2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
19 #include <linux/stddef.h>
20 #include <linux/errno.h>
21 #include <linux/gfp.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/vmalloc.h>
25 #include <linux/bio.h>
26 #include <linux/sysctl.h>
27 #include <linux/proc_fs.h>
28 #include <linux/workqueue.h>
29 #include <linux/percpu.h>
30 #include <linux/blkdev.h>
31 #include <linux/hash.h>
32 #include <linux/kthread.h>
33 #include <linux/migrate.h>
34 #include <linux/backing-dev.h>
35 #include <linux/freezer.h>
37 #include "xfs_log_format.h"
38 #include "xfs_trans_resv.h"
41 #include "xfs_mount.h"
42 #include "xfs_trace.h"
45 static kmem_zone_t *xfs_buf_zone;
47 static struct workqueue_struct *xfslogd_workqueue;
49 #ifdef XFS_BUF_LOCK_TRACKING
50 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
51 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
52 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
54 # define XB_SET_OWNER(bp) do { } while (0)
55 # define XB_CLEAR_OWNER(bp) do { } while (0)
56 # define XB_GET_OWNER(bp) do { } while (0)
59 #define xb_to_gfp(flags) \
60 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN)
68 * Return true if the buffer is vmapped.
70 * b_addr is null if the buffer is not mapped, but the code is clever
71 * enough to know it doesn't have to map a single page, so the check has
72 * to be both for b_addr and bp->b_page_count > 1.
74 return bp->b_addr && bp->b_page_count > 1;
81 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
85 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
86 * b_lru_ref count so that the buffer is freed immediately when the buffer
87 * reference count falls to zero. If the buffer is already on the LRU, we need
88 * to remove the reference that LRU holds on the buffer.
90 * This prevents build-up of stale buffers on the LRU.
96 ASSERT(xfs_buf_islocked(bp));
98 bp->b_flags |= XBF_STALE;
101 * Clear the delwri status so that a delwri queue walker will not
102 * flush this buffer to disk now that it is stale. The delwri queue has
103 * a reference to the buffer, so this is safe to do.
105 bp->b_flags &= ~_XBF_DELWRI_Q;
107 spin_lock(&bp->b_lock);
108 atomic_set(&bp->b_lru_ref, 0);
109 if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
110 (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru)))
111 atomic_dec(&bp->b_hold);
113 ASSERT(atomic_read(&bp->b_hold) >= 1);
114 spin_unlock(&bp->b_lock);
122 ASSERT(bp->b_maps == NULL);
123 bp->b_map_count = map_count;
125 if (map_count == 1) {
126 bp->b_maps = &bp->__b_map;
130 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
138 * Frees b_pages if it was allocated.
144 if (bp->b_maps != &bp->__b_map) {
145 kmem_free(bp->b_maps);
152 struct xfs_buftarg *target,
153 struct xfs_buf_map *map,
155 xfs_buf_flags_t flags)
161 bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS);
166 * We don't want certain flags to appear in b_flags unless they are
167 * specifically set by later operations on the buffer.
169 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
171 atomic_set(&bp->b_hold, 1);
172 atomic_set(&bp->b_lru_ref, 1);
173 init_completion(&bp->b_iowait);
174 INIT_LIST_HEAD(&bp->b_lru);
175 INIT_LIST_HEAD(&bp->b_list);
176 RB_CLEAR_NODE(&bp->b_rbnode);
177 sema_init(&bp->b_sema, 0); /* held, no waiters */
178 spin_lock_init(&bp->b_lock);
180 bp->b_target = target;
184 * Set length and io_length to the same value initially.
185 * I/O routines should use io_length, which will be the same in
186 * most cases but may be reset (e.g. XFS recovery).
188 error = xfs_buf_get_maps(bp, nmaps);
190 kmem_zone_free(xfs_buf_zone, bp);
194 bp->b_bn = map[0].bm_bn;
196 for (i = 0; i < nmaps; i++) {
197 bp->b_maps[i].bm_bn = map[i].bm_bn;
198 bp->b_maps[i].bm_len = map[i].bm_len;
199 bp->b_length += map[i].bm_len;
201 bp->b_io_length = bp->b_length;
203 atomic_set(&bp->b_pin_count, 0);
204 init_waitqueue_head(&bp->b_waiters);
206 XFS_STATS_INC(xb_create);
207 trace_xfs_buf_init(bp, _RET_IP_);
213 * Allocate a page array capable of holding a specified number
214 * of pages, and point the page buf at it.
220 xfs_buf_flags_t flags)
222 /* Make sure that we have a page list */
223 if (bp->b_pages == NULL) {
224 bp->b_page_count = page_count;
225 if (page_count <= XB_PAGES) {
226 bp->b_pages = bp->b_page_array;
228 bp->b_pages = kmem_alloc(sizeof(struct page *) *
229 page_count, KM_NOFS);
230 if (bp->b_pages == NULL)
233 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
239 * Frees b_pages if it was allocated.
245 if (bp->b_pages != bp->b_page_array) {
246 kmem_free(bp->b_pages);
252 * Releases the specified buffer.
254 * The modification state of any associated pages is left unchanged.
255 * The buffer must not be on any hash - use xfs_buf_rele instead for
256 * hashed and refcounted buffers
262 trace_xfs_buf_free(bp, _RET_IP_);
264 ASSERT(list_empty(&bp->b_lru));
266 if (bp->b_flags & _XBF_PAGES) {
269 if (xfs_buf_is_vmapped(bp))
270 vm_unmap_ram(bp->b_addr - bp->b_offset,
273 for (i = 0; i < bp->b_page_count; i++) {
274 struct page *page = bp->b_pages[i];
278 } else if (bp->b_flags & _XBF_KMEM)
279 kmem_free(bp->b_addr);
280 _xfs_buf_free_pages(bp);
281 xfs_buf_free_maps(bp);
282 kmem_zone_free(xfs_buf_zone, bp);
286 * Allocates all the pages for buffer in question and builds it's page list.
289 xfs_buf_allocate_memory(
294 size_t nbytes, offset;
295 gfp_t gfp_mask = xb_to_gfp(flags);
296 unsigned short page_count, i;
297 xfs_off_t start, end;
301 * for buffers that are contained within a single page, just allocate
302 * the memory from the heap - there's no need for the complexity of
303 * page arrays to keep allocation down to order 0.
305 size = BBTOB(bp->b_length);
306 if (size < PAGE_SIZE) {
307 bp->b_addr = kmem_alloc(size, KM_NOFS);
309 /* low memory - use alloc_page loop instead */
313 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
314 ((unsigned long)bp->b_addr & PAGE_MASK)) {
315 /* b_addr spans two pages - use alloc_page instead */
316 kmem_free(bp->b_addr);
320 bp->b_offset = offset_in_page(bp->b_addr);
321 bp->b_pages = bp->b_page_array;
322 bp->b_pages[0] = virt_to_page(bp->b_addr);
323 bp->b_page_count = 1;
324 bp->b_flags |= _XBF_KMEM;
329 start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT;
330 end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1)
332 page_count = end - start;
333 error = _xfs_buf_get_pages(bp, page_count, flags);
337 offset = bp->b_offset;
338 bp->b_flags |= _XBF_PAGES;
340 for (i = 0; i < bp->b_page_count; i++) {
344 page = alloc_page(gfp_mask);
345 if (unlikely(page == NULL)) {
346 if (flags & XBF_READ_AHEAD) {
347 bp->b_page_count = i;
353 * This could deadlock.
355 * But until all the XFS lowlevel code is revamped to
356 * handle buffer allocation failures we can't do much.
358 if (!(++retries % 100))
360 "possible memory allocation deadlock in %s (mode:0x%x)",
363 XFS_STATS_INC(xb_page_retries);
364 congestion_wait(BLK_RW_ASYNC, HZ/50);
368 XFS_STATS_INC(xb_page_found);
370 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
372 bp->b_pages[i] = page;
378 for (i = 0; i < bp->b_page_count; i++)
379 __free_page(bp->b_pages[i]);
384 * Map buffer into kernel address-space if necessary.
391 ASSERT(bp->b_flags & _XBF_PAGES);
392 if (bp->b_page_count == 1) {
393 /* A single page buffer is always mappable */
394 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
395 } else if (flags & XBF_UNMAPPED) {
402 * vm_map_ram() will allocate auxillary structures (e.g.
403 * pagetables) with GFP_KERNEL, yet we are likely to be under
404 * GFP_NOFS context here. Hence we need to tell memory reclaim
405 * that we are in such a context via PF_MEMALLOC_NOIO to prevent
406 * memory reclaim re-entering the filesystem here and
407 * potentially deadlocking.
409 noio_flag = memalloc_noio_save();
411 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
416 } while (retried++ <= 1);
417 memalloc_noio_restore(noio_flag);
421 bp->b_addr += bp->b_offset;
428 * Finding and Reading Buffers
432 * Look up, and creates if absent, a lockable buffer for
433 * a given range of an inode. The buffer is returned
434 * locked. No I/O is implied by this call.
438 struct xfs_buftarg *btp,
439 struct xfs_buf_map *map,
441 xfs_buf_flags_t flags,
445 struct xfs_perag *pag;
446 struct rb_node **rbp;
447 struct rb_node *parent;
449 xfs_daddr_t blkno = map[0].bm_bn;
454 for (i = 0; i < nmaps; i++)
455 numblks += map[i].bm_len;
456 numbytes = BBTOB(numblks);
458 /* Check for IOs smaller than the sector size / not sector aligned */
459 ASSERT(!(numbytes < btp->bt_meta_sectorsize));
460 ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_meta_sectormask));
463 * Corrupted block numbers can get through to here, unfortunately, so we
464 * have to check that the buffer falls within the filesystem bounds.
466 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
469 * XXX (dgc): we should really be returning EFSCORRUPTED here,
470 * but none of the higher level infrastructure supports
471 * returning a specific error on buffer lookup failures.
473 xfs_alert(btp->bt_mount,
474 "%s: Block out of range: block 0x%llx, EOFS 0x%llx ",
475 __func__, blkno, eofs);
481 pag = xfs_perag_get(btp->bt_mount,
482 xfs_daddr_to_agno(btp->bt_mount, blkno));
485 spin_lock(&pag->pag_buf_lock);
486 rbp = &pag->pag_buf_tree.rb_node;
491 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
493 if (blkno < bp->b_bn)
494 rbp = &(*rbp)->rb_left;
495 else if (blkno > bp->b_bn)
496 rbp = &(*rbp)->rb_right;
499 * found a block number match. If the range doesn't
500 * match, the only way this is allowed is if the buffer
501 * in the cache is stale and the transaction that made
502 * it stale has not yet committed. i.e. we are
503 * reallocating a busy extent. Skip this buffer and
504 * continue searching to the right for an exact match.
506 if (bp->b_length != numblks) {
507 ASSERT(bp->b_flags & XBF_STALE);
508 rbp = &(*rbp)->rb_right;
511 atomic_inc(&bp->b_hold);
518 rb_link_node(&new_bp->b_rbnode, parent, rbp);
519 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
520 /* the buffer keeps the perag reference until it is freed */
522 spin_unlock(&pag->pag_buf_lock);
524 XFS_STATS_INC(xb_miss_locked);
525 spin_unlock(&pag->pag_buf_lock);
531 spin_unlock(&pag->pag_buf_lock);
534 if (!xfs_buf_trylock(bp)) {
535 if (flags & XBF_TRYLOCK) {
537 XFS_STATS_INC(xb_busy_locked);
541 XFS_STATS_INC(xb_get_locked_waited);
545 * if the buffer is stale, clear all the external state associated with
546 * it. We need to keep flags such as how we allocated the buffer memory
549 if (bp->b_flags & XBF_STALE) {
550 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
551 ASSERT(bp->b_iodone == NULL);
552 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
556 trace_xfs_buf_find(bp, flags, _RET_IP_);
557 XFS_STATS_INC(xb_get_locked);
562 * Assembles a buffer covering the specified range. The code is optimised for
563 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
564 * more hits than misses.
568 struct xfs_buftarg *target,
569 struct xfs_buf_map *map,
571 xfs_buf_flags_t flags)
574 struct xfs_buf *new_bp;
577 bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
581 new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
582 if (unlikely(!new_bp))
585 error = xfs_buf_allocate_memory(new_bp, flags);
587 xfs_buf_free(new_bp);
591 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
593 xfs_buf_free(new_bp);
598 xfs_buf_free(new_bp);
602 error = _xfs_buf_map_pages(bp, flags);
603 if (unlikely(error)) {
604 xfs_warn(target->bt_mount,
605 "%s: failed to map pagesn", __func__);
611 XFS_STATS_INC(xb_get);
612 trace_xfs_buf_get(bp, flags, _RET_IP_);
619 xfs_buf_flags_t flags)
621 ASSERT(!(flags & XBF_WRITE));
622 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
624 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
625 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
627 xfs_buf_iorequest(bp);
628 if (flags & XBF_ASYNC)
630 return xfs_buf_iowait(bp);
635 struct xfs_buftarg *target,
636 struct xfs_buf_map *map,
638 xfs_buf_flags_t flags,
639 const struct xfs_buf_ops *ops)
645 bp = xfs_buf_get_map(target, map, nmaps, flags);
647 trace_xfs_buf_read(bp, flags, _RET_IP_);
649 if (!XFS_BUF_ISDONE(bp)) {
650 XFS_STATS_INC(xb_get_read);
652 _xfs_buf_read(bp, flags);
653 } else if (flags & XBF_ASYNC) {
655 * Read ahead call which is already satisfied,
661 /* We do not want read in the flags */
662 bp->b_flags &= ~XBF_READ;
670 * If we are not low on memory then do the readahead in a deadlock
674 xfs_buf_readahead_map(
675 struct xfs_buftarg *target,
676 struct xfs_buf_map *map,
678 const struct xfs_buf_ops *ops)
680 if (bdi_read_congested(target->bt_bdi))
683 xfs_buf_read_map(target, map, nmaps,
684 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops);
688 * Read an uncached buffer from disk. Allocates and returns a locked
689 * buffer containing the disk contents or nothing.
692 xfs_buf_read_uncached(
693 struct xfs_buftarg *target,
697 const struct xfs_buf_ops *ops)
701 bp = xfs_buf_get_uncached(target, numblks, flags);
705 /* set up the buffer for a read IO */
706 ASSERT(bp->b_map_count == 1);
708 bp->b_maps[0].bm_bn = daddr;
709 bp->b_flags |= XBF_READ;
712 if (XFS_FORCED_SHUTDOWN(target->bt_mount)) {
716 xfs_buf_iorequest(bp);
722 * Return a buffer allocated as an empty buffer and associated to external
723 * memory via xfs_buf_associate_memory() back to it's empty state.
731 _xfs_buf_free_pages(bp);
734 bp->b_page_count = 0;
736 bp->b_length = numblks;
737 bp->b_io_length = numblks;
739 ASSERT(bp->b_map_count == 1);
740 bp->b_bn = XFS_BUF_DADDR_NULL;
741 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
742 bp->b_maps[0].bm_len = bp->b_length;
745 static inline struct page *
749 if ((!is_vmalloc_addr(addr))) {
750 return virt_to_page(addr);
752 return vmalloc_to_page(addr);
757 xfs_buf_associate_memory(
764 unsigned long pageaddr;
765 unsigned long offset;
769 pageaddr = (unsigned long)mem & PAGE_MASK;
770 offset = (unsigned long)mem - pageaddr;
771 buflen = PAGE_ALIGN(len + offset);
772 page_count = buflen >> PAGE_SHIFT;
774 /* Free any previous set of page pointers */
776 _xfs_buf_free_pages(bp);
781 rval = _xfs_buf_get_pages(bp, page_count, 0);
785 bp->b_offset = offset;
787 for (i = 0; i < bp->b_page_count; i++) {
788 bp->b_pages[i] = mem_to_page((void *)pageaddr);
789 pageaddr += PAGE_SIZE;
792 bp->b_io_length = BTOBB(len);
793 bp->b_length = BTOBB(buflen);
799 xfs_buf_get_uncached(
800 struct xfs_buftarg *target,
804 unsigned long page_count;
807 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
809 bp = _xfs_buf_alloc(target, &map, 1, 0);
810 if (unlikely(bp == NULL))
813 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
814 error = _xfs_buf_get_pages(bp, page_count, 0);
818 for (i = 0; i < page_count; i++) {
819 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
823 bp->b_flags |= _XBF_PAGES;
825 error = _xfs_buf_map_pages(bp, 0);
826 if (unlikely(error)) {
827 xfs_warn(target->bt_mount,
828 "%s: failed to map pages", __func__);
832 trace_xfs_buf_get_uncached(bp, _RET_IP_);
837 __free_page(bp->b_pages[i]);
838 _xfs_buf_free_pages(bp);
840 xfs_buf_free_maps(bp);
841 kmem_zone_free(xfs_buf_zone, bp);
847 * Increment reference count on buffer, to hold the buffer concurrently
848 * with another thread which may release (free) the buffer asynchronously.
849 * Must hold the buffer already to call this function.
855 trace_xfs_buf_hold(bp, _RET_IP_);
856 atomic_inc(&bp->b_hold);
860 * Releases a hold on the specified buffer. If the
861 * the hold count is 1, calls xfs_buf_free.
867 struct xfs_perag *pag = bp->b_pag;
869 trace_xfs_buf_rele(bp, _RET_IP_);
872 ASSERT(list_empty(&bp->b_lru));
873 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
874 if (atomic_dec_and_test(&bp->b_hold))
879 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
881 ASSERT(atomic_read(&bp->b_hold) > 0);
882 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
883 spin_lock(&bp->b_lock);
884 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
886 * If the buffer is added to the LRU take a new
887 * reference to the buffer for the LRU and clear the
888 * (now stale) dispose list state flag
890 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) {
891 bp->b_state &= ~XFS_BSTATE_DISPOSE;
892 atomic_inc(&bp->b_hold);
894 spin_unlock(&bp->b_lock);
895 spin_unlock(&pag->pag_buf_lock);
898 * most of the time buffers will already be removed from
899 * the LRU, so optimise that case by checking for the
900 * XFS_BSTATE_DISPOSE flag indicating the last list the
901 * buffer was on was the disposal list
903 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
904 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru);
906 ASSERT(list_empty(&bp->b_lru));
908 spin_unlock(&bp->b_lock);
910 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
911 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
912 spin_unlock(&pag->pag_buf_lock);
921 * Lock a buffer object, if it is not already locked.
923 * If we come across a stale, pinned, locked buffer, we know that we are
924 * being asked to lock a buffer that has been reallocated. Because it is
925 * pinned, we know that the log has not been pushed to disk and hence it
926 * will still be locked. Rather than continuing to have trylock attempts
927 * fail until someone else pushes the log, push it ourselves before
928 * returning. This means that the xfsaild will not get stuck trying
929 * to push on stale inode buffers.
937 locked = down_trylock(&bp->b_sema) == 0;
941 trace_xfs_buf_trylock(bp, _RET_IP_);
946 * Lock a buffer object.
948 * If we come across a stale, pinned, locked buffer, we know that we
949 * are being asked to lock a buffer that has been reallocated. Because
950 * it is pinned, we know that the log has not been pushed to disk and
951 * hence it will still be locked. Rather than sleeping until someone
952 * else pushes the log, push it ourselves before trying to get the lock.
958 trace_xfs_buf_lock(bp, _RET_IP_);
960 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
961 xfs_log_force(bp->b_target->bt_mount, 0);
965 trace_xfs_buf_lock_done(bp, _RET_IP_);
975 trace_xfs_buf_unlock(bp, _RET_IP_);
982 DECLARE_WAITQUEUE (wait, current);
984 if (atomic_read(&bp->b_pin_count) == 0)
987 add_wait_queue(&bp->b_waiters, &wait);
989 set_current_state(TASK_UNINTERRUPTIBLE);
990 if (atomic_read(&bp->b_pin_count) == 0)
994 remove_wait_queue(&bp->b_waiters, &wait);
995 set_current_state(TASK_RUNNING);
999 * Buffer Utility Routines
1003 xfs_buf_iodone_work(
1004 struct work_struct *work)
1006 struct xfs_buf *bp =
1007 container_of(work, xfs_buf_t, b_iodone_work);
1008 bool read = !!(bp->b_flags & XBF_READ);
1010 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1012 /* only validate buffers that were read without errors */
1013 if (read && bp->b_ops && !bp->b_error && (bp->b_flags & XBF_DONE))
1014 bp->b_ops->verify_read(bp);
1017 (*(bp->b_iodone))(bp);
1018 else if (bp->b_flags & XBF_ASYNC)
1021 ASSERT(read && bp->b_ops);
1022 complete(&bp->b_iowait);
1031 bool read = !!(bp->b_flags & XBF_READ);
1033 trace_xfs_buf_iodone(bp, _RET_IP_);
1035 if (bp->b_error == 0)
1036 bp->b_flags |= XBF_DONE;
1038 if (bp->b_iodone || (read && bp->b_ops) || (bp->b_flags & XBF_ASYNC)) {
1040 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1041 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1043 xfs_buf_iodone_work(&bp->b_iodone_work);
1046 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1047 complete(&bp->b_iowait);
1056 ASSERT(error >= 0 && error <= 0xffff);
1057 bp->b_error = (unsigned short)error;
1058 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1062 xfs_buf_ioerror_alert(
1066 xfs_alert(bp->b_target->bt_mount,
1067 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1068 (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length);
1072 * Called when we want to stop a buffer from getting written or read.
1073 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1074 * so that the proper iodone callbacks get called.
1080 #ifdef XFSERRORDEBUG
1081 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1085 * No need to wait until the buffer is unpinned, we aren't flushing it.
1087 xfs_buf_ioerror(bp, EIO);
1090 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1096 xfs_buf_ioend(bp, 0);
1102 * Same as xfs_bioerror, except that we are releasing the buffer
1103 * here ourselves, and avoiding the xfs_buf_ioend call.
1104 * This is meant for userdata errors; metadata bufs come with
1105 * iodone functions attached, so that we can track down errors.
1111 int64_t fl = bp->b_flags;
1113 * No need to wait until the buffer is unpinned.
1114 * We aren't flushing it.
1116 * chunkhold expects B_DONE to be set, whether
1117 * we actually finish the I/O or not. We don't want to
1118 * change that interface.
1123 bp->b_iodone = NULL;
1124 if (!(fl & XBF_ASYNC)) {
1126 * Mark b_error and B_ERROR _both_.
1127 * Lot's of chunkcache code assumes that.
1128 * There's no reason to mark error for
1131 xfs_buf_ioerror(bp, EIO);
1132 complete(&bp->b_iowait);
1144 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1145 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1147 * Metadata write that didn't get logged but
1148 * written delayed anyway. These aren't associated
1149 * with a transaction, and can be ignored.
1151 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1152 return xfs_bioerror_relse(bp);
1154 return xfs_bioerror(bp);
1157 xfs_buf_iorequest(bp);
1167 ASSERT(xfs_buf_islocked(bp));
1169 bp->b_flags |= XBF_WRITE;
1170 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q | XBF_WRITE_FAIL);
1174 error = xfs_buf_iowait(bp);
1176 xfs_force_shutdown(bp->b_target->bt_mount,
1177 SHUTDOWN_META_IO_ERROR);
1187 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1188 xfs_buf_ioend(bp, schedule);
1196 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1199 * don't overwrite existing errors - otherwise we can lose errors on
1200 * buffers that require multiple bios to complete.
1203 xfs_buf_ioerror(bp, -error);
1205 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1206 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1208 _xfs_buf_ioend(bp, 1);
1213 xfs_buf_ioapply_map(
1221 int total_nr_pages = bp->b_page_count;
1224 sector_t sector = bp->b_maps[map].bm_bn;
1228 total_nr_pages = bp->b_page_count;
1230 /* skip the pages in the buffer before the start offset */
1232 offset = *buf_offset;
1233 while (offset >= PAGE_SIZE) {
1235 offset -= PAGE_SIZE;
1239 * Limit the IO size to the length of the current vector, and update the
1240 * remaining IO count for the next time around.
1242 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1244 *buf_offset += size;
1247 atomic_inc(&bp->b_io_remaining);
1248 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1249 if (nr_pages > total_nr_pages)
1250 nr_pages = total_nr_pages;
1252 bio = bio_alloc(GFP_NOIO, nr_pages);
1253 bio->bi_bdev = bp->b_target->bt_bdev;
1254 bio->bi_iter.bi_sector = sector;
1255 bio->bi_end_io = xfs_buf_bio_end_io;
1256 bio->bi_private = bp;
1259 for (; size && nr_pages; nr_pages--, page_index++) {
1260 int rbytes, nbytes = PAGE_SIZE - offset;
1265 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1267 if (rbytes < nbytes)
1271 sector += BTOBB(nbytes);
1276 if (likely(bio->bi_iter.bi_size)) {
1277 if (xfs_buf_is_vmapped(bp)) {
1278 flush_kernel_vmap_range(bp->b_addr,
1279 xfs_buf_vmap_len(bp));
1281 submit_bio(rw, bio);
1286 * This is guaranteed not to be the last io reference count
1287 * because the caller (xfs_buf_iorequest) holds a count itself.
1289 atomic_dec(&bp->b_io_remaining);
1290 xfs_buf_ioerror(bp, EIO);
1300 struct blk_plug plug;
1307 * Make sure we capture only current IO errors rather than stale errors
1308 * left over from previous use of the buffer (e.g. failed readahead).
1312 if (bp->b_flags & XBF_WRITE) {
1313 if (bp->b_flags & XBF_SYNCIO)
1317 if (bp->b_flags & XBF_FUA)
1319 if (bp->b_flags & XBF_FLUSH)
1323 * Run the write verifier callback function if it exists. If
1324 * this function fails it will mark the buffer with an error and
1325 * the IO should not be dispatched.
1328 bp->b_ops->verify_write(bp);
1330 xfs_force_shutdown(bp->b_target->bt_mount,
1331 SHUTDOWN_CORRUPT_INCORE);
1335 } else if (bp->b_flags & XBF_READ_AHEAD) {
1341 /* we only use the buffer cache for meta-data */
1345 * Walk all the vectors issuing IO on them. Set up the initial offset
1346 * into the buffer and the desired IO size before we start -
1347 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1350 offset = bp->b_offset;
1351 size = BBTOB(bp->b_io_length);
1352 blk_start_plug(&plug);
1353 for (i = 0; i < bp->b_map_count; i++) {
1354 xfs_buf_ioapply_map(bp, i, &offset, &size, rw);
1358 break; /* all done */
1360 blk_finish_plug(&plug);
1367 trace_xfs_buf_iorequest(bp, _RET_IP_);
1369 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1371 if (bp->b_flags & XBF_WRITE)
1372 xfs_buf_wait_unpin(bp);
1376 * Set the count to 1 initially, this will stop an I/O
1377 * completion callout which happens before we have started
1378 * all the I/O from calling xfs_buf_ioend too early.
1380 atomic_set(&bp->b_io_remaining, 1);
1381 _xfs_buf_ioapply(bp);
1383 * If _xfs_buf_ioapply failed, we'll get back here with
1384 * only the reference we took above. _xfs_buf_ioend will
1385 * drop it to zero, so we'd better not queue it for later,
1386 * or we'll free it before it's done.
1388 _xfs_buf_ioend(bp, bp->b_error ? 0 : 1);
1394 * Waits for I/O to complete on the buffer supplied. It returns immediately if
1395 * no I/O is pending or there is already a pending error on the buffer, in which
1396 * case nothing will ever complete. It returns the I/O error code, if any, or
1397 * 0 if there was no error.
1403 trace_xfs_buf_iowait(bp, _RET_IP_);
1406 wait_for_completion(&bp->b_iowait);
1408 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1420 return bp->b_addr + offset;
1422 offset += bp->b_offset;
1423 page = bp->b_pages[offset >> PAGE_SHIFT];
1424 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1428 * Move data into or out of a buffer.
1432 xfs_buf_t *bp, /* buffer to process */
1433 size_t boff, /* starting buffer offset */
1434 size_t bsize, /* length to copy */
1435 void *data, /* data address */
1436 xfs_buf_rw_t mode) /* read/write/zero flag */
1440 bend = boff + bsize;
1441 while (boff < bend) {
1443 int page_index, page_offset, csize;
1445 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1446 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1447 page = bp->b_pages[page_index];
1448 csize = min_t(size_t, PAGE_SIZE - page_offset,
1449 BBTOB(bp->b_io_length) - boff);
1451 ASSERT((csize + page_offset) <= PAGE_SIZE);
1455 memset(page_address(page) + page_offset, 0, csize);
1458 memcpy(data, page_address(page) + page_offset, csize);
1461 memcpy(page_address(page) + page_offset, data, csize);
1470 * Handling of buffer targets (buftargs).
1474 * Wait for any bufs with callbacks that have been submitted but have not yet
1475 * returned. These buffers will have an elevated hold count, so wait on those
1476 * while freeing all the buffers only held by the LRU.
1478 static enum lru_status
1479 xfs_buftarg_wait_rele(
1480 struct list_head *item,
1481 spinlock_t *lru_lock,
1485 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1486 struct list_head *dispose = arg;
1488 if (atomic_read(&bp->b_hold) > 1) {
1489 /* need to wait, so skip it this pass */
1490 trace_xfs_buf_wait_buftarg(bp, _RET_IP_);
1493 if (!spin_trylock(&bp->b_lock))
1497 * clear the LRU reference count so the buffer doesn't get
1498 * ignored in xfs_buf_rele().
1500 atomic_set(&bp->b_lru_ref, 0);
1501 bp->b_state |= XFS_BSTATE_DISPOSE;
1502 list_move(item, dispose);
1503 spin_unlock(&bp->b_lock);
1509 struct xfs_buftarg *btp)
1514 /* loop until there is nothing left on the lru list. */
1515 while (list_lru_count(&btp->bt_lru)) {
1516 list_lru_walk(&btp->bt_lru, xfs_buftarg_wait_rele,
1517 &dispose, LONG_MAX);
1519 while (!list_empty(&dispose)) {
1521 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1522 list_del_init(&bp->b_lru);
1523 if (bp->b_flags & XBF_WRITE_FAIL) {
1524 xfs_alert(btp->bt_mount,
1525 "Corruption Alert: Buffer at block 0x%llx had permanent write failures!\n"
1526 "Please run xfs_repair to determine the extent of the problem.",
1527 (long long)bp->b_bn);
1536 static enum lru_status
1537 xfs_buftarg_isolate(
1538 struct list_head *item,
1539 spinlock_t *lru_lock,
1542 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1543 struct list_head *dispose = arg;
1546 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1547 * If we fail to get the lock, just skip it.
1549 if (!spin_trylock(&bp->b_lock))
1552 * Decrement the b_lru_ref count unless the value is already
1553 * zero. If the value is already zero, we need to reclaim the
1554 * buffer, otherwise it gets another trip through the LRU.
1556 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1557 spin_unlock(&bp->b_lock);
1561 bp->b_state |= XFS_BSTATE_DISPOSE;
1562 list_move(item, dispose);
1563 spin_unlock(&bp->b_lock);
1567 static unsigned long
1568 xfs_buftarg_shrink_scan(
1569 struct shrinker *shrink,
1570 struct shrink_control *sc)
1572 struct xfs_buftarg *btp = container_of(shrink,
1573 struct xfs_buftarg, bt_shrinker);
1575 unsigned long freed;
1576 unsigned long nr_to_scan = sc->nr_to_scan;
1578 freed = list_lru_walk_node(&btp->bt_lru, sc->nid, xfs_buftarg_isolate,
1579 &dispose, &nr_to_scan);
1581 while (!list_empty(&dispose)) {
1583 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1584 list_del_init(&bp->b_lru);
1591 static unsigned long
1592 xfs_buftarg_shrink_count(
1593 struct shrinker *shrink,
1594 struct shrink_control *sc)
1596 struct xfs_buftarg *btp = container_of(shrink,
1597 struct xfs_buftarg, bt_shrinker);
1598 return list_lru_count_node(&btp->bt_lru, sc->nid);
1603 struct xfs_mount *mp,
1604 struct xfs_buftarg *btp)
1606 unregister_shrinker(&btp->bt_shrinker);
1607 list_lru_destroy(&btp->bt_lru);
1609 if (mp->m_flags & XFS_MOUNT_BARRIER)
1610 xfs_blkdev_issue_flush(btp);
1616 xfs_setsize_buftarg(
1618 unsigned int blocksize,
1619 unsigned int sectorsize)
1621 /* Set up metadata sector size info */
1622 btp->bt_meta_sectorsize = sectorsize;
1623 btp->bt_meta_sectormask = sectorsize - 1;
1625 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1626 char name[BDEVNAME_SIZE];
1628 bdevname(btp->bt_bdev, name);
1630 xfs_warn(btp->bt_mount,
1631 "Cannot set_blocksize to %u on device %s",
1636 /* Set up device logical sector size mask */
1637 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1638 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1644 * When allocating the initial buffer target we have not yet
1645 * read in the superblock, so don't know what sized sectors
1646 * are being used at this early stage. Play safe.
1649 xfs_setsize_buftarg_early(
1651 struct block_device *bdev)
1653 return xfs_setsize_buftarg(btp, PAGE_SIZE,
1654 bdev_logical_block_size(bdev));
1659 struct xfs_mount *mp,
1660 struct block_device *bdev,
1666 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS);
1669 btp->bt_dev = bdev->bd_dev;
1670 btp->bt_bdev = bdev;
1671 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1675 if (xfs_setsize_buftarg_early(btp, bdev))
1678 if (list_lru_init(&btp->bt_lru))
1681 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
1682 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
1683 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1684 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
1685 register_shrinker(&btp->bt_shrinker);
1694 * Add a buffer to the delayed write list.
1696 * This queues a buffer for writeout if it hasn't already been. Note that
1697 * neither this routine nor the buffer list submission functions perform
1698 * any internal synchronization. It is expected that the lists are thread-local
1701 * Returns true if we queued up the buffer, or false if it already had
1702 * been on the buffer list.
1705 xfs_buf_delwri_queue(
1707 struct list_head *list)
1709 ASSERT(xfs_buf_islocked(bp));
1710 ASSERT(!(bp->b_flags & XBF_READ));
1713 * If the buffer is already marked delwri it already is queued up
1714 * by someone else for imediate writeout. Just ignore it in that
1717 if (bp->b_flags & _XBF_DELWRI_Q) {
1718 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1722 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1725 * If a buffer gets written out synchronously or marked stale while it
1726 * is on a delwri list we lazily remove it. To do this, the other party
1727 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1728 * It remains referenced and on the list. In a rare corner case it
1729 * might get readded to a delwri list after the synchronous writeout, in
1730 * which case we need just need to re-add the flag here.
1732 bp->b_flags |= _XBF_DELWRI_Q;
1733 if (list_empty(&bp->b_list)) {
1734 atomic_inc(&bp->b_hold);
1735 list_add_tail(&bp->b_list, list);
1742 * Compare function is more complex than it needs to be because
1743 * the return value is only 32 bits and we are doing comparisons
1749 struct list_head *a,
1750 struct list_head *b)
1752 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1753 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1756 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
1765 __xfs_buf_delwri_submit(
1766 struct list_head *buffer_list,
1767 struct list_head *io_list,
1770 struct blk_plug plug;
1771 struct xfs_buf *bp, *n;
1774 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1776 if (xfs_buf_ispinned(bp)) {
1780 if (!xfs_buf_trylock(bp))
1787 * Someone else might have written the buffer synchronously or
1788 * marked it stale in the meantime. In that case only the
1789 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1790 * reference and remove it from the list here.
1792 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1793 list_del_init(&bp->b_list);
1798 list_move_tail(&bp->b_list, io_list);
1799 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1802 list_sort(NULL, io_list, xfs_buf_cmp);
1804 blk_start_plug(&plug);
1805 list_for_each_entry_safe(bp, n, io_list, b_list) {
1806 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC | XBF_WRITE_FAIL);
1807 bp->b_flags |= XBF_WRITE;
1810 bp->b_flags |= XBF_ASYNC;
1811 list_del_init(&bp->b_list);
1815 blk_finish_plug(&plug);
1821 * Write out a buffer list asynchronously.
1823 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1824 * out and not wait for I/O completion on any of the buffers. This interface
1825 * is only safely useable for callers that can track I/O completion by higher
1826 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1830 xfs_buf_delwri_submit_nowait(
1831 struct list_head *buffer_list)
1833 LIST_HEAD (io_list);
1834 return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
1838 * Write out a buffer list synchronously.
1840 * This will take the @buffer_list, write all buffers out and wait for I/O
1841 * completion on all of the buffers. @buffer_list is consumed by the function,
1842 * so callers must have some other way of tracking buffers if they require such
1846 xfs_buf_delwri_submit(
1847 struct list_head *buffer_list)
1849 LIST_HEAD (io_list);
1850 int error = 0, error2;
1853 __xfs_buf_delwri_submit(buffer_list, &io_list, true);
1855 /* Wait for IO to complete. */
1856 while (!list_empty(&io_list)) {
1857 bp = list_first_entry(&io_list, struct xfs_buf, b_list);
1859 list_del_init(&bp->b_list);
1860 error2 = xfs_buf_iowait(bp);
1872 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1873 KM_ZONE_HWALIGN, NULL);
1877 xfslogd_workqueue = alloc_workqueue("xfslogd",
1878 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1879 if (!xfslogd_workqueue)
1880 goto out_free_buf_zone;
1885 kmem_zone_destroy(xfs_buf_zone);
1891 xfs_buf_terminate(void)
1893 destroy_workqueue(xfslogd_workqueue);
1894 kmem_zone_destroy(xfs_buf_zone);