reference counts have been dropped before we start walking the
1794 while (percpu_counter_sum(&btp->bt_io_count))
1796 flush_workqueue(btp->bt_mount->m_buf_workqueue);
1801 struct xfs_buftarg *btp)
1805 bool write_fail = false;
1807 xfs_buftarg_wait(btp);
1809 /* loop until there is nothing left on the lru list. */
1810 while (list_lru_count(&btp->bt_lru)) {
1811 list_lru_walk(&btp->bt_lru, xfs_buftarg_drain_rele,
1812 &dispose, LONG_MAX);
1814 while (!list_empty(&dispose)) {
1816 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1817 list_del_init(&bp->b_lru);
1818 if (bp->b_flags & XBF_WRITE_FAIL) {
1820 xfs_buf_alert_ratelimited(bp,
1821 "XFS: Corruption Alert",
1822 "Corruption Alert: Buffer at daddr 0x%llx had permanent write failures!",
1823 (long long)xfs_buf_daddr(bp));
1832 * If one or more failed buffers were freed, that means dirty metadata
1833 * was thrown away. This should only ever happen after I/O completion
1834 * handling has elevated I/O error(s) to permanent failures and shuts
1838 ASSERT(xlog_is_shutdown(btp->bt_mount->m_log));
1839 xfs_alert(btp->bt_mount,
1840 "Please run xfs_repair to determine the extent of the problem.");
1844 static enum lru_status
1845 xfs_buftarg_isolate(
1846 struct list_head *item,
1847 struct list_lru_one *lru,
1848 spinlock_t *lru_lock,
1851 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1852 struct list_head *dispose = arg;
1855 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1856 * If we fail to get the lock, just skip it.
1858 if (!spin_trylock(&bp->b_lock))
1861 * Decrement the b_lru_ref count unless the value is already
1862 * zero. If the value is already zero, we need to reclaim the
1863 * buffer, otherwise it gets another trip through the LRU.
1865 if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1866 spin_unlock(&bp->b_lock);
1870 bp->b_state |= XFS_BSTATE_DISPOSE;
1871 list_lru_isolate_move(lru, item, dispose);
1872 spin_unlock(&bp->b_lock);
1876 static unsigned long
1877 xfs_buftarg_shrink_scan(
1878 struct shrinker *shrink,
1879 struct shrink_control *sc)
1881 struct xfs_buftarg *btp = container_of(shrink,
1882 struct xfs_buftarg, bt_shrinker);
1884 unsigned long freed;
1886 freed = list_lru_shrink_walk(&btp->bt_lru, sc,
1887 xfs_buftarg_isolate, &dispose);
1889 while (!list_empty(&dispose)) {
1891 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1892 list_del_init(&bp->b_lru);
1899 static unsigned long
1900 xfs_buftarg_shrink_count(
1901 struct shrinker *shrink,
1902 struct shrink_control *sc)
1904 struct xfs_buftarg *btp = container_of(shrink,
1905 struct xfs_buftarg, bt_shrinker);
1906 return list_lru_shrink_count(&btp->bt_lru, sc);
1911 struct xfs_buftarg *btp)
1913 unregister_shrinker(&btp->bt_shrinker);
1914 ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
1915 percpu_counter_destroy(&btp->bt_io_count);
1916 list_lru_destroy(&btp->bt_lru);
1918 blkdev_issue_flush(btp->bt_bdev);
1919 fs_put_dax(btp->bt_daxdev);
1925 xfs_setsize_buftarg(
1927 unsigned int sectorsize)
1929 /* Set up metadata sector size info */
1930 btp->bt_meta_sectorsize = sectorsize;
1931 btp->bt_meta_sectormask = sectorsize - 1;
1933 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1934 xfs_warn(btp->bt_mount,
1935 "Cannot set_blocksize to %u on device %pg",
1936 sectorsize, btp->bt_bdev);
1940 /* Set up device logical sector size mask */
1941 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1942 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1948 * When allocating the initial buffer target we have not yet
1949 * read in the superblock, so don't know what sized sectors
1950 * are being used at this early stage. Play safe.
1953 xfs_setsize_buftarg_early(
1955 struct block_device *bdev)
1957 return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev));
1960 struct xfs_buftarg *
1962 struct xfs_mount *mp,
1963 struct block_device *bdev)
1967 btp = kmem_zalloc(sizeof(*btp), KM_NOFS);
1970 btp->bt_dev = bdev->bd_dev;
1971 btp->bt_bdev = bdev;
1972 btp->bt_daxdev = fs_dax_get_by_bdev(bdev, &btp->bt_dax_part_off);
1975 * Buffer IO error rate limiting. Limit it to no more than 10 messages
1976 * per 30 seconds so as to not spam logs too much on repeated errors.
1978 ratelimit_state_init(&btp->bt_ioerror_rl, 30 * HZ,
1979 DEFAULT_RATELIMIT_BURST);
1981 if (xfs_setsize_buftarg_early(btp, bdev))
1984 if (list_lru_init(&btp->bt_lru))
1987 if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL))
1990 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
1991 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
1992 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1993 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
1994 if (register_shrinker(&btp->bt_shrinker))
1999 percpu_counter_destroy(&btp->bt_io_count);
2001 list_lru_destroy(&btp->bt_lru);
2008 * Cancel a delayed write list.
2010 * Remove each buffer from the list, clear the delwri queue flag and drop the
2011 * associated buffer reference.
2014 xfs_buf_delwri_cancel(
2015 struct list_head *list)
2019 while (!list_empty(list)) {
2020 bp = list_first_entry(list, struct xfs_buf, b_list);
2023 bp->b_flags &= ~_XBF_DELWRI_Q;
2024 list_del_init(&bp->b_list);
2030 * Add a buffer to the delayed write list.
2032 * This queues a buffer for writeout if it hasn't already been. Note that
2033 * neither this routine nor the buffer list submission functions perform
2034 * any internal synchronization. It is expected that the lists are thread-local
2037 * Returns true if we queued up the buffer, or false if it already had
2038 * been on the buffer list.
2041 xfs_buf_delwri_queue(
2043 struct list_head *list)
2045 ASSERT(xfs_buf_islocked(bp));
2046 ASSERT(!(bp->b_flags & XBF_READ));
2049 * If the buffer is already marked delwri it already is queued up
2050 * by someone else for imediate writeout. Just ignore it in that
2053 if (bp->b_flags & _XBF_DELWRI_Q) {
2054 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
2058 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
2061 * If a buffer gets written out synchronously or marked stale while it
2062 * is on a delwri list we lazily remove it. To do this, the other party
2063 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
2064 * It remains referenced and on the list. In a rare corner case it
2065 * might get readded to a delwri list after the synchronous writeout, in
2066 * which case we need just need to re-add the flag here.
2068 bp->b_flags |= _XBF_DELWRI_Q;
2069 if (list_empty(&bp->b_list)) {
2070 atomic_inc(&bp->b_hold);
2071 list_add_tail(&bp->b_list, list);
2078 * Compare function is more complex than it needs to be because
2079 * the return value is only 32 bits and we are doing comparisons
2085 const struct list_head *a,
2086 const struct list_head *b)
2088 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
2089 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
2092 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
2101 * Submit buffers for write. If wait_list is specified, the buffers are
2102 * submitted using sync I/O and placed on the wait list such that the caller can
2103 * iowait each buffer. Otherwise async I/O is used and the buffers are released
2104 * at I/O completion time. In either case, buffers remain locked until I/O
2105 * completes and the buffer is released from the queue.
2108 xfs_buf_delwri_submit_buffers(
2109 struct list_head *buffer_list,
2110 struct list_head *wait_list)
2112 struct xfs_buf *bp, *n;
2114 struct blk_plug plug;
2116 list_sort(NULL, buffer_list, xfs_buf_cmp);
2118 blk_start_plug(&plug);
2119 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
2121 if (!xfs_buf_trylock(bp))
2123 if (xfs_buf_ispinned(bp)) {
2133 * Someone else might have written the buffer synchronously or
2134 * marked it stale in the meantime. In that case only the
2135 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
2136 * reference and remove it from the list here.
2138 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
2139 list_del_init(&bp->b_list);
2144 trace_xfs_buf_delwri_split(bp, _RET_IP_);
2147 * If we have a wait list, each buffer (and associated delwri
2148 * queue reference) transfers to it and is submitted
2149 * synchronously. Otherwise, drop the buffer from the delwri
2150 * queue and submit async.
2152 bp->b_flags &= ~_XBF_DELWRI_Q;
2153 bp->b_flags |= XBF_WRITE;
2155 bp->b_flags &= ~XBF_ASYNC;
2156 list_move_tail(&bp->b_list, wait_list);
2158 bp->b_flags |= XBF_ASYNC;
2159 list_del_init(&bp->b_list);
2161 __xfs_buf_submit(bp, false);
2163 blk_finish_plug(&plug);
2169 * Write out a buffer list asynchronously.
2171 * This will take the @buffer_list, write all non-locked and non-pinned buffers
2172 * out and not wait for I/O completion on any of the buffers. This interface
2173 * is only safely useable for callers that can track I/O completion by higher
2174 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
2177 * Note: this function will skip buffers it would block on, and in doing so
2178 * leaves them on @buffer_list so they can be retried on a later pass. As such,
2179 * it is up to the caller to ensure that the buffer list is fully submitted or
2180 * cancelled appropriately when they are finished with the list. Failure to
2181 * cancel or resubmit the list until it is empty will result in leaked buffers
2185 xfs_buf_delwri_submit_nowait(
2186 struct list_head *buffer_list)
2188 return xfs_buf_delwri_submit_buffers(buffer_list, NULL);
2192 * Write out a buffer list synchronously.
2194 * This will take the @buffer_list, write all buffers out and wait for I/O
2195 * completion on all of the buffers. @buffer_list is consumed by the function,
2196 * so callers must have some other way of tracking buffers if they require such
2200 xfs_buf_delwri_submit(
2201 struct list_head *buffer_list)
2203 LIST_HEAD (wait_list);
2204 int error = 0, error2;
2207 xfs_buf_delwri_submit_buffers(buffer_list, &wait_list);
2209 /* Wait for IO to complete. */
2210 while (!list_empty(&wait_list)) {
2211 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
2213 list_del_init(&bp->b_list);
2216 * Wait on the locked buffer, check for errors and unlock and
2217 * release the delwri queue reference.
2219 error2 = xfs_buf_iowait(bp);
2229 * Push a single buffer on a delwri queue.
2231 * The purpose of this function is to submit a single buffer of a delwri queue
2232 * and return with the buffer still on the original queue. The waiting delwri
2233 * buffer submission infrastructure guarantees transfer of the delwri queue
2234 * buffer reference to a temporary wait list. We reuse this infrastructure to
2235 * transfer the buffer back to the original queue.
2237 * Note the buffer transitions from the queued state, to the submitted and wait
2238 * listed state and back to the queued state during this call. The buffer
2239 * locking and queue management logic between _delwri_pushbuf() and
2240 * _delwri_queue() guarantee that the buffer cannot be queued to another list
2244 xfs_buf_delwri_pushbuf(
2246 struct list_head *buffer_list)
2248 LIST_HEAD (submit_list);
2251 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
2253 trace_xfs_buf_delwri_pushbuf(bp, _RET_IP_);
2256 * Isolate the buffer to a new local list so we can submit it for I/O
2257 * independently from the rest of the original list.
2260 list_move(&bp->b_list, &submit_list);
2264 * Delwri submission clears the DELWRI_Q buffer flag and returns with
2265 * the buffer on the wait list with the original reference. Rather than
2266 * bounce the buffer from a local wait list back to the original list
2267 * after I/O completion, reuse the original list as the wait list.
2269 xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
2272 * The buffer is now locked, under I/O and wait listed on the original
2273 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
2274 * return with the buffer unlocked and on the original queue.
2276 error = xfs_buf_iowait(bp);
2277 bp->b_flags |= _XBF_DELWRI_Q;
2286 xfs_buf_cache = kmem_cache_create("xfs_buf", sizeof(struct xfs_buf), 0,
2287 SLAB_HWCACHE_ALIGN |
2288 SLAB_RECLAIM_ACCOUNT |
2301 xfs_buf_terminate(void)
2303 kmem_cache_destroy(xfs_buf_cache);
2306 void xfs_buf_set_ref(struct xfs_buf *bp, int lru_ref)
2309 * Set the lru reference count to 0 based on the error injection tag.
2310 * This allows userspace to disrupt buffer caching for debug/testing
2313 if (XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_LRU_REF))
2316 atomic_set(&bp->b_lru_ref, lru_ref);
2320 * Verify an on-disk magic value against the magic value specified in the
2321 * verifier structure. The verifier magic is in disk byte order so the caller is
2322 * expected to pass the value directly from disk.
2329 struct xfs_mount *mp = bp->b_mount;
2332 idx = xfs_has_crc(mp);
2333 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic[idx]))
2335 return dmagic == bp->b_ops->magic[idx];
2338 * Verify an on-disk magic value against the magic value specified in the
2339 * verifier structure. The verifier magic is in disk byte order so the caller is
2340 * expected to pass the value directly from disk.
2347 struct xfs_mount *mp = bp->b_mount;
2350 idx = xfs_has_crc(mp);
2351 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic16[idx]))
2353 return dmagic == bp->b_ops->magic16[idx];