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>
40 #include "xfs_mount.h"
41 #include "xfs_trace.h"
43 static kmem_zone_t *xfs_buf_zone;
45 static struct workqueue_struct *xfslogd_workqueue;
47 #ifdef XFS_BUF_LOCK_TRACKING
48 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
49 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
50 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
52 # define XB_SET_OWNER(bp) do { } while (0)
53 # define XB_CLEAR_OWNER(bp) do { } while (0)
54 # define XB_GET_OWNER(bp) do { } while (0)
57 #define xb_to_gfp(flags) \
58 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN)
66 * Return true if the buffer is vmapped.
68 * b_addr is null if the buffer is not mapped, but the code is clever
69 * enough to know it doesn't have to map a single page, so the check has
70 * to be both for b_addr and bp->b_page_count > 1.
72 return bp->b_addr && bp->b_page_count > 1;
79 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
83 * xfs_buf_lru_add - add a buffer to the LRU.
85 * The LRU takes a new reference to the buffer so that it will only be freed
86 * once the shrinker takes the buffer off the LRU.
92 struct xfs_buftarg *btp = bp->b_target;
94 spin_lock(&btp->bt_lru_lock);
95 if (list_empty(&bp->b_lru)) {
96 atomic_inc(&bp->b_hold);
97 list_add_tail(&bp->b_lru, &btp->bt_lru);
100 spin_unlock(&btp->bt_lru_lock);
104 * xfs_buf_lru_del - remove a buffer from the LRU
106 * The unlocked check is safe here because it only occurs when there are not
107 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
108 * to optimise the shrinker removing the buffer from the LRU and calling
109 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
116 struct xfs_buftarg *btp = bp->b_target;
118 if (list_empty(&bp->b_lru))
121 spin_lock(&btp->bt_lru_lock);
122 if (!list_empty(&bp->b_lru)) {
123 list_del_init(&bp->b_lru);
126 spin_unlock(&btp->bt_lru_lock);
130 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
131 * b_lru_ref count so that the buffer is freed immediately when the buffer
132 * reference count falls to zero. If the buffer is already on the LRU, we need
133 * to remove the reference that LRU holds on the buffer.
135 * This prevents build-up of stale buffers on the LRU.
141 ASSERT(xfs_buf_islocked(bp));
143 bp->b_flags |= XBF_STALE;
146 * Clear the delwri status so that a delwri queue walker will not
147 * flush this buffer to disk now that it is stale. The delwri queue has
148 * a reference to the buffer, so this is safe to do.
150 bp->b_flags &= ~_XBF_DELWRI_Q;
152 atomic_set(&(bp)->b_lru_ref, 0);
153 if (!list_empty(&bp->b_lru)) {
154 struct xfs_buftarg *btp = bp->b_target;
156 spin_lock(&btp->bt_lru_lock);
157 if (!list_empty(&bp->b_lru)) {
158 list_del_init(&bp->b_lru);
160 atomic_dec(&bp->b_hold);
162 spin_unlock(&btp->bt_lru_lock);
164 ASSERT(atomic_read(&bp->b_hold) >= 1);
172 ASSERT(bp->b_maps == NULL);
173 bp->b_map_count = map_count;
175 if (map_count == 1) {
176 bp->b_maps = &bp->b_map;
180 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
188 * Frees b_pages if it was allocated.
194 if (bp->b_maps != &bp->b_map) {
195 kmem_free(bp->b_maps);
202 struct xfs_buftarg *target,
203 struct xfs_buf_map *map,
205 xfs_buf_flags_t flags)
211 bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS);
216 * We don't want certain flags to appear in b_flags unless they are
217 * specifically set by later operations on the buffer.
219 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
221 atomic_set(&bp->b_hold, 1);
222 atomic_set(&bp->b_lru_ref, 1);
223 init_completion(&bp->b_iowait);
224 INIT_LIST_HEAD(&bp->b_lru);
225 INIT_LIST_HEAD(&bp->b_list);
226 RB_CLEAR_NODE(&bp->b_rbnode);
227 sema_init(&bp->b_sema, 0); /* held, no waiters */
229 bp->b_target = target;
233 * Set length and io_length to the same value initially.
234 * I/O routines should use io_length, which will be the same in
235 * most cases but may be reset (e.g. XFS recovery).
237 error = xfs_buf_get_maps(bp, nmaps);
239 kmem_zone_free(xfs_buf_zone, bp);
243 bp->b_bn = map[0].bm_bn;
245 for (i = 0; i < nmaps; i++) {
246 bp->b_maps[i].bm_bn = map[i].bm_bn;
247 bp->b_maps[i].bm_len = map[i].bm_len;
248 bp->b_length += map[i].bm_len;
250 bp->b_io_length = bp->b_length;
252 atomic_set(&bp->b_pin_count, 0);
253 init_waitqueue_head(&bp->b_waiters);
255 XFS_STATS_INC(xb_create);
256 trace_xfs_buf_init(bp, _RET_IP_);
262 * Allocate a page array capable of holding a specified number
263 * of pages, and point the page buf at it.
269 xfs_buf_flags_t flags)
271 /* Make sure that we have a page list */
272 if (bp->b_pages == NULL) {
273 bp->b_page_count = page_count;
274 if (page_count <= XB_PAGES) {
275 bp->b_pages = bp->b_page_array;
277 bp->b_pages = kmem_alloc(sizeof(struct page *) *
278 page_count, KM_NOFS);
279 if (bp->b_pages == NULL)
282 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
288 * Frees b_pages if it was allocated.
294 if (bp->b_pages != bp->b_page_array) {
295 kmem_free(bp->b_pages);
301 * Releases the specified buffer.
303 * The modification state of any associated pages is left unchanged.
304 * The buffer most not be on any hash - use xfs_buf_rele instead for
305 * hashed and refcounted buffers
311 trace_xfs_buf_free(bp, _RET_IP_);
313 ASSERT(list_empty(&bp->b_lru));
315 if (bp->b_flags & _XBF_PAGES) {
318 if (xfs_buf_is_vmapped(bp))
319 vm_unmap_ram(bp->b_addr - bp->b_offset,
322 for (i = 0; i < bp->b_page_count; i++) {
323 struct page *page = bp->b_pages[i];
327 } else if (bp->b_flags & _XBF_KMEM)
328 kmem_free(bp->b_addr);
329 _xfs_buf_free_pages(bp);
330 xfs_buf_free_maps(bp);
331 kmem_zone_free(xfs_buf_zone, bp);
335 * Allocates all the pages for buffer in question and builds it's page list.
338 xfs_buf_allocate_memory(
343 size_t nbytes, offset;
344 gfp_t gfp_mask = xb_to_gfp(flags);
345 unsigned short page_count, i;
346 xfs_off_t start, end;
350 * for buffers that are contained within a single page, just allocate
351 * the memory from the heap - there's no need for the complexity of
352 * page arrays to keep allocation down to order 0.
354 size = BBTOB(bp->b_length);
355 if (size < PAGE_SIZE) {
356 bp->b_addr = kmem_alloc(size, KM_NOFS);
358 /* low memory - use alloc_page loop instead */
362 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
363 ((unsigned long)bp->b_addr & PAGE_MASK)) {
364 /* b_addr spans two pages - use alloc_page instead */
365 kmem_free(bp->b_addr);
369 bp->b_offset = offset_in_page(bp->b_addr);
370 bp->b_pages = bp->b_page_array;
371 bp->b_pages[0] = virt_to_page(bp->b_addr);
372 bp->b_page_count = 1;
373 bp->b_flags |= _XBF_KMEM;
378 start = BBTOB(bp->b_map.bm_bn) >> PAGE_SHIFT;
379 end = (BBTOB(bp->b_map.bm_bn + bp->b_length) + PAGE_SIZE - 1)
381 page_count = end - start;
382 error = _xfs_buf_get_pages(bp, page_count, flags);
386 offset = bp->b_offset;
387 bp->b_flags |= _XBF_PAGES;
389 for (i = 0; i < bp->b_page_count; i++) {
393 page = alloc_page(gfp_mask);
394 if (unlikely(page == NULL)) {
395 if (flags & XBF_READ_AHEAD) {
396 bp->b_page_count = i;
402 * This could deadlock.
404 * But until all the XFS lowlevel code is revamped to
405 * handle buffer allocation failures we can't do much.
407 if (!(++retries % 100))
409 "possible memory allocation deadlock in %s (mode:0x%x)",
412 XFS_STATS_INC(xb_page_retries);
413 congestion_wait(BLK_RW_ASYNC, HZ/50);
417 XFS_STATS_INC(xb_page_found);
419 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
421 bp->b_pages[i] = page;
427 for (i = 0; i < bp->b_page_count; i++)
428 __free_page(bp->b_pages[i]);
433 * Map buffer into kernel address-space if necessary.
440 ASSERT(bp->b_flags & _XBF_PAGES);
441 if (bp->b_page_count == 1) {
442 /* A single page buffer is always mappable */
443 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
444 } else if (flags & XBF_UNMAPPED) {
450 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
455 } while (retried++ <= 1);
459 bp->b_addr += bp->b_offset;
466 * Finding and Reading Buffers
470 * Look up, and creates if absent, a lockable buffer for
471 * a given range of an inode. The buffer is returned
472 * locked. No I/O is implied by this call.
476 struct xfs_buftarg *btp,
477 struct xfs_buf_map *map,
479 xfs_buf_flags_t flags,
483 struct xfs_perag *pag;
484 struct rb_node **rbp;
485 struct rb_node *parent;
487 xfs_daddr_t blkno = map[0].bm_bn;
491 for (i = 0; i < nmaps; i++)
492 numblks += map[i].bm_len;
493 numbytes = BBTOB(numblks);
495 /* Check for IOs smaller than the sector size / not sector aligned */
496 ASSERT(!(numbytes < (1 << btp->bt_sshift)));
497 ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_smask));
500 pag = xfs_perag_get(btp->bt_mount,
501 xfs_daddr_to_agno(btp->bt_mount, blkno));
504 spin_lock(&pag->pag_buf_lock);
505 rbp = &pag->pag_buf_tree.rb_node;
510 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
512 if (blkno < bp->b_bn)
513 rbp = &(*rbp)->rb_left;
514 else if (blkno > bp->b_bn)
515 rbp = &(*rbp)->rb_right;
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 to the right for an exact match.
525 if (bp->b_length != numblks) {
526 ASSERT(bp->b_flags & XBF_STALE);
527 rbp = &(*rbp)->rb_right;
530 atomic_inc(&bp->b_hold);
537 rb_link_node(&new_bp->b_rbnode, parent, rbp);
538 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
539 /* the buffer keeps the perag reference until it is freed */
541 spin_unlock(&pag->pag_buf_lock);
543 XFS_STATS_INC(xb_miss_locked);
544 spin_unlock(&pag->pag_buf_lock);
550 spin_unlock(&pag->pag_buf_lock);
553 if (!xfs_buf_trylock(bp)) {
554 if (flags & XBF_TRYLOCK) {
556 XFS_STATS_INC(xb_busy_locked);
560 XFS_STATS_INC(xb_get_locked_waited);
564 * if the buffer is stale, clear all the external state associated with
565 * it. We need to keep flags such as how we allocated the buffer memory
568 if (bp->b_flags & XBF_STALE) {
569 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
570 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
573 trace_xfs_buf_find(bp, flags, _RET_IP_);
574 XFS_STATS_INC(xb_get_locked);
579 * Assembles a buffer covering the specified range. The code is optimised for
580 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
581 * more hits than misses.
585 struct xfs_buftarg *target,
586 struct xfs_buf_map *map,
588 xfs_buf_flags_t flags)
591 struct xfs_buf *new_bp;
594 bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
598 new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
599 if (unlikely(!new_bp))
602 error = xfs_buf_allocate_memory(new_bp, flags);
604 xfs_buf_free(new_bp);
608 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
610 xfs_buf_free(new_bp);
615 xfs_buf_free(new_bp);
619 error = _xfs_buf_map_pages(bp, flags);
620 if (unlikely(error)) {
621 xfs_warn(target->bt_mount,
622 "%s: failed to map pages\n", __func__);
628 XFS_STATS_INC(xb_get);
629 trace_xfs_buf_get(bp, flags, _RET_IP_);
636 xfs_buf_flags_t flags)
638 ASSERT(!(flags & XBF_WRITE));
639 ASSERT(bp->b_map.bm_bn != XFS_BUF_DADDR_NULL);
641 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
642 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
644 xfs_buf_iorequest(bp);
645 if (flags & XBF_ASYNC)
647 return xfs_buf_iowait(bp);
652 struct xfs_buftarg *target,
653 struct xfs_buf_map *map,
655 xfs_buf_flags_t flags)
661 bp = xfs_buf_get_map(target, map, nmaps, flags);
663 trace_xfs_buf_read(bp, flags, _RET_IP_);
665 if (!XFS_BUF_ISDONE(bp)) {
666 XFS_STATS_INC(xb_get_read);
667 _xfs_buf_read(bp, flags);
668 } else if (flags & XBF_ASYNC) {
670 * Read ahead call which is already satisfied,
676 /* We do not want read in the flags */
677 bp->b_flags &= ~XBF_READ;
685 * If we are not low on memory then do the readahead in a deadlock
689 xfs_buf_readahead_map(
690 struct xfs_buftarg *target,
691 struct xfs_buf_map *map,
694 if (bdi_read_congested(target->bt_bdi))
697 xfs_buf_read_map(target, map, nmaps,
698 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
702 * Read an uncached buffer from disk. Allocates and returns a locked
703 * buffer containing the disk contents or nothing.
706 xfs_buf_read_uncached(
707 struct xfs_buftarg *target,
715 bp = xfs_buf_get_uncached(target, numblks, flags);
719 /* set up the buffer for a read IO */
720 ASSERT(bp->b_map_count == 1);
722 bp->b_maps[0].bm_bn = daddr;
723 bp->b_flags |= XBF_READ;
725 xfsbdstrat(target->bt_mount, bp);
726 error = xfs_buf_iowait(bp);
735 * Return a buffer allocated as an empty buffer and associated to external
736 * memory via xfs_buf_associate_memory() back to it's empty state.
744 _xfs_buf_free_pages(bp);
747 bp->b_page_count = 0;
749 bp->b_length = numblks;
750 bp->b_io_length = numblks;
752 ASSERT(bp->b_map_count == 1);
753 bp->b_bn = XFS_BUF_DADDR_NULL;
754 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
755 bp->b_maps[0].bm_len = bp->b_length;
758 static inline struct page *
762 if ((!is_vmalloc_addr(addr))) {
763 return virt_to_page(addr);
765 return vmalloc_to_page(addr);
770 xfs_buf_associate_memory(
777 unsigned long pageaddr;
778 unsigned long offset;
782 pageaddr = (unsigned long)mem & PAGE_MASK;
783 offset = (unsigned long)mem - pageaddr;
784 buflen = PAGE_ALIGN(len + offset);
785 page_count = buflen >> PAGE_SHIFT;
787 /* Free any previous set of page pointers */
789 _xfs_buf_free_pages(bp);
794 rval = _xfs_buf_get_pages(bp, page_count, 0);
798 bp->b_offset = offset;
800 for (i = 0; i < bp->b_page_count; i++) {
801 bp->b_pages[i] = mem_to_page((void *)pageaddr);
802 pageaddr += PAGE_SIZE;
805 bp->b_io_length = BTOBB(len);
806 bp->b_length = BTOBB(buflen);
812 xfs_buf_get_uncached(
813 struct xfs_buftarg *target,
817 unsigned long page_count;
820 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
822 bp = _xfs_buf_alloc(target, &map, 1, 0);
823 if (unlikely(bp == NULL))
826 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
827 error = _xfs_buf_get_pages(bp, page_count, 0);
831 for (i = 0; i < page_count; i++) {
832 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
836 bp->b_flags |= _XBF_PAGES;
838 error = _xfs_buf_map_pages(bp, 0);
839 if (unlikely(error)) {
840 xfs_warn(target->bt_mount,
841 "%s: failed to map pages\n", __func__);
845 trace_xfs_buf_get_uncached(bp, _RET_IP_);
850 __free_page(bp->b_pages[i]);
851 _xfs_buf_free_pages(bp);
853 xfs_buf_free_maps(bp);
854 kmem_zone_free(xfs_buf_zone, bp);
860 * Increment reference count on buffer, to hold the buffer concurrently
861 * with another thread which may release (free) the buffer asynchronously.
862 * Must hold the buffer already to call this function.
868 trace_xfs_buf_hold(bp, _RET_IP_);
869 atomic_inc(&bp->b_hold);
873 * Releases a hold on the specified buffer. If the
874 * the hold count is 1, calls xfs_buf_free.
880 struct xfs_perag *pag = bp->b_pag;
882 trace_xfs_buf_rele(bp, _RET_IP_);
885 ASSERT(list_empty(&bp->b_lru));
886 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
887 if (atomic_dec_and_test(&bp->b_hold))
892 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
894 ASSERT(atomic_read(&bp->b_hold) > 0);
895 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
896 if (!(bp->b_flags & XBF_STALE) &&
897 atomic_read(&bp->b_lru_ref)) {
899 spin_unlock(&pag->pag_buf_lock);
902 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
903 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
904 spin_unlock(&pag->pag_buf_lock);
913 * Lock a buffer object, if it is not already locked.
915 * If we come across a stale, pinned, locked buffer, we know that we are
916 * being asked to lock a buffer that has been reallocated. Because it is
917 * pinned, we know that the log has not been pushed to disk and hence it
918 * will still be locked. Rather than continuing to have trylock attempts
919 * fail until someone else pushes the log, push it ourselves before
920 * returning. This means that the xfsaild will not get stuck trying
921 * to push on stale inode buffers.
929 locked = down_trylock(&bp->b_sema) == 0;
932 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
933 xfs_log_force(bp->b_target->bt_mount, 0);
935 trace_xfs_buf_trylock(bp, _RET_IP_);
940 * Lock a buffer object.
942 * If we come across a stale, pinned, locked buffer, we know that we
943 * are being asked to lock a buffer that has been reallocated. Because
944 * it is pinned, we know that the log has not been pushed to disk and
945 * hence it will still be locked. Rather than sleeping until someone
946 * else pushes the log, push it ourselves before trying to get the lock.
952 trace_xfs_buf_lock(bp, _RET_IP_);
954 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
955 xfs_log_force(bp->b_target->bt_mount, 0);
959 trace_xfs_buf_lock_done(bp, _RET_IP_);
969 trace_xfs_buf_unlock(bp, _RET_IP_);
976 DECLARE_WAITQUEUE (wait, current);
978 if (atomic_read(&bp->b_pin_count) == 0)
981 add_wait_queue(&bp->b_waiters, &wait);
983 set_current_state(TASK_UNINTERRUPTIBLE);
984 if (atomic_read(&bp->b_pin_count) == 0)
988 remove_wait_queue(&bp->b_waiters, &wait);
989 set_current_state(TASK_RUNNING);
993 * Buffer Utility Routines
998 struct work_struct *work)
1001 container_of(work, xfs_buf_t, b_iodone_work);
1004 (*(bp->b_iodone))(bp);
1005 else if (bp->b_flags & XBF_ASYNC)
1014 trace_xfs_buf_iodone(bp, _RET_IP_);
1016 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1017 if (bp->b_error == 0)
1018 bp->b_flags |= XBF_DONE;
1020 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1022 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1023 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1025 xfs_buf_iodone_work(&bp->b_iodone_work);
1028 complete(&bp->b_iowait);
1037 ASSERT(error >= 0 && error <= 0xffff);
1038 bp->b_error = (unsigned short)error;
1039 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1043 xfs_buf_ioerror_alert(
1047 xfs_alert(bp->b_target->bt_mount,
1048 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1049 (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length);
1053 * Called when we want to stop a buffer from getting written or read.
1054 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1055 * so that the proper iodone callbacks get called.
1061 #ifdef XFSERRORDEBUG
1062 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1066 * No need to wait until the buffer is unpinned, we aren't flushing it.
1068 xfs_buf_ioerror(bp, EIO);
1071 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1077 xfs_buf_ioend(bp, 0);
1083 * Same as xfs_bioerror, except that we are releasing the buffer
1084 * here ourselves, and avoiding the xfs_buf_ioend call.
1085 * This is meant for userdata errors; metadata bufs come with
1086 * iodone functions attached, so that we can track down errors.
1092 int64_t fl = bp->b_flags;
1094 * No need to wait until the buffer is unpinned.
1095 * We aren't flushing it.
1097 * chunkhold expects B_DONE to be set, whether
1098 * we actually finish the I/O or not. We don't want to
1099 * change that interface.
1104 bp->b_iodone = NULL;
1105 if (!(fl & XBF_ASYNC)) {
1107 * Mark b_error and B_ERROR _both_.
1108 * Lot's of chunkcache code assumes that.
1109 * There's no reason to mark error for
1112 xfs_buf_ioerror(bp, EIO);
1113 complete(&bp->b_iowait);
1125 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1126 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1128 * Metadata write that didn't get logged but
1129 * written delayed anyway. These aren't associated
1130 * with a transaction, and can be ignored.
1132 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1133 return xfs_bioerror_relse(bp);
1135 return xfs_bioerror(bp);
1138 xfs_buf_iorequest(bp);
1148 ASSERT(xfs_buf_islocked(bp));
1150 bp->b_flags |= XBF_WRITE;
1151 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q);
1155 error = xfs_buf_iowait(bp);
1157 xfs_force_shutdown(bp->b_target->bt_mount,
1158 SHUTDOWN_META_IO_ERROR);
1164 * Wrapper around bdstrat so that we can stop data from going to disk in case
1165 * we are shutting down the filesystem. Typically user data goes thru this
1166 * path; one of the exceptions is the superblock.
1170 struct xfs_mount *mp,
1173 if (XFS_FORCED_SHUTDOWN(mp)) {
1174 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1175 xfs_bioerror_relse(bp);
1179 xfs_buf_iorequest(bp);
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;
1198 xfs_buf_ioerror(bp, -error);
1200 if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1201 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1203 _xfs_buf_ioend(bp, 1);
1208 xfs_buf_ioapply_map(
1216 int total_nr_pages = bp->b_page_count;
1219 sector_t sector = bp->b_maps[map].bm_bn;
1223 total_nr_pages = bp->b_page_count;
1225 /* skip the pages in the buffer before the start offset */
1227 offset = *buf_offset;
1228 while (offset >= PAGE_SIZE) {
1230 offset -= PAGE_SIZE;
1234 * Limit the IO size to the length of the current vector, and update the
1235 * remaining IO count for the next time around.
1237 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1239 *buf_offset += size;
1242 atomic_inc(&bp->b_io_remaining);
1243 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1244 if (nr_pages > total_nr_pages)
1245 nr_pages = total_nr_pages;
1247 bio = bio_alloc(GFP_NOIO, nr_pages);
1248 bio->bi_bdev = bp->b_target->bt_bdev;
1249 bio->bi_sector = sector;
1250 bio->bi_end_io = xfs_buf_bio_end_io;
1251 bio->bi_private = bp;
1254 for (; size && nr_pages; nr_pages--, page_index++) {
1255 int rbytes, nbytes = PAGE_SIZE - offset;
1260 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1262 if (rbytes < nbytes)
1266 sector += BTOBB(nbytes);
1271 if (likely(bio->bi_size)) {
1272 if (xfs_buf_is_vmapped(bp)) {
1273 flush_kernel_vmap_range(bp->b_addr,
1274 xfs_buf_vmap_len(bp));
1276 submit_bio(rw, bio);
1280 xfs_buf_ioerror(bp, EIO);
1290 struct blk_plug plug;
1296 if (bp->b_flags & XBF_WRITE) {
1297 if (bp->b_flags & XBF_SYNCIO)
1301 if (bp->b_flags & XBF_FUA)
1303 if (bp->b_flags & XBF_FLUSH)
1305 } else if (bp->b_flags & XBF_READ_AHEAD) {
1311 /* we only use the buffer cache for meta-data */
1315 * Walk all the vectors issuing IO on them. Set up the initial offset
1316 * into the buffer and the desired IO size before we start -
1317 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1320 offset = bp->b_offset;
1321 size = BBTOB(bp->b_io_length);
1322 blk_start_plug(&plug);
1323 for (i = 0; i < bp->b_map_count; i++) {
1324 xfs_buf_ioapply_map(bp, i, &offset, &size, rw);
1328 break; /* all done */
1330 blk_finish_plug(&plug);
1337 trace_xfs_buf_iorequest(bp, _RET_IP_);
1339 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1341 if (bp->b_flags & XBF_WRITE)
1342 xfs_buf_wait_unpin(bp);
1345 /* Set the count to 1 initially, this will stop an I/O
1346 * completion callout which happens before we have started
1347 * all the I/O from calling xfs_buf_ioend too early.
1349 atomic_set(&bp->b_io_remaining, 1);
1350 _xfs_buf_ioapply(bp);
1351 _xfs_buf_ioend(bp, 1);
1357 * Waits for I/O to complete on the buffer supplied. It returns immediately if
1358 * no I/O is pending or there is already a pending error on the buffer. It
1359 * returns the I/O error code, if any, or 0 if there was no error.
1365 trace_xfs_buf_iowait(bp, _RET_IP_);
1368 wait_for_completion(&bp->b_iowait);
1370 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1382 return bp->b_addr + offset;
1384 offset += bp->b_offset;
1385 page = bp->b_pages[offset >> PAGE_SHIFT];
1386 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1390 * Move data into or out of a buffer.
1394 xfs_buf_t *bp, /* buffer to process */
1395 size_t boff, /* starting buffer offset */
1396 size_t bsize, /* length to copy */
1397 void *data, /* data address */
1398 xfs_buf_rw_t mode) /* read/write/zero flag */
1402 bend = boff + bsize;
1403 while (boff < bend) {
1405 int page_index, page_offset, csize;
1407 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1408 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1409 page = bp->b_pages[page_index];
1410 csize = min_t(size_t, PAGE_SIZE - page_offset,
1411 BBTOB(bp->b_io_length) - boff);
1413 ASSERT((csize + page_offset) <= PAGE_SIZE);
1417 memset(page_address(page) + page_offset, 0, csize);
1420 memcpy(data, page_address(page) + page_offset, csize);
1423 memcpy(page_address(page) + page_offset, data, csize);
1432 * Handling of buffer targets (buftargs).
1436 * Wait for any bufs with callbacks that have been submitted but have not yet
1437 * returned. These buffers will have an elevated hold count, so wait on those
1438 * while freeing all the buffers only held by the LRU.
1442 struct xfs_buftarg *btp)
1447 spin_lock(&btp->bt_lru_lock);
1448 while (!list_empty(&btp->bt_lru)) {
1449 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1450 if (atomic_read(&bp->b_hold) > 1) {
1451 spin_unlock(&btp->bt_lru_lock);
1456 * clear the LRU reference count so the buffer doesn't get
1457 * ignored in xfs_buf_rele().
1459 atomic_set(&bp->b_lru_ref, 0);
1460 spin_unlock(&btp->bt_lru_lock);
1462 spin_lock(&btp->bt_lru_lock);
1464 spin_unlock(&btp->bt_lru_lock);
1469 struct shrinker *shrink,
1470 struct shrink_control *sc)
1472 struct xfs_buftarg *btp = container_of(shrink,
1473 struct xfs_buftarg, bt_shrinker);
1475 int nr_to_scan = sc->nr_to_scan;
1479 return btp->bt_lru_nr;
1481 spin_lock(&btp->bt_lru_lock);
1482 while (!list_empty(&btp->bt_lru)) {
1483 if (nr_to_scan-- <= 0)
1486 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1489 * Decrement the b_lru_ref count unless the value is already
1490 * zero. If the value is already zero, we need to reclaim the
1491 * buffer, otherwise it gets another trip through the LRU.
1493 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1494 list_move_tail(&bp->b_lru, &btp->bt_lru);
1499 * remove the buffer from the LRU now to avoid needing another
1500 * lock round trip inside xfs_buf_rele().
1502 list_move(&bp->b_lru, &dispose);
1505 spin_unlock(&btp->bt_lru_lock);
1507 while (!list_empty(&dispose)) {
1508 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1509 list_del_init(&bp->b_lru);
1513 return btp->bt_lru_nr;
1518 struct xfs_mount *mp,
1519 struct xfs_buftarg *btp)
1521 unregister_shrinker(&btp->bt_shrinker);
1523 if (mp->m_flags & XFS_MOUNT_BARRIER)
1524 xfs_blkdev_issue_flush(btp);
1530 xfs_setsize_buftarg_flags(
1532 unsigned int blocksize,
1533 unsigned int sectorsize,
1536 btp->bt_bsize = blocksize;
1537 btp->bt_sshift = ffs(sectorsize) - 1;
1538 btp->bt_smask = sectorsize - 1;
1540 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1541 char name[BDEVNAME_SIZE];
1543 bdevname(btp->bt_bdev, name);
1545 xfs_warn(btp->bt_mount,
1546 "Cannot set_blocksize to %u on device %s\n",
1555 * When allocating the initial buffer target we have not yet
1556 * read in the superblock, so don't know what sized sectors
1557 * are being used is at this early stage. Play safe.
1560 xfs_setsize_buftarg_early(
1562 struct block_device *bdev)
1564 return xfs_setsize_buftarg_flags(btp,
1565 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1569 xfs_setsize_buftarg(
1571 unsigned int blocksize,
1572 unsigned int sectorsize)
1574 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1579 struct xfs_mount *mp,
1580 struct block_device *bdev,
1586 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1589 btp->bt_dev = bdev->bd_dev;
1590 btp->bt_bdev = bdev;
1591 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1595 INIT_LIST_HEAD(&btp->bt_lru);
1596 spin_lock_init(&btp->bt_lru_lock);
1597 if (xfs_setsize_buftarg_early(btp, bdev))
1599 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1600 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1601 register_shrinker(&btp->bt_shrinker);
1610 * Add a buffer to the delayed write list.
1612 * This queues a buffer for writeout if it hasn't already been. Note that
1613 * neither this routine nor the buffer list submission functions perform
1614 * any internal synchronization. It is expected that the lists are thread-local
1617 * Returns true if we queued up the buffer, or false if it already had
1618 * been on the buffer list.
1621 xfs_buf_delwri_queue(
1623 struct list_head *list)
1625 ASSERT(xfs_buf_islocked(bp));
1626 ASSERT(!(bp->b_flags & XBF_READ));
1629 * If the buffer is already marked delwri it already is queued up
1630 * by someone else for imediate writeout. Just ignore it in that
1633 if (bp->b_flags & _XBF_DELWRI_Q) {
1634 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1638 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1641 * If a buffer gets written out synchronously or marked stale while it
1642 * is on a delwri list we lazily remove it. To do this, the other party
1643 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1644 * It remains referenced and on the list. In a rare corner case it
1645 * might get readded to a delwri list after the synchronous writeout, in
1646 * which case we need just need to re-add the flag here.
1648 bp->b_flags |= _XBF_DELWRI_Q;
1649 if (list_empty(&bp->b_list)) {
1650 atomic_inc(&bp->b_hold);
1651 list_add_tail(&bp->b_list, list);
1658 * Compare function is more complex than it needs to be because
1659 * the return value is only 32 bits and we are doing comparisons
1665 struct list_head *a,
1666 struct list_head *b)
1668 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1669 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1672 diff = ap->b_map.bm_bn - bp->b_map.bm_bn;
1681 __xfs_buf_delwri_submit(
1682 struct list_head *buffer_list,
1683 struct list_head *io_list,
1686 struct blk_plug plug;
1687 struct xfs_buf *bp, *n;
1690 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1692 if (xfs_buf_ispinned(bp)) {
1696 if (!xfs_buf_trylock(bp))
1703 * Someone else might have written the buffer synchronously or
1704 * marked it stale in the meantime. In that case only the
1705 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1706 * reference and remove it from the list here.
1708 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1709 list_del_init(&bp->b_list);
1714 list_move_tail(&bp->b_list, io_list);
1715 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1718 list_sort(NULL, io_list, xfs_buf_cmp);
1720 blk_start_plug(&plug);
1721 list_for_each_entry_safe(bp, n, io_list, b_list) {
1722 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC);
1723 bp->b_flags |= XBF_WRITE;
1726 bp->b_flags |= XBF_ASYNC;
1727 list_del_init(&bp->b_list);
1731 blk_finish_plug(&plug);
1737 * Write out a buffer list asynchronously.
1739 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1740 * out and not wait for I/O completion on any of the buffers. This interface
1741 * is only safely useable for callers that can track I/O completion by higher
1742 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1746 xfs_buf_delwri_submit_nowait(
1747 struct list_head *buffer_list)
1749 LIST_HEAD (io_list);
1750 return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
1754 * Write out a buffer list synchronously.
1756 * This will take the @buffer_list, write all buffers out and wait for I/O
1757 * completion on all of the buffers. @buffer_list is consumed by the function,
1758 * so callers must have some other way of tracking buffers if they require such
1762 xfs_buf_delwri_submit(
1763 struct list_head *buffer_list)
1765 LIST_HEAD (io_list);
1766 int error = 0, error2;
1769 __xfs_buf_delwri_submit(buffer_list, &io_list, true);
1771 /* Wait for IO to complete. */
1772 while (!list_empty(&io_list)) {
1773 bp = list_first_entry(&io_list, struct xfs_buf, b_list);
1775 list_del_init(&bp->b_list);
1776 error2 = xfs_buf_iowait(bp);
1788 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1789 KM_ZONE_HWALIGN, NULL);
1793 xfslogd_workqueue = alloc_workqueue("xfslogd",
1794 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1795 if (!xfslogd_workqueue)
1796 goto out_free_buf_zone;
1801 kmem_zone_destroy(xfs_buf_zone);
1807 xfs_buf_terminate(void)
1809 destroy_workqueue(xfslogd_workqueue);
1810 kmem_zone_destroy(xfs_buf_zone);