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;
45 STATIC int xfsbufd(void *);
47 static struct workqueue_struct *xfslogd_workqueue;
48 struct workqueue_struct *xfsdatad_workqueue;
49 struct workqueue_struct *xfsconvertd_workqueue;
51 #ifdef XFS_BUF_LOCK_TRACKING
52 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
53 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
54 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
56 # define XB_SET_OWNER(bp) do { } while (0)
57 # define XB_CLEAR_OWNER(bp) do { } while (0)
58 # define XB_GET_OWNER(bp) do { } while (0)
61 #define xb_to_gfp(flags) \
62 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
63 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
65 #define xb_to_km(flags) \
66 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
68 #define xfs_buf_allocate(flags) \
69 kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
70 #define xfs_buf_deallocate(bp) \
71 kmem_zone_free(xfs_buf_zone, (bp));
78 * Return true if the buffer is vmapped.
80 * The XBF_MAPPED flag is set if the buffer should be mapped, but the
81 * code is clever enough to know it doesn't have to map a single page,
82 * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
84 return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
91 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
95 * xfs_buf_lru_add - add a buffer to the LRU.
97 * The LRU takes a new reference to the buffer so that it will only be freed
98 * once the shrinker takes the buffer off the LRU.
104 struct xfs_buftarg *btp = bp->b_target;
106 spin_lock(&btp->bt_lru_lock);
107 if (list_empty(&bp->b_lru)) {
108 atomic_inc(&bp->b_hold);
109 list_add_tail(&bp->b_lru, &btp->bt_lru);
112 spin_unlock(&btp->bt_lru_lock);
116 * xfs_buf_lru_del - remove a buffer from the LRU
118 * The unlocked check is safe here because it only occurs when there are not
119 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
120 * to optimise the shrinker removing the buffer from the LRU and calling
121 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
128 struct xfs_buftarg *btp = bp->b_target;
130 if (list_empty(&bp->b_lru))
133 spin_lock(&btp->bt_lru_lock);
134 if (!list_empty(&bp->b_lru)) {
135 list_del_init(&bp->b_lru);
138 spin_unlock(&btp->bt_lru_lock);
142 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
143 * b_lru_ref count so that the buffer is freed immediately when the buffer
144 * reference count falls to zero. If the buffer is already on the LRU, we need
145 * to remove the reference that LRU holds on the buffer.
147 * This prevents build-up of stale buffers on the LRU.
153 bp->b_flags |= XBF_STALE;
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 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);
172 xfs_buftarg_t *target,
173 xfs_off_t range_base,
175 xfs_buf_flags_t flags)
178 * We don't want certain flags to appear in b_flags.
180 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
182 memset(bp, 0, sizeof(xfs_buf_t));
183 atomic_set(&bp->b_hold, 1);
184 atomic_set(&bp->b_lru_ref, 1);
185 init_completion(&bp->b_iowait);
186 INIT_LIST_HEAD(&bp->b_lru);
187 INIT_LIST_HEAD(&bp->b_list);
188 RB_CLEAR_NODE(&bp->b_rbnode);
189 sema_init(&bp->b_sema, 0); /* held, no waiters */
191 bp->b_target = target;
192 bp->b_file_offset = range_base;
194 * Set buffer_length and count_desired to the same value initially.
195 * I/O routines should use count_desired, which will be the same in
196 * most cases but may be reset (e.g. XFS recovery).
198 bp->b_buffer_length = bp->b_count_desired = range_length;
200 bp->b_bn = XFS_BUF_DADDR_NULL;
201 atomic_set(&bp->b_pin_count, 0);
202 init_waitqueue_head(&bp->b_waiters);
204 XFS_STATS_INC(xb_create);
206 trace_xfs_buf_init(bp, _RET_IP_);
210 * Allocate a page array capable of holding a specified number
211 * of pages, and point the page buf at it.
217 xfs_buf_flags_t flags)
219 /* Make sure that we have a page list */
220 if (bp->b_pages == NULL) {
221 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
222 bp->b_page_count = page_count;
223 if (page_count <= XB_PAGES) {
224 bp->b_pages = bp->b_page_array;
226 bp->b_pages = kmem_alloc(sizeof(struct page *) *
227 page_count, xb_to_km(flags));
228 if (bp->b_pages == NULL)
231 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
237 * Frees b_pages if it was allocated.
243 if (bp->b_pages != bp->b_page_array) {
244 kmem_free(bp->b_pages);
250 * Releases the specified buffer.
252 * The modification state of any associated pages is left unchanged.
253 * The buffer most not be on any hash - use xfs_buf_rele instead for
254 * hashed and refcounted buffers
260 trace_xfs_buf_free(bp, _RET_IP_);
262 ASSERT(list_empty(&bp->b_lru));
264 if (bp->b_flags & _XBF_PAGES) {
267 if (xfs_buf_is_vmapped(bp))
268 vm_unmap_ram(bp->b_addr - bp->b_offset,
271 for (i = 0; i < bp->b_page_count; i++) {
272 struct page *page = bp->b_pages[i];
276 } else if (bp->b_flags & _XBF_KMEM)
277 kmem_free(bp->b_addr);
278 _xfs_buf_free_pages(bp);
279 xfs_buf_deallocate(bp);
283 * Allocates all the pages for buffer in question and builds it's page list.
286 xfs_buf_allocate_memory(
290 size_t size = bp->b_count_desired;
291 size_t nbytes, offset;
292 gfp_t gfp_mask = xb_to_gfp(flags);
293 unsigned short page_count, i;
298 * for buffers that are contained within a single page, just allocate
299 * the memory from the heap - there's no need for the complexity of
300 * page arrays to keep allocation down to order 0.
302 if (bp->b_buffer_length < PAGE_SIZE) {
303 bp->b_addr = kmem_alloc(bp->b_buffer_length, xb_to_km(flags));
305 /* low memory - use alloc_page loop instead */
309 if (((unsigned long)(bp->b_addr + bp->b_buffer_length - 1) &
311 ((unsigned long)bp->b_addr & PAGE_MASK)) {
312 /* b_addr spans two pages - use alloc_page instead */
313 kmem_free(bp->b_addr);
317 bp->b_offset = offset_in_page(bp->b_addr);
318 bp->b_pages = bp->b_page_array;
319 bp->b_pages[0] = virt_to_page(bp->b_addr);
320 bp->b_page_count = 1;
321 bp->b_flags |= XBF_MAPPED | _XBF_KMEM;
326 end = bp->b_file_offset + bp->b_buffer_length;
327 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
328 error = _xfs_buf_get_pages(bp, page_count, flags);
332 offset = bp->b_offset;
333 bp->b_flags |= _XBF_PAGES;
335 for (i = 0; i < bp->b_page_count; i++) {
339 page = alloc_page(gfp_mask);
340 if (unlikely(page == NULL)) {
341 if (flags & XBF_READ_AHEAD) {
342 bp->b_page_count = i;
348 * This could deadlock.
350 * But until all the XFS lowlevel code is revamped to
351 * handle buffer allocation failures we can't do much.
353 if (!(++retries % 100))
355 "possible memory allocation deadlock in %s (mode:0x%x)",
358 XFS_STATS_INC(xb_page_retries);
359 congestion_wait(BLK_RW_ASYNC, HZ/50);
363 XFS_STATS_INC(xb_page_found);
365 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
367 bp->b_pages[i] = page;
373 for (i = 0; i < bp->b_page_count; i++)
374 __free_page(bp->b_pages[i]);
379 * Map buffer into kernel address-space if necessary.
386 ASSERT(bp->b_flags & _XBF_PAGES);
387 if (bp->b_page_count == 1) {
388 /* A single page buffer is always mappable */
389 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
390 bp->b_flags |= XBF_MAPPED;
391 } else if (flags & XBF_MAPPED) {
395 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
400 } while (retried++ <= 1);
404 bp->b_addr += bp->b_offset;
405 bp->b_flags |= XBF_MAPPED;
412 * Finding and Reading Buffers
416 * Look up, and creates if absent, a lockable buffer for
417 * a given range of an inode. The buffer is returned
418 * locked. No I/O is implied by this call.
422 xfs_buftarg_t *btp, /* block device target */
423 xfs_off_t ioff, /* starting offset of range */
424 size_t isize, /* length of range */
425 xfs_buf_flags_t flags,
428 xfs_off_t range_base;
430 struct xfs_perag *pag;
431 struct rb_node **rbp;
432 struct rb_node *parent;
435 range_base = (ioff << BBSHIFT);
436 range_length = (isize << BBSHIFT);
438 /* Check for IOs smaller than the sector size / not sector aligned */
439 ASSERT(!(range_length < (1 << btp->bt_sshift)));
440 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
443 pag = xfs_perag_get(btp->bt_mount,
444 xfs_daddr_to_agno(btp->bt_mount, ioff));
447 spin_lock(&pag->pag_buf_lock);
448 rbp = &pag->pag_buf_tree.rb_node;
453 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
455 if (range_base < bp->b_file_offset)
456 rbp = &(*rbp)->rb_left;
457 else if (range_base > bp->b_file_offset)
458 rbp = &(*rbp)->rb_right;
461 * found a block offset match. If the range doesn't
462 * match, the only way this is allowed is if the buffer
463 * in the cache is stale and the transaction that made
464 * it stale has not yet committed. i.e. we are
465 * reallocating a busy extent. Skip this buffer and
466 * continue searching to the right for an exact match.
468 if (bp->b_buffer_length != range_length) {
469 ASSERT(bp->b_flags & XBF_STALE);
470 rbp = &(*rbp)->rb_right;
473 atomic_inc(&bp->b_hold);
480 _xfs_buf_initialize(new_bp, btp, range_base,
481 range_length, flags);
482 rb_link_node(&new_bp->b_rbnode, parent, rbp);
483 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
484 /* the buffer keeps the perag reference until it is freed */
486 spin_unlock(&pag->pag_buf_lock);
488 XFS_STATS_INC(xb_miss_locked);
489 spin_unlock(&pag->pag_buf_lock);
495 spin_unlock(&pag->pag_buf_lock);
498 if (!xfs_buf_trylock(bp)) {
499 if (flags & XBF_TRYLOCK) {
501 XFS_STATS_INC(xb_busy_locked);
505 XFS_STATS_INC(xb_get_locked_waited);
509 * if the buffer is stale, clear all the external state associated with
510 * it. We need to keep flags such as how we allocated the buffer memory
513 if (bp->b_flags & XBF_STALE) {
514 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
515 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
518 trace_xfs_buf_find(bp, flags, _RET_IP_);
519 XFS_STATS_INC(xb_get_locked);
524 * Assembles a buffer covering the specified range.
525 * Storage in memory for all portions of the buffer will be allocated,
526 * although backing storage may not be.
530 xfs_buftarg_t *target,/* target for buffer */
531 xfs_off_t ioff, /* starting offset of range */
532 size_t isize, /* length of range */
533 xfs_buf_flags_t flags)
535 xfs_buf_t *bp, *new_bp;
538 new_bp = xfs_buf_allocate(flags);
539 if (unlikely(!new_bp))
542 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
544 error = xfs_buf_allocate_memory(bp, flags);
548 xfs_buf_deallocate(new_bp);
549 if (unlikely(bp == NULL))
553 if (!(bp->b_flags & XBF_MAPPED)) {
554 error = _xfs_buf_map_pages(bp, flags);
555 if (unlikely(error)) {
556 xfs_warn(target->bt_mount,
557 "%s: failed to map pages\n", __func__);
562 XFS_STATS_INC(xb_get);
565 * Always fill in the block number now, the mapped cases can do
566 * their own overlay of this later.
569 bp->b_count_desired = bp->b_buffer_length;
571 trace_xfs_buf_get(bp, flags, _RET_IP_);
575 if (flags & (XBF_LOCK | XBF_TRYLOCK))
584 xfs_buf_flags_t flags)
588 ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
589 ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
591 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | XBF_READ_AHEAD);
592 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
594 status = xfs_buf_iorequest(bp);
595 if (status || bp->b_error || (flags & XBF_ASYNC))
597 return xfs_buf_iowait(bp);
602 xfs_buftarg_t *target,
605 xfs_buf_flags_t flags)
611 bp = xfs_buf_get(target, ioff, isize, flags);
613 trace_xfs_buf_read(bp, flags, _RET_IP_);
615 if (!XFS_BUF_ISDONE(bp)) {
616 XFS_STATS_INC(xb_get_read);
617 _xfs_buf_read(bp, flags);
618 } else if (flags & XBF_ASYNC) {
620 * Read ahead call which is already satisfied,
625 /* We do not want read in the flags */
626 bp->b_flags &= ~XBF_READ;
633 if (flags & (XBF_LOCK | XBF_TRYLOCK))
640 * If we are not low on memory then do the readahead in a deadlock
645 xfs_buftarg_t *target,
649 if (bdi_read_congested(target->bt_bdi))
652 xfs_buf_read(target, ioff, isize,
653 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
657 * Read an uncached buffer from disk. Allocates and returns a locked
658 * buffer containing the disk contents or nothing.
661 xfs_buf_read_uncached(
662 struct xfs_mount *mp,
663 struct xfs_buftarg *target,
671 bp = xfs_buf_get_uncached(target, length, flags);
675 /* set up the buffer for a read IO */
676 XFS_BUF_SET_ADDR(bp, daddr);
680 error = xfs_buf_iowait(bp);
681 if (error || bp->b_error) {
691 xfs_buftarg_t *target)
695 bp = xfs_buf_allocate(0);
697 _xfs_buf_initialize(bp, target, 0, len, 0);
702 * Return a buffer allocated as an empty buffer and associated to external
703 * memory via xfs_buf_associate_memory() back to it's empty state.
711 _xfs_buf_free_pages(bp);
714 bp->b_page_count = 0;
716 bp->b_file_offset = 0;
717 bp->b_buffer_length = bp->b_count_desired = len;
718 bp->b_bn = XFS_BUF_DADDR_NULL;
719 bp->b_flags &= ~XBF_MAPPED;
722 static inline struct page *
726 if ((!is_vmalloc_addr(addr))) {
727 return virt_to_page(addr);
729 return vmalloc_to_page(addr);
734 xfs_buf_associate_memory(
741 unsigned long pageaddr;
742 unsigned long offset;
746 pageaddr = (unsigned long)mem & PAGE_MASK;
747 offset = (unsigned long)mem - pageaddr;
748 buflen = PAGE_ALIGN(len + offset);
749 page_count = buflen >> PAGE_SHIFT;
751 /* Free any previous set of page pointers */
753 _xfs_buf_free_pages(bp);
758 rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
762 bp->b_offset = offset;
764 for (i = 0; i < bp->b_page_count; i++) {
765 bp->b_pages[i] = mem_to_page((void *)pageaddr);
766 pageaddr += PAGE_SIZE;
769 bp->b_count_desired = len;
770 bp->b_buffer_length = buflen;
771 bp->b_flags |= XBF_MAPPED;
777 xfs_buf_get_uncached(
778 struct xfs_buftarg *target,
782 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
786 bp = xfs_buf_allocate(0);
787 if (unlikely(bp == NULL))
789 _xfs_buf_initialize(bp, target, 0, len, 0);
791 error = _xfs_buf_get_pages(bp, page_count, 0);
795 for (i = 0; i < page_count; i++) {
796 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
800 bp->b_flags |= _XBF_PAGES;
802 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
803 if (unlikely(error)) {
804 xfs_warn(target->bt_mount,
805 "%s: failed to map pages\n", __func__);
809 trace_xfs_buf_get_uncached(bp, _RET_IP_);
814 __free_page(bp->b_pages[i]);
815 _xfs_buf_free_pages(bp);
817 xfs_buf_deallocate(bp);
823 * Increment reference count on buffer, to hold the buffer concurrently
824 * with another thread which may release (free) the buffer asynchronously.
825 * Must hold the buffer already to call this function.
831 trace_xfs_buf_hold(bp, _RET_IP_);
832 atomic_inc(&bp->b_hold);
836 * Releases a hold on the specified buffer. If the
837 * the hold count is 1, calls xfs_buf_free.
843 struct xfs_perag *pag = bp->b_pag;
845 trace_xfs_buf_rele(bp, _RET_IP_);
848 ASSERT(list_empty(&bp->b_lru));
849 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
850 if (atomic_dec_and_test(&bp->b_hold))
855 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
857 ASSERT(atomic_read(&bp->b_hold) > 0);
858 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
859 if (!(bp->b_flags & XBF_STALE) &&
860 atomic_read(&bp->b_lru_ref)) {
862 spin_unlock(&pag->pag_buf_lock);
865 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
866 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
867 spin_unlock(&pag->pag_buf_lock);
876 * Lock a buffer object, if it is not already locked.
878 * If we come across a stale, pinned, locked buffer, we know that we are
879 * being asked to lock a buffer that has been reallocated. Because it is
880 * pinned, we know that the log has not been pushed to disk and hence it
881 * will still be locked. Rather than continuing to have trylock attempts
882 * fail until someone else pushes the log, push it ourselves before
883 * returning. This means that the xfsaild will not get stuck trying
884 * to push on stale inode buffers.
892 locked = down_trylock(&bp->b_sema) == 0;
895 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
896 xfs_log_force(bp->b_target->bt_mount, 0);
898 trace_xfs_buf_trylock(bp, _RET_IP_);
903 * Lock a buffer object.
905 * If we come across a stale, pinned, locked buffer, we know that we
906 * are being asked to lock a buffer that has been reallocated. Because
907 * it is pinned, we know that the log has not been pushed to disk and
908 * hence it will still be locked. Rather than sleeping until someone
909 * else pushes the log, push it ourselves before trying to get the lock.
915 trace_xfs_buf_lock(bp, _RET_IP_);
917 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
918 xfs_log_force(bp->b_target->bt_mount, 0);
922 trace_xfs_buf_lock_done(bp, _RET_IP_);
926 * Releases the lock on the buffer object.
927 * If the buffer is marked delwri but is not queued, do so before we
928 * unlock the buffer as we need to set flags correctly. We also need to
929 * take a reference for the delwri queue because the unlocker is going to
930 * drop their's and they don't know we just queued it.
939 trace_xfs_buf_unlock(bp, _RET_IP_);
946 DECLARE_WAITQUEUE (wait, current);
948 if (atomic_read(&bp->b_pin_count) == 0)
951 add_wait_queue(&bp->b_waiters, &wait);
953 set_current_state(TASK_UNINTERRUPTIBLE);
954 if (atomic_read(&bp->b_pin_count) == 0)
958 remove_wait_queue(&bp->b_waiters, &wait);
959 set_current_state(TASK_RUNNING);
963 * Buffer Utility Routines
968 struct work_struct *work)
971 container_of(work, xfs_buf_t, b_iodone_work);
974 (*(bp->b_iodone))(bp);
975 else if (bp->b_flags & XBF_ASYNC)
984 trace_xfs_buf_iodone(bp, _RET_IP_);
986 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
987 if (bp->b_error == 0)
988 bp->b_flags |= XBF_DONE;
990 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
992 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
993 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
995 xfs_buf_iodone_work(&bp->b_iodone_work);
998 complete(&bp->b_iowait);
1007 ASSERT(error >= 0 && error <= 0xffff);
1008 bp->b_error = (unsigned short)error;
1009 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1018 bp->b_flags |= XBF_WRITE;
1019 bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1021 xfs_buf_delwri_dequeue(bp);
1024 error = xfs_buf_iowait(bp);
1026 xfs_force_shutdown(bp->b_target->bt_mount,
1027 SHUTDOWN_META_IO_ERROR);
1033 * Called when we want to stop a buffer from getting written or read.
1034 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1035 * so that the proper iodone callbacks get called.
1041 #ifdef XFSERRORDEBUG
1042 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1046 * No need to wait until the buffer is unpinned, we aren't flushing it.
1048 xfs_buf_ioerror(bp, EIO);
1051 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1054 xfs_buf_delwri_dequeue(bp);
1058 xfs_buf_ioend(bp, 0);
1064 * Same as xfs_bioerror, except that we are releasing the buffer
1065 * here ourselves, and avoiding the xfs_buf_ioend call.
1066 * This is meant for userdata errors; metadata bufs come with
1067 * iodone functions attached, so that we can track down errors.
1073 int64_t fl = bp->b_flags;
1075 * No need to wait until the buffer is unpinned.
1076 * We aren't flushing it.
1078 * chunkhold expects B_DONE to be set, whether
1079 * we actually finish the I/O or not. We don't want to
1080 * change that interface.
1083 xfs_buf_delwri_dequeue(bp);
1086 bp->b_iodone = NULL;
1087 if (!(fl & XBF_ASYNC)) {
1089 * Mark b_error and B_ERROR _both_.
1090 * Lot's of chunkcache code assumes that.
1091 * There's no reason to mark error for
1094 xfs_buf_ioerror(bp, EIO);
1095 XFS_BUF_FINISH_IOWAIT(bp);
1105 * All xfs metadata buffers except log state machine buffers
1106 * get this attached as their b_bdstrat callback function.
1107 * This is so that we can catch a buffer
1108 * after prematurely unpinning it to forcibly shutdown the filesystem.
1114 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1115 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1117 * Metadata write that didn't get logged but
1118 * written delayed anyway. These aren't associated
1119 * with a transaction, and can be ignored.
1121 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1122 return xfs_bioerror_relse(bp);
1124 return xfs_bioerror(bp);
1127 xfs_buf_iorequest(bp);
1132 * Wrapper around bdstrat so that we can stop data from going to disk in case
1133 * we are shutting down the filesystem. Typically user data goes thru this
1134 * path; one of the exceptions is the superblock.
1138 struct xfs_mount *mp,
1141 if (XFS_FORCED_SHUTDOWN(mp)) {
1142 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1143 xfs_bioerror_relse(bp);
1147 xfs_buf_iorequest(bp);
1155 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1156 xfs_buf_ioend(bp, schedule);
1164 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1166 xfs_buf_ioerror(bp, -error);
1168 if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1169 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1171 _xfs_buf_ioend(bp, 1);
1179 int rw, map_i, total_nr_pages, nr_pages;
1181 int offset = bp->b_offset;
1182 int size = bp->b_count_desired;
1183 sector_t sector = bp->b_bn;
1185 total_nr_pages = bp->b_page_count;
1188 if (bp->b_flags & XBF_WRITE) {
1189 if (bp->b_flags & XBF_SYNCIO)
1193 if (bp->b_flags & XBF_FUA)
1195 if (bp->b_flags & XBF_FLUSH)
1197 } else if (bp->b_flags & XBF_READ_AHEAD) {
1203 /* we only use the buffer cache for meta-data */
1207 atomic_inc(&bp->b_io_remaining);
1208 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1209 if (nr_pages > total_nr_pages)
1210 nr_pages = total_nr_pages;
1212 bio = bio_alloc(GFP_NOIO, nr_pages);
1213 bio->bi_bdev = bp->b_target->bt_bdev;
1214 bio->bi_sector = sector;
1215 bio->bi_end_io = xfs_buf_bio_end_io;
1216 bio->bi_private = bp;
1219 for (; size && nr_pages; nr_pages--, map_i++) {
1220 int rbytes, nbytes = PAGE_SIZE - offset;
1225 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1226 if (rbytes < nbytes)
1230 sector += nbytes >> BBSHIFT;
1235 if (likely(bio->bi_size)) {
1236 if (xfs_buf_is_vmapped(bp)) {
1237 flush_kernel_vmap_range(bp->b_addr,
1238 xfs_buf_vmap_len(bp));
1240 submit_bio(rw, bio);
1244 xfs_buf_ioerror(bp, EIO);
1253 trace_xfs_buf_iorequest(bp, _RET_IP_);
1255 ASSERT(!(bp->b_flags & XBF_DELWRI));
1257 if (bp->b_flags & XBF_WRITE)
1258 xfs_buf_wait_unpin(bp);
1261 /* Set the count to 1 initially, this will stop an I/O
1262 * completion callout which happens before we have started
1263 * all the I/O from calling xfs_buf_ioend too early.
1265 atomic_set(&bp->b_io_remaining, 1);
1266 _xfs_buf_ioapply(bp);
1267 _xfs_buf_ioend(bp, 0);
1274 * Waits for I/O to complete on the buffer supplied.
1275 * It returns immediately if no I/O is pending.
1276 * It returns the I/O error code, if any, or 0 if there was no error.
1282 trace_xfs_buf_iowait(bp, _RET_IP_);
1284 wait_for_completion(&bp->b_iowait);
1286 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1297 if (bp->b_flags & XBF_MAPPED)
1298 return bp->b_addr + offset;
1300 offset += bp->b_offset;
1301 page = bp->b_pages[offset >> PAGE_SHIFT];
1302 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1306 * Move data into or out of a buffer.
1310 xfs_buf_t *bp, /* buffer to process */
1311 size_t boff, /* starting buffer offset */
1312 size_t bsize, /* length to copy */
1313 void *data, /* data address */
1314 xfs_buf_rw_t mode) /* read/write/zero flag */
1316 size_t bend, cpoff, csize;
1319 bend = boff + bsize;
1320 while (boff < bend) {
1321 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1322 cpoff = xfs_buf_poff(boff + bp->b_offset);
1323 csize = min_t(size_t,
1324 PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1326 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1330 memset(page_address(page) + cpoff, 0, csize);
1333 memcpy(data, page_address(page) + cpoff, csize);
1336 memcpy(page_address(page) + cpoff, data, csize);
1345 * Handling of buffer targets (buftargs).
1349 * Wait for any bufs with callbacks that have been submitted but have not yet
1350 * returned. These buffers will have an elevated hold count, so wait on those
1351 * while freeing all the buffers only held by the LRU.
1355 struct xfs_buftarg *btp)
1360 spin_lock(&btp->bt_lru_lock);
1361 while (!list_empty(&btp->bt_lru)) {
1362 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1363 if (atomic_read(&bp->b_hold) > 1) {
1364 spin_unlock(&btp->bt_lru_lock);
1369 * clear the LRU reference count so the bufer doesn't get
1370 * ignored in xfs_buf_rele().
1372 atomic_set(&bp->b_lru_ref, 0);
1373 spin_unlock(&btp->bt_lru_lock);
1375 spin_lock(&btp->bt_lru_lock);
1377 spin_unlock(&btp->bt_lru_lock);
1382 struct shrinker *shrink,
1383 struct shrink_control *sc)
1385 struct xfs_buftarg *btp = container_of(shrink,
1386 struct xfs_buftarg, bt_shrinker);
1388 int nr_to_scan = sc->nr_to_scan;
1392 return btp->bt_lru_nr;
1394 spin_lock(&btp->bt_lru_lock);
1395 while (!list_empty(&btp->bt_lru)) {
1396 if (nr_to_scan-- <= 0)
1399 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1402 * Decrement the b_lru_ref count unless the value is already
1403 * zero. If the value is already zero, we need to reclaim the
1404 * buffer, otherwise it gets another trip through the LRU.
1406 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1407 list_move_tail(&bp->b_lru, &btp->bt_lru);
1412 * remove the buffer from the LRU now to avoid needing another
1413 * lock round trip inside xfs_buf_rele().
1415 list_move(&bp->b_lru, &dispose);
1418 spin_unlock(&btp->bt_lru_lock);
1420 while (!list_empty(&dispose)) {
1421 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1422 list_del_init(&bp->b_lru);
1426 return btp->bt_lru_nr;
1431 struct xfs_mount *mp,
1432 struct xfs_buftarg *btp)
1434 unregister_shrinker(&btp->bt_shrinker);
1436 xfs_flush_buftarg(btp, 1);
1437 if (mp->m_flags & XFS_MOUNT_BARRIER)
1438 xfs_blkdev_issue_flush(btp);
1440 kthread_stop(btp->bt_task);
1445 xfs_setsize_buftarg_flags(
1447 unsigned int blocksize,
1448 unsigned int sectorsize,
1451 btp->bt_bsize = blocksize;
1452 btp->bt_sshift = ffs(sectorsize) - 1;
1453 btp->bt_smask = sectorsize - 1;
1455 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1456 xfs_warn(btp->bt_mount,
1457 "Cannot set_blocksize to %u on device %s\n",
1458 sectorsize, xfs_buf_target_name(btp));
1466 * When allocating the initial buffer target we have not yet
1467 * read in the superblock, so don't know what sized sectors
1468 * are being used is at this early stage. Play safe.
1471 xfs_setsize_buftarg_early(
1473 struct block_device *bdev)
1475 return xfs_setsize_buftarg_flags(btp,
1476 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1480 xfs_setsize_buftarg(
1482 unsigned int blocksize,
1483 unsigned int sectorsize)
1485 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1489 xfs_alloc_delwri_queue(
1493 INIT_LIST_HEAD(&btp->bt_delwri_queue);
1494 spin_lock_init(&btp->bt_delwri_lock);
1496 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1497 if (IS_ERR(btp->bt_task))
1498 return PTR_ERR(btp->bt_task);
1504 struct xfs_mount *mp,
1505 struct block_device *bdev,
1511 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1514 btp->bt_dev = bdev->bd_dev;
1515 btp->bt_bdev = bdev;
1516 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1520 INIT_LIST_HEAD(&btp->bt_lru);
1521 spin_lock_init(&btp->bt_lru_lock);
1522 if (xfs_setsize_buftarg_early(btp, bdev))
1524 if (xfs_alloc_delwri_queue(btp, fsname))
1526 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1527 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1528 register_shrinker(&btp->bt_shrinker);
1538 * Delayed write buffer handling
1541 xfs_buf_delwri_queue(
1544 struct xfs_buftarg *btp = bp->b_target;
1546 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1548 ASSERT(!(bp->b_flags & XBF_READ));
1550 spin_lock(&btp->bt_delwri_lock);
1551 if (!list_empty(&bp->b_list)) {
1552 /* if already in the queue, move it to the tail */
1553 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1554 list_move_tail(&bp->b_list, &btp->bt_delwri_queue);
1556 /* start xfsbufd as it is about to have something to do */
1557 if (list_empty(&btp->bt_delwri_queue))
1558 wake_up_process(bp->b_target->bt_task);
1560 atomic_inc(&bp->b_hold);
1561 bp->b_flags |= XBF_DELWRI | _XBF_DELWRI_Q | XBF_ASYNC;
1562 list_add_tail(&bp->b_list, &btp->bt_delwri_queue);
1564 bp->b_queuetime = jiffies;
1565 spin_unlock(&btp->bt_delwri_lock);
1569 xfs_buf_delwri_dequeue(
1574 spin_lock(&bp->b_target->bt_delwri_lock);
1575 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1576 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1577 list_del_init(&bp->b_list);
1580 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1581 spin_unlock(&bp->b_target->bt_delwri_lock);
1586 trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1590 * If a delwri buffer needs to be pushed before it has aged out, then promote
1591 * it to the head of the delwri queue so that it will be flushed on the next
1592 * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1593 * than the age currently needed to flush the buffer. Hence the next time the
1594 * xfsbufd sees it is guaranteed to be considered old enough to flush.
1597 xfs_buf_delwri_promote(
1600 struct xfs_buftarg *btp = bp->b_target;
1601 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1603 ASSERT(bp->b_flags & XBF_DELWRI);
1604 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1607 * Check the buffer age before locking the delayed write queue as we
1608 * don't need to promote buffers that are already past the flush age.
1610 if (bp->b_queuetime < jiffies - age)
1612 bp->b_queuetime = jiffies - age;
1613 spin_lock(&btp->bt_delwri_lock);
1614 list_move(&bp->b_list, &btp->bt_delwri_queue);
1615 spin_unlock(&btp->bt_delwri_lock);
1619 xfs_buf_runall_queues(
1620 struct workqueue_struct *queue)
1622 flush_workqueue(queue);
1626 * Move as many buffers as specified to the supplied list
1627 * idicating if we skipped any buffers to prevent deadlocks.
1630 xfs_buf_delwri_split(
1631 xfs_buftarg_t *target,
1632 struct list_head *list,
1639 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1640 INIT_LIST_HEAD(list);
1641 spin_lock(&target->bt_delwri_lock);
1642 list_for_each_entry_safe(bp, n, &target->bt_delwri_queue, b_list) {
1643 ASSERT(bp->b_flags & XBF_DELWRI);
1645 if (!xfs_buf_ispinned(bp) && xfs_buf_trylock(bp)) {
1647 time_before(jiffies, bp->b_queuetime + age)) {
1652 bp->b_flags &= ~(XBF_DELWRI | _XBF_DELWRI_Q);
1653 bp->b_flags |= XBF_WRITE;
1654 list_move_tail(&bp->b_list, list);
1655 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1660 spin_unlock(&target->bt_delwri_lock);
1665 * Compare function is more complex than it needs to be because
1666 * the return value is only 32 bits and we are doing comparisons
1672 struct list_head *a,
1673 struct list_head *b)
1675 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1676 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1679 diff = ap->b_bn - bp->b_bn;
1691 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1693 current->flags |= PF_MEMALLOC;
1698 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1699 long tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1700 struct list_head tmp;
1701 struct blk_plug plug;
1703 if (unlikely(freezing(current))) {
1704 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1707 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1710 /* sleep for a long time if there is nothing to do. */
1711 if (list_empty(&target->bt_delwri_queue))
1712 tout = MAX_SCHEDULE_TIMEOUT;
1713 schedule_timeout_interruptible(tout);
1715 xfs_buf_delwri_split(target, &tmp, age);
1716 list_sort(NULL, &tmp, xfs_buf_cmp);
1718 blk_start_plug(&plug);
1719 while (!list_empty(&tmp)) {
1721 bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1722 list_del_init(&bp->b_list);
1725 blk_finish_plug(&plug);
1726 } while (!kthread_should_stop());
1732 * Go through all incore buffers, and release buffers if they belong to
1733 * the given device. This is used in filesystem error handling to
1734 * preserve the consistency of its metadata.
1738 xfs_buftarg_t *target,
1743 LIST_HEAD(tmp_list);
1744 LIST_HEAD(wait_list);
1745 struct blk_plug plug;
1747 xfs_buf_runall_queues(xfsconvertd_workqueue);
1748 xfs_buf_runall_queues(xfsdatad_workqueue);
1749 xfs_buf_runall_queues(xfslogd_workqueue);
1751 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1752 pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1755 * Dropped the delayed write list lock, now walk the temporary list.
1756 * All I/O is issued async and then if we need to wait for completion
1757 * we do that after issuing all the IO.
1759 list_sort(NULL, &tmp_list, xfs_buf_cmp);
1761 blk_start_plug(&plug);
1762 while (!list_empty(&tmp_list)) {
1763 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1764 ASSERT(target == bp->b_target);
1765 list_del_init(&bp->b_list);
1767 bp->b_flags &= ~XBF_ASYNC;
1768 list_add(&bp->b_list, &wait_list);
1772 blk_finish_plug(&plug);
1775 /* Wait for IO to complete. */
1776 while (!list_empty(&wait_list)) {
1777 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1779 list_del_init(&bp->b_list);
1791 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1792 KM_ZONE_HWALIGN, NULL);
1796 xfslogd_workqueue = alloc_workqueue("xfslogd",
1797 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1798 if (!xfslogd_workqueue)
1799 goto out_free_buf_zone;
1801 xfsdatad_workqueue = alloc_workqueue("xfsdatad", WQ_MEM_RECLAIM, 1);
1802 if (!xfsdatad_workqueue)
1803 goto out_destroy_xfslogd_workqueue;
1805 xfsconvertd_workqueue = alloc_workqueue("xfsconvertd",
1807 if (!xfsconvertd_workqueue)
1808 goto out_destroy_xfsdatad_workqueue;
1812 out_destroy_xfsdatad_workqueue:
1813 destroy_workqueue(xfsdatad_workqueue);
1814 out_destroy_xfslogd_workqueue:
1815 destroy_workqueue(xfslogd_workqueue);
1817 kmem_zone_destroy(xfs_buf_zone);
1823 xfs_buf_terminate(void)
1825 destroy_workqueue(xfsconvertd_workqueue);
1826 destroy_workqueue(xfsdatad_workqueue);
1827 destroy_workqueue(xfslogd_workqueue);
1828 kmem_zone_destroy(xfs_buf_zone);
1831 #ifdef CONFIG_KDB_MODULES
1833 xfs_get_buftarg_list(void)
1835 return &xfs_buftarg_list;