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"
14 #include "xfs_mount.h"
15 #include "xfs_trace.h"
17 #include "xfs_log_recover.h"
18 #include "xfs_trans.h"
19 #include "xfs_buf_item.h"
20 #include "xfs_errortag.h"
21 #include "xfs_error.h"
23 static kmem_zone_t *xfs_buf_zone;
25 #define xb_to_gfp(flags) \
26 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN)
33 * b_sema (caller holds)
37 * b_sema (caller holds)
46 * xfs_buftarg_wait_rele
48 * b_lock (trylock due to inversion)
52 * b_lock (trylock due to inversion)
60 * Return true if the buffer is vmapped.
62 * b_addr is null if the buffer is not mapped, but the code is clever
63 * enough to know it doesn't have to map a single page, so the check has
64 * to be both for b_addr and bp->b_page_count > 1.
66 return bp->b_addr && bp->b_page_count > 1;
73 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
77 * Bump the I/O in flight count on the buftarg if we haven't yet done so for
78 * this buffer. The count is incremented once per buffer (per hold cycle)
79 * because the corresponding decrement is deferred to buffer release. Buffers
80 * can undergo I/O multiple times in a hold-release cycle and per buffer I/O
81 * tracking adds unnecessary overhead. This is used for sychronization purposes
82 * with unmount (see xfs_wait_buftarg()), so all we really need is a count of
85 * Buffers that are never released (e.g., superblock, iclog buffers) must set
86 * the XBF_NO_IOACCT flag before I/O submission. Otherwise, the buftarg count
87 * never reaches zero and unmount hangs indefinitely.
93 if (bp->b_flags & XBF_NO_IOACCT)
96 ASSERT(bp->b_flags & XBF_ASYNC);
97 spin_lock(&bp->b_lock);
98 if (!(bp->b_state & XFS_BSTATE_IN_FLIGHT)) {
99 bp->b_state |= XFS_BSTATE_IN_FLIGHT;
100 percpu_counter_inc(&bp->b_target->bt_io_count);
102 spin_unlock(&bp->b_lock);
106 * Clear the in-flight state on a buffer about to be released to the LRU or
107 * freed and unaccount from the buftarg.
110 __xfs_buf_ioacct_dec(
113 lockdep_assert_held(&bp->b_lock);
115 if (bp->b_state & XFS_BSTATE_IN_FLIGHT) {
116 bp->b_state &= ~XFS_BSTATE_IN_FLIGHT;
117 percpu_counter_dec(&bp->b_target->bt_io_count);
125 spin_lock(&bp->b_lock);
126 __xfs_buf_ioacct_dec(bp);
127 spin_unlock(&bp->b_lock);
131 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
132 * b_lru_ref count so that the buffer is freed immediately when the buffer
133 * reference count falls to zero. If the buffer is already on the LRU, we need
134 * to remove the reference that LRU holds on the buffer.
136 * This prevents build-up of stale buffers on the LRU.
142 ASSERT(xfs_buf_islocked(bp));
144 bp->b_flags |= XBF_STALE;
147 * Clear the delwri status so that a delwri queue walker will not
148 * flush this buffer to disk now that it is stale. The delwri queue has
149 * a reference to the buffer, so this is safe to do.
151 bp->b_flags &= ~_XBF_DELWRI_Q;
154 * Once the buffer is marked stale and unlocked, a subsequent lookup
155 * could reset b_flags. There is no guarantee that the buffer is
156 * unaccounted (released to LRU) before that occurs. Drop in-flight
157 * status now to preserve accounting consistency.
159 spin_lock(&bp->b_lock);
160 __xfs_buf_ioacct_dec(bp);
162 atomic_set(&bp->b_lru_ref, 0);
163 if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
164 (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru)))
165 atomic_dec(&bp->b_hold);
167 ASSERT(atomic_read(&bp->b_hold) >= 1);
168 spin_unlock(&bp->b_lock);
176 ASSERT(bp->b_maps == NULL);
177 bp->b_map_count = map_count;
179 if (map_count == 1) {
180 bp->b_maps = &bp->__b_map;
184 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
192 * Frees b_pages if it was allocated.
198 if (bp->b_maps != &bp->__b_map) {
199 kmem_free(bp->b_maps);
206 struct xfs_buftarg *target,
207 struct xfs_buf_map *map,
209 xfs_buf_flags_t flags,
210 struct xfs_buf **bpp)
217 bp = kmem_cache_zalloc(xfs_buf_zone, GFP_NOFS | __GFP_NOFAIL);
220 * We don't want certain flags to appear in b_flags unless they are
221 * specifically set by later operations on the buffer.
223 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
225 atomic_set(&bp->b_hold, 1);
226 atomic_set(&bp->b_lru_ref, 1);
227 init_completion(&bp->b_iowait);
228 INIT_LIST_HEAD(&bp->b_lru);
229 INIT_LIST_HEAD(&bp->b_list);
230 INIT_LIST_HEAD(&bp->b_li_list);
231 sema_init(&bp->b_sema, 0); /* held, no waiters */
232 spin_lock_init(&bp->b_lock);
233 bp->b_target = target;
234 bp->b_mount = target->bt_mount;
238 * Set length and io_length to the same value initially.
239 * I/O routines should use io_length, which will be the same in
240 * most cases but may be reset (e.g. XFS recovery).
242 error = xfs_buf_get_maps(bp, nmaps);
244 kmem_cache_free(xfs_buf_zone, bp);
248 bp->b_bn = map[0].bm_bn;
250 for (i = 0; i < nmaps; i++) {
251 bp->b_maps[i].bm_bn = map[i].bm_bn;
252 bp->b_maps[i].bm_len = map[i].bm_len;
253 bp->b_length += map[i].bm_len;
256 atomic_set(&bp->b_pin_count, 0);
257 init_waitqueue_head(&bp->b_waiters);
259 XFS_STATS_INC(bp->b_mount, xb_create);
260 trace_xfs_buf_init(bp, _RET_IP_);
267 * Allocate a page array capable of holding a specified number
268 * of pages, and point the page buf at it.
275 /* Make sure that we have a page list */
276 if (bp->b_pages == NULL) {
277 bp->b_page_count = page_count;
278 if (page_count <= XB_PAGES) {
279 bp->b_pages = bp->b_page_array;
281 bp->b_pages = kmem_alloc(sizeof(struct page *) *
282 page_count, KM_NOFS);
283 if (bp->b_pages == NULL)
286 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
292 * Frees b_pages if it was allocated.
298 if (bp->b_pages != bp->b_page_array) {
299 kmem_free(bp->b_pages);
305 * Releases the specified buffer.
307 * The modification state of any associated pages is left unchanged.
308 * The buffer must not be on any hash - use xfs_buf_rele instead for
309 * hashed and refcounted buffers
315 trace_xfs_buf_free(bp, _RET_IP_);
317 ASSERT(list_empty(&bp->b_lru));
319 if (bp->b_flags & _XBF_PAGES) {
322 if (xfs_buf_is_vmapped(bp))
323 vm_unmap_ram(bp->b_addr - bp->b_offset,
326 for (i = 0; i < bp->b_page_count; i++) {
327 struct page *page = bp->b_pages[i];
331 if (current->reclaim_state)
332 current->reclaim_state->reclaimed_slab +=
334 } else if (bp->b_flags & _XBF_KMEM)
335 kmem_free(bp->b_addr);
336 _xfs_buf_free_pages(bp);
337 xfs_buf_free_maps(bp);
338 kmem_cache_free(xfs_buf_zone, bp);
342 * Allocates all the pages for buffer in question and builds it's page list.
345 xfs_buf_allocate_memory(
350 size_t nbytes, offset;
351 gfp_t gfp_mask = xb_to_gfp(flags);
352 unsigned short page_count, i;
353 xfs_off_t start, end;
355 xfs_km_flags_t kmflag_mask = 0;
358 * assure zeroed buffer for non-read cases.
360 if (!(flags & XBF_READ)) {
361 kmflag_mask |= KM_ZERO;
362 gfp_mask |= __GFP_ZERO;
366 * for buffers that are contained within a single page, just allocate
367 * the memory from the heap - there's no need for the complexity of
368 * page arrays to keep allocation down to order 0.
370 size = BBTOB(bp->b_length);
371 if (size < PAGE_SIZE) {
372 int align_mask = xfs_buftarg_dma_alignment(bp->b_target);
373 bp->b_addr = kmem_alloc_io(size, align_mask,
374 KM_NOFS | kmflag_mask);
376 /* low memory - use alloc_page loop instead */
380 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
381 ((unsigned long)bp->b_addr & PAGE_MASK)) {
382 /* b_addr spans two pages - use alloc_page instead */
383 kmem_free(bp->b_addr);
387 bp->b_offset = offset_in_page(bp->b_addr);
388 bp->b_pages = bp->b_page_array;
389 bp->b_pages[0] = kmem_to_page(bp->b_addr);
390 bp->b_page_count = 1;
391 bp->b_flags |= _XBF_KMEM;
396 start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT;
397 end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1)
399 page_count = end - start;
400 error = _xfs_buf_get_pages(bp, page_count);
404 offset = bp->b_offset;
405 bp->b_flags |= _XBF_PAGES;
407 for (i = 0; i < bp->b_page_count; i++) {
411 page = alloc_page(gfp_mask);
412 if (unlikely(page == NULL)) {
413 if (flags & XBF_READ_AHEAD) {
414 bp->b_page_count = i;
420 * This could deadlock.
422 * But until all the XFS lowlevel code is revamped to
423 * handle buffer allocation failures we can't do much.
425 if (!(++retries % 100))
427 "%s(%u) possible memory allocation deadlock in %s (mode:0x%x)",
428 current->comm, current->pid,
431 XFS_STATS_INC(bp->b_mount, xb_page_retries);
432 congestion_wait(BLK_RW_ASYNC, HZ/50);
436 XFS_STATS_INC(bp->b_mount, xb_page_found);
438 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
440 bp->b_pages[i] = page;
446 for (i = 0; i < bp->b_page_count; i++)
447 __free_page(bp->b_pages[i]);
448 bp->b_flags &= ~_XBF_PAGES;
453 * Map buffer into kernel address-space if necessary.
460 ASSERT(bp->b_flags & _XBF_PAGES);
461 if (bp->b_page_count == 1) {
462 /* A single page buffer is always mappable */
463 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
464 } else if (flags & XBF_UNMAPPED) {
471 * vm_map_ram() will allocate auxiliary structures (e.g.
472 * pagetables) with GFP_KERNEL, yet we are likely to be under
473 * GFP_NOFS context here. Hence we need to tell memory reclaim
474 * that we are in such a context via PF_MEMALLOC_NOFS to prevent
475 * memory reclaim re-entering the filesystem here and
476 * potentially deadlocking.
478 nofs_flag = memalloc_nofs_save();
480 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
485 } while (retried++ <= 1);
486 memalloc_nofs_restore(nofs_flag);
490 bp->b_addr += bp->b_offset;
497 * Finding and Reading Buffers
501 struct rhashtable_compare_arg *arg,
504 const struct xfs_buf_map *map = arg->key;
505 const struct xfs_buf *bp = obj;
508 * The key hashing in the lookup path depends on the key being the
509 * first element of the compare_arg, make sure to assert this.
511 BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0);
513 if (bp->b_bn != map->bm_bn)
516 if (unlikely(bp->b_length != map->bm_len)) {
518 * found a block number match. If the range doesn't
519 * match, the only way this is allowed is if the buffer
520 * in the cache is stale and the transaction that made
521 * it stale has not yet committed. i.e. we are
522 * reallocating a busy extent. Skip this buffer and
523 * continue searching for an exact match.
525 ASSERT(bp->b_flags & XBF_STALE);
531 static const struct rhashtable_params xfs_buf_hash_params = {
532 .min_size = 32, /* empty AGs have minimal footprint */
534 .key_len = sizeof(xfs_daddr_t),
535 .key_offset = offsetof(struct xfs_buf, b_bn),
536 .head_offset = offsetof(struct xfs_buf, b_rhash_head),
537 .automatic_shrinking = true,
538 .obj_cmpfn = _xfs_buf_obj_cmp,
543 struct xfs_perag *pag)
545 spin_lock_init(&pag->pag_buf_lock);
546 return rhashtable_init(&pag->pag_buf_hash, &xfs_buf_hash_params);
550 xfs_buf_hash_destroy(
551 struct xfs_perag *pag)
553 rhashtable_destroy(&pag->pag_buf_hash);
557 * Look up a buffer in the buffer cache and return it referenced and locked
560 * If @new_bp is supplied and we have a lookup miss, insert @new_bp into the
563 * If XBF_TRYLOCK is set in @flags, only try to lock the buffer and return
564 * -EAGAIN if we fail to lock it.
567 * -EFSCORRUPTED if have been supplied with an invalid address
568 * -EAGAIN on trylock failure
569 * -ENOENT if we fail to find a match and @new_bp was NULL
571 * - @new_bp if we inserted it into the cache
572 * - the buffer we found and locked.
576 struct xfs_buftarg *btp,
577 struct xfs_buf_map *map,
579 xfs_buf_flags_t flags,
580 struct xfs_buf *new_bp,
581 struct xfs_buf **found_bp)
583 struct xfs_perag *pag;
585 struct xfs_buf_map cmap = { .bm_bn = map[0].bm_bn };
591 for (i = 0; i < nmaps; i++)
592 cmap.bm_len += map[i].bm_len;
594 /* Check for IOs smaller than the sector size / not sector aligned */
595 ASSERT(!(BBTOB(cmap.bm_len) < btp->bt_meta_sectorsize));
596 ASSERT(!(BBTOB(cmap.bm_bn) & (xfs_off_t)btp->bt_meta_sectormask));
599 * Corrupted block numbers can get through to here, unfortunately, so we
600 * have to check that the buffer falls within the filesystem bounds.
602 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
603 if (cmap.bm_bn < 0 || cmap.bm_bn >= eofs) {
604 xfs_alert(btp->bt_mount,
605 "%s: daddr 0x%llx out of range, EOFS 0x%llx",
606 __func__, cmap.bm_bn, eofs);
608 return -EFSCORRUPTED;
611 pag = xfs_perag_get(btp->bt_mount,
612 xfs_daddr_to_agno(btp->bt_mount, cmap.bm_bn));
614 spin_lock(&pag->pag_buf_lock);
615 bp = rhashtable_lookup_fast(&pag->pag_buf_hash, &cmap,
616 xfs_buf_hash_params);
618 atomic_inc(&bp->b_hold);
624 XFS_STATS_INC(btp->bt_mount, xb_miss_locked);
625 spin_unlock(&pag->pag_buf_lock);
630 /* the buffer keeps the perag reference until it is freed */
632 rhashtable_insert_fast(&pag->pag_buf_hash, &new_bp->b_rhash_head,
633 xfs_buf_hash_params);
634 spin_unlock(&pag->pag_buf_lock);
639 spin_unlock(&pag->pag_buf_lock);
642 if (!xfs_buf_trylock(bp)) {
643 if (flags & XBF_TRYLOCK) {
645 XFS_STATS_INC(btp->bt_mount, xb_busy_locked);
649 XFS_STATS_INC(btp->bt_mount, xb_get_locked_waited);
653 * if the buffer is stale, clear all the external state associated with
654 * it. We need to keep flags such as how we allocated the buffer memory
657 if (bp->b_flags & XBF_STALE) {
658 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
659 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
663 trace_xfs_buf_find(bp, flags, _RET_IP_);
664 XFS_STATS_INC(btp->bt_mount, xb_get_locked);
671 struct xfs_buftarg *target,
674 xfs_buf_flags_t flags)
678 DEFINE_SINGLE_BUF_MAP(map, blkno, numblks);
680 error = xfs_buf_find(target, &map, 1, flags, NULL, &bp);
687 * Assembles a buffer covering the specified range. The code is optimised for
688 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
689 * more hits than misses.
693 struct xfs_buftarg *target,
694 struct xfs_buf_map *map,
696 xfs_buf_flags_t flags,
697 struct xfs_buf **bpp)
700 struct xfs_buf *new_bp;
704 error = xfs_buf_find(target, map, nmaps, flags, NULL, &bp);
707 if (error != -ENOENT)
710 error = _xfs_buf_alloc(target, map, nmaps, flags, &new_bp);
714 error = xfs_buf_allocate_memory(new_bp, flags);
716 xfs_buf_free(new_bp);
720 error = xfs_buf_find(target, map, nmaps, flags, new_bp, &bp);
722 xfs_buf_free(new_bp);
727 xfs_buf_free(new_bp);
731 error = _xfs_buf_map_pages(bp, flags);
732 if (unlikely(error)) {
733 xfs_warn_ratelimited(target->bt_mount,
734 "%s: failed to map %u pages", __func__,
742 * Clear b_error if this is a lookup from a caller that doesn't expect
743 * valid data to be found in the buffer.
745 if (!(flags & XBF_READ))
746 xfs_buf_ioerror(bp, 0);
748 XFS_STATS_INC(target->bt_mount, xb_get);
749 trace_xfs_buf_get(bp, flags, _RET_IP_);
757 xfs_buf_flags_t flags)
759 ASSERT(!(flags & XBF_WRITE));
760 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
762 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
763 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
765 return xfs_buf_submit(bp);
769 * Reverify a buffer found in cache without an attached ->b_ops.
771 * If the caller passed an ops structure and the buffer doesn't have ops
772 * assigned, set the ops and use it to verify the contents. If verification
773 * fails, clear XBF_DONE. We assume the buffer has no recorded errors and is
774 * already in XBF_DONE state on entry.
776 * Under normal operations, every in-core buffer is verified on read I/O
777 * completion. There are two scenarios that can lead to in-core buffers without
778 * an assigned ->b_ops. The first is during log recovery of buffers on a V4
779 * filesystem, though these buffers are purged at the end of recovery. The
780 * other is online repair, which intentionally reads with a NULL buffer ops to
781 * run several verifiers across an in-core buffer in order to establish buffer
782 * type. If repair can't establish that, the buffer will be left in memory
783 * with NULL buffer ops.
788 const struct xfs_buf_ops *ops)
790 ASSERT(bp->b_flags & XBF_DONE);
791 ASSERT(bp->b_error == 0);
793 if (!ops || bp->b_ops)
797 bp->b_ops->verify_read(bp);
799 bp->b_flags &= ~XBF_DONE;
805 struct xfs_buftarg *target,
806 struct xfs_buf_map *map,
808 xfs_buf_flags_t flags,
809 struct xfs_buf **bpp,
810 const struct xfs_buf_ops *ops,
819 error = xfs_buf_get_map(target, map, nmaps, flags, &bp);
823 trace_xfs_buf_read(bp, flags, _RET_IP_);
825 if (!(bp->b_flags & XBF_DONE)) {
826 /* Initiate the buffer read and wait. */
827 XFS_STATS_INC(target->bt_mount, xb_get_read);
829 error = _xfs_buf_read(bp, flags);
831 /* Readahead iodone already dropped the buffer, so exit. */
832 if (flags & XBF_ASYNC)
835 /* Buffer already read; all we need to do is check it. */
836 error = xfs_buf_reverify(bp, ops);
838 /* Readahead already finished; drop the buffer and exit. */
839 if (flags & XBF_ASYNC) {
844 /* We do not want read in the flags */
845 bp->b_flags &= ~XBF_READ;
846 ASSERT(bp->b_ops != NULL || ops == NULL);
850 * If we've had a read error, then the contents of the buffer are
851 * invalid and should not be used. To ensure that a followup read tries
852 * to pull the buffer from disk again, we clear the XBF_DONE flag and
853 * mark the buffer stale. This ensures that anyone who has a current
854 * reference to the buffer will interpret it's contents correctly and
855 * future cache lookups will also treat it as an empty, uninitialised
859 if (!XFS_FORCED_SHUTDOWN(target->bt_mount))
860 xfs_buf_ioerror_alert(bp, fa);
862 bp->b_flags &= ~XBF_DONE;
866 /* bad CRC means corrupted metadata */
867 if (error == -EFSBADCRC)
868 error = -EFSCORRUPTED;
877 * If we are not low on memory then do the readahead in a deadlock
881 xfs_buf_readahead_map(
882 struct xfs_buftarg *target,
883 struct xfs_buf_map *map,
885 const struct xfs_buf_ops *ops)
889 if (bdi_read_congested(target->bt_bdev->bd_bdi))
892 xfs_buf_read_map(target, map, nmaps,
893 XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD, &bp, ops,
898 * Read an uncached buffer from disk. Allocates and returns a locked
899 * buffer containing the disk contents or nothing.
902 xfs_buf_read_uncached(
903 struct xfs_buftarg *target,
907 struct xfs_buf **bpp,
908 const struct xfs_buf_ops *ops)
915 error = xfs_buf_get_uncached(target, numblks, flags, &bp);
919 /* set up the buffer for a read IO */
920 ASSERT(bp->b_map_count == 1);
921 bp->b_bn = XFS_BUF_DADDR_NULL; /* always null for uncached buffers */
922 bp->b_maps[0].bm_bn = daddr;
923 bp->b_flags |= XBF_READ;
938 xfs_buf_get_uncached(
939 struct xfs_buftarg *target,
942 struct xfs_buf **bpp)
944 unsigned long page_count;
947 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
951 /* flags might contain irrelevant bits, pass only what we care about */
952 error = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT, &bp);
956 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
957 error = _xfs_buf_get_pages(bp, page_count);
961 for (i = 0; i < page_count; i++) {
962 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
963 if (!bp->b_pages[i]) {
968 bp->b_flags |= _XBF_PAGES;
970 error = _xfs_buf_map_pages(bp, 0);
971 if (unlikely(error)) {
972 xfs_warn(target->bt_mount,
973 "%s: failed to map pages", __func__);
977 trace_xfs_buf_get_uncached(bp, _RET_IP_);
983 __free_page(bp->b_pages[i]);
984 _xfs_buf_free_pages(bp);
986 xfs_buf_free_maps(bp);
987 kmem_cache_free(xfs_buf_zone, bp);
993 * Increment reference count on buffer, to hold the buffer concurrently
994 * with another thread which may release (free) the buffer asynchronously.
995 * Must hold the buffer already to call this function.
1001 trace_xfs_buf_hold(bp, _RET_IP_);
1002 atomic_inc(&bp->b_hold);
1006 * Release a hold on the specified buffer. If the hold count is 1, the buffer is
1007 * placed on LRU or freed (depending on b_lru_ref).
1013 struct xfs_perag *pag = bp->b_pag;
1015 bool freebuf = false;
1017 trace_xfs_buf_rele(bp, _RET_IP_);
1020 ASSERT(list_empty(&bp->b_lru));
1021 if (atomic_dec_and_test(&bp->b_hold)) {
1022 xfs_buf_ioacct_dec(bp);
1028 ASSERT(atomic_read(&bp->b_hold) > 0);
1031 * We grab the b_lock here first to serialise racing xfs_buf_rele()
1032 * calls. The pag_buf_lock being taken on the last reference only
1033 * serialises against racing lookups in xfs_buf_find(). IOWs, the second
1034 * to last reference we drop here is not serialised against the last
1035 * reference until we take bp->b_lock. Hence if we don't grab b_lock
1036 * first, the last "release" reference can win the race to the lock and
1037 * free the buffer before the second-to-last reference is processed,
1038 * leading to a use-after-free scenario.
1040 spin_lock(&bp->b_lock);
1041 release = atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock);
1044 * Drop the in-flight state if the buffer is already on the LRU
1045 * and it holds the only reference. This is racy because we
1046 * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT
1047 * ensures the decrement occurs only once per-buf.
1049 if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru))
1050 __xfs_buf_ioacct_dec(bp);
1054 /* the last reference has been dropped ... */
1055 __xfs_buf_ioacct_dec(bp);
1056 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
1058 * If the buffer is added to the LRU take a new reference to the
1059 * buffer for the LRU and clear the (now stale) dispose list
1062 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) {
1063 bp->b_state &= ~XFS_BSTATE_DISPOSE;
1064 atomic_inc(&bp->b_hold);
1066 spin_unlock(&pag->pag_buf_lock);
1069 * most of the time buffers will already be removed from the
1070 * LRU, so optimise that case by checking for the
1071 * XFS_BSTATE_DISPOSE flag indicating the last list the buffer
1072 * was on was the disposal list
1074 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
1075 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru);
1077 ASSERT(list_empty(&bp->b_lru));
1080 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1081 rhashtable_remove_fast(&pag->pag_buf_hash, &bp->b_rhash_head,
1082 xfs_buf_hash_params);
1083 spin_unlock(&pag->pag_buf_lock);
1089 spin_unlock(&bp->b_lock);
1097 * Lock a buffer object, if it is not already locked.
1099 * If we come across a stale, pinned, locked buffer, we know that we are
1100 * being asked to lock a buffer that has been reallocated. Because it is
1101 * pinned, we know that the log has not been pushed to disk and hence it
1102 * will still be locked. Rather than continuing to have trylock attempts
1103 * fail until someone else pushes the log, push it ourselves before
1104 * returning. This means that the xfsaild will not get stuck trying
1105 * to push on stale inode buffers.
1113 locked = down_trylock(&bp->b_sema) == 0;
1115 trace_xfs_buf_trylock(bp, _RET_IP_);
1117 trace_xfs_buf_trylock_fail(bp, _RET_IP_);
1122 * Lock a buffer object.
1124 * If we come across a stale, pinned, locked buffer, we know that we
1125 * are being asked to lock a buffer that has been reallocated. Because
1126 * it is pinned, we know that the log has not been pushed to disk and
1127 * hence it will still be locked. Rather than sleeping until someone
1128 * else pushes the log, push it ourselves before trying to get the lock.
1134 trace_xfs_buf_lock(bp, _RET_IP_);
1136 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
1137 xfs_log_force(bp->b_mount, 0);
1140 trace_xfs_buf_lock_done(bp, _RET_IP_);
1147 ASSERT(xfs_buf_islocked(bp));
1150 trace_xfs_buf_unlock(bp, _RET_IP_);
1157 DECLARE_WAITQUEUE (wait, current);
1159 if (atomic_read(&bp->b_pin_count) == 0)
1162 add_wait_queue(&bp->b_waiters, &wait);
1164 set_current_state(TASK_UNINTERRUPTIBLE);
1165 if (atomic_read(&bp->b_pin_count) == 0)
1169 remove_wait_queue(&bp->b_waiters, &wait);
1170 set_current_state(TASK_RUNNING);
1174 * Buffer Utility Routines
1181 bool read = bp->b_flags & XBF_READ;
1183 trace_xfs_buf_iodone(bp, _RET_IP_);
1185 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1188 * Pull in IO completion errors now. We are guaranteed to be running
1189 * single threaded, so we don't need the lock to read b_io_error.
1191 if (!bp->b_error && bp->b_io_error)
1192 xfs_buf_ioerror(bp, bp->b_io_error);
1195 if (!bp->b_error && bp->b_ops)
1196 bp->b_ops->verify_read(bp);
1198 bp->b_flags |= XBF_DONE;
1199 xfs_buf_ioend_finish(bp);
1204 bp->b_flags &= ~XBF_WRITE_FAIL;
1205 bp->b_flags |= XBF_DONE;
1209 * If this is a log recovery buffer, we aren't doing transactional IO
1210 * yet so we need to let it handle IO completions.
1212 if (bp->b_flags & _XBF_LOGRECOVERY) {
1213 xlog_recover_iodone(bp);
1217 if (bp->b_flags & _XBF_INODES) {
1218 xfs_buf_inode_iodone(bp);
1222 if (bp->b_flags & _XBF_DQUOTS) {
1223 xfs_buf_dquot_iodone(bp);
1231 struct work_struct *work)
1233 struct xfs_buf *bp =
1234 container_of(work, xfs_buf_t, b_ioend_work);
1240 xfs_buf_ioend_async(
1243 INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
1244 queue_work(bp->b_mount->m_buf_workqueue, &bp->b_ioend_work);
1251 xfs_failaddr_t failaddr)
1253 ASSERT(error <= 0 && error >= -1000);
1254 bp->b_error = error;
1255 trace_xfs_buf_ioerror(bp, error, failaddr);
1259 xfs_buf_ioerror_alert(
1261 xfs_failaddr_t func)
1263 xfs_buf_alert_ratelimited(bp, "XFS: metadata IO error",
1264 "metadata I/O error in \"%pS\" at daddr 0x%llx len %d error %d",
1265 func, (uint64_t)XFS_BUF_ADDR(bp),
1266 bp->b_length, -bp->b_error);
1270 * To simulate an I/O failure, the buffer must be locked and held with at least
1271 * three references. The LRU reference is dropped by the stale call. The buf
1272 * item reference is dropped via ioend processing. The third reference is owned
1273 * by the caller and is dropped on I/O completion if the buffer is XBF_ASYNC.
1279 bp->b_flags &= ~XBF_DONE;
1281 xfs_buf_ioerror(bp, -EIO);
1291 ASSERT(xfs_buf_islocked(bp));
1293 bp->b_flags |= XBF_WRITE;
1294 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1297 error = xfs_buf_submit(bp);
1299 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
1307 struct xfs_buf *bp = (struct xfs_buf *)bio->bi_private;
1309 if (!bio->bi_status &&
1310 (bp->b_flags & XBF_WRITE) && (bp->b_flags & XBF_ASYNC) &&
1311 XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_IOERROR))
1312 bio->bi_status = BLK_STS_IOERR;
1315 * don't overwrite existing errors - otherwise we can lose errors on
1316 * buffers that require multiple bios to complete.
1318 if (bio->bi_status) {
1319 int error = blk_status_to_errno(bio->bi_status);
1321 cmpxchg(&bp->b_io_error, 0, error);
1324 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1325 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1327 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1328 xfs_buf_ioend_async(bp);
1333 xfs_buf_ioapply_map(
1341 int total_nr_pages = bp->b_page_count;
1344 sector_t sector = bp->b_maps[map].bm_bn;
1348 /* skip the pages in the buffer before the start offset */
1350 offset = *buf_offset;
1351 while (offset >= PAGE_SIZE) {
1353 offset -= PAGE_SIZE;
1357 * Limit the IO size to the length of the current vector, and update the
1358 * remaining IO count for the next time around.
1360 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1362 *buf_offset += size;
1365 atomic_inc(&bp->b_io_remaining);
1366 nr_pages = min(total_nr_pages, BIO_MAX_PAGES);
1368 bio = bio_alloc(GFP_NOIO, nr_pages);
1369 bio_set_dev(bio, bp->b_target->bt_bdev);
1370 bio->bi_iter.bi_sector = sector;
1371 bio->bi_end_io = xfs_buf_bio_end_io;
1372 bio->bi_private = bp;
1375 for (; size && nr_pages; nr_pages--, page_index++) {
1376 int rbytes, nbytes = PAGE_SIZE - offset;
1381 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1383 if (rbytes < nbytes)
1387 sector += BTOBB(nbytes);
1392 if (likely(bio->bi_iter.bi_size)) {
1393 if (xfs_buf_is_vmapped(bp)) {
1394 flush_kernel_vmap_range(bp->b_addr,
1395 xfs_buf_vmap_len(bp));
1402 * This is guaranteed not to be the last io reference count
1403 * because the caller (xfs_buf_submit) holds a count itself.
1405 atomic_dec(&bp->b_io_remaining);
1406 xfs_buf_ioerror(bp, -EIO);
1416 struct blk_plug plug;
1423 * Make sure we capture only current IO errors rather than stale errors
1424 * left over from previous use of the buffer (e.g. failed readahead).
1428 if (bp->b_flags & XBF_WRITE) {
1432 * Run the write verifier callback function if it exists. If
1433 * this function fails it will mark the buffer with an error and
1434 * the IO should not be dispatched.
1437 bp->b_ops->verify_write(bp);
1439 xfs_force_shutdown(bp->b_mount,
1440 SHUTDOWN_CORRUPT_INCORE);
1443 } else if (bp->b_bn != XFS_BUF_DADDR_NULL) {
1444 struct xfs_mount *mp = bp->b_mount;
1447 * non-crc filesystems don't attach verifiers during
1448 * log recovery, so don't warn for such filesystems.
1450 if (xfs_sb_version_hascrc(&mp->m_sb)) {
1452 "%s: no buf ops on daddr 0x%llx len %d",
1453 __func__, bp->b_bn, bp->b_length);
1454 xfs_hex_dump(bp->b_addr,
1455 XFS_CORRUPTION_DUMP_LEN);
1461 if (bp->b_flags & XBF_READ_AHEAD)
1465 /* we only use the buffer cache for meta-data */
1469 * Walk all the vectors issuing IO on them. Set up the initial offset
1470 * into the buffer and the desired IO size before we start -
1471 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1474 offset = bp->b_offset;
1475 size = BBTOB(bp->b_length);
1476 blk_start_plug(&plug);
1477 for (i = 0; i < bp->b_map_count; i++) {
1478 xfs_buf_ioapply_map(bp, i, &offset, &size, op);
1482 break; /* all done */
1484 blk_finish_plug(&plug);
1488 * Wait for I/O completion of a sync buffer and return the I/O error code.
1494 ASSERT(!(bp->b_flags & XBF_ASYNC));
1496 trace_xfs_buf_iowait(bp, _RET_IP_);
1497 wait_for_completion(&bp->b_iowait);
1498 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1504 * Buffer I/O submission path, read or write. Asynchronous submission transfers
1505 * the buffer lock ownership and the current reference to the IO. It is not
1506 * safe to reference the buffer after a call to this function unless the caller
1507 * holds an additional reference itself.
1516 trace_xfs_buf_submit(bp, _RET_IP_);
1518 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1520 /* on shutdown we stale and complete the buffer immediately */
1521 if (XFS_FORCED_SHUTDOWN(bp->b_mount)) {
1522 xfs_buf_ioend_fail(bp);
1527 * Grab a reference so the buffer does not go away underneath us. For
1528 * async buffers, I/O completion drops the callers reference, which
1529 * could occur before submission returns.
1533 if (bp->b_flags & XBF_WRITE)
1534 xfs_buf_wait_unpin(bp);
1536 /* clear the internal error state to avoid spurious errors */
1540 * Set the count to 1 initially, this will stop an I/O completion
1541 * callout which happens before we have started all the I/O from calling
1542 * xfs_buf_ioend too early.
1544 atomic_set(&bp->b_io_remaining, 1);
1545 if (bp->b_flags & XBF_ASYNC)
1546 xfs_buf_ioacct_inc(bp);
1547 _xfs_buf_ioapply(bp);
1550 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1551 * reference we took above. If we drop it to zero, run completion so
1552 * that we don't return to the caller with completion still pending.
1554 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1555 if (bp->b_error || !(bp->b_flags & XBF_ASYNC))
1558 xfs_buf_ioend_async(bp);
1562 error = xfs_buf_iowait(bp);
1565 * Release the hold that keeps the buffer referenced for the entire
1566 * I/O. Note that if the buffer is async, it is not safe to reference
1567 * after this release.
1581 return bp->b_addr + offset;
1583 offset += bp->b_offset;
1584 page = bp->b_pages[offset >> PAGE_SHIFT];
1585 return page_address(page) + (offset & (PAGE_SIZE-1));
1596 bend = boff + bsize;
1597 while (boff < bend) {
1599 int page_index, page_offset, csize;
1601 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1602 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1603 page = bp->b_pages[page_index];
1604 csize = min_t(size_t, PAGE_SIZE - page_offset,
1605 BBTOB(bp->b_length) - boff);
1607 ASSERT((csize + page_offset) <= PAGE_SIZE);
1609 memset(page_address(page) + page_offset, 0, csize);
1616 * Log a message about and stale a buffer that a caller has decided is corrupt.
1618 * This function should be called for the kinds of metadata corruption that
1619 * cannot be detect from a verifier, such as incorrect inter-block relationship
1620 * data. Do /not/ call this function from a verifier function.
1622 * The buffer must be XBF_DONE prior to the call. Afterwards, the buffer will
1623 * be marked stale, but b_error will not be set. The caller is responsible for
1624 * releasing the buffer or fixing it.
1627 __xfs_buf_mark_corrupt(
1631 ASSERT(bp->b_flags & XBF_DONE);
1633 xfs_buf_corruption_error(bp, fa);
1638 * Handling of buffer targets (buftargs).
1642 * Wait for any bufs with callbacks that have been submitted but have not yet
1643 * returned. These buffers will have an elevated hold count, so wait on those
1644 * while freeing all the buffers only held by the LRU.
1646 static enum lru_status
1647 xfs_buftarg_wait_rele(
1648 struct list_head *item,
1649 struct list_lru_one *lru,
1650 spinlock_t *lru_lock,
1654 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1655 struct list_head *dispose = arg;
1657 if (atomic_read(&bp->b_hold) > 1) {
1658 /* need to wait, so skip it this pass */
1659 trace_xfs_buf_wait_buftarg(bp, _RET_IP_);
1662 if (!spin_trylock(&bp->b_lock))
1666 * clear the LRU reference count so the buffer doesn't get
1667 * ignored in xfs_buf_rele().
1669 atomic_set(&bp->b_lru_ref, 0);
1670 bp->b_state |= XFS_BSTATE_DISPOSE;
1671 list_lru_isolate_move(lru, item, dispose);
1672 spin_unlock(&bp->b_lock);
1678 struct xfs_buftarg *btp)
1682 bool write_fail = false;
1685 * First wait on the buftarg I/O count for all in-flight buffers to be
1686 * released. This is critical as new buffers do not make the LRU until
1687 * they are released.
1689 * Next, flush the buffer workqueue to ensure all completion processing
1690 * has finished. Just waiting on buffer locks is not sufficient for
1691 * async IO as the reference count held over IO is not released until
1692 * after the buffer lock is dropped. Hence we need to ensure here that
1693 * all reference counts have been dropped before we start walking the
1696 while (percpu_counter_sum(&btp->bt_io_count))
1698 flush_workqueue(btp->bt_mount->m_buf_workqueue);
1700 /* loop until there is nothing left on the lru list. */
1701 while (list_lru_count(&btp->bt_lru)) {
1702 list_lru_walk(&btp->bt_lru, xfs_buftarg_wait_rele,
1703 &dispose, LONG_MAX);
1705 while (!list_empty(&dispose)) {
1707 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1708 list_del_init(&bp->b_lru);
1709 if (bp->b_flags & XBF_WRITE_FAIL) {
1711 xfs_buf_alert_ratelimited(bp,
1712 "XFS: Corruption Alert",
1713 "Corruption Alert: Buffer at daddr 0x%llx had permanent write failures!",
1714 (long long)bp->b_bn);
1723 * If one or more failed buffers were freed, that means dirty metadata
1724 * was thrown away. This should only ever happen after I/O completion
1725 * handling has elevated I/O error(s) to permanent failures and shuts
1729 ASSERT(XFS_FORCED_SHUTDOWN(btp->bt_mount));
1730 xfs_alert(btp->bt_mount,
1731 "Please run xfs_repair to determine the extent of the problem.");
1735 static enum lru_status
1736 xfs_buftarg_isolate(
1737 struct list_head *item,
1738 struct list_lru_one *lru,
1739 spinlock_t *lru_lock,
1742 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1743 struct list_head *dispose = arg;
1746 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1747 * If we fail to get the lock, just skip it.
1749 if (!spin_trylock(&bp->b_lock))
1752 * Decrement the b_lru_ref count unless the value is already
1753 * zero. If the value is already zero, we need to reclaim the
1754 * buffer, otherwise it gets another trip through the LRU.
1756 if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1757 spin_unlock(&bp->b_lock);
1761 bp->b_state |= XFS_BSTATE_DISPOSE;
1762 list_lru_isolate_move(lru, item, dispose);
1763 spin_unlock(&bp->b_lock);
1767 static unsigned long
1768 xfs_buftarg_shrink_scan(
1769 struct shrinker *shrink,
1770 struct shrink_control *sc)
1772 struct xfs_buftarg *btp = container_of(shrink,
1773 struct xfs_buftarg, bt_shrinker);
1775 unsigned long freed;
1777 freed = list_lru_shrink_walk(&btp->bt_lru, sc,
1778 xfs_buftarg_isolate, &dispose);
1780 while (!list_empty(&dispose)) {
1782 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1783 list_del_init(&bp->b_lru);
1790 static unsigned long
1791 xfs_buftarg_shrink_count(
1792 struct shrinker *shrink,
1793 struct shrink_control *sc)
1795 struct xfs_buftarg *btp = container_of(shrink,
1796 struct xfs_buftarg, bt_shrinker);
1797 return list_lru_shrink_count(&btp->bt_lru, sc);
1802 struct xfs_buftarg *btp)
1804 unregister_shrinker(&btp->bt_shrinker);
1805 ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
1806 percpu_counter_destroy(&btp->bt_io_count);
1807 list_lru_destroy(&btp->bt_lru);
1809 xfs_blkdev_issue_flush(btp);
1815 xfs_setsize_buftarg(
1817 unsigned int sectorsize)
1819 /* Set up metadata sector size info */
1820 btp->bt_meta_sectorsize = sectorsize;
1821 btp->bt_meta_sectormask = sectorsize - 1;
1823 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1824 xfs_warn(btp->bt_mount,
1825 "Cannot set_blocksize to %u on device %pg",
1826 sectorsize, btp->bt_bdev);
1830 /* Set up device logical sector size mask */
1831 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1832 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1838 * When allocating the initial buffer target we have not yet
1839 * read in the superblock, so don't know what sized sectors
1840 * are being used at this early stage. Play safe.
1843 xfs_setsize_buftarg_early(
1845 struct block_device *bdev)
1847 return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev));
1852 struct xfs_mount *mp,
1853 struct block_device *bdev,
1854 struct dax_device *dax_dev)
1858 btp = kmem_zalloc(sizeof(*btp), KM_NOFS);
1861 btp->bt_dev = bdev->bd_dev;
1862 btp->bt_bdev = bdev;
1863 btp->bt_daxdev = dax_dev;
1866 * Buffer IO error rate limiting. Limit it to no more than 10 messages
1867 * per 30 seconds so as to not spam logs too much on repeated errors.
1869 ratelimit_state_init(&btp->bt_ioerror_rl, 30 * HZ,
1870 DEFAULT_RATELIMIT_BURST);
1872 if (xfs_setsize_buftarg_early(btp, bdev))
1875 if (list_lru_init(&btp->bt_lru))
1878 if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL))
1881 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
1882 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
1883 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1884 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
1885 if (register_shrinker(&btp->bt_shrinker))
1890 percpu_counter_destroy(&btp->bt_io_count);
1892 list_lru_destroy(&btp->bt_lru);
1899 * Cancel a delayed write list.
1901 * Remove each buffer from the list, clear the delwri queue flag and drop the
1902 * associated buffer reference.
1905 xfs_buf_delwri_cancel(
1906 struct list_head *list)
1910 while (!list_empty(list)) {
1911 bp = list_first_entry(list, struct xfs_buf, b_list);
1914 bp->b_flags &= ~_XBF_DELWRI_Q;
1915 list_del_init(&bp->b_list);
1921 * Add a buffer to the delayed write list.
1923 * This queues a buffer for writeout if it hasn't already been. Note that
1924 * neither this routine nor the buffer list submission functions perform
1925 * any internal synchronization. It is expected that the lists are thread-local
1928 * Returns true if we queued up the buffer, or false if it already had
1929 * been on the buffer list.
1932 xfs_buf_delwri_queue(
1934 struct list_head *list)
1936 ASSERT(xfs_buf_islocked(bp));
1937 ASSERT(!(bp->b_flags & XBF_READ));
1940 * If the buffer is already marked delwri it already is queued up
1941 * by someone else for imediate writeout. Just ignore it in that
1944 if (bp->b_flags & _XBF_DELWRI_Q) {
1945 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1949 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1952 * If a buffer gets written out synchronously or marked stale while it
1953 * is on a delwri list we lazily remove it. To do this, the other party
1954 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1955 * It remains referenced and on the list. In a rare corner case it
1956 * might get readded to a delwri list after the synchronous writeout, in
1957 * which case we need just need to re-add the flag here.
1959 bp->b_flags |= _XBF_DELWRI_Q;
1960 if (list_empty(&bp->b_list)) {
1961 atomic_inc(&bp->b_hold);
1962 list_add_tail(&bp->b_list, list);
1969 * Compare function is more complex than it needs to be because
1970 * the return value is only 32 bits and we are doing comparisons
1976 struct list_head *a,
1977 struct list_head *b)
1979 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1980 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1983 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
1992 * Submit buffers for write. If wait_list is specified, the buffers are
1993 * submitted using sync I/O and placed on the wait list such that the caller can
1994 * iowait each buffer. Otherwise async I/O is used and the buffers are released
1995 * at I/O completion time. In either case, buffers remain locked until I/O
1996 * completes and the buffer is released from the queue.
1999 xfs_buf_delwri_submit_buffers(
2000 struct list_head *buffer_list,
2001 struct list_head *wait_list)
2003 struct xfs_buf *bp, *n;
2005 struct blk_plug plug;
2007 list_sort(NULL, buffer_list, xfs_buf_cmp);
2009 blk_start_plug(&plug);
2010 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
2012 if (xfs_buf_ispinned(bp)) {
2016 if (!xfs_buf_trylock(bp))
2023 * Someone else might have written the buffer synchronously or
2024 * marked it stale in the meantime. In that case only the
2025 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
2026 * reference and remove it from the list here.
2028 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
2029 list_del_init(&bp->b_list);
2034 trace_xfs_buf_delwri_split(bp, _RET_IP_);
2037 * If we have a wait list, each buffer (and associated delwri
2038 * queue reference) transfers to it and is submitted
2039 * synchronously. Otherwise, drop the buffer from the delwri
2040 * queue and submit async.
2042 bp->b_flags &= ~_XBF_DELWRI_Q;
2043 bp->b_flags |= XBF_WRITE;
2045 bp->b_flags &= ~XBF_ASYNC;
2046 list_move_tail(&bp->b_list, wait_list);
2048 bp->b_flags |= XBF_ASYNC;
2049 list_del_init(&bp->b_list);
2051 __xfs_buf_submit(bp, false);
2053 blk_finish_plug(&plug);
2059 * Write out a buffer list asynchronously.
2061 * This will take the @buffer_list, write all non-locked and non-pinned buffers
2062 * out and not wait for I/O completion on any of the buffers. This interface
2063 * is only safely useable for callers that can track I/O completion by higher
2064 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
2067 * Note: this function will skip buffers it would block on, and in doing so
2068 * leaves them on @buffer_list so they can be retried on a later pass. As such,
2069 * it is up to the caller to ensure that the buffer list is fully submitted or
2070 * cancelled appropriately when they are finished with the list. Failure to
2071 * cancel or resubmit the list until it is empty will result in leaked buffers
2075 xfs_buf_delwri_submit_nowait(
2076 struct list_head *buffer_list)
2078 return xfs_buf_delwri_submit_buffers(buffer_list, NULL);
2082 * Write out a buffer list synchronously.
2084 * This will take the @buffer_list, write all buffers out and wait for I/O
2085 * completion on all of the buffers. @buffer_list is consumed by the function,
2086 * so callers must have some other way of tracking buffers if they require such
2090 xfs_buf_delwri_submit(
2091 struct list_head *buffer_list)
2093 LIST_HEAD (wait_list);
2094 int error = 0, error2;
2097 xfs_buf_delwri_submit_buffers(buffer_list, &wait_list);
2099 /* Wait for IO to complete. */
2100 while (!list_empty(&wait_list)) {
2101 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
2103 list_del_init(&bp->b_list);
2106 * Wait on the locked buffer, check for errors and unlock and
2107 * release the delwri queue reference.
2109 error2 = xfs_buf_iowait(bp);
2119 * Push a single buffer on a delwri queue.
2121 * The purpose of this function is to submit a single buffer of a delwri queue
2122 * and return with the buffer still on the original queue. The waiting delwri
2123 * buffer submission infrastructure guarantees transfer of the delwri queue
2124 * buffer reference to a temporary wait list. We reuse this infrastructure to
2125 * transfer the buffer back to the original queue.
2127 * Note the buffer transitions from the queued state, to the submitted and wait
2128 * listed state and back to the queued state during this call. The buffer
2129 * locking and queue management logic between _delwri_pushbuf() and
2130 * _delwri_queue() guarantee that the buffer cannot be queued to another list
2134 xfs_buf_delwri_pushbuf(
2136 struct list_head *buffer_list)
2138 LIST_HEAD (submit_list);
2141 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
2143 trace_xfs_buf_delwri_pushbuf(bp, _RET_IP_);
2146 * Isolate the buffer to a new local list so we can submit it for I/O
2147 * independently from the rest of the original list.
2150 list_move(&bp->b_list, &submit_list);
2154 * Delwri submission clears the DELWRI_Q buffer flag and returns with
2155 * the buffer on the wait list with the original reference. Rather than
2156 * bounce the buffer from a local wait list back to the original list
2157 * after I/O completion, reuse the original list as the wait list.
2159 xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
2162 * The buffer is now locked, under I/O and wait listed on the original
2163 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
2164 * return with the buffer unlocked and on the original queue.
2166 error = xfs_buf_iowait(bp);
2167 bp->b_flags |= _XBF_DELWRI_Q;
2176 xfs_buf_zone = kmem_cache_create("xfs_buf", sizeof(struct xfs_buf), 0,
2177 SLAB_HWCACHE_ALIGN |
2178 SLAB_RECLAIM_ACCOUNT |
2191 xfs_buf_terminate(void)
2193 kmem_cache_destroy(xfs_buf_zone);
2196 void xfs_buf_set_ref(struct xfs_buf *bp, int lru_ref)
2199 * Set the lru reference count to 0 based on the error injection tag.
2200 * This allows userspace to disrupt buffer caching for debug/testing
2203 if (XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_LRU_REF))
2206 atomic_set(&bp->b_lru_ref, lru_ref);
2210 * Verify an on-disk magic value against the magic value specified in the
2211 * verifier structure. The verifier magic is in disk byte order so the caller is
2212 * expected to pass the value directly from disk.
2219 struct xfs_mount *mp = bp->b_mount;
2222 idx = xfs_sb_version_hascrc(&mp->m_sb);
2223 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic[idx]))
2225 return dmagic == bp->b_ops->magic[idx];
2228 * Verify an on-disk magic value against the magic value specified in the
2229 * verifier structure. The verifier magic is in disk byte order so the caller is
2230 * expected to pass the value directly from disk.
2237 struct xfs_mount *mp = bp->b_mount;
2240 idx = xfs_sb_version_hascrc(&mp->m_sb);
2241 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic16[idx]))
2243 return dmagic == bp->b_ops->magic16[idx];