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>
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 : \
60 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
62 #define xb_to_km(flags) \
63 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
71 * Return true if the buffer is vmapped.
73 * The XBF_MAPPED flag is set if the buffer should be mapped, but the
74 * code is clever enough to know it doesn't have to map a single page,
75 * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
77 return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
84 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
88 * xfs_buf_lru_add - add a buffer to the LRU.
90 * The LRU takes a new reference to the buffer so that it will only be freed
91 * once the shrinker takes the buffer off the LRU.
97 struct xfs_buftarg *btp = bp->b_target;
99 spin_lock(&btp->bt_lru_lock);
100 if (list_empty(&bp->b_lru)) {
101 atomic_inc(&bp->b_hold);
102 list_add_tail(&bp->b_lru, &btp->bt_lru);
105 spin_unlock(&btp->bt_lru_lock);
109 * xfs_buf_lru_del - remove a buffer from the LRU
111 * The unlocked check is safe here because it only occurs when there are not
112 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
113 * to optimise the shrinker removing the buffer from the LRU and calling
114 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
121 struct xfs_buftarg *btp = bp->b_target;
123 if (list_empty(&bp->b_lru))
126 spin_lock(&btp->bt_lru_lock);
127 if (!list_empty(&bp->b_lru)) {
128 list_del_init(&bp->b_lru);
131 spin_unlock(&btp->bt_lru_lock);
135 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
136 * b_lru_ref count so that the buffer is freed immediately when the buffer
137 * reference count falls to zero. If the buffer is already on the LRU, we need
138 * to remove the reference that LRU holds on the buffer.
140 * This prevents build-up of stale buffers on the LRU.
146 ASSERT(xfs_buf_islocked(bp));
148 bp->b_flags |= XBF_STALE;
151 * Clear the delwri status so that a delwri queue walker will not
152 * flush this buffer to disk now that it is stale. The delwri queue has
153 * a reference to the buffer, so this is safe to do.
155 bp->b_flags &= ~_XBF_DELWRI_Q;
157 atomic_set(&(bp)->b_lru_ref, 0);
158 if (!list_empty(&bp->b_lru)) {
159 struct xfs_buftarg *btp = bp->b_target;
161 spin_lock(&btp->bt_lru_lock);
162 if (!list_empty(&bp->b_lru)) {
163 list_del_init(&bp->b_lru);
165 atomic_dec(&bp->b_hold);
167 spin_unlock(&btp->bt_lru_lock);
169 ASSERT(atomic_read(&bp->b_hold) >= 1);
174 struct xfs_buftarg *target,
175 xfs_off_t range_base,
177 xfs_buf_flags_t flags)
181 bp = kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags));
186 * We don't want certain flags to appear in b_flags.
188 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
190 memset(bp, 0, sizeof(xfs_buf_t));
191 atomic_set(&bp->b_hold, 1);
192 atomic_set(&bp->b_lru_ref, 1);
193 init_completion(&bp->b_iowait);
194 INIT_LIST_HEAD(&bp->b_lru);
195 INIT_LIST_HEAD(&bp->b_list);
196 RB_CLEAR_NODE(&bp->b_rbnode);
197 sema_init(&bp->b_sema, 0); /* held, no waiters */
199 bp->b_target = target;
200 bp->b_file_offset = range_base;
202 * Set buffer_length and count_desired to the same value initially.
203 * I/O routines should use count_desired, which will be the same in
204 * most cases but may be reset (e.g. XFS recovery).
206 bp->b_buffer_length = bp->b_count_desired = range_length;
208 bp->b_bn = XFS_BUF_DADDR_NULL;
209 atomic_set(&bp->b_pin_count, 0);
210 init_waitqueue_head(&bp->b_waiters);
212 XFS_STATS_INC(xb_create);
213 trace_xfs_buf_init(bp, _RET_IP_);
219 * Allocate a page array capable of holding a specified number
220 * of pages, and point the page buf at it.
226 xfs_buf_flags_t flags)
228 /* Make sure that we have a page list */
229 if (bp->b_pages == NULL) {
230 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
231 bp->b_page_count = page_count;
232 if (page_count <= XB_PAGES) {
233 bp->b_pages = bp->b_page_array;
235 bp->b_pages = kmem_alloc(sizeof(struct page *) *
236 page_count, xb_to_km(flags));
237 if (bp->b_pages == NULL)
240 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
246 * Frees b_pages if it was allocated.
252 if (bp->b_pages != bp->b_page_array) {
253 kmem_free(bp->b_pages);
259 * Releases the specified buffer.
261 * The modification state of any associated pages is left unchanged.
262 * The buffer most not be on any hash - use xfs_buf_rele instead for
263 * hashed and refcounted buffers
269 trace_xfs_buf_free(bp, _RET_IP_);
271 ASSERT(list_empty(&bp->b_lru));
273 if (bp->b_flags & _XBF_PAGES) {
276 if (xfs_buf_is_vmapped(bp))
277 vm_unmap_ram(bp->b_addr - bp->b_offset,
280 for (i = 0; i < bp->b_page_count; i++) {
281 struct page *page = bp->b_pages[i];
285 } else if (bp->b_flags & _XBF_KMEM)
286 kmem_free(bp->b_addr);
287 _xfs_buf_free_pages(bp);
288 kmem_zone_free(xfs_buf_zone, bp);
292 * Allocates all the pages for buffer in question and builds it's page list.
295 xfs_buf_allocate_memory(
299 size_t size = bp->b_count_desired;
300 size_t nbytes, offset;
301 gfp_t gfp_mask = xb_to_gfp(flags);
302 unsigned short page_count, i;
307 * for buffers that are contained within a single page, just allocate
308 * the memory from the heap - there's no need for the complexity of
309 * page arrays to keep allocation down to order 0.
311 if (bp->b_buffer_length < PAGE_SIZE) {
312 bp->b_addr = kmem_alloc(bp->b_buffer_length, xb_to_km(flags));
314 /* low memory - use alloc_page loop instead */
318 if (((unsigned long)(bp->b_addr + bp->b_buffer_length - 1) &
320 ((unsigned long)bp->b_addr & PAGE_MASK)) {
321 /* b_addr spans two pages - use alloc_page instead */
322 kmem_free(bp->b_addr);
326 bp->b_offset = offset_in_page(bp->b_addr);
327 bp->b_pages = bp->b_page_array;
328 bp->b_pages[0] = virt_to_page(bp->b_addr);
329 bp->b_page_count = 1;
330 bp->b_flags |= XBF_MAPPED | _XBF_KMEM;
335 end = bp->b_file_offset + bp->b_buffer_length;
336 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
337 error = _xfs_buf_get_pages(bp, page_count, flags);
341 offset = bp->b_offset;
342 bp->b_flags |= _XBF_PAGES;
344 for (i = 0; i < bp->b_page_count; i++) {
348 page = alloc_page(gfp_mask);
349 if (unlikely(page == NULL)) {
350 if (flags & XBF_READ_AHEAD) {
351 bp->b_page_count = i;
357 * This could deadlock.
359 * But until all the XFS lowlevel code is revamped to
360 * handle buffer allocation failures we can't do much.
362 if (!(++retries % 100))
364 "possible memory allocation deadlock in %s (mode:0x%x)",
367 XFS_STATS_INC(xb_page_retries);
368 congestion_wait(BLK_RW_ASYNC, HZ/50);
372 XFS_STATS_INC(xb_page_found);
374 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
376 bp->b_pages[i] = page;
382 for (i = 0; i < bp->b_page_count; i++)
383 __free_page(bp->b_pages[i]);
388 * Map buffer into kernel address-space if necessary.
395 ASSERT(bp->b_flags & _XBF_PAGES);
396 if (bp->b_page_count == 1) {
397 /* A single page buffer is always mappable */
398 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
399 bp->b_flags |= XBF_MAPPED;
400 } else if (flags & XBF_MAPPED) {
404 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
409 } while (retried++ <= 1);
413 bp->b_addr += bp->b_offset;
414 bp->b_flags |= XBF_MAPPED;
421 * Finding and Reading Buffers
425 * Look up, and creates if absent, a lockable buffer for
426 * a given range of an inode. The buffer is returned
427 * locked. No I/O is implied by this call.
431 xfs_buftarg_t *btp, /* block device target */
432 xfs_off_t ioff, /* starting offset of range */
433 size_t isize, /* length of range */
434 xfs_buf_flags_t flags,
437 xfs_off_t range_base;
439 struct xfs_perag *pag;
440 struct rb_node **rbp;
441 struct rb_node *parent;
444 range_base = (ioff << BBSHIFT);
445 range_length = (isize << BBSHIFT);
447 /* Check for IOs smaller than the sector size / not sector aligned */
448 ASSERT(!(range_length < (1 << btp->bt_sshift)));
449 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
452 pag = xfs_perag_get(btp->bt_mount,
453 xfs_daddr_to_agno(btp->bt_mount, ioff));
456 spin_lock(&pag->pag_buf_lock);
457 rbp = &pag->pag_buf_tree.rb_node;
462 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
464 if (range_base < bp->b_file_offset)
465 rbp = &(*rbp)->rb_left;
466 else if (range_base > bp->b_file_offset)
467 rbp = &(*rbp)->rb_right;
470 * found a block offset match. If the range doesn't
471 * match, the only way this is allowed is if the buffer
472 * in the cache is stale and the transaction that made
473 * it stale has not yet committed. i.e. we are
474 * reallocating a busy extent. Skip this buffer and
475 * continue searching to the right for an exact match.
477 if (bp->b_buffer_length != range_length) {
478 ASSERT(bp->b_flags & XBF_STALE);
479 rbp = &(*rbp)->rb_right;
482 atomic_inc(&bp->b_hold);
489 rb_link_node(&new_bp->b_rbnode, parent, rbp);
490 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
491 /* the buffer keeps the perag reference until it is freed */
493 spin_unlock(&pag->pag_buf_lock);
495 XFS_STATS_INC(xb_miss_locked);
496 spin_unlock(&pag->pag_buf_lock);
502 spin_unlock(&pag->pag_buf_lock);
505 if (!xfs_buf_trylock(bp)) {
506 if (flags & XBF_TRYLOCK) {
508 XFS_STATS_INC(xb_busy_locked);
512 XFS_STATS_INC(xb_get_locked_waited);
516 * if the buffer is stale, clear all the external state associated with
517 * it. We need to keep flags such as how we allocated the buffer memory
520 if (bp->b_flags & XBF_STALE) {
521 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
522 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
525 trace_xfs_buf_find(bp, flags, _RET_IP_);
526 XFS_STATS_INC(xb_get_locked);
531 * Assembles a buffer covering the specified range. The code is optimised for
532 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
533 * more hits than misses.
537 xfs_buftarg_t *target,/* target for buffer */
538 xfs_off_t ioff, /* starting offset of range */
539 size_t isize, /* length of range */
540 xfs_buf_flags_t flags)
543 struct xfs_buf *new_bp;
546 bp = _xfs_buf_find(target, ioff, isize, flags, NULL);
550 new_bp = xfs_buf_alloc(target, ioff << BBSHIFT, isize << BBSHIFT,
552 if (unlikely(!new_bp))
555 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
557 kmem_zone_free(xfs_buf_zone, new_bp);
562 error = xfs_buf_allocate_memory(bp, flags);
566 kmem_zone_free(xfs_buf_zone, new_bp);
569 * Now we have a workable buffer, fill in the block number so
570 * that we can do IO on it.
573 bp->b_count_desired = bp->b_buffer_length;
576 if (!(bp->b_flags & XBF_MAPPED)) {
577 error = _xfs_buf_map_pages(bp, flags);
578 if (unlikely(error)) {
579 xfs_warn(target->bt_mount,
580 "%s: failed to map pages\n", __func__);
585 XFS_STATS_INC(xb_get);
586 trace_xfs_buf_get(bp, flags, _RET_IP_);
590 if (flags & (XBF_LOCK | XBF_TRYLOCK))
599 xfs_buf_flags_t flags)
603 ASSERT(!(flags & XBF_WRITE));
604 ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
606 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
607 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
609 status = xfs_buf_iorequest(bp);
610 if (status || bp->b_error || (flags & XBF_ASYNC))
612 return xfs_buf_iowait(bp);
617 xfs_buftarg_t *target,
620 xfs_buf_flags_t flags)
626 bp = xfs_buf_get(target, ioff, isize, flags);
628 trace_xfs_buf_read(bp, flags, _RET_IP_);
630 if (!XFS_BUF_ISDONE(bp)) {
631 XFS_STATS_INC(xb_get_read);
632 _xfs_buf_read(bp, flags);
633 } else if (flags & XBF_ASYNC) {
635 * Read ahead call which is already satisfied,
640 /* We do not want read in the flags */
641 bp->b_flags &= ~XBF_READ;
648 if (flags & (XBF_LOCK | XBF_TRYLOCK))
655 * If we are not low on memory then do the readahead in a deadlock
660 xfs_buftarg_t *target,
664 if (bdi_read_congested(target->bt_bdi))
667 xfs_buf_read(target, ioff, isize,
668 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
672 * Read an uncached buffer from disk. Allocates and returns a locked
673 * buffer containing the disk contents or nothing.
676 xfs_buf_read_uncached(
677 struct xfs_mount *mp,
678 struct xfs_buftarg *target,
686 bp = xfs_buf_get_uncached(target, length, flags);
690 /* set up the buffer for a read IO */
691 XFS_BUF_SET_ADDR(bp, daddr);
695 error = xfs_buf_iowait(bp);
696 if (error || bp->b_error) {
704 * Return a buffer allocated as an empty buffer and associated to external
705 * memory via xfs_buf_associate_memory() back to it's empty state.
713 _xfs_buf_free_pages(bp);
716 bp->b_page_count = 0;
718 bp->b_file_offset = 0;
719 bp->b_buffer_length = bp->b_count_desired = len;
720 bp->b_bn = XFS_BUF_DADDR_NULL;
721 bp->b_flags &= ~XBF_MAPPED;
724 static inline struct page *
728 if ((!is_vmalloc_addr(addr))) {
729 return virt_to_page(addr);
731 return vmalloc_to_page(addr);
736 xfs_buf_associate_memory(
743 unsigned long pageaddr;
744 unsigned long offset;
748 pageaddr = (unsigned long)mem & PAGE_MASK;
749 offset = (unsigned long)mem - pageaddr;
750 buflen = PAGE_ALIGN(len + offset);
751 page_count = buflen >> PAGE_SHIFT;
753 /* Free any previous set of page pointers */
755 _xfs_buf_free_pages(bp);
760 rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
764 bp->b_offset = offset;
766 for (i = 0; i < bp->b_page_count; i++) {
767 bp->b_pages[i] = mem_to_page((void *)pageaddr);
768 pageaddr += PAGE_SIZE;
771 bp->b_count_desired = len;
772 bp->b_buffer_length = buflen;
773 bp->b_flags |= XBF_MAPPED;
779 xfs_buf_get_uncached(
780 struct xfs_buftarg *target,
784 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
788 bp = xfs_buf_alloc(target, 0, len, 0);
789 if (unlikely(bp == NULL))
792 error = _xfs_buf_get_pages(bp, page_count, 0);
796 for (i = 0; i < page_count; i++) {
797 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
801 bp->b_flags |= _XBF_PAGES;
803 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
804 if (unlikely(error)) {
805 xfs_warn(target->bt_mount,
806 "%s: failed to map pages\n", __func__);
810 trace_xfs_buf_get_uncached(bp, _RET_IP_);
815 __free_page(bp->b_pages[i]);
816 _xfs_buf_free_pages(bp);
818 kmem_zone_free(xfs_buf_zone, bp);
824 * Increment reference count on buffer, to hold the buffer concurrently
825 * with another thread which may release (free) the buffer asynchronously.
826 * Must hold the buffer already to call this function.
832 trace_xfs_buf_hold(bp, _RET_IP_);
833 atomic_inc(&bp->b_hold);
837 * Releases a hold on the specified buffer. If the
838 * the hold count is 1, calls xfs_buf_free.
844 struct xfs_perag *pag = bp->b_pag;
846 trace_xfs_buf_rele(bp, _RET_IP_);
849 ASSERT(list_empty(&bp->b_lru));
850 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
851 if (atomic_dec_and_test(&bp->b_hold))
856 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
858 ASSERT(atomic_read(&bp->b_hold) > 0);
859 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
860 if (!(bp->b_flags & XBF_STALE) &&
861 atomic_read(&bp->b_lru_ref)) {
863 spin_unlock(&pag->pag_buf_lock);
866 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
867 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
868 spin_unlock(&pag->pag_buf_lock);
877 * Lock a buffer object, if it is not already locked.
879 * If we come across a stale, pinned, locked buffer, we know that we are
880 * being asked to lock a buffer that has been reallocated. Because it is
881 * pinned, we know that the log has not been pushed to disk and hence it
882 * will still be locked. Rather than continuing to have trylock attempts
883 * fail until someone else pushes the log, push it ourselves before
884 * returning. This means that the xfsaild will not get stuck trying
885 * to push on stale inode buffers.
893 locked = down_trylock(&bp->b_sema) == 0;
896 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
897 xfs_log_force(bp->b_target->bt_mount, 0);
899 trace_xfs_buf_trylock(bp, _RET_IP_);
904 * Lock a buffer object.
906 * If we come across a stale, pinned, locked buffer, we know that we
907 * are being asked to lock a buffer that has been reallocated. Because
908 * it is pinned, we know that the log has not been pushed to disk and
909 * hence it will still be locked. Rather than sleeping until someone
910 * else pushes the log, push it ourselves before trying to get the lock.
916 trace_xfs_buf_lock(bp, _RET_IP_);
918 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
919 xfs_log_force(bp->b_target->bt_mount, 0);
923 trace_xfs_buf_lock_done(bp, _RET_IP_);
933 trace_xfs_buf_unlock(bp, _RET_IP_);
940 DECLARE_WAITQUEUE (wait, current);
942 if (atomic_read(&bp->b_pin_count) == 0)
945 add_wait_queue(&bp->b_waiters, &wait);
947 set_current_state(TASK_UNINTERRUPTIBLE);
948 if (atomic_read(&bp->b_pin_count) == 0)
952 remove_wait_queue(&bp->b_waiters, &wait);
953 set_current_state(TASK_RUNNING);
957 * Buffer Utility Routines
962 struct work_struct *work)
965 container_of(work, xfs_buf_t, b_iodone_work);
968 (*(bp->b_iodone))(bp);
969 else if (bp->b_flags & XBF_ASYNC)
978 trace_xfs_buf_iodone(bp, _RET_IP_);
980 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
981 if (bp->b_error == 0)
982 bp->b_flags |= XBF_DONE;
984 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
986 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
987 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
989 xfs_buf_iodone_work(&bp->b_iodone_work);
992 complete(&bp->b_iowait);
1001 ASSERT(error >= 0 && error <= 0xffff);
1002 bp->b_error = (unsigned short)error;
1003 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1007 xfs_buf_ioerror_alert(
1011 xfs_alert(bp->b_target->bt_mount,
1012 "metadata I/O error: block 0x%llx (\"%s\") error %d buf count %zd",
1013 (__uint64_t)XFS_BUF_ADDR(bp), func,
1014 bp->b_error, XFS_BUF_COUNT(bp));
1023 ASSERT(xfs_buf_islocked(bp));
1025 bp->b_flags |= XBF_WRITE;
1026 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q);
1030 error = xfs_buf_iowait(bp);
1032 xfs_force_shutdown(bp->b_target->bt_mount,
1033 SHUTDOWN_META_IO_ERROR);
1039 * Called when we want to stop a buffer from getting written or read.
1040 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1041 * so that the proper iodone callbacks get called.
1047 #ifdef XFSERRORDEBUG
1048 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1052 * No need to wait until the buffer is unpinned, we aren't flushing it.
1054 xfs_buf_ioerror(bp, EIO);
1057 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1063 xfs_buf_ioend(bp, 0);
1069 * Same as xfs_bioerror, except that we are releasing the buffer
1070 * here ourselves, and avoiding the xfs_buf_ioend call.
1071 * This is meant for userdata errors; metadata bufs come with
1072 * iodone functions attached, so that we can track down errors.
1078 int64_t fl = bp->b_flags;
1080 * No need to wait until the buffer is unpinned.
1081 * We aren't flushing it.
1083 * chunkhold expects B_DONE to be set, whether
1084 * we actually finish the I/O or not. We don't want to
1085 * change that interface.
1090 bp->b_iodone = NULL;
1091 if (!(fl & XBF_ASYNC)) {
1093 * Mark b_error and B_ERROR _both_.
1094 * Lot's of chunkcache code assumes that.
1095 * There's no reason to mark error for
1098 xfs_buf_ioerror(bp, EIO);
1099 complete(&bp->b_iowait);
1109 * All xfs metadata buffers except log state machine buffers
1110 * get this attached as their b_bdstrat callback function.
1111 * This is so that we can catch a buffer
1112 * after prematurely unpinning it to forcibly shutdown the filesystem.
1118 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1119 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1121 * Metadata write that didn't get logged but
1122 * written delayed anyway. These aren't associated
1123 * with a transaction, and can be ignored.
1125 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1126 return xfs_bioerror_relse(bp);
1128 return xfs_bioerror(bp);
1131 xfs_buf_iorequest(bp);
1136 * Wrapper around bdstrat so that we can stop data from going to disk in case
1137 * we are shutting down the filesystem. Typically user data goes thru this
1138 * path; one of the exceptions is the superblock.
1142 struct xfs_mount *mp,
1145 if (XFS_FORCED_SHUTDOWN(mp)) {
1146 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1147 xfs_bioerror_relse(bp);
1151 xfs_buf_iorequest(bp);
1159 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1160 xfs_buf_ioend(bp, schedule);
1168 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1170 xfs_buf_ioerror(bp, -error);
1172 if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1173 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1175 _xfs_buf_ioend(bp, 1);
1183 int rw, map_i, total_nr_pages, nr_pages;
1185 int offset = bp->b_offset;
1186 int size = bp->b_count_desired;
1187 sector_t sector = bp->b_bn;
1189 total_nr_pages = bp->b_page_count;
1192 if (bp->b_flags & XBF_WRITE) {
1193 if (bp->b_flags & XBF_SYNCIO)
1197 if (bp->b_flags & XBF_FUA)
1199 if (bp->b_flags & XBF_FLUSH)
1201 } else if (bp->b_flags & XBF_READ_AHEAD) {
1207 /* we only use the buffer cache for meta-data */
1211 atomic_inc(&bp->b_io_remaining);
1212 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1213 if (nr_pages > total_nr_pages)
1214 nr_pages = total_nr_pages;
1216 bio = bio_alloc(GFP_NOIO, nr_pages);
1217 bio->bi_bdev = bp->b_target->bt_bdev;
1218 bio->bi_sector = sector;
1219 bio->bi_end_io = xfs_buf_bio_end_io;
1220 bio->bi_private = bp;
1223 for (; size && nr_pages; nr_pages--, map_i++) {
1224 int rbytes, nbytes = PAGE_SIZE - offset;
1229 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1230 if (rbytes < nbytes)
1234 sector += nbytes >> BBSHIFT;
1239 if (likely(bio->bi_size)) {
1240 if (xfs_buf_is_vmapped(bp)) {
1241 flush_kernel_vmap_range(bp->b_addr,
1242 xfs_buf_vmap_len(bp));
1244 submit_bio(rw, bio);
1248 xfs_buf_ioerror(bp, EIO);
1257 trace_xfs_buf_iorequest(bp, _RET_IP_);
1259 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1261 if (bp->b_flags & XBF_WRITE)
1262 xfs_buf_wait_unpin(bp);
1265 /* Set the count to 1 initially, this will stop an I/O
1266 * completion callout which happens before we have started
1267 * all the I/O from calling xfs_buf_ioend too early.
1269 atomic_set(&bp->b_io_remaining, 1);
1270 _xfs_buf_ioapply(bp);
1271 _xfs_buf_ioend(bp, 0);
1278 * Waits for I/O to complete on the buffer supplied.
1279 * It returns immediately if no I/O is pending.
1280 * It returns the I/O error code, if any, or 0 if there was no error.
1286 trace_xfs_buf_iowait(bp, _RET_IP_);
1288 wait_for_completion(&bp->b_iowait);
1290 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1301 if (bp->b_flags & XBF_MAPPED)
1302 return bp->b_addr + offset;
1304 offset += bp->b_offset;
1305 page = bp->b_pages[offset >> PAGE_SHIFT];
1306 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1310 * Move data into or out of a buffer.
1314 xfs_buf_t *bp, /* buffer to process */
1315 size_t boff, /* starting buffer offset */
1316 size_t bsize, /* length to copy */
1317 void *data, /* data address */
1318 xfs_buf_rw_t mode) /* read/write/zero flag */
1320 size_t bend, cpoff, csize;
1323 bend = boff + bsize;
1324 while (boff < bend) {
1325 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1326 cpoff = xfs_buf_poff(boff + bp->b_offset);
1327 csize = min_t(size_t,
1328 PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1330 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1334 memset(page_address(page) + cpoff, 0, csize);
1337 memcpy(data, page_address(page) + cpoff, csize);
1340 memcpy(page_address(page) + cpoff, data, csize);
1349 * Handling of buffer targets (buftargs).
1353 * Wait for any bufs with callbacks that have been submitted but have not yet
1354 * returned. These buffers will have an elevated hold count, so wait on those
1355 * while freeing all the buffers only held by the LRU.
1359 struct xfs_buftarg *btp)
1364 spin_lock(&btp->bt_lru_lock);
1365 while (!list_empty(&btp->bt_lru)) {
1366 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1367 if (atomic_read(&bp->b_hold) > 1) {
1368 spin_unlock(&btp->bt_lru_lock);
1373 * clear the LRU reference count so the buffer doesn't get
1374 * ignored in xfs_buf_rele().
1376 atomic_set(&bp->b_lru_ref, 0);
1377 spin_unlock(&btp->bt_lru_lock);
1379 spin_lock(&btp->bt_lru_lock);
1381 spin_unlock(&btp->bt_lru_lock);
1386 struct shrinker *shrink,
1387 struct shrink_control *sc)
1389 struct xfs_buftarg *btp = container_of(shrink,
1390 struct xfs_buftarg, bt_shrinker);
1392 int nr_to_scan = sc->nr_to_scan;
1396 return btp->bt_lru_nr;
1398 spin_lock(&btp->bt_lru_lock);
1399 while (!list_empty(&btp->bt_lru)) {
1400 if (nr_to_scan-- <= 0)
1403 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1406 * Decrement the b_lru_ref count unless the value is already
1407 * zero. If the value is already zero, we need to reclaim the
1408 * buffer, otherwise it gets another trip through the LRU.
1410 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1411 list_move_tail(&bp->b_lru, &btp->bt_lru);
1416 * remove the buffer from the LRU now to avoid needing another
1417 * lock round trip inside xfs_buf_rele().
1419 list_move(&bp->b_lru, &dispose);
1422 spin_unlock(&btp->bt_lru_lock);
1424 while (!list_empty(&dispose)) {
1425 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1426 list_del_init(&bp->b_lru);
1430 return btp->bt_lru_nr;
1435 struct xfs_mount *mp,
1436 struct xfs_buftarg *btp)
1438 unregister_shrinker(&btp->bt_shrinker);
1440 if (mp->m_flags & XFS_MOUNT_BARRIER)
1441 xfs_blkdev_issue_flush(btp);
1447 xfs_setsize_buftarg_flags(
1449 unsigned int blocksize,
1450 unsigned int sectorsize,
1453 btp->bt_bsize = blocksize;
1454 btp->bt_sshift = ffs(sectorsize) - 1;
1455 btp->bt_smask = sectorsize - 1;
1457 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1458 char name[BDEVNAME_SIZE];
1460 bdevname(btp->bt_bdev, name);
1462 xfs_warn(btp->bt_mount,
1463 "Cannot set_blocksize to %u on device %s\n",
1472 * When allocating the initial buffer target we have not yet
1473 * read in the superblock, so don't know what sized sectors
1474 * are being used is at this early stage. Play safe.
1477 xfs_setsize_buftarg_early(
1479 struct block_device *bdev)
1481 return xfs_setsize_buftarg_flags(btp,
1482 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1486 xfs_setsize_buftarg(
1488 unsigned int blocksize,
1489 unsigned int sectorsize)
1491 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1496 struct xfs_mount *mp,
1497 struct block_device *bdev,
1503 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1506 btp->bt_dev = bdev->bd_dev;
1507 btp->bt_bdev = bdev;
1508 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1512 INIT_LIST_HEAD(&btp->bt_lru);
1513 spin_lock_init(&btp->bt_lru_lock);
1514 if (xfs_setsize_buftarg_early(btp, bdev))
1516 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1517 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1518 register_shrinker(&btp->bt_shrinker);
1527 * Add a buffer to the delayed write list.
1529 * This queues a buffer for writeout if it hasn't already been. Note that
1530 * neither this routine nor the buffer list submission functions perform
1531 * any internal synchronization. It is expected that the lists are thread-local
1534 * Returns true if we queued up the buffer, or false if it already had
1535 * been on the buffer list.
1538 xfs_buf_delwri_queue(
1540 struct list_head *list)
1542 ASSERT(xfs_buf_islocked(bp));
1543 ASSERT(!(bp->b_flags & XBF_READ));
1546 * If the buffer is already marked delwri it already is queued up
1547 * by someone else for imediate writeout. Just ignore it in that
1550 if (bp->b_flags & _XBF_DELWRI_Q) {
1551 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1555 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1558 * If a buffer gets written out synchronously or marked stale while it
1559 * is on a delwri list we lazily remove it. To do this, the other party
1560 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1561 * It remains referenced and on the list. In a rare corner case it
1562 * might get readded to a delwri list after the synchronous writeout, in
1563 * which case we need just need to re-add the flag here.
1565 bp->b_flags |= _XBF_DELWRI_Q;
1566 if (list_empty(&bp->b_list)) {
1567 atomic_inc(&bp->b_hold);
1568 list_add_tail(&bp->b_list, list);
1575 * Compare function is more complex than it needs to be because
1576 * the return value is only 32 bits and we are doing comparisons
1582 struct list_head *a,
1583 struct list_head *b)
1585 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1586 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1589 diff = ap->b_bn - bp->b_bn;
1598 __xfs_buf_delwri_submit(
1599 struct list_head *buffer_list,
1600 struct list_head *io_list,
1603 struct blk_plug plug;
1604 struct xfs_buf *bp, *n;
1607 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1609 if (xfs_buf_ispinned(bp)) {
1613 if (!xfs_buf_trylock(bp))
1620 * Someone else might have written the buffer synchronously or
1621 * marked it stale in the meantime. In that case only the
1622 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1623 * reference and remove it from the list here.
1625 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1626 list_del_init(&bp->b_list);
1631 list_move_tail(&bp->b_list, io_list);
1632 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1635 list_sort(NULL, io_list, xfs_buf_cmp);
1637 blk_start_plug(&plug);
1638 list_for_each_entry_safe(bp, n, io_list, b_list) {
1639 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC);
1640 bp->b_flags |= XBF_WRITE;
1643 bp->b_flags |= XBF_ASYNC;
1644 list_del_init(&bp->b_list);
1648 blk_finish_plug(&plug);
1654 * Write out a buffer list asynchronously.
1656 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1657 * out and not wait for I/O completion on any of the buffers. This interface
1658 * is only safely useable for callers that can track I/O completion by higher
1659 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1663 xfs_buf_delwri_submit_nowait(
1664 struct list_head *buffer_list)
1666 LIST_HEAD (io_list);
1667 return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
1671 * Write out a buffer list synchronously.
1673 * This will take the @buffer_list, write all buffers out and wait for I/O
1674 * completion on all of the buffers. @buffer_list is consumed by the function,
1675 * so callers must have some other way of tracking buffers if they require such
1679 xfs_buf_delwri_submit(
1680 struct list_head *buffer_list)
1682 LIST_HEAD (io_list);
1683 int error = 0, error2;
1686 __xfs_buf_delwri_submit(buffer_list, &io_list, true);
1688 /* Wait for IO to complete. */
1689 while (!list_empty(&io_list)) {
1690 bp = list_first_entry(&io_list, struct xfs_buf, b_list);
1692 list_del_init(&bp->b_list);
1693 error2 = xfs_buf_iowait(bp);
1705 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1706 KM_ZONE_HWALIGN, NULL);
1710 xfslogd_workqueue = alloc_workqueue("xfslogd",
1711 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1712 if (!xfslogd_workqueue)
1713 goto out_free_buf_zone;
1718 kmem_zone_destroy(xfs_buf_zone);
1724 xfs_buf_terminate(void)
1726 destroy_workqueue(xfslogd_workqueue);
1727 kmem_zone_destroy(xfs_buf_zone);