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>
38 #include "xfs_trans_resv.h"
41 #include "xfs_mount.h"
42 #include "xfs_trace.h"
44 static kmem_zone_t *xfs_buf_zone;
46 static struct workqueue_struct *xfslogd_workqueue;
48 #ifdef XFS_BUF_LOCK_TRACKING
49 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
50 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
51 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
53 # define XB_SET_OWNER(bp) do { } while (0)
54 # define XB_CLEAR_OWNER(bp) do { } while (0)
55 # define XB_GET_OWNER(bp) do { } while (0)
58 #define xb_to_gfp(flags) \
59 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN)
67 * Return true if the buffer is vmapped.
69 * b_addr is null if the buffer is not mapped, but the code is clever
70 * enough to know it doesn't have to map a single page, so the check has
71 * to be both for b_addr and bp->b_page_count > 1.
73 return bp->b_addr && bp->b_page_count > 1;
80 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
84 * xfs_buf_lru_add - add a buffer to the LRU.
86 * The LRU takes a new reference to the buffer so that it will only be freed
87 * once the shrinker takes the buffer off the LRU.
93 struct xfs_buftarg *btp = bp->b_target;
95 spin_lock(&btp->bt_lru_lock);
96 if (list_empty(&bp->b_lru)) {
97 atomic_inc(&bp->b_hold);
98 list_add_tail(&bp->b_lru, &btp->bt_lru);
100 bp->b_lru_flags &= ~_XBF_LRU_DISPOSE;
102 spin_unlock(&btp->bt_lru_lock);
106 * xfs_buf_lru_del - remove a buffer from the LRU
108 * The unlocked check is safe here because it only occurs when there are not
109 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
110 * to optimise the shrinker removing the buffer from the LRU and calling
111 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
118 struct xfs_buftarg *btp = bp->b_target;
120 if (list_empty(&bp->b_lru))
123 spin_lock(&btp->bt_lru_lock);
124 if (!list_empty(&bp->b_lru)) {
125 list_del_init(&bp->b_lru);
128 spin_unlock(&btp->bt_lru_lock);
132 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
133 * b_lru_ref count so that the buffer is freed immediately when the buffer
134 * reference count falls to zero. If the buffer is already on the LRU, we need
135 * to remove the reference that LRU holds on the buffer.
137 * This prevents build-up of stale buffers on the LRU.
143 ASSERT(xfs_buf_islocked(bp));
145 bp->b_flags |= XBF_STALE;
148 * Clear the delwri status so that a delwri queue walker will not
149 * flush this buffer to disk now that it is stale. The delwri queue has
150 * a reference to the buffer, so this is safe to do.
152 bp->b_flags &= ~_XBF_DELWRI_Q;
154 atomic_set(&(bp)->b_lru_ref, 0);
155 if (!list_empty(&bp->b_lru)) {
156 struct xfs_buftarg *btp = bp->b_target;
158 spin_lock(&btp->bt_lru_lock);
159 if (!list_empty(&bp->b_lru) &&
160 !(bp->b_lru_flags & _XBF_LRU_DISPOSE)) {
161 list_del_init(&bp->b_lru);
163 atomic_dec(&bp->b_hold);
165 spin_unlock(&btp->bt_lru_lock);
167 ASSERT(atomic_read(&bp->b_hold) >= 1);
175 ASSERT(bp->b_maps == NULL);
176 bp->b_map_count = map_count;
178 if (map_count == 1) {
179 bp->b_maps = &bp->__b_map;
183 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
191 * Frees b_pages if it was allocated.
197 if (bp->b_maps != &bp->__b_map) {
198 kmem_free(bp->b_maps);
205 struct xfs_buftarg *target,
206 struct xfs_buf_map *map,
208 xfs_buf_flags_t flags)
214 bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS);
219 * We don't want certain flags to appear in b_flags unless they are
220 * specifically set by later operations on the buffer.
222 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
224 atomic_set(&bp->b_hold, 1);
225 atomic_set(&bp->b_lru_ref, 1);
226 init_completion(&bp->b_iowait);
227 INIT_LIST_HEAD(&bp->b_lru);
228 INIT_LIST_HEAD(&bp->b_list);
229 RB_CLEAR_NODE(&bp->b_rbnode);
230 sema_init(&bp->b_sema, 0); /* held, no waiters */
232 bp->b_target = target;
236 * Set length and io_length to the same value initially.
237 * I/O routines should use io_length, which will be the same in
238 * most cases but may be reset (e.g. XFS recovery).
240 error = xfs_buf_get_maps(bp, nmaps);
242 kmem_zone_free(xfs_buf_zone, bp);
246 bp->b_bn = map[0].bm_bn;
248 for (i = 0; i < nmaps; i++) {
249 bp->b_maps[i].bm_bn = map[i].bm_bn;
250 bp->b_maps[i].bm_len = map[i].bm_len;
251 bp->b_length += map[i].bm_len;
253 bp->b_io_length = bp->b_length;
255 atomic_set(&bp->b_pin_count, 0);
256 init_waitqueue_head(&bp->b_waiters);
258 XFS_STATS_INC(xb_create);
259 trace_xfs_buf_init(bp, _RET_IP_);
265 * Allocate a page array capable of holding a specified number
266 * of pages, and point the page buf at it.
272 xfs_buf_flags_t flags)
274 /* Make sure that we have a page list */
275 if (bp->b_pages == NULL) {
276 bp->b_page_count = page_count;
277 if (page_count <= XB_PAGES) {
278 bp->b_pages = bp->b_page_array;
280 bp->b_pages = kmem_alloc(sizeof(struct page *) *
281 page_count, KM_NOFS);
282 if (bp->b_pages == NULL)
285 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
291 * Frees b_pages if it was allocated.
297 if (bp->b_pages != bp->b_page_array) {
298 kmem_free(bp->b_pages);
304 * Releases the specified buffer.
306 * The modification state of any associated pages is left unchanged.
307 * The buffer must not be on any hash - use xfs_buf_rele instead for
308 * hashed and refcounted buffers
314 trace_xfs_buf_free(bp, _RET_IP_);
316 ASSERT(list_empty(&bp->b_lru));
318 if (bp->b_flags & _XBF_PAGES) {
321 if (xfs_buf_is_vmapped(bp))
322 vm_unmap_ram(bp->b_addr - bp->b_offset,
325 for (i = 0; i < bp->b_page_count; i++) {
326 struct page *page = bp->b_pages[i];
330 } else if (bp->b_flags & _XBF_KMEM)
331 kmem_free(bp->b_addr);
332 _xfs_buf_free_pages(bp);
333 xfs_buf_free_maps(bp);
334 kmem_zone_free(xfs_buf_zone, bp);
338 * Allocates all the pages for buffer in question and builds it's page list.
341 xfs_buf_allocate_memory(
346 size_t nbytes, offset;
347 gfp_t gfp_mask = xb_to_gfp(flags);
348 unsigned short page_count, i;
349 xfs_off_t start, end;
353 * for buffers that are contained within a single page, just allocate
354 * the memory from the heap - there's no need for the complexity of
355 * page arrays to keep allocation down to order 0.
357 size = BBTOB(bp->b_length);
358 if (size < PAGE_SIZE) {
359 bp->b_addr = kmem_alloc(size, KM_NOFS);
361 /* low memory - use alloc_page loop instead */
365 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
366 ((unsigned long)bp->b_addr & PAGE_MASK)) {
367 /* b_addr spans two pages - use alloc_page instead */
368 kmem_free(bp->b_addr);
372 bp->b_offset = offset_in_page(bp->b_addr);
373 bp->b_pages = bp->b_page_array;
374 bp->b_pages[0] = virt_to_page(bp->b_addr);
375 bp->b_page_count = 1;
376 bp->b_flags |= _XBF_KMEM;
381 start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT;
382 end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1)
384 page_count = end - start;
385 error = _xfs_buf_get_pages(bp, page_count, flags);
389 offset = bp->b_offset;
390 bp->b_flags |= _XBF_PAGES;
392 for (i = 0; i < bp->b_page_count; i++) {
396 page = alloc_page(gfp_mask);
397 if (unlikely(page == NULL)) {
398 if (flags & XBF_READ_AHEAD) {
399 bp->b_page_count = i;
405 * This could deadlock.
407 * But until all the XFS lowlevel code is revamped to
408 * handle buffer allocation failures we can't do much.
410 if (!(++retries % 100))
412 "possible memory allocation deadlock in %s (mode:0x%x)",
415 XFS_STATS_INC(xb_page_retries);
416 congestion_wait(BLK_RW_ASYNC, HZ/50);
420 XFS_STATS_INC(xb_page_found);
422 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
424 bp->b_pages[i] = page;
430 for (i = 0; i < bp->b_page_count; i++)
431 __free_page(bp->b_pages[i]);
436 * Map buffer into kernel address-space if necessary.
443 ASSERT(bp->b_flags & _XBF_PAGES);
444 if (bp->b_page_count == 1) {
445 /* A single page buffer is always mappable */
446 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
447 } else if (flags & XBF_UNMAPPED) {
453 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
458 } while (retried++ <= 1);
462 bp->b_addr += bp->b_offset;
469 * Finding and Reading Buffers
473 * Look up, and creates if absent, a lockable buffer for
474 * a given range of an inode. The buffer is returned
475 * locked. No I/O is implied by this call.
479 struct xfs_buftarg *btp,
480 struct xfs_buf_map *map,
482 xfs_buf_flags_t flags,
486 struct xfs_perag *pag;
487 struct rb_node **rbp;
488 struct rb_node *parent;
490 xfs_daddr_t blkno = map[0].bm_bn;
495 for (i = 0; i < nmaps; i++)
496 numblks += map[i].bm_len;
497 numbytes = BBTOB(numblks);
499 /* Check for IOs smaller than the sector size / not sector aligned */
500 ASSERT(!(numbytes < (1 << btp->bt_sshift)));
501 ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_smask));
504 * Corrupted block numbers can get through to here, unfortunately, so we
505 * have to check that the buffer falls within the filesystem bounds.
507 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
510 * XXX (dgc): we should really be returning EFSCORRUPTED here,
511 * but none of the higher level infrastructure supports
512 * returning a specific error on buffer lookup failures.
514 xfs_alert(btp->bt_mount,
515 "%s: Block out of range: block 0x%llx, EOFS 0x%llx ",
516 __func__, blkno, eofs);
522 pag = xfs_perag_get(btp->bt_mount,
523 xfs_daddr_to_agno(btp->bt_mount, blkno));
526 spin_lock(&pag->pag_buf_lock);
527 rbp = &pag->pag_buf_tree.rb_node;
532 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
534 if (blkno < bp->b_bn)
535 rbp = &(*rbp)->rb_left;
536 else if (blkno > bp->b_bn)
537 rbp = &(*rbp)->rb_right;
540 * found a block number match. If the range doesn't
541 * match, the only way this is allowed is if the buffer
542 * in the cache is stale and the transaction that made
543 * it stale has not yet committed. i.e. we are
544 * reallocating a busy extent. Skip this buffer and
545 * continue searching to the right for an exact match.
547 if (bp->b_length != numblks) {
548 ASSERT(bp->b_flags & XBF_STALE);
549 rbp = &(*rbp)->rb_right;
552 atomic_inc(&bp->b_hold);
559 rb_link_node(&new_bp->b_rbnode, parent, rbp);
560 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
561 /* the buffer keeps the perag reference until it is freed */
563 spin_unlock(&pag->pag_buf_lock);
565 XFS_STATS_INC(xb_miss_locked);
566 spin_unlock(&pag->pag_buf_lock);
572 spin_unlock(&pag->pag_buf_lock);
575 if (!xfs_buf_trylock(bp)) {
576 if (flags & XBF_TRYLOCK) {
578 XFS_STATS_INC(xb_busy_locked);
582 XFS_STATS_INC(xb_get_locked_waited);
586 * if the buffer is stale, clear all the external state associated with
587 * it. We need to keep flags such as how we allocated the buffer memory
590 if (bp->b_flags & XBF_STALE) {
591 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
592 ASSERT(bp->b_iodone == NULL);
593 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
597 trace_xfs_buf_find(bp, flags, _RET_IP_);
598 XFS_STATS_INC(xb_get_locked);
603 * Assembles a buffer covering the specified range. The code is optimised for
604 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
605 * more hits than misses.
609 struct xfs_buftarg *target,
610 struct xfs_buf_map *map,
612 xfs_buf_flags_t flags)
615 struct xfs_buf *new_bp;
618 bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
622 new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
623 if (unlikely(!new_bp))
626 error = xfs_buf_allocate_memory(new_bp, flags);
628 xfs_buf_free(new_bp);
632 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
634 xfs_buf_free(new_bp);
639 xfs_buf_free(new_bp);
643 error = _xfs_buf_map_pages(bp, flags);
644 if (unlikely(error)) {
645 xfs_warn(target->bt_mount,
646 "%s: failed to map pages\n", __func__);
652 XFS_STATS_INC(xb_get);
653 trace_xfs_buf_get(bp, flags, _RET_IP_);
660 xfs_buf_flags_t flags)
662 ASSERT(!(flags & XBF_WRITE));
663 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
665 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
666 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
668 xfs_buf_iorequest(bp);
669 if (flags & XBF_ASYNC)
671 return xfs_buf_iowait(bp);
676 struct xfs_buftarg *target,
677 struct xfs_buf_map *map,
679 xfs_buf_flags_t flags,
680 const struct xfs_buf_ops *ops)
686 bp = xfs_buf_get_map(target, map, nmaps, flags);
688 trace_xfs_buf_read(bp, flags, _RET_IP_);
690 if (!XFS_BUF_ISDONE(bp)) {
691 XFS_STATS_INC(xb_get_read);
693 _xfs_buf_read(bp, flags);
694 } else if (flags & XBF_ASYNC) {
696 * Read ahead call which is already satisfied,
702 /* We do not want read in the flags */
703 bp->b_flags &= ~XBF_READ;
711 * If we are not low on memory then do the readahead in a deadlock
715 xfs_buf_readahead_map(
716 struct xfs_buftarg *target,
717 struct xfs_buf_map *map,
719 const struct xfs_buf_ops *ops)
721 if (bdi_read_congested(target->bt_bdi))
724 xfs_buf_read_map(target, map, nmaps,
725 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops);
729 * Read an uncached buffer from disk. Allocates and returns a locked
730 * buffer containing the disk contents or nothing.
733 xfs_buf_read_uncached(
734 struct xfs_buftarg *target,
738 const struct xfs_buf_ops *ops)
742 bp = xfs_buf_get_uncached(target, numblks, flags);
746 /* set up the buffer for a read IO */
747 ASSERT(bp->b_map_count == 1);
749 bp->b_maps[0].bm_bn = daddr;
750 bp->b_flags |= XBF_READ;
753 xfsbdstrat(target->bt_mount, bp);
759 * Return a buffer allocated as an empty buffer and associated to external
760 * memory via xfs_buf_associate_memory() back to it's empty state.
768 _xfs_buf_free_pages(bp);
771 bp->b_page_count = 0;
773 bp->b_length = numblks;
774 bp->b_io_length = numblks;
776 ASSERT(bp->b_map_count == 1);
777 bp->b_bn = XFS_BUF_DADDR_NULL;
778 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
779 bp->b_maps[0].bm_len = bp->b_length;
782 static inline struct page *
786 if ((!is_vmalloc_addr(addr))) {
787 return virt_to_page(addr);
789 return vmalloc_to_page(addr);
794 xfs_buf_associate_memory(
801 unsigned long pageaddr;
802 unsigned long offset;
806 pageaddr = (unsigned long)mem & PAGE_MASK;
807 offset = (unsigned long)mem - pageaddr;
808 buflen = PAGE_ALIGN(len + offset);
809 page_count = buflen >> PAGE_SHIFT;
811 /* Free any previous set of page pointers */
813 _xfs_buf_free_pages(bp);
818 rval = _xfs_buf_get_pages(bp, page_count, 0);
822 bp->b_offset = offset;
824 for (i = 0; i < bp->b_page_count; i++) {
825 bp->b_pages[i] = mem_to_page((void *)pageaddr);
826 pageaddr += PAGE_SIZE;
829 bp->b_io_length = BTOBB(len);
830 bp->b_length = BTOBB(buflen);
836 xfs_buf_get_uncached(
837 struct xfs_buftarg *target,
841 unsigned long page_count;
844 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
846 bp = _xfs_buf_alloc(target, &map, 1, 0);
847 if (unlikely(bp == NULL))
850 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
851 error = _xfs_buf_get_pages(bp, page_count, 0);
855 for (i = 0; i < page_count; i++) {
856 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
860 bp->b_flags |= _XBF_PAGES;
862 error = _xfs_buf_map_pages(bp, 0);
863 if (unlikely(error)) {
864 xfs_warn(target->bt_mount,
865 "%s: failed to map pages\n", __func__);
869 trace_xfs_buf_get_uncached(bp, _RET_IP_);
874 __free_page(bp->b_pages[i]);
875 _xfs_buf_free_pages(bp);
877 xfs_buf_free_maps(bp);
878 kmem_zone_free(xfs_buf_zone, bp);
884 * Increment reference count on buffer, to hold the buffer concurrently
885 * with another thread which may release (free) the buffer asynchronously.
886 * Must hold the buffer already to call this function.
892 trace_xfs_buf_hold(bp, _RET_IP_);
893 atomic_inc(&bp->b_hold);
897 * Releases a hold on the specified buffer. If the
898 * the hold count is 1, calls xfs_buf_free.
904 struct xfs_perag *pag = bp->b_pag;
906 trace_xfs_buf_rele(bp, _RET_IP_);
909 ASSERT(list_empty(&bp->b_lru));
910 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
911 if (atomic_dec_and_test(&bp->b_hold))
916 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
918 ASSERT(atomic_read(&bp->b_hold) > 0);
919 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
920 if (!(bp->b_flags & XBF_STALE) &&
921 atomic_read(&bp->b_lru_ref)) {
923 spin_unlock(&pag->pag_buf_lock);
926 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
927 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
928 spin_unlock(&pag->pag_buf_lock);
937 * Lock a buffer object, if it is not already locked.
939 * If we come across a stale, pinned, locked buffer, we know that we are
940 * being asked to lock a buffer that has been reallocated. Because it is
941 * pinned, we know that the log has not been pushed to disk and hence it
942 * will still be locked. Rather than continuing to have trylock attempts
943 * fail until someone else pushes the log, push it ourselves before
944 * returning. This means that the xfsaild will not get stuck trying
945 * to push on stale inode buffers.
953 locked = down_trylock(&bp->b_sema) == 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);
1028 /* only validate buffers that were read without errors */
1029 if (read && bp->b_ops && !bp->b_error && (bp->b_flags & XBF_DONE))
1030 bp->b_ops->verify_read(bp);
1033 (*(bp->b_iodone))(bp);
1034 else if (bp->b_flags & XBF_ASYNC)
1037 ASSERT(read && bp->b_ops);
1038 complete(&bp->b_iowait);
1047 bool read = !!(bp->b_flags & XBF_READ);
1049 trace_xfs_buf_iodone(bp, _RET_IP_);
1051 if (bp->b_error == 0)
1052 bp->b_flags |= XBF_DONE;
1054 if (bp->b_iodone || (read && bp->b_ops) || (bp->b_flags & XBF_ASYNC)) {
1056 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1057 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1059 xfs_buf_iodone_work(&bp->b_iodone_work);
1062 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1063 complete(&bp->b_iowait);
1072 ASSERT(error >= 0 && error <= 0xffff);
1073 bp->b_error = (unsigned short)error;
1074 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1078 xfs_buf_ioerror_alert(
1082 xfs_alert(bp->b_target->bt_mount,
1083 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1084 (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length);
1088 * Called when we want to stop a buffer from getting written or read.
1089 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1090 * so that the proper iodone callbacks get called.
1096 #ifdef XFSERRORDEBUG
1097 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1101 * No need to wait until the buffer is unpinned, we aren't flushing it.
1103 xfs_buf_ioerror(bp, EIO);
1106 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1112 xfs_buf_ioend(bp, 0);
1118 * Same as xfs_bioerror, except that we are releasing the buffer
1119 * here ourselves, and avoiding the xfs_buf_ioend call.
1120 * This is meant for userdata errors; metadata bufs come with
1121 * iodone functions attached, so that we can track down errors.
1127 int64_t fl = bp->b_flags;
1129 * No need to wait until the buffer is unpinned.
1130 * We aren't flushing it.
1132 * chunkhold expects B_DONE to be set, whether
1133 * we actually finish the I/O or not. We don't want to
1134 * change that interface.
1139 bp->b_iodone = NULL;
1140 if (!(fl & XBF_ASYNC)) {
1142 * Mark b_error and B_ERROR _both_.
1143 * Lot's of chunkcache code assumes that.
1144 * There's no reason to mark error for
1147 xfs_buf_ioerror(bp, EIO);
1148 complete(&bp->b_iowait);
1160 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1161 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1163 * Metadata write that didn't get logged but
1164 * written delayed anyway. These aren't associated
1165 * with a transaction, and can be ignored.
1167 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1168 return xfs_bioerror_relse(bp);
1170 return xfs_bioerror(bp);
1173 xfs_buf_iorequest(bp);
1183 ASSERT(xfs_buf_islocked(bp));
1185 bp->b_flags |= XBF_WRITE;
1186 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q);
1190 error = xfs_buf_iowait(bp);
1192 xfs_force_shutdown(bp->b_target->bt_mount,
1193 SHUTDOWN_META_IO_ERROR);
1199 * Wrapper around bdstrat so that we can stop data from going to disk in case
1200 * we are shutting down the filesystem. Typically user data goes thru this
1201 * path; one of the exceptions is the superblock.
1205 struct xfs_mount *mp,
1208 if (XFS_FORCED_SHUTDOWN(mp)) {
1209 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1210 xfs_bioerror_relse(bp);
1214 xfs_buf_iorequest(bp);
1222 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1223 xfs_buf_ioend(bp, schedule);
1231 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1234 * don't overwrite existing errors - otherwise we can lose errors on
1235 * buffers that require multiple bios to complete.
1238 xfs_buf_ioerror(bp, -error);
1240 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1241 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1243 _xfs_buf_ioend(bp, 1);
1248 xfs_buf_ioapply_map(
1256 int total_nr_pages = bp->b_page_count;
1259 sector_t sector = bp->b_maps[map].bm_bn;
1263 total_nr_pages = bp->b_page_count;
1265 /* skip the pages in the buffer before the start offset */
1267 offset = *buf_offset;
1268 while (offset >= PAGE_SIZE) {
1270 offset -= PAGE_SIZE;
1274 * Limit the IO size to the length of the current vector, and update the
1275 * remaining IO count for the next time around.
1277 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1279 *buf_offset += size;
1282 atomic_inc(&bp->b_io_remaining);
1283 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1284 if (nr_pages > total_nr_pages)
1285 nr_pages = total_nr_pages;
1287 bio = bio_alloc(GFP_NOIO, nr_pages);
1288 bio->bi_bdev = bp->b_target->bt_bdev;
1289 bio->bi_sector = sector;
1290 bio->bi_end_io = xfs_buf_bio_end_io;
1291 bio->bi_private = bp;
1294 for (; size && nr_pages; nr_pages--, page_index++) {
1295 int rbytes, nbytes = PAGE_SIZE - offset;
1300 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1302 if (rbytes < nbytes)
1306 sector += BTOBB(nbytes);
1311 if (likely(bio->bi_size)) {
1312 if (xfs_buf_is_vmapped(bp)) {
1313 flush_kernel_vmap_range(bp->b_addr,
1314 xfs_buf_vmap_len(bp));
1316 submit_bio(rw, bio);
1321 * This is guaranteed not to be the last io reference count
1322 * because the caller (xfs_buf_iorequest) holds a count itself.
1324 atomic_dec(&bp->b_io_remaining);
1325 xfs_buf_ioerror(bp, EIO);
1335 struct blk_plug plug;
1342 * Make sure we capture only current IO errors rather than stale errors
1343 * left over from previous use of the buffer (e.g. failed readahead).
1347 if (bp->b_flags & XBF_WRITE) {
1348 if (bp->b_flags & XBF_SYNCIO)
1352 if (bp->b_flags & XBF_FUA)
1354 if (bp->b_flags & XBF_FLUSH)
1358 * Run the write verifier callback function if it exists. If
1359 * this function fails it will mark the buffer with an error and
1360 * the IO should not be dispatched.
1363 bp->b_ops->verify_write(bp);
1365 xfs_force_shutdown(bp->b_target->bt_mount,
1366 SHUTDOWN_CORRUPT_INCORE);
1370 } else if (bp->b_flags & XBF_READ_AHEAD) {
1376 /* we only use the buffer cache for meta-data */
1380 * Walk all the vectors issuing IO on them. Set up the initial offset
1381 * into the buffer and the desired IO size before we start -
1382 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1385 offset = bp->b_offset;
1386 size = BBTOB(bp->b_io_length);
1387 blk_start_plug(&plug);
1388 for (i = 0; i < bp->b_map_count; i++) {
1389 xfs_buf_ioapply_map(bp, i, &offset, &size, rw);
1393 break; /* all done */
1395 blk_finish_plug(&plug);
1402 trace_xfs_buf_iorequest(bp, _RET_IP_);
1404 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1406 if (bp->b_flags & XBF_WRITE)
1407 xfs_buf_wait_unpin(bp);
1410 /* Set the count to 1 initially, this will stop an I/O
1411 * completion callout which happens before we have started
1412 * all the I/O from calling xfs_buf_ioend too early.
1414 atomic_set(&bp->b_io_remaining, 1);
1415 _xfs_buf_ioapply(bp);
1416 _xfs_buf_ioend(bp, 1);
1422 * Waits for I/O to complete on the buffer supplied. It returns immediately if
1423 * no I/O is pending or there is already a pending error on the buffer. It
1424 * returns the I/O error code, if any, or 0 if there was no error.
1430 trace_xfs_buf_iowait(bp, _RET_IP_);
1433 wait_for_completion(&bp->b_iowait);
1435 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1447 return bp->b_addr + offset;
1449 offset += bp->b_offset;
1450 page = bp->b_pages[offset >> PAGE_SHIFT];
1451 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1455 * Move data into or out of a buffer.
1459 xfs_buf_t *bp, /* buffer to process */
1460 size_t boff, /* starting buffer offset */
1461 size_t bsize, /* length to copy */
1462 void *data, /* data address */
1463 xfs_buf_rw_t mode) /* read/write/zero flag */
1467 bend = boff + bsize;
1468 while (boff < bend) {
1470 int page_index, page_offset, csize;
1472 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1473 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1474 page = bp->b_pages[page_index];
1475 csize = min_t(size_t, PAGE_SIZE - page_offset,
1476 BBTOB(bp->b_io_length) - boff);
1478 ASSERT((csize + page_offset) <= PAGE_SIZE);
1482 memset(page_address(page) + page_offset, 0, csize);
1485 memcpy(data, page_address(page) + page_offset, csize);
1488 memcpy(page_address(page) + page_offset, data, csize);
1497 * Handling of buffer targets (buftargs).
1501 * Wait for any bufs with callbacks that have been submitted but have not yet
1502 * returned. These buffers will have an elevated hold count, so wait on those
1503 * while freeing all the buffers only held by the LRU.
1507 struct xfs_buftarg *btp)
1512 spin_lock(&btp->bt_lru_lock);
1513 while (!list_empty(&btp->bt_lru)) {
1514 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1515 if (atomic_read(&bp->b_hold) > 1) {
1516 trace_xfs_buf_wait_buftarg(bp, _RET_IP_);
1517 list_move_tail(&bp->b_lru, &btp->bt_lru);
1518 spin_unlock(&btp->bt_lru_lock);
1523 * clear the LRU reference count so the buffer doesn't get
1524 * ignored in xfs_buf_rele().
1526 atomic_set(&bp->b_lru_ref, 0);
1527 spin_unlock(&btp->bt_lru_lock);
1529 spin_lock(&btp->bt_lru_lock);
1531 spin_unlock(&btp->bt_lru_lock);
1536 struct shrinker *shrink,
1537 struct shrink_control *sc)
1539 struct xfs_buftarg *btp = container_of(shrink,
1540 struct xfs_buftarg, bt_shrinker);
1542 int nr_to_scan = sc->nr_to_scan;
1546 return btp->bt_lru_nr;
1548 spin_lock(&btp->bt_lru_lock);
1549 while (!list_empty(&btp->bt_lru)) {
1550 if (nr_to_scan-- <= 0)
1553 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1556 * Decrement the b_lru_ref count unless the value is already
1557 * zero. If the value is already zero, we need to reclaim the
1558 * buffer, otherwise it gets another trip through the LRU.
1560 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1561 list_move_tail(&bp->b_lru, &btp->bt_lru);
1566 * remove the buffer from the LRU now to avoid needing another
1567 * lock round trip inside xfs_buf_rele().
1569 list_move(&bp->b_lru, &dispose);
1571 bp->b_lru_flags |= _XBF_LRU_DISPOSE;
1573 spin_unlock(&btp->bt_lru_lock);
1575 while (!list_empty(&dispose)) {
1576 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1577 list_del_init(&bp->b_lru);
1581 return btp->bt_lru_nr;
1586 struct xfs_mount *mp,
1587 struct xfs_buftarg *btp)
1589 unregister_shrinker(&btp->bt_shrinker);
1591 if (mp->m_flags & XFS_MOUNT_BARRIER)
1592 xfs_blkdev_issue_flush(btp);
1598 xfs_setsize_buftarg_flags(
1600 unsigned int blocksize,
1601 unsigned int sectorsize,
1604 btp->bt_bsize = blocksize;
1605 btp->bt_sshift = ffs(sectorsize) - 1;
1606 btp->bt_smask = sectorsize - 1;
1608 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1609 char name[BDEVNAME_SIZE];
1611 bdevname(btp->bt_bdev, name);
1613 xfs_warn(btp->bt_mount,
1614 "Cannot set_blocksize to %u on device %s\n",
1623 * When allocating the initial buffer target we have not yet
1624 * read in the superblock, so don't know what sized sectors
1625 * are being used at this early stage. Play safe.
1628 xfs_setsize_buftarg_early(
1630 struct block_device *bdev)
1632 return xfs_setsize_buftarg_flags(btp,
1633 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1637 xfs_setsize_buftarg(
1639 unsigned int blocksize,
1640 unsigned int sectorsize)
1642 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1647 struct xfs_mount *mp,
1648 struct block_device *bdev,
1654 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS);
1657 btp->bt_dev = bdev->bd_dev;
1658 btp->bt_bdev = bdev;
1659 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1663 INIT_LIST_HEAD(&btp->bt_lru);
1664 spin_lock_init(&btp->bt_lru_lock);
1665 if (xfs_setsize_buftarg_early(btp, bdev))
1667 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1668 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1669 register_shrinker(&btp->bt_shrinker);
1678 * Add a buffer to the delayed write list.
1680 * This queues a buffer for writeout if it hasn't already been. Note that
1681 * neither this routine nor the buffer list submission functions perform
1682 * any internal synchronization. It is expected that the lists are thread-local
1685 * Returns true if we queued up the buffer, or false if it already had
1686 * been on the buffer list.
1689 xfs_buf_delwri_queue(
1691 struct list_head *list)
1693 ASSERT(xfs_buf_islocked(bp));
1694 ASSERT(!(bp->b_flags & XBF_READ));
1697 * If the buffer is already marked delwri it already is queued up
1698 * by someone else for imediate writeout. Just ignore it in that
1701 if (bp->b_flags & _XBF_DELWRI_Q) {
1702 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1706 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1709 * If a buffer gets written out synchronously or marked stale while it
1710 * is on a delwri list we lazily remove it. To do this, the other party
1711 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1712 * It remains referenced and on the list. In a rare corner case it
1713 * might get readded to a delwri list after the synchronous writeout, in
1714 * which case we need just need to re-add the flag here.
1716 bp->b_flags |= _XBF_DELWRI_Q;
1717 if (list_empty(&bp->b_list)) {
1718 atomic_inc(&bp->b_hold);
1719 list_add_tail(&bp->b_list, list);
1726 * Compare function is more complex than it needs to be because
1727 * the return value is only 32 bits and we are doing comparisons
1733 struct list_head *a,
1734 struct list_head *b)
1736 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1737 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1740 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
1749 __xfs_buf_delwri_submit(
1750 struct list_head *buffer_list,
1751 struct list_head *io_list,
1754 struct blk_plug plug;
1755 struct xfs_buf *bp, *n;
1758 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1760 if (xfs_buf_ispinned(bp)) {
1764 if (!xfs_buf_trylock(bp))
1771 * Someone else might have written the buffer synchronously or
1772 * marked it stale in the meantime. In that case only the
1773 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1774 * reference and remove it from the list here.
1776 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1777 list_del_init(&bp->b_list);
1782 list_move_tail(&bp->b_list, io_list);
1783 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1786 list_sort(NULL, io_list, xfs_buf_cmp);
1788 blk_start_plug(&plug);
1789 list_for_each_entry_safe(bp, n, io_list, b_list) {
1790 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC);
1791 bp->b_flags |= XBF_WRITE;
1794 bp->b_flags |= XBF_ASYNC;
1795 list_del_init(&bp->b_list);
1799 blk_finish_plug(&plug);
1805 * Write out a buffer list asynchronously.
1807 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1808 * out and not wait for I/O completion on any of the buffers. This interface
1809 * is only safely useable for callers that can track I/O completion by higher
1810 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1814 xfs_buf_delwri_submit_nowait(
1815 struct list_head *buffer_list)
1817 LIST_HEAD (io_list);
1818 return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
1822 * Write out a buffer list synchronously.
1824 * This will take the @buffer_list, write all buffers out and wait for I/O
1825 * completion on all of the buffers. @buffer_list is consumed by the function,
1826 * so callers must have some other way of tracking buffers if they require such
1830 xfs_buf_delwri_submit(
1831 struct list_head *buffer_list)
1833 LIST_HEAD (io_list);
1834 int error = 0, error2;
1837 __xfs_buf_delwri_submit(buffer_list, &io_list, true);
1839 /* Wait for IO to complete. */
1840 while (!list_empty(&io_list)) {
1841 bp = list_first_entry(&io_list, struct xfs_buf, b_list);
1843 list_del_init(&bp->b_list);
1844 error2 = xfs_buf_iowait(bp);
1856 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1857 KM_ZONE_HWALIGN, NULL);
1861 xfslogd_workqueue = alloc_workqueue("xfslogd",
1862 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1863 if (!xfslogd_workqueue)
1864 goto out_free_buf_zone;
1869 kmem_zone_destroy(xfs_buf_zone);
1875 xfs_buf_terminate(void)
1877 destroy_workqueue(xfslogd_workqueue);
1878 kmem_zone_destroy(xfs_buf_zone);