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);
99 bp->b_lru_flags &= ~_XBF_LRU_DISPOSE;
101 spin_unlock(&btp->bt_lru_lock);
105 * xfs_buf_lru_del - remove a buffer from the LRU
107 * The unlocked check is safe here because it only occurs when there are not
108 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
109 * to optimise the shrinker removing the buffer from the LRU and calling
110 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
117 struct xfs_buftarg *btp = bp->b_target;
119 if (list_empty(&bp->b_lru))
122 spin_lock(&btp->bt_lru_lock);
123 if (!list_empty(&bp->b_lru)) {
124 list_del_init(&bp->b_lru);
127 spin_unlock(&btp->bt_lru_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;
153 atomic_set(&(bp)->b_lru_ref, 0);
154 if (!list_empty(&bp->b_lru)) {
155 struct xfs_buftarg *btp = bp->b_target;
157 spin_lock(&btp->bt_lru_lock);
158 if (!list_empty(&bp->b_lru) &&
159 !(bp->b_lru_flags & _XBF_LRU_DISPOSE)) {
160 list_del_init(&bp->b_lru);
162 atomic_dec(&bp->b_hold);
164 spin_unlock(&btp->bt_lru_lock);
166 ASSERT(atomic_read(&bp->b_hold) >= 1);
174 ASSERT(bp->b_maps == NULL);
175 bp->b_map_count = map_count;
177 if (map_count == 1) {
178 bp->b_maps = &bp->b_map;
182 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
190 * Frees b_pages if it was allocated.
196 if (bp->b_maps != &bp->b_map) {
197 kmem_free(bp->b_maps);
204 struct xfs_buftarg *target,
205 struct xfs_buf_map *map,
207 xfs_buf_flags_t flags)
213 bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS);
218 * We don't want certain flags to appear in b_flags unless they are
219 * specifically set by later operations on the buffer.
221 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
223 atomic_set(&bp->b_hold, 1);
224 atomic_set(&bp->b_lru_ref, 1);
225 init_completion(&bp->b_iowait);
226 INIT_LIST_HEAD(&bp->b_lru);
227 INIT_LIST_HEAD(&bp->b_list);
228 RB_CLEAR_NODE(&bp->b_rbnode);
229 sema_init(&bp->b_sema, 0); /* held, no waiters */
231 bp->b_target = target;
235 * Set length and io_length to the same value initially.
236 * I/O routines should use io_length, which will be the same in
237 * most cases but may be reset (e.g. XFS recovery).
239 error = xfs_buf_get_maps(bp, nmaps);
241 kmem_zone_free(xfs_buf_zone, bp);
245 bp->b_bn = map[0].bm_bn;
247 for (i = 0; i < nmaps; i++) {
248 bp->b_maps[i].bm_bn = map[i].bm_bn;
249 bp->b_maps[i].bm_len = map[i].bm_len;
250 bp->b_length += map[i].bm_len;
252 bp->b_io_length = bp->b_length;
254 atomic_set(&bp->b_pin_count, 0);
255 init_waitqueue_head(&bp->b_waiters);
257 XFS_STATS_INC(xb_create);
258 trace_xfs_buf_init(bp, _RET_IP_);
264 * Allocate a page array capable of holding a specified number
265 * of pages, and point the page buf at it.
271 xfs_buf_flags_t flags)
273 /* Make sure that we have a page list */
274 if (bp->b_pages == NULL) {
275 bp->b_page_count = page_count;
276 if (page_count <= XB_PAGES) {
277 bp->b_pages = bp->b_page_array;
279 bp->b_pages = kmem_alloc(sizeof(struct page *) *
280 page_count, KM_NOFS);
281 if (bp->b_pages == NULL)
284 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
290 * Frees b_pages if it was allocated.
296 if (bp->b_pages != bp->b_page_array) {
297 kmem_free(bp->b_pages);
303 * Releases the specified buffer.
305 * The modification state of any associated pages is left unchanged.
306 * The buffer most not be on any hash - use xfs_buf_rele instead for
307 * hashed and refcounted buffers
313 trace_xfs_buf_free(bp, _RET_IP_);
315 ASSERT(list_empty(&bp->b_lru));
317 if (bp->b_flags & _XBF_PAGES) {
320 if (xfs_buf_is_vmapped(bp))
321 vm_unmap_ram(bp->b_addr - bp->b_offset,
324 for (i = 0; i < bp->b_page_count; i++) {
325 struct page *page = bp->b_pages[i];
329 } else if (bp->b_flags & _XBF_KMEM)
330 kmem_free(bp->b_addr);
331 _xfs_buf_free_pages(bp);
332 xfs_buf_free_maps(bp);
333 kmem_zone_free(xfs_buf_zone, bp);
337 * Allocates all the pages for buffer in question and builds it's page list.
340 xfs_buf_allocate_memory(
345 size_t nbytes, offset;
346 gfp_t gfp_mask = xb_to_gfp(flags);
347 unsigned short page_count, i;
348 xfs_off_t start, end;
352 * for buffers that are contained within a single page, just allocate
353 * the memory from the heap - there's no need for the complexity of
354 * page arrays to keep allocation down to order 0.
356 size = BBTOB(bp->b_length);
357 if (size < PAGE_SIZE) {
358 bp->b_addr = kmem_alloc(size, KM_NOFS);
360 /* low memory - use alloc_page loop instead */
364 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
365 ((unsigned long)bp->b_addr & PAGE_MASK)) {
366 /* b_addr spans two pages - use alloc_page instead */
367 kmem_free(bp->b_addr);
371 bp->b_offset = offset_in_page(bp->b_addr);
372 bp->b_pages = bp->b_page_array;
373 bp->b_pages[0] = virt_to_page(bp->b_addr);
374 bp->b_page_count = 1;
375 bp->b_flags |= _XBF_KMEM;
380 start = BBTOB(bp->b_map.bm_bn) >> PAGE_SHIFT;
381 end = (BBTOB(bp->b_map.bm_bn + bp->b_length) + PAGE_SIZE - 1)
383 page_count = end - start;
384 error = _xfs_buf_get_pages(bp, page_count, flags);
388 offset = bp->b_offset;
389 bp->b_flags |= _XBF_PAGES;
391 for (i = 0; i < bp->b_page_count; i++) {
395 page = alloc_page(gfp_mask);
396 if (unlikely(page == NULL)) {
397 if (flags & XBF_READ_AHEAD) {
398 bp->b_page_count = i;
404 * This could deadlock.
406 * But until all the XFS lowlevel code is revamped to
407 * handle buffer allocation failures we can't do much.
409 if (!(++retries % 100))
411 "possible memory allocation deadlock in %s (mode:0x%x)",
414 XFS_STATS_INC(xb_page_retries);
415 congestion_wait(BLK_RW_ASYNC, HZ/50);
419 XFS_STATS_INC(xb_page_found);
421 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
423 bp->b_pages[i] = page;
429 for (i = 0; i < bp->b_page_count; i++)
430 __free_page(bp->b_pages[i]);
435 * Map buffer into kernel address-space if necessary.
442 ASSERT(bp->b_flags & _XBF_PAGES);
443 if (bp->b_page_count == 1) {
444 /* A single page buffer is always mappable */
445 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
446 } else if (flags & XBF_UNMAPPED) {
452 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
457 } while (retried++ <= 1);
461 bp->b_addr += bp->b_offset;
468 * Finding and Reading Buffers
472 * Look up, and creates if absent, a lockable buffer for
473 * a given range of an inode. The buffer is returned
474 * locked. No I/O is implied by this call.
478 struct xfs_buftarg *btp,
479 struct xfs_buf_map *map,
481 xfs_buf_flags_t flags,
485 struct xfs_perag *pag;
486 struct rb_node **rbp;
487 struct rb_node *parent;
489 xfs_daddr_t blkno = map[0].bm_bn;
493 for (i = 0; i < nmaps; i++)
494 numblks += map[i].bm_len;
495 numbytes = BBTOB(numblks);
497 /* Check for IOs smaller than the sector size / not sector aligned */
498 ASSERT(!(numbytes < (1 << btp->bt_sshift)));
499 ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_smask));
502 pag = xfs_perag_get(btp->bt_mount,
503 xfs_daddr_to_agno(btp->bt_mount, blkno));
506 spin_lock(&pag->pag_buf_lock);
507 rbp = &pag->pag_buf_tree.rb_node;
512 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
514 if (blkno < bp->b_bn)
515 rbp = &(*rbp)->rb_left;
516 else if (blkno > bp->b_bn)
517 rbp = &(*rbp)->rb_right;
520 * found a block number match. If the range doesn't
521 * match, the only way this is allowed is if the buffer
522 * in the cache is stale and the transaction that made
523 * it stale has not yet committed. i.e. we are
524 * reallocating a busy extent. Skip this buffer and
525 * continue searching to the right for an exact match.
527 if (bp->b_length != numblks) {
528 ASSERT(bp->b_flags & XBF_STALE);
529 rbp = &(*rbp)->rb_right;
532 atomic_inc(&bp->b_hold);
539 rb_link_node(&new_bp->b_rbnode, parent, rbp);
540 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
541 /* the buffer keeps the perag reference until it is freed */
543 spin_unlock(&pag->pag_buf_lock);
545 XFS_STATS_INC(xb_miss_locked);
546 spin_unlock(&pag->pag_buf_lock);
552 spin_unlock(&pag->pag_buf_lock);
555 if (!xfs_buf_trylock(bp)) {
556 if (flags & XBF_TRYLOCK) {
558 XFS_STATS_INC(xb_busy_locked);
562 XFS_STATS_INC(xb_get_locked_waited);
566 * if the buffer is stale, clear all the external state associated with
567 * it. We need to keep flags such as how we allocated the buffer memory
570 if (bp->b_flags & XBF_STALE) {
571 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
572 ASSERT(bp->b_iodone == NULL);
573 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
577 trace_xfs_buf_find(bp, flags, _RET_IP_);
578 XFS_STATS_INC(xb_get_locked);
583 * Assembles a buffer covering the specified range. The code is optimised for
584 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
585 * more hits than misses.
589 struct xfs_buftarg *target,
590 struct xfs_buf_map *map,
592 xfs_buf_flags_t flags)
595 struct xfs_buf *new_bp;
598 bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
602 new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
603 if (unlikely(!new_bp))
606 error = xfs_buf_allocate_memory(new_bp, flags);
608 xfs_buf_free(new_bp);
612 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
614 xfs_buf_free(new_bp);
619 xfs_buf_free(new_bp);
623 error = _xfs_buf_map_pages(bp, flags);
624 if (unlikely(error)) {
625 xfs_warn(target->bt_mount,
626 "%s: failed to map pages\n", __func__);
632 XFS_STATS_INC(xb_get);
633 trace_xfs_buf_get(bp, flags, _RET_IP_);
640 xfs_buf_flags_t flags)
642 ASSERT(!(flags & XBF_WRITE));
643 ASSERT(bp->b_map.bm_bn != XFS_BUF_DADDR_NULL);
645 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
646 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
648 xfs_buf_iorequest(bp);
649 if (flags & XBF_ASYNC)
651 return xfs_buf_iowait(bp);
656 struct xfs_buftarg *target,
657 struct xfs_buf_map *map,
659 xfs_buf_flags_t flags,
660 const struct xfs_buf_ops *ops)
666 bp = xfs_buf_get_map(target, map, nmaps, flags);
668 trace_xfs_buf_read(bp, flags, _RET_IP_);
670 if (!XFS_BUF_ISDONE(bp)) {
671 XFS_STATS_INC(xb_get_read);
673 _xfs_buf_read(bp, flags);
674 } else if (flags & XBF_ASYNC) {
676 * Read ahead call which is already satisfied,
682 /* We do not want read in the flags */
683 bp->b_flags &= ~XBF_READ;
691 * If we are not low on memory then do the readahead in a deadlock
695 xfs_buf_readahead_map(
696 struct xfs_buftarg *target,
697 struct xfs_buf_map *map,
699 const struct xfs_buf_ops *ops)
701 if (bdi_read_congested(target->bt_bdi))
704 xfs_buf_read_map(target, map, nmaps,
705 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops);
709 * Read an uncached buffer from disk. Allocates and returns a locked
710 * buffer containing the disk contents or nothing.
713 xfs_buf_read_uncached(
714 struct xfs_buftarg *target,
718 const struct xfs_buf_ops *ops)
722 bp = xfs_buf_get_uncached(target, numblks, flags);
726 /* set up the buffer for a read IO */
727 ASSERT(bp->b_map_count == 1);
729 bp->b_maps[0].bm_bn = daddr;
730 bp->b_flags |= XBF_READ;
733 xfsbdstrat(target->bt_mount, bp);
739 * Return a buffer allocated as an empty buffer and associated to external
740 * memory via xfs_buf_associate_memory() back to it's empty state.
748 _xfs_buf_free_pages(bp);
751 bp->b_page_count = 0;
753 bp->b_length = numblks;
754 bp->b_io_length = numblks;
756 ASSERT(bp->b_map_count == 1);
757 bp->b_bn = XFS_BUF_DADDR_NULL;
758 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
759 bp->b_maps[0].bm_len = bp->b_length;
762 static inline struct page *
766 if ((!is_vmalloc_addr(addr))) {
767 return virt_to_page(addr);
769 return vmalloc_to_page(addr);
774 xfs_buf_associate_memory(
781 unsigned long pageaddr;
782 unsigned long offset;
786 pageaddr = (unsigned long)mem & PAGE_MASK;
787 offset = (unsigned long)mem - pageaddr;
788 buflen = PAGE_ALIGN(len + offset);
789 page_count = buflen >> PAGE_SHIFT;
791 /* Free any previous set of page pointers */
793 _xfs_buf_free_pages(bp);
798 rval = _xfs_buf_get_pages(bp, page_count, 0);
802 bp->b_offset = offset;
804 for (i = 0; i < bp->b_page_count; i++) {
805 bp->b_pages[i] = mem_to_page((void *)pageaddr);
806 pageaddr += PAGE_SIZE;
809 bp->b_io_length = BTOBB(len);
810 bp->b_length = BTOBB(buflen);
816 xfs_buf_get_uncached(
817 struct xfs_buftarg *target,
821 unsigned long page_count;
824 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
826 bp = _xfs_buf_alloc(target, &map, 1, 0);
827 if (unlikely(bp == NULL))
830 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
831 error = _xfs_buf_get_pages(bp, page_count, 0);
835 for (i = 0; i < page_count; i++) {
836 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
840 bp->b_flags |= _XBF_PAGES;
842 error = _xfs_buf_map_pages(bp, 0);
843 if (unlikely(error)) {
844 xfs_warn(target->bt_mount,
845 "%s: failed to map pages\n", __func__);
849 trace_xfs_buf_get_uncached(bp, _RET_IP_);
854 __free_page(bp->b_pages[i]);
855 _xfs_buf_free_pages(bp);
857 xfs_buf_free_maps(bp);
858 kmem_zone_free(xfs_buf_zone, bp);
864 * Increment reference count on buffer, to hold the buffer concurrently
865 * with another thread which may release (free) the buffer asynchronously.
866 * Must hold the buffer already to call this function.
872 trace_xfs_buf_hold(bp, _RET_IP_);
873 atomic_inc(&bp->b_hold);
877 * Releases a hold on the specified buffer. If the
878 * the hold count is 1, calls xfs_buf_free.
884 struct xfs_perag *pag = bp->b_pag;
886 trace_xfs_buf_rele(bp, _RET_IP_);
889 ASSERT(list_empty(&bp->b_lru));
890 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
891 if (atomic_dec_and_test(&bp->b_hold))
896 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
898 ASSERT(atomic_read(&bp->b_hold) > 0);
899 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
900 if (!(bp->b_flags & XBF_STALE) &&
901 atomic_read(&bp->b_lru_ref)) {
903 spin_unlock(&pag->pag_buf_lock);
906 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
907 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
908 spin_unlock(&pag->pag_buf_lock);
917 * Lock a buffer object, if it is not already locked.
919 * If we come across a stale, pinned, locked buffer, we know that we are
920 * being asked to lock a buffer that has been reallocated. Because it is
921 * pinned, we know that the log has not been pushed to disk and hence it
922 * will still be locked. Rather than continuing to have trylock attempts
923 * fail until someone else pushes the log, push it ourselves before
924 * returning. This means that the xfsaild will not get stuck trying
925 * to push on stale inode buffers.
933 locked = down_trylock(&bp->b_sema) == 0;
936 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
937 xfs_log_force(bp->b_target->bt_mount, 0);
939 trace_xfs_buf_trylock(bp, _RET_IP_);
944 * Lock a buffer object.
946 * If we come across a stale, pinned, locked buffer, we know that we
947 * are being asked to lock a buffer that has been reallocated. Because
948 * it is pinned, we know that the log has not been pushed to disk and
949 * hence it will still be locked. Rather than sleeping until someone
950 * else pushes the log, push it ourselves before trying to get the lock.
956 trace_xfs_buf_lock(bp, _RET_IP_);
958 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
959 xfs_log_force(bp->b_target->bt_mount, 0);
963 trace_xfs_buf_lock_done(bp, _RET_IP_);
973 trace_xfs_buf_unlock(bp, _RET_IP_);
980 DECLARE_WAITQUEUE (wait, current);
982 if (atomic_read(&bp->b_pin_count) == 0)
985 add_wait_queue(&bp->b_waiters, &wait);
987 set_current_state(TASK_UNINTERRUPTIBLE);
988 if (atomic_read(&bp->b_pin_count) == 0)
992 remove_wait_queue(&bp->b_waiters, &wait);
993 set_current_state(TASK_RUNNING);
997 * Buffer Utility Routines
1001 xfs_buf_iodone_work(
1002 struct work_struct *work)
1004 struct xfs_buf *bp =
1005 container_of(work, xfs_buf_t, b_iodone_work);
1006 bool read = !!(bp->b_flags & XBF_READ);
1008 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1009 if (read && bp->b_ops)
1010 bp->b_ops->verify_read(bp);
1013 (*(bp->b_iodone))(bp);
1014 else if (bp->b_flags & XBF_ASYNC)
1017 ASSERT(read && bp->b_ops);
1018 complete(&bp->b_iowait);
1027 bool read = !!(bp->b_flags & XBF_READ);
1029 trace_xfs_buf_iodone(bp, _RET_IP_);
1031 if (bp->b_error == 0)
1032 bp->b_flags |= XBF_DONE;
1034 if (bp->b_iodone || (read && bp->b_ops) || (bp->b_flags & XBF_ASYNC)) {
1036 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1037 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1039 xfs_buf_iodone_work(&bp->b_iodone_work);
1042 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1043 complete(&bp->b_iowait);
1052 ASSERT(error >= 0 && error <= 0xffff);
1053 bp->b_error = (unsigned short)error;
1054 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1058 xfs_buf_ioerror_alert(
1062 xfs_alert(bp->b_target->bt_mount,
1063 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1064 (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length);
1068 * Called when we want to stop a buffer from getting written or read.
1069 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1070 * so that the proper iodone callbacks get called.
1076 #ifdef XFSERRORDEBUG
1077 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1081 * No need to wait until the buffer is unpinned, we aren't flushing it.
1083 xfs_buf_ioerror(bp, EIO);
1086 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1092 xfs_buf_ioend(bp, 0);
1098 * Same as xfs_bioerror, except that we are releasing the buffer
1099 * here ourselves, and avoiding the xfs_buf_ioend call.
1100 * This is meant for userdata errors; metadata bufs come with
1101 * iodone functions attached, so that we can track down errors.
1107 int64_t fl = bp->b_flags;
1109 * No need to wait until the buffer is unpinned.
1110 * We aren't flushing it.
1112 * chunkhold expects B_DONE to be set, whether
1113 * we actually finish the I/O or not. We don't want to
1114 * change that interface.
1119 bp->b_iodone = NULL;
1120 if (!(fl & XBF_ASYNC)) {
1122 * Mark b_error and B_ERROR _both_.
1123 * Lot's of chunkcache code assumes that.
1124 * There's no reason to mark error for
1127 xfs_buf_ioerror(bp, EIO);
1128 complete(&bp->b_iowait);
1140 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1141 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1143 * Metadata write that didn't get logged but
1144 * written delayed anyway. These aren't associated
1145 * with a transaction, and can be ignored.
1147 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1148 return xfs_bioerror_relse(bp);
1150 return xfs_bioerror(bp);
1153 xfs_buf_iorequest(bp);
1163 ASSERT(xfs_buf_islocked(bp));
1165 bp->b_flags |= XBF_WRITE;
1166 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q);
1170 error = xfs_buf_iowait(bp);
1172 xfs_force_shutdown(bp->b_target->bt_mount,
1173 SHUTDOWN_META_IO_ERROR);
1179 * Wrapper around bdstrat so that we can stop data from going to disk in case
1180 * we are shutting down the filesystem. Typically user data goes thru this
1181 * path; one of the exceptions is the superblock.
1185 struct xfs_mount *mp,
1188 if (XFS_FORCED_SHUTDOWN(mp)) {
1189 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1190 xfs_bioerror_relse(bp);
1194 xfs_buf_iorequest(bp);
1202 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1203 xfs_buf_ioend(bp, schedule);
1211 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1214 * don't overwrite existing errors - otherwise we can lose errors on
1215 * buffers that require multiple bios to complete.
1218 xfs_buf_ioerror(bp, -error);
1220 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1221 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1223 _xfs_buf_ioend(bp, 1);
1228 xfs_buf_ioapply_map(
1236 int total_nr_pages = bp->b_page_count;
1239 sector_t sector = bp->b_maps[map].bm_bn;
1243 total_nr_pages = bp->b_page_count;
1245 /* skip the pages in the buffer before the start offset */
1247 offset = *buf_offset;
1248 while (offset >= PAGE_SIZE) {
1250 offset -= PAGE_SIZE;
1254 * Limit the IO size to the length of the current vector, and update the
1255 * remaining IO count for the next time around.
1257 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1259 *buf_offset += size;
1262 atomic_inc(&bp->b_io_remaining);
1263 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1264 if (nr_pages > total_nr_pages)
1265 nr_pages = total_nr_pages;
1267 bio = bio_alloc(GFP_NOIO, nr_pages);
1268 bio->bi_bdev = bp->b_target->bt_bdev;
1269 bio->bi_sector = sector;
1270 bio->bi_end_io = xfs_buf_bio_end_io;
1271 bio->bi_private = bp;
1274 for (; size && nr_pages; nr_pages--, page_index++) {
1275 int rbytes, nbytes = PAGE_SIZE - offset;
1280 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1282 if (rbytes < nbytes)
1286 sector += BTOBB(nbytes);
1291 if (likely(bio->bi_size)) {
1292 if (xfs_buf_is_vmapped(bp)) {
1293 flush_kernel_vmap_range(bp->b_addr,
1294 xfs_buf_vmap_len(bp));
1296 submit_bio(rw, bio);
1301 * This is guaranteed not to be the last io reference count
1302 * because the caller (xfs_buf_iorequest) holds a count itself.
1304 atomic_dec(&bp->b_io_remaining);
1305 xfs_buf_ioerror(bp, EIO);
1315 struct blk_plug plug;
1321 if (bp->b_flags & XBF_WRITE) {
1322 if (bp->b_flags & XBF_SYNCIO)
1326 if (bp->b_flags & XBF_FUA)
1328 if (bp->b_flags & XBF_FLUSH)
1332 * Run the write verifier callback function if it exists. If
1333 * this function fails it will mark the buffer with an error and
1334 * the IO should not be dispatched.
1337 bp->b_ops->verify_write(bp);
1339 xfs_force_shutdown(bp->b_target->bt_mount,
1340 SHUTDOWN_CORRUPT_INCORE);
1344 } else if (bp->b_flags & XBF_READ_AHEAD) {
1350 /* we only use the buffer cache for meta-data */
1354 * Walk all the vectors issuing IO on them. Set up the initial offset
1355 * into the buffer and the desired IO size before we start -
1356 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1359 offset = bp->b_offset;
1360 size = BBTOB(bp->b_io_length);
1361 blk_start_plug(&plug);
1362 for (i = 0; i < bp->b_map_count; i++) {
1363 xfs_buf_ioapply_map(bp, i, &offset, &size, rw);
1367 break; /* all done */
1369 blk_finish_plug(&plug);
1376 trace_xfs_buf_iorequest(bp, _RET_IP_);
1378 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1380 if (bp->b_flags & XBF_WRITE)
1381 xfs_buf_wait_unpin(bp);
1384 /* Set the count to 1 initially, this will stop an I/O
1385 * completion callout which happens before we have started
1386 * all the I/O from calling xfs_buf_ioend too early.
1388 atomic_set(&bp->b_io_remaining, 1);
1389 _xfs_buf_ioapply(bp);
1390 _xfs_buf_ioend(bp, 1);
1396 * Waits for I/O to complete on the buffer supplied. It returns immediately if
1397 * no I/O is pending or there is already a pending error on the buffer. It
1398 * returns the I/O error code, if any, or 0 if there was no error.
1404 trace_xfs_buf_iowait(bp, _RET_IP_);
1407 wait_for_completion(&bp->b_iowait);
1409 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1421 return bp->b_addr + offset;
1423 offset += bp->b_offset;
1424 page = bp->b_pages[offset >> PAGE_SHIFT];
1425 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1429 * Move data into or out of a buffer.
1433 xfs_buf_t *bp, /* buffer to process */
1434 size_t boff, /* starting buffer offset */
1435 size_t bsize, /* length to copy */
1436 void *data, /* data address */
1437 xfs_buf_rw_t mode) /* read/write/zero flag */
1441 bend = boff + bsize;
1442 while (boff < bend) {
1444 int page_index, page_offset, csize;
1446 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1447 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1448 page = bp->b_pages[page_index];
1449 csize = min_t(size_t, PAGE_SIZE - page_offset,
1450 BBTOB(bp->b_io_length) - boff);
1452 ASSERT((csize + page_offset) <= PAGE_SIZE);
1456 memset(page_address(page) + page_offset, 0, csize);
1459 memcpy(data, page_address(page) + page_offset, csize);
1462 memcpy(page_address(page) + page_offset, data, csize);
1471 * Handling of buffer targets (buftargs).
1475 * Wait for any bufs with callbacks that have been submitted but have not yet
1476 * returned. These buffers will have an elevated hold count, so wait on those
1477 * while freeing all the buffers only held by the LRU.
1481 struct xfs_buftarg *btp)
1486 spin_lock(&btp->bt_lru_lock);
1487 while (!list_empty(&btp->bt_lru)) {
1488 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1489 if (atomic_read(&bp->b_hold) > 1) {
1490 spin_unlock(&btp->bt_lru_lock);
1495 * clear the LRU reference count so the buffer doesn't get
1496 * ignored in xfs_buf_rele().
1498 atomic_set(&bp->b_lru_ref, 0);
1499 spin_unlock(&btp->bt_lru_lock);
1501 spin_lock(&btp->bt_lru_lock);
1503 spin_unlock(&btp->bt_lru_lock);
1508 struct shrinker *shrink,
1509 struct shrink_control *sc)
1511 struct xfs_buftarg *btp = container_of(shrink,
1512 struct xfs_buftarg, bt_shrinker);
1514 int nr_to_scan = sc->nr_to_scan;
1518 return btp->bt_lru_nr;
1520 spin_lock(&btp->bt_lru_lock);
1521 while (!list_empty(&btp->bt_lru)) {
1522 if (nr_to_scan-- <= 0)
1525 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1528 * Decrement the b_lru_ref count unless the value is already
1529 * zero. If the value is already zero, we need to reclaim the
1530 * buffer, otherwise it gets another trip through the LRU.
1532 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1533 list_move_tail(&bp->b_lru, &btp->bt_lru);
1538 * remove the buffer from the LRU now to avoid needing another
1539 * lock round trip inside xfs_buf_rele().
1541 list_move(&bp->b_lru, &dispose);
1543 bp->b_lru_flags |= _XBF_LRU_DISPOSE;
1545 spin_unlock(&btp->bt_lru_lock);
1547 while (!list_empty(&dispose)) {
1548 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1549 list_del_init(&bp->b_lru);
1553 return btp->bt_lru_nr;
1558 struct xfs_mount *mp,
1559 struct xfs_buftarg *btp)
1561 unregister_shrinker(&btp->bt_shrinker);
1563 if (mp->m_flags & XFS_MOUNT_BARRIER)
1564 xfs_blkdev_issue_flush(btp);
1570 xfs_setsize_buftarg_flags(
1572 unsigned int blocksize,
1573 unsigned int sectorsize,
1576 btp->bt_bsize = blocksize;
1577 btp->bt_sshift = ffs(sectorsize) - 1;
1578 btp->bt_smask = sectorsize - 1;
1580 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1581 char name[BDEVNAME_SIZE];
1583 bdevname(btp->bt_bdev, name);
1585 xfs_warn(btp->bt_mount,
1586 "Cannot set_blocksize to %u on device %s\n",
1595 * When allocating the initial buffer target we have not yet
1596 * read in the superblock, so don't know what sized sectors
1597 * are being used is at this early stage. Play safe.
1600 xfs_setsize_buftarg_early(
1602 struct block_device *bdev)
1604 return xfs_setsize_buftarg_flags(btp,
1605 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1609 xfs_setsize_buftarg(
1611 unsigned int blocksize,
1612 unsigned int sectorsize)
1614 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1619 struct xfs_mount *mp,
1620 struct block_device *bdev,
1626 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1629 btp->bt_dev = bdev->bd_dev;
1630 btp->bt_bdev = bdev;
1631 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1635 INIT_LIST_HEAD(&btp->bt_lru);
1636 spin_lock_init(&btp->bt_lru_lock);
1637 if (xfs_setsize_buftarg_early(btp, bdev))
1639 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1640 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1641 register_shrinker(&btp->bt_shrinker);
1650 * Add a buffer to the delayed write list.
1652 * This queues a buffer for writeout if it hasn't already been. Note that
1653 * neither this routine nor the buffer list submission functions perform
1654 * any internal synchronization. It is expected that the lists are thread-local
1657 * Returns true if we queued up the buffer, or false if it already had
1658 * been on the buffer list.
1661 xfs_buf_delwri_queue(
1663 struct list_head *list)
1665 ASSERT(xfs_buf_islocked(bp));
1666 ASSERT(!(bp->b_flags & XBF_READ));
1669 * If the buffer is already marked delwri it already is queued up
1670 * by someone else for imediate writeout. Just ignore it in that
1673 if (bp->b_flags & _XBF_DELWRI_Q) {
1674 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1678 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1681 * If a buffer gets written out synchronously or marked stale while it
1682 * is on a delwri list we lazily remove it. To do this, the other party
1683 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1684 * It remains referenced and on the list. In a rare corner case it
1685 * might get readded to a delwri list after the synchronous writeout, in
1686 * which case we need just need to re-add the flag here.
1688 bp->b_flags |= _XBF_DELWRI_Q;
1689 if (list_empty(&bp->b_list)) {
1690 atomic_inc(&bp->b_hold);
1691 list_add_tail(&bp->b_list, list);
1698 * Compare function is more complex than it needs to be because
1699 * the return value is only 32 bits and we are doing comparisons
1705 struct list_head *a,
1706 struct list_head *b)
1708 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1709 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1712 diff = ap->b_map.bm_bn - bp->b_map.bm_bn;
1721 __xfs_buf_delwri_submit(
1722 struct list_head *buffer_list,
1723 struct list_head *io_list,
1726 struct blk_plug plug;
1727 struct xfs_buf *bp, *n;
1730 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1732 if (xfs_buf_ispinned(bp)) {
1736 if (!xfs_buf_trylock(bp))
1743 * Someone else might have written the buffer synchronously or
1744 * marked it stale in the meantime. In that case only the
1745 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1746 * reference and remove it from the list here.
1748 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1749 list_del_init(&bp->b_list);
1754 list_move_tail(&bp->b_list, io_list);
1755 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1758 list_sort(NULL, io_list, xfs_buf_cmp);
1760 blk_start_plug(&plug);
1761 list_for_each_entry_safe(bp, n, io_list, b_list) {
1762 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC);
1763 bp->b_flags |= XBF_WRITE;
1766 bp->b_flags |= XBF_ASYNC;
1767 list_del_init(&bp->b_list);
1771 blk_finish_plug(&plug);
1777 * Write out a buffer list asynchronously.
1779 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1780 * out and not wait for I/O completion on any of the buffers. This interface
1781 * is only safely useable for callers that can track I/O completion by higher
1782 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1786 xfs_buf_delwri_submit_nowait(
1787 struct list_head *buffer_list)
1789 LIST_HEAD (io_list);
1790 return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
1794 * Write out a buffer list synchronously.
1796 * This will take the @buffer_list, write all buffers out and wait for I/O
1797 * completion on all of the buffers. @buffer_list is consumed by the function,
1798 * so callers must have some other way of tracking buffers if they require such
1802 xfs_buf_delwri_submit(
1803 struct list_head *buffer_list)
1805 LIST_HEAD (io_list);
1806 int error = 0, error2;
1809 __xfs_buf_delwri_submit(buffer_list, &io_list, true);
1811 /* Wait for IO to complete. */
1812 while (!list_empty(&io_list)) {
1813 bp = list_first_entry(&io_list, struct xfs_buf, b_list);
1815 list_del_init(&bp->b_list);
1816 error2 = xfs_buf_iowait(bp);
1828 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1829 KM_ZONE_HWALIGN, NULL);
1833 xfslogd_workqueue = alloc_workqueue("xfslogd",
1834 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1835 if (!xfslogd_workqueue)
1836 goto out_free_buf_zone;
1841 kmem_zone_destroy(xfs_buf_zone);
1847 xfs_buf_terminate(void)
1849 destroy_workqueue(xfslogd_workqueue);
1850 kmem_zone_destroy(xfs_buf_zone);