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_maps[0].bm_bn) >> PAGE_SHIFT;
381 end = (BBTOB(bp->b_maps[0].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;
494 for (i = 0; i < nmaps; i++)
495 numblks += map[i].bm_len;
496 numbytes = BBTOB(numblks);
498 /* Check for IOs smaller than the sector size / not sector aligned */
499 ASSERT(!(numbytes < (1 << btp->bt_sshift)));
500 ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_smask));
503 * Corrupted block numbers can get through to here, unfortunately, so we
504 * have to check that the buffer falls within the filesystem bounds.
506 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
509 * XXX (dgc): we should really be returning EFSCORRUPTED here,
510 * but none of the higher level infrastructure supports
511 * returning a specific error on buffer lookup failures.
513 xfs_alert(btp->bt_mount,
514 "%s: Block out of range: block 0x%llx, EOFS 0x%llx ",
515 __func__, blkno, eofs);
520 pag = xfs_perag_get(btp->bt_mount,
521 xfs_daddr_to_agno(btp->bt_mount, blkno));
524 spin_lock(&pag->pag_buf_lock);
525 rbp = &pag->pag_buf_tree.rb_node;
530 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
532 if (blkno < bp->b_bn)
533 rbp = &(*rbp)->rb_left;
534 else if (blkno > bp->b_bn)
535 rbp = &(*rbp)->rb_right;
538 * found a block number match. If the range doesn't
539 * match, the only way this is allowed is if the buffer
540 * in the cache is stale and the transaction that made
541 * it stale has not yet committed. i.e. we are
542 * reallocating a busy extent. Skip this buffer and
543 * continue searching to the right for an exact match.
545 if (bp->b_length != numblks) {
546 ASSERT(bp->b_flags & XBF_STALE);
547 rbp = &(*rbp)->rb_right;
550 atomic_inc(&bp->b_hold);
557 rb_link_node(&new_bp->b_rbnode, parent, rbp);
558 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
559 /* the buffer keeps the perag reference until it is freed */
561 spin_unlock(&pag->pag_buf_lock);
563 XFS_STATS_INC(xb_miss_locked);
564 spin_unlock(&pag->pag_buf_lock);
570 spin_unlock(&pag->pag_buf_lock);
573 if (!xfs_buf_trylock(bp)) {
574 if (flags & XBF_TRYLOCK) {
576 XFS_STATS_INC(xb_busy_locked);
580 XFS_STATS_INC(xb_get_locked_waited);
584 * if the buffer is stale, clear all the external state associated with
585 * it. We need to keep flags such as how we allocated the buffer memory
588 if (bp->b_flags & XBF_STALE) {
589 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
590 ASSERT(bp->b_iodone == NULL);
591 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
595 trace_xfs_buf_find(bp, flags, _RET_IP_);
596 XFS_STATS_INC(xb_get_locked);
601 * Assembles a buffer covering the specified range. The code is optimised for
602 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
603 * more hits than misses.
607 struct xfs_buftarg *target,
608 struct xfs_buf_map *map,
610 xfs_buf_flags_t flags)
613 struct xfs_buf *new_bp;
616 bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
620 new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
621 if (unlikely(!new_bp))
624 error = xfs_buf_allocate_memory(new_bp, flags);
626 xfs_buf_free(new_bp);
630 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
632 xfs_buf_free(new_bp);
637 xfs_buf_free(new_bp);
641 error = _xfs_buf_map_pages(bp, flags);
642 if (unlikely(error)) {
643 xfs_warn(target->bt_mount,
644 "%s: failed to map pages\n", __func__);
650 XFS_STATS_INC(xb_get);
651 trace_xfs_buf_get(bp, flags, _RET_IP_);
658 xfs_buf_flags_t flags)
660 ASSERT(!(flags & XBF_WRITE));
661 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
663 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
664 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
666 xfs_buf_iorequest(bp);
667 if (flags & XBF_ASYNC)
669 return xfs_buf_iowait(bp);
674 struct xfs_buftarg *target,
675 struct xfs_buf_map *map,
677 xfs_buf_flags_t flags,
678 const struct xfs_buf_ops *ops)
684 bp = xfs_buf_get_map(target, map, nmaps, flags);
686 trace_xfs_buf_read(bp, flags, _RET_IP_);
688 if (!XFS_BUF_ISDONE(bp)) {
689 XFS_STATS_INC(xb_get_read);
691 _xfs_buf_read(bp, flags);
692 } else if (flags & XBF_ASYNC) {
694 * Read ahead call which is already satisfied,
700 /* We do not want read in the flags */
701 bp->b_flags &= ~XBF_READ;
709 * If we are not low on memory then do the readahead in a deadlock
713 xfs_buf_readahead_map(
714 struct xfs_buftarg *target,
715 struct xfs_buf_map *map,
717 const struct xfs_buf_ops *ops)
719 if (bdi_read_congested(target->bt_bdi))
722 xfs_buf_read_map(target, map, nmaps,
723 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops);
727 * Read an uncached buffer from disk. Allocates and returns a locked
728 * buffer containing the disk contents or nothing.
731 xfs_buf_read_uncached(
732 struct xfs_buftarg *target,
736 const struct xfs_buf_ops *ops)
740 bp = xfs_buf_get_uncached(target, numblks, flags);
744 /* set up the buffer for a read IO */
745 ASSERT(bp->b_map_count == 1);
747 bp->b_maps[0].bm_bn = daddr;
748 bp->b_flags |= XBF_READ;
751 xfsbdstrat(target->bt_mount, bp);
757 * Return a buffer allocated as an empty buffer and associated to external
758 * memory via xfs_buf_associate_memory() back to it's empty state.
766 _xfs_buf_free_pages(bp);
769 bp->b_page_count = 0;
771 bp->b_length = numblks;
772 bp->b_io_length = numblks;
774 ASSERT(bp->b_map_count == 1);
775 bp->b_bn = XFS_BUF_DADDR_NULL;
776 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
777 bp->b_maps[0].bm_len = bp->b_length;
780 static inline struct page *
784 if ((!is_vmalloc_addr(addr))) {
785 return virt_to_page(addr);
787 return vmalloc_to_page(addr);
792 xfs_buf_associate_memory(
799 unsigned long pageaddr;
800 unsigned long offset;
804 pageaddr = (unsigned long)mem & PAGE_MASK;
805 offset = (unsigned long)mem - pageaddr;
806 buflen = PAGE_ALIGN(len + offset);
807 page_count = buflen >> PAGE_SHIFT;
809 /* Free any previous set of page pointers */
811 _xfs_buf_free_pages(bp);
816 rval = _xfs_buf_get_pages(bp, page_count, 0);
820 bp->b_offset = offset;
822 for (i = 0; i < bp->b_page_count; i++) {
823 bp->b_pages[i] = mem_to_page((void *)pageaddr);
824 pageaddr += PAGE_SIZE;
827 bp->b_io_length = BTOBB(len);
828 bp->b_length = BTOBB(buflen);
834 xfs_buf_get_uncached(
835 struct xfs_buftarg *target,
839 unsigned long page_count;
842 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
844 bp = _xfs_buf_alloc(target, &map, 1, 0);
845 if (unlikely(bp == NULL))
848 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
849 error = _xfs_buf_get_pages(bp, page_count, 0);
853 for (i = 0; i < page_count; i++) {
854 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
858 bp->b_flags |= _XBF_PAGES;
860 error = _xfs_buf_map_pages(bp, 0);
861 if (unlikely(error)) {
862 xfs_warn(target->bt_mount,
863 "%s: failed to map pages\n", __func__);
867 trace_xfs_buf_get_uncached(bp, _RET_IP_);
872 __free_page(bp->b_pages[i]);
873 _xfs_buf_free_pages(bp);
875 xfs_buf_free_maps(bp);
876 kmem_zone_free(xfs_buf_zone, bp);
882 * Increment reference count on buffer, to hold the buffer concurrently
883 * with another thread which may release (free) the buffer asynchronously.
884 * Must hold the buffer already to call this function.
890 trace_xfs_buf_hold(bp, _RET_IP_);
891 atomic_inc(&bp->b_hold);
895 * Releases a hold on the specified buffer. If the
896 * the hold count is 1, calls xfs_buf_free.
902 struct xfs_perag *pag = bp->b_pag;
904 trace_xfs_buf_rele(bp, _RET_IP_);
907 ASSERT(list_empty(&bp->b_lru));
908 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
909 if (atomic_dec_and_test(&bp->b_hold))
914 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
916 ASSERT(atomic_read(&bp->b_hold) > 0);
917 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
918 if (!(bp->b_flags & XBF_STALE) &&
919 atomic_read(&bp->b_lru_ref)) {
921 spin_unlock(&pag->pag_buf_lock);
924 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
925 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
926 spin_unlock(&pag->pag_buf_lock);
935 * Lock a buffer object, if it is not already locked.
937 * If we come across a stale, pinned, locked buffer, we know that we are
938 * being asked to lock a buffer that has been reallocated. Because it is
939 * pinned, we know that the log has not been pushed to disk and hence it
940 * will still be locked. Rather than continuing to have trylock attempts
941 * fail until someone else pushes the log, push it ourselves before
942 * returning. This means that the xfsaild will not get stuck trying
943 * to push on stale inode buffers.
951 locked = down_trylock(&bp->b_sema) == 0;
954 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
955 xfs_log_force(bp->b_target->bt_mount, 0);
957 trace_xfs_buf_trylock(bp, _RET_IP_);
962 * Lock a buffer object.
964 * If we come across a stale, pinned, locked buffer, we know that we
965 * are being asked to lock a buffer that has been reallocated. Because
966 * it is pinned, we know that the log has not been pushed to disk and
967 * hence it will still be locked. Rather than sleeping until someone
968 * else pushes the log, push it ourselves before trying to get the lock.
974 trace_xfs_buf_lock(bp, _RET_IP_);
976 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
977 xfs_log_force(bp->b_target->bt_mount, 0);
981 trace_xfs_buf_lock_done(bp, _RET_IP_);
991 trace_xfs_buf_unlock(bp, _RET_IP_);
998 DECLARE_WAITQUEUE (wait, current);
1000 if (atomic_read(&bp->b_pin_count) == 0)
1003 add_wait_queue(&bp->b_waiters, &wait);
1005 set_current_state(TASK_UNINTERRUPTIBLE);
1006 if (atomic_read(&bp->b_pin_count) == 0)
1010 remove_wait_queue(&bp->b_waiters, &wait);
1011 set_current_state(TASK_RUNNING);
1015 * Buffer Utility Routines
1019 xfs_buf_iodone_work(
1020 struct work_struct *work)
1022 struct xfs_buf *bp =
1023 container_of(work, xfs_buf_t, b_iodone_work);
1024 bool read = !!(bp->b_flags & XBF_READ);
1026 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1027 if (read && bp->b_ops)
1028 bp->b_ops->verify_read(bp);
1031 (*(bp->b_iodone))(bp);
1032 else if (bp->b_flags & XBF_ASYNC)
1035 ASSERT(read && bp->b_ops);
1036 complete(&bp->b_iowait);
1045 bool read = !!(bp->b_flags & XBF_READ);
1047 trace_xfs_buf_iodone(bp, _RET_IP_);
1049 if (bp->b_error == 0)
1050 bp->b_flags |= XBF_DONE;
1052 if (bp->b_iodone || (read && bp->b_ops) || (bp->b_flags & XBF_ASYNC)) {
1054 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1055 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1057 xfs_buf_iodone_work(&bp->b_iodone_work);
1060 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1061 complete(&bp->b_iowait);
1070 ASSERT(error >= 0 && error <= 0xffff);
1071 bp->b_error = (unsigned short)error;
1072 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1076 xfs_buf_ioerror_alert(
1080 xfs_alert(bp->b_target->bt_mount,
1081 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1082 (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length);
1086 * Called when we want to stop a buffer from getting written or read.
1087 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1088 * so that the proper iodone callbacks get called.
1094 #ifdef XFSERRORDEBUG
1095 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1099 * No need to wait until the buffer is unpinned, we aren't flushing it.
1101 xfs_buf_ioerror(bp, EIO);
1104 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1110 xfs_buf_ioend(bp, 0);
1116 * Same as xfs_bioerror, except that we are releasing the buffer
1117 * here ourselves, and avoiding the xfs_buf_ioend call.
1118 * This is meant for userdata errors; metadata bufs come with
1119 * iodone functions attached, so that we can track down errors.
1125 int64_t fl = bp->b_flags;
1127 * No need to wait until the buffer is unpinned.
1128 * We aren't flushing it.
1130 * chunkhold expects B_DONE to be set, whether
1131 * we actually finish the I/O or not. We don't want to
1132 * change that interface.
1137 bp->b_iodone = NULL;
1138 if (!(fl & XBF_ASYNC)) {
1140 * Mark b_error and B_ERROR _both_.
1141 * Lot's of chunkcache code assumes that.
1142 * There's no reason to mark error for
1145 xfs_buf_ioerror(bp, EIO);
1146 complete(&bp->b_iowait);
1158 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1159 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1161 * Metadata write that didn't get logged but
1162 * written delayed anyway. These aren't associated
1163 * with a transaction, and can be ignored.
1165 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1166 return xfs_bioerror_relse(bp);
1168 return xfs_bioerror(bp);
1171 xfs_buf_iorequest(bp);
1181 ASSERT(xfs_buf_islocked(bp));
1183 bp->b_flags |= XBF_WRITE;
1184 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q);
1188 error = xfs_buf_iowait(bp);
1190 xfs_force_shutdown(bp->b_target->bt_mount,
1191 SHUTDOWN_META_IO_ERROR);
1197 * Wrapper around bdstrat so that we can stop data from going to disk in case
1198 * we are shutting down the filesystem. Typically user data goes thru this
1199 * path; one of the exceptions is the superblock.
1203 struct xfs_mount *mp,
1206 if (XFS_FORCED_SHUTDOWN(mp)) {
1207 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1208 xfs_bioerror_relse(bp);
1212 xfs_buf_iorequest(bp);
1220 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1221 xfs_buf_ioend(bp, schedule);
1229 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1232 * don't overwrite existing errors - otherwise we can lose errors on
1233 * buffers that require multiple bios to complete.
1236 xfs_buf_ioerror(bp, -error);
1238 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1239 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1241 _xfs_buf_ioend(bp, 1);
1246 xfs_buf_ioapply_map(
1254 int total_nr_pages = bp->b_page_count;
1257 sector_t sector = bp->b_maps[map].bm_bn;
1261 total_nr_pages = bp->b_page_count;
1263 /* skip the pages in the buffer before the start offset */
1265 offset = *buf_offset;
1266 while (offset >= PAGE_SIZE) {
1268 offset -= PAGE_SIZE;
1272 * Limit the IO size to the length of the current vector, and update the
1273 * remaining IO count for the next time around.
1275 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1277 *buf_offset += size;
1280 atomic_inc(&bp->b_io_remaining);
1281 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1282 if (nr_pages > total_nr_pages)
1283 nr_pages = total_nr_pages;
1285 bio = bio_alloc(GFP_NOIO, nr_pages);
1286 bio->bi_bdev = bp->b_target->bt_bdev;
1287 bio->bi_sector = sector;
1288 bio->bi_end_io = xfs_buf_bio_end_io;
1289 bio->bi_private = bp;
1292 for (; size && nr_pages; nr_pages--, page_index++) {
1293 int rbytes, nbytes = PAGE_SIZE - offset;
1298 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1300 if (rbytes < nbytes)
1304 sector += BTOBB(nbytes);
1309 if (likely(bio->bi_size)) {
1310 if (xfs_buf_is_vmapped(bp)) {
1311 flush_kernel_vmap_range(bp->b_addr,
1312 xfs_buf_vmap_len(bp));
1314 submit_bio(rw, bio);
1319 * This is guaranteed not to be the last io reference count
1320 * because the caller (xfs_buf_iorequest) holds a count itself.
1322 atomic_dec(&bp->b_io_remaining);
1323 xfs_buf_ioerror(bp, EIO);
1333 struct blk_plug plug;
1339 if (bp->b_flags & XBF_WRITE) {
1340 if (bp->b_flags & XBF_SYNCIO)
1344 if (bp->b_flags & XBF_FUA)
1346 if (bp->b_flags & XBF_FLUSH)
1350 * Run the write verifier callback function if it exists. If
1351 * this function fails it will mark the buffer with an error and
1352 * the IO should not be dispatched.
1355 bp->b_ops->verify_write(bp);
1357 xfs_force_shutdown(bp->b_target->bt_mount,
1358 SHUTDOWN_CORRUPT_INCORE);
1362 } else if (bp->b_flags & XBF_READ_AHEAD) {
1368 /* we only use the buffer cache for meta-data */
1372 * Walk all the vectors issuing IO on them. Set up the initial offset
1373 * into the buffer and the desired IO size before we start -
1374 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1377 offset = bp->b_offset;
1378 size = BBTOB(bp->b_io_length);
1379 blk_start_plug(&plug);
1380 for (i = 0; i < bp->b_map_count; i++) {
1381 xfs_buf_ioapply_map(bp, i, &offset, &size, rw);
1385 break; /* all done */
1387 blk_finish_plug(&plug);
1394 trace_xfs_buf_iorequest(bp, _RET_IP_);
1396 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1398 if (bp->b_flags & XBF_WRITE)
1399 xfs_buf_wait_unpin(bp);
1402 /* Set the count to 1 initially, this will stop an I/O
1403 * completion callout which happens before we have started
1404 * all the I/O from calling xfs_buf_ioend too early.
1406 atomic_set(&bp->b_io_remaining, 1);
1407 _xfs_buf_ioapply(bp);
1408 _xfs_buf_ioend(bp, 1);
1414 * Waits for I/O to complete on the buffer supplied. It returns immediately if
1415 * no I/O is pending or there is already a pending error on the buffer. It
1416 * returns the I/O error code, if any, or 0 if there was no error.
1422 trace_xfs_buf_iowait(bp, _RET_IP_);
1425 wait_for_completion(&bp->b_iowait);
1427 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1439 return bp->b_addr + offset;
1441 offset += bp->b_offset;
1442 page = bp->b_pages[offset >> PAGE_SHIFT];
1443 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1447 * Move data into or out of a buffer.
1451 xfs_buf_t *bp, /* buffer to process */
1452 size_t boff, /* starting buffer offset */
1453 size_t bsize, /* length to copy */
1454 void *data, /* data address */
1455 xfs_buf_rw_t mode) /* read/write/zero flag */
1459 bend = boff + bsize;
1460 while (boff < bend) {
1462 int page_index, page_offset, csize;
1464 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1465 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1466 page = bp->b_pages[page_index];
1467 csize = min_t(size_t, PAGE_SIZE - page_offset,
1468 BBTOB(bp->b_io_length) - boff);
1470 ASSERT((csize + page_offset) <= PAGE_SIZE);
1474 memset(page_address(page) + page_offset, 0, csize);
1477 memcpy(data, page_address(page) + page_offset, csize);
1480 memcpy(page_address(page) + page_offset, data, csize);
1489 * Handling of buffer targets (buftargs).
1493 * Wait for any bufs with callbacks that have been submitted but have not yet
1494 * returned. These buffers will have an elevated hold count, so wait on those
1495 * while freeing all the buffers only held by the LRU.
1499 struct xfs_buftarg *btp)
1504 spin_lock(&btp->bt_lru_lock);
1505 while (!list_empty(&btp->bt_lru)) {
1506 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1507 if (atomic_read(&bp->b_hold) > 1) {
1508 trace_xfs_buf_wait_buftarg(bp, _RET_IP_);
1509 list_move_tail(&bp->b_lru, &btp->bt_lru);
1510 spin_unlock(&btp->bt_lru_lock);
1515 * clear the LRU reference count so the buffer doesn't get
1516 * ignored in xfs_buf_rele().
1518 atomic_set(&bp->b_lru_ref, 0);
1519 spin_unlock(&btp->bt_lru_lock);
1521 spin_lock(&btp->bt_lru_lock);
1523 spin_unlock(&btp->bt_lru_lock);
1528 struct shrinker *shrink,
1529 struct shrink_control *sc)
1531 struct xfs_buftarg *btp = container_of(shrink,
1532 struct xfs_buftarg, bt_shrinker);
1534 int nr_to_scan = sc->nr_to_scan;
1538 return btp->bt_lru_nr;
1540 spin_lock(&btp->bt_lru_lock);
1541 while (!list_empty(&btp->bt_lru)) {
1542 if (nr_to_scan-- <= 0)
1545 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1548 * Decrement the b_lru_ref count unless the value is already
1549 * zero. If the value is already zero, we need to reclaim the
1550 * buffer, otherwise it gets another trip through the LRU.
1552 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1553 list_move_tail(&bp->b_lru, &btp->bt_lru);
1558 * remove the buffer from the LRU now to avoid needing another
1559 * lock round trip inside xfs_buf_rele().
1561 list_move(&bp->b_lru, &dispose);
1563 bp->b_lru_flags |= _XBF_LRU_DISPOSE;
1565 spin_unlock(&btp->bt_lru_lock);
1567 while (!list_empty(&dispose)) {
1568 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1569 list_del_init(&bp->b_lru);
1573 return btp->bt_lru_nr;
1578 struct xfs_mount *mp,
1579 struct xfs_buftarg *btp)
1581 unregister_shrinker(&btp->bt_shrinker);
1583 if (mp->m_flags & XFS_MOUNT_BARRIER)
1584 xfs_blkdev_issue_flush(btp);
1590 xfs_setsize_buftarg_flags(
1592 unsigned int blocksize,
1593 unsigned int sectorsize,
1596 btp->bt_bsize = blocksize;
1597 btp->bt_sshift = ffs(sectorsize) - 1;
1598 btp->bt_smask = sectorsize - 1;
1600 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1601 char name[BDEVNAME_SIZE];
1603 bdevname(btp->bt_bdev, name);
1605 xfs_warn(btp->bt_mount,
1606 "Cannot set_blocksize to %u on device %s\n",
1615 * When allocating the initial buffer target we have not yet
1616 * read in the superblock, so don't know what sized sectors
1617 * are being used is at this early stage. Play safe.
1620 xfs_setsize_buftarg_early(
1622 struct block_device *bdev)
1624 return xfs_setsize_buftarg_flags(btp,
1625 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1629 xfs_setsize_buftarg(
1631 unsigned int blocksize,
1632 unsigned int sectorsize)
1634 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1639 struct xfs_mount *mp,
1640 struct block_device *bdev,
1646 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1649 btp->bt_dev = bdev->bd_dev;
1650 btp->bt_bdev = bdev;
1651 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1655 INIT_LIST_HEAD(&btp->bt_lru);
1656 spin_lock_init(&btp->bt_lru_lock);
1657 if (xfs_setsize_buftarg_early(btp, bdev))
1659 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1660 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1661 register_shrinker(&btp->bt_shrinker);
1670 * Add a buffer to the delayed write list.
1672 * This queues a buffer for writeout if it hasn't already been. Note that
1673 * neither this routine nor the buffer list submission functions perform
1674 * any internal synchronization. It is expected that the lists are thread-local
1677 * Returns true if we queued up the buffer, or false if it already had
1678 * been on the buffer list.
1681 xfs_buf_delwri_queue(
1683 struct list_head *list)
1685 ASSERT(xfs_buf_islocked(bp));
1686 ASSERT(!(bp->b_flags & XBF_READ));
1689 * If the buffer is already marked delwri it already is queued up
1690 * by someone else for imediate writeout. Just ignore it in that
1693 if (bp->b_flags & _XBF_DELWRI_Q) {
1694 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1698 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1701 * If a buffer gets written out synchronously or marked stale while it
1702 * is on a delwri list we lazily remove it. To do this, the other party
1703 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1704 * It remains referenced and on the list. In a rare corner case it
1705 * might get readded to a delwri list after the synchronous writeout, in
1706 * which case we need just need to re-add the flag here.
1708 bp->b_flags |= _XBF_DELWRI_Q;
1709 if (list_empty(&bp->b_list)) {
1710 atomic_inc(&bp->b_hold);
1711 list_add_tail(&bp->b_list, list);
1718 * Compare function is more complex than it needs to be because
1719 * the return value is only 32 bits and we are doing comparisons
1725 struct list_head *a,
1726 struct list_head *b)
1728 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1729 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1732 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
1741 __xfs_buf_delwri_submit(
1742 struct list_head *buffer_list,
1743 struct list_head *io_list,
1746 struct blk_plug plug;
1747 struct xfs_buf *bp, *n;
1750 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1752 if (xfs_buf_ispinned(bp)) {
1756 if (!xfs_buf_trylock(bp))
1763 * Someone else might have written the buffer synchronously or
1764 * marked it stale in the meantime. In that case only the
1765 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1766 * reference and remove it from the list here.
1768 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1769 list_del_init(&bp->b_list);
1774 list_move_tail(&bp->b_list, io_list);
1775 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1778 list_sort(NULL, io_list, xfs_buf_cmp);
1780 blk_start_plug(&plug);
1781 list_for_each_entry_safe(bp, n, io_list, b_list) {
1782 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC);
1783 bp->b_flags |= XBF_WRITE;
1786 bp->b_flags |= XBF_ASYNC;
1787 list_del_init(&bp->b_list);
1791 blk_finish_plug(&plug);
1797 * Write out a buffer list asynchronously.
1799 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1800 * out and not wait for I/O completion on any of the buffers. This interface
1801 * is only safely useable for callers that can track I/O completion by higher
1802 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1806 xfs_buf_delwri_submit_nowait(
1807 struct list_head *buffer_list)
1809 LIST_HEAD (io_list);
1810 return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
1814 * Write out a buffer list synchronously.
1816 * This will take the @buffer_list, write all buffers out and wait for I/O
1817 * completion on all of the buffers. @buffer_list is consumed by the function,
1818 * so callers must have some other way of tracking buffers if they require such
1822 xfs_buf_delwri_submit(
1823 struct list_head *buffer_list)
1825 LIST_HEAD (io_list);
1826 int error = 0, error2;
1829 __xfs_buf_delwri_submit(buffer_list, &io_list, true);
1831 /* Wait for IO to complete. */
1832 while (!list_empty(&io_list)) {
1833 bp = list_first_entry(&io_list, struct xfs_buf, b_list);
1835 list_del_init(&bp->b_list);
1836 error2 = xfs_buf_iowait(bp);
1848 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1849 KM_ZONE_HWALIGN, NULL);
1853 xfslogd_workqueue = alloc_workqueue("xfslogd",
1854 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1855 if (!xfslogd_workqueue)
1856 goto out_free_buf_zone;
1861 kmem_zone_destroy(xfs_buf_zone);
1867 xfs_buf_terminate(void)
1869 destroy_workqueue(xfslogd_workqueue);
1870 kmem_zone_destroy(xfs_buf_zone);