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 xfs_buf_delwri_dequeue(bp);
155 atomic_set(&(bp)->b_lru_ref, 0);
156 if (!list_empty(&bp->b_lru)) {
157 struct xfs_buftarg *btp = bp->b_target;
159 spin_lock(&btp->bt_lru_lock);
160 if (!list_empty(&bp->b_lru)) {
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);
173 xfs_buftarg_t *target,
174 xfs_off_t range_base,
176 xfs_buf_flags_t flags)
179 * We don't want certain flags to appear in b_flags.
181 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
183 memset(bp, 0, sizeof(xfs_buf_t));
184 atomic_set(&bp->b_hold, 1);
185 atomic_set(&bp->b_lru_ref, 1);
186 init_completion(&bp->b_iowait);
187 INIT_LIST_HEAD(&bp->b_lru);
188 INIT_LIST_HEAD(&bp->b_list);
189 RB_CLEAR_NODE(&bp->b_rbnode);
190 sema_init(&bp->b_sema, 0); /* held, no waiters */
192 bp->b_target = target;
193 bp->b_file_offset = range_base;
195 * Set buffer_length and count_desired to the same value initially.
196 * I/O routines should use count_desired, which will be the same in
197 * most cases but may be reset (e.g. XFS recovery).
199 bp->b_buffer_length = bp->b_count_desired = range_length;
201 bp->b_bn = XFS_BUF_DADDR_NULL;
202 atomic_set(&bp->b_pin_count, 0);
203 init_waitqueue_head(&bp->b_waiters);
205 XFS_STATS_INC(xb_create);
207 trace_xfs_buf_init(bp, _RET_IP_);
211 * Allocate a page array capable of holding a specified number
212 * of pages, and point the page buf at it.
218 xfs_buf_flags_t flags)
220 /* Make sure that we have a page list */
221 if (bp->b_pages == NULL) {
222 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
223 bp->b_page_count = page_count;
224 if (page_count <= XB_PAGES) {
225 bp->b_pages = bp->b_page_array;
227 bp->b_pages = kmem_alloc(sizeof(struct page *) *
228 page_count, xb_to_km(flags));
229 if (bp->b_pages == NULL)
232 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
238 * Frees b_pages if it was allocated.
244 if (bp->b_pages != bp->b_page_array) {
245 kmem_free(bp->b_pages);
251 * Releases the specified buffer.
253 * The modification state of any associated pages is left unchanged.
254 * The buffer most not be on any hash - use xfs_buf_rele instead for
255 * hashed and refcounted buffers
261 trace_xfs_buf_free(bp, _RET_IP_);
263 ASSERT(list_empty(&bp->b_lru));
265 if (bp->b_flags & _XBF_PAGES) {
268 if (xfs_buf_is_vmapped(bp))
269 vm_unmap_ram(bp->b_addr - bp->b_offset,
272 for (i = 0; i < bp->b_page_count; i++) {
273 struct page *page = bp->b_pages[i];
277 } else if (bp->b_flags & _XBF_KMEM)
278 kmem_free(bp->b_addr);
279 _xfs_buf_free_pages(bp);
280 xfs_buf_deallocate(bp);
284 * Allocates all the pages for buffer in question and builds it's page list.
287 xfs_buf_allocate_memory(
291 size_t size = bp->b_count_desired;
292 size_t nbytes, offset;
293 gfp_t gfp_mask = xb_to_gfp(flags);
294 unsigned short page_count, i;
299 * for buffers that are contained within a single page, just allocate
300 * the memory from the heap - there's no need for the complexity of
301 * page arrays to keep allocation down to order 0.
303 if (bp->b_buffer_length < PAGE_SIZE) {
304 bp->b_addr = kmem_alloc(bp->b_buffer_length, xb_to_km(flags));
306 /* low memory - use alloc_page loop instead */
310 if (((unsigned long)(bp->b_addr + bp->b_buffer_length - 1) &
312 ((unsigned long)bp->b_addr & PAGE_MASK)) {
313 /* b_addr spans two pages - use alloc_page instead */
314 kmem_free(bp->b_addr);
318 bp->b_offset = offset_in_page(bp->b_addr);
319 bp->b_pages = bp->b_page_array;
320 bp->b_pages[0] = virt_to_page(bp->b_addr);
321 bp->b_page_count = 1;
322 bp->b_flags |= XBF_MAPPED | _XBF_KMEM;
327 end = bp->b_file_offset + bp->b_buffer_length;
328 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
329 error = _xfs_buf_get_pages(bp, page_count, flags);
333 offset = bp->b_offset;
334 bp->b_flags |= _XBF_PAGES;
336 for (i = 0; i < bp->b_page_count; i++) {
340 page = alloc_page(gfp_mask);
341 if (unlikely(page == NULL)) {
342 if (flags & XBF_READ_AHEAD) {
343 bp->b_page_count = i;
349 * This could deadlock.
351 * But until all the XFS lowlevel code is revamped to
352 * handle buffer allocation failures we can't do much.
354 if (!(++retries % 100))
356 "possible memory allocation deadlock in %s (mode:0x%x)",
359 XFS_STATS_INC(xb_page_retries);
360 congestion_wait(BLK_RW_ASYNC, HZ/50);
364 XFS_STATS_INC(xb_page_found);
366 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
368 bp->b_pages[i] = page;
374 for (i = 0; i < bp->b_page_count; i++)
375 __free_page(bp->b_pages[i]);
380 * Map buffer into kernel address-space if necessary.
387 ASSERT(bp->b_flags & _XBF_PAGES);
388 if (bp->b_page_count == 1) {
389 /* A single page buffer is always mappable */
390 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
391 bp->b_flags |= XBF_MAPPED;
392 } else if (flags & XBF_MAPPED) {
396 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
401 } while (retried++ <= 1);
405 bp->b_addr += bp->b_offset;
406 bp->b_flags |= XBF_MAPPED;
413 * Finding and Reading Buffers
417 * Look up, and creates if absent, a lockable buffer for
418 * a given range of an inode. The buffer is returned
419 * locked. No I/O is implied by this call.
423 xfs_buftarg_t *btp, /* block device target */
424 xfs_off_t ioff, /* starting offset of range */
425 size_t isize, /* length of range */
426 xfs_buf_flags_t flags,
429 xfs_off_t range_base;
431 struct xfs_perag *pag;
432 struct rb_node **rbp;
433 struct rb_node *parent;
436 range_base = (ioff << BBSHIFT);
437 range_length = (isize << BBSHIFT);
439 /* Check for IOs smaller than the sector size / not sector aligned */
440 ASSERT(!(range_length < (1 << btp->bt_sshift)));
441 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
444 pag = xfs_perag_get(btp->bt_mount,
445 xfs_daddr_to_agno(btp->bt_mount, ioff));
448 spin_lock(&pag->pag_buf_lock);
449 rbp = &pag->pag_buf_tree.rb_node;
454 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
456 if (range_base < bp->b_file_offset)
457 rbp = &(*rbp)->rb_left;
458 else if (range_base > bp->b_file_offset)
459 rbp = &(*rbp)->rb_right;
462 * found a block offset match. If the range doesn't
463 * match, the only way this is allowed is if the buffer
464 * in the cache is stale and the transaction that made
465 * it stale has not yet committed. i.e. we are
466 * reallocating a busy extent. Skip this buffer and
467 * continue searching to the right for an exact match.
469 if (bp->b_buffer_length != range_length) {
470 ASSERT(bp->b_flags & XBF_STALE);
471 rbp = &(*rbp)->rb_right;
474 atomic_inc(&bp->b_hold);
481 rb_link_node(&new_bp->b_rbnode, parent, rbp);
482 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
483 /* the buffer keeps the perag reference until it is freed */
485 spin_unlock(&pag->pag_buf_lock);
487 XFS_STATS_INC(xb_miss_locked);
488 spin_unlock(&pag->pag_buf_lock);
494 spin_unlock(&pag->pag_buf_lock);
497 if (!xfs_buf_trylock(bp)) {
498 if (flags & XBF_TRYLOCK) {
500 XFS_STATS_INC(xb_busy_locked);
504 XFS_STATS_INC(xb_get_locked_waited);
508 * if the buffer is stale, clear all the external state associated with
509 * it. We need to keep flags such as how we allocated the buffer memory
512 if (bp->b_flags & XBF_STALE) {
513 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
514 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
517 trace_xfs_buf_find(bp, flags, _RET_IP_);
518 XFS_STATS_INC(xb_get_locked);
523 * Assembles a buffer covering the specified range. The code is optimised for
524 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
525 * more hits than misses.
529 xfs_buftarg_t *target,/* target for buffer */
530 xfs_off_t ioff, /* starting offset of range */
531 size_t isize, /* length of range */
532 xfs_buf_flags_t flags)
535 struct xfs_buf *new_bp;
538 bp = _xfs_buf_find(target, ioff, isize, flags, NULL);
542 new_bp = xfs_buf_allocate(flags);
543 if (unlikely(!new_bp))
546 _xfs_buf_initialize(new_bp, target,
547 ioff << BBSHIFT, isize << BBSHIFT, flags);
549 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
551 xfs_buf_deallocate(new_bp);
556 error = xfs_buf_allocate_memory(bp, flags);
560 xfs_buf_deallocate(new_bp);
563 * Now we have a workable buffer, fill in the block number so
564 * that we can do IO on it.
567 bp->b_count_desired = bp->b_buffer_length;
570 if (!(bp->b_flags & XBF_MAPPED)) {
571 error = _xfs_buf_map_pages(bp, flags);
572 if (unlikely(error)) {
573 xfs_warn(target->bt_mount,
574 "%s: failed to map pages\n", __func__);
579 XFS_STATS_INC(xb_get);
580 trace_xfs_buf_get(bp, flags, _RET_IP_);
584 if (flags & (XBF_LOCK | XBF_TRYLOCK))
593 xfs_buf_flags_t flags)
597 ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
598 ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
600 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | XBF_READ_AHEAD);
601 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
603 status = xfs_buf_iorequest(bp);
604 if (status || bp->b_error || (flags & XBF_ASYNC))
606 return xfs_buf_iowait(bp);
611 xfs_buftarg_t *target,
614 xfs_buf_flags_t flags)
620 bp = xfs_buf_get(target, ioff, isize, flags);
622 trace_xfs_buf_read(bp, flags, _RET_IP_);
624 if (!XFS_BUF_ISDONE(bp)) {
625 XFS_STATS_INC(xb_get_read);
626 _xfs_buf_read(bp, flags);
627 } else if (flags & XBF_ASYNC) {
629 * Read ahead call which is already satisfied,
634 /* We do not want read in the flags */
635 bp->b_flags &= ~XBF_READ;
642 if (flags & (XBF_LOCK | XBF_TRYLOCK))
649 * If we are not low on memory then do the readahead in a deadlock
654 xfs_buftarg_t *target,
658 if (bdi_read_congested(target->bt_bdi))
661 xfs_buf_read(target, ioff, isize,
662 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
666 * Read an uncached buffer from disk. Allocates and returns a locked
667 * buffer containing the disk contents or nothing.
670 xfs_buf_read_uncached(
671 struct xfs_mount *mp,
672 struct xfs_buftarg *target,
680 bp = xfs_buf_get_uncached(target, length, flags);
684 /* set up the buffer for a read IO */
685 XFS_BUF_SET_ADDR(bp, daddr);
689 error = xfs_buf_iowait(bp);
690 if (error || bp->b_error) {
700 xfs_buftarg_t *target)
704 bp = xfs_buf_allocate(0);
706 _xfs_buf_initialize(bp, target, 0, len, 0);
711 * Return a buffer allocated as an empty buffer and associated to external
712 * memory via xfs_buf_associate_memory() back to it's empty state.
720 _xfs_buf_free_pages(bp);
723 bp->b_page_count = 0;
725 bp->b_file_offset = 0;
726 bp->b_buffer_length = bp->b_count_desired = len;
727 bp->b_bn = XFS_BUF_DADDR_NULL;
728 bp->b_flags &= ~XBF_MAPPED;
731 static inline struct page *
735 if ((!is_vmalloc_addr(addr))) {
736 return virt_to_page(addr);
738 return vmalloc_to_page(addr);
743 xfs_buf_associate_memory(
750 unsigned long pageaddr;
751 unsigned long offset;
755 pageaddr = (unsigned long)mem & PAGE_MASK;
756 offset = (unsigned long)mem - pageaddr;
757 buflen = PAGE_ALIGN(len + offset);
758 page_count = buflen >> PAGE_SHIFT;
760 /* Free any previous set of page pointers */
762 _xfs_buf_free_pages(bp);
767 rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
771 bp->b_offset = offset;
773 for (i = 0; i < bp->b_page_count; i++) {
774 bp->b_pages[i] = mem_to_page((void *)pageaddr);
775 pageaddr += PAGE_SIZE;
778 bp->b_count_desired = len;
779 bp->b_buffer_length = buflen;
780 bp->b_flags |= XBF_MAPPED;
786 xfs_buf_get_uncached(
787 struct xfs_buftarg *target,
791 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
795 bp = xfs_buf_allocate(0);
796 if (unlikely(bp == NULL))
798 _xfs_buf_initialize(bp, target, 0, len, 0);
800 error = _xfs_buf_get_pages(bp, page_count, 0);
804 for (i = 0; i < page_count; i++) {
805 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
809 bp->b_flags |= _XBF_PAGES;
811 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
812 if (unlikely(error)) {
813 xfs_warn(target->bt_mount,
814 "%s: failed to map pages\n", __func__);
818 trace_xfs_buf_get_uncached(bp, _RET_IP_);
823 __free_page(bp->b_pages[i]);
824 _xfs_buf_free_pages(bp);
826 xfs_buf_deallocate(bp);
832 * Increment reference count on buffer, to hold the buffer concurrently
833 * with another thread which may release (free) the buffer asynchronously.
834 * Must hold the buffer already to call this function.
840 trace_xfs_buf_hold(bp, _RET_IP_);
841 atomic_inc(&bp->b_hold);
845 * Releases a hold on the specified buffer. If the
846 * the hold count is 1, calls xfs_buf_free.
852 struct xfs_perag *pag = bp->b_pag;
854 trace_xfs_buf_rele(bp, _RET_IP_);
857 ASSERT(list_empty(&bp->b_lru));
858 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
859 if (atomic_dec_and_test(&bp->b_hold))
864 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
866 ASSERT(atomic_read(&bp->b_hold) > 0);
867 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
868 if (!(bp->b_flags & XBF_STALE) &&
869 atomic_read(&bp->b_lru_ref)) {
871 spin_unlock(&pag->pag_buf_lock);
874 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
875 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
876 spin_unlock(&pag->pag_buf_lock);
885 * Lock a buffer object, if it is not already locked.
887 * If we come across a stale, pinned, locked buffer, we know that we are
888 * being asked to lock a buffer that has been reallocated. Because it is
889 * pinned, we know that the log has not been pushed to disk and hence it
890 * will still be locked. Rather than continuing to have trylock attempts
891 * fail until someone else pushes the log, push it ourselves before
892 * returning. This means that the xfsaild will not get stuck trying
893 * to push on stale inode buffers.
901 locked = down_trylock(&bp->b_sema) == 0;
904 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
905 xfs_log_force(bp->b_target->bt_mount, 0);
907 trace_xfs_buf_trylock(bp, _RET_IP_);
912 * Lock a buffer object.
914 * If we come across a stale, pinned, locked buffer, we know that we
915 * are being asked to lock a buffer that has been reallocated. Because
916 * it is pinned, we know that the log has not been pushed to disk and
917 * hence it will still be locked. Rather than sleeping until someone
918 * else pushes the log, push it ourselves before trying to get the lock.
924 trace_xfs_buf_lock(bp, _RET_IP_);
926 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
927 xfs_log_force(bp->b_target->bt_mount, 0);
931 trace_xfs_buf_lock_done(bp, _RET_IP_);
935 * Releases the lock on the buffer object.
936 * If the buffer is marked delwri but is not queued, do so before we
937 * unlock the buffer as we need to set flags correctly. We also need to
938 * take a reference for the delwri queue because the unlocker is going to
939 * drop their's and they don't know we just queued it.
948 trace_xfs_buf_unlock(bp, _RET_IP_);
955 DECLARE_WAITQUEUE (wait, current);
957 if (atomic_read(&bp->b_pin_count) == 0)
960 add_wait_queue(&bp->b_waiters, &wait);
962 set_current_state(TASK_UNINTERRUPTIBLE);
963 if (atomic_read(&bp->b_pin_count) == 0)
967 remove_wait_queue(&bp->b_waiters, &wait);
968 set_current_state(TASK_RUNNING);
972 * Buffer Utility Routines
977 struct work_struct *work)
980 container_of(work, xfs_buf_t, b_iodone_work);
983 (*(bp->b_iodone))(bp);
984 else if (bp->b_flags & XBF_ASYNC)
993 trace_xfs_buf_iodone(bp, _RET_IP_);
995 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
996 if (bp->b_error == 0)
997 bp->b_flags |= XBF_DONE;
999 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1001 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1002 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1004 xfs_buf_iodone_work(&bp->b_iodone_work);
1007 complete(&bp->b_iowait);
1016 ASSERT(error >= 0 && error <= 0xffff);
1017 bp->b_error = (unsigned short)error;
1018 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1027 bp->b_flags |= XBF_WRITE;
1028 bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1030 xfs_buf_delwri_dequeue(bp);
1033 error = xfs_buf_iowait(bp);
1035 xfs_force_shutdown(bp->b_target->bt_mount,
1036 SHUTDOWN_META_IO_ERROR);
1042 * Called when we want to stop a buffer from getting written or read.
1043 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1044 * so that the proper iodone callbacks get called.
1050 #ifdef XFSERRORDEBUG
1051 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1055 * No need to wait until the buffer is unpinned, we aren't flushing it.
1057 xfs_buf_ioerror(bp, EIO);
1060 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1066 xfs_buf_ioend(bp, 0);
1072 * Same as xfs_bioerror, except that we are releasing the buffer
1073 * here ourselves, and avoiding the xfs_buf_ioend call.
1074 * This is meant for userdata errors; metadata bufs come with
1075 * iodone functions attached, so that we can track down errors.
1081 int64_t fl = bp->b_flags;
1083 * No need to wait until the buffer is unpinned.
1084 * We aren't flushing it.
1086 * chunkhold expects B_DONE to be set, whether
1087 * we actually finish the I/O or not. We don't want to
1088 * change that interface.
1093 bp->b_iodone = NULL;
1094 if (!(fl & XBF_ASYNC)) {
1096 * Mark b_error and B_ERROR _both_.
1097 * Lot's of chunkcache code assumes that.
1098 * There's no reason to mark error for
1101 xfs_buf_ioerror(bp, EIO);
1102 complete(&bp->b_iowait);
1112 * All xfs metadata buffers except log state machine buffers
1113 * get this attached as their b_bdstrat callback function.
1114 * This is so that we can catch a buffer
1115 * after prematurely unpinning it to forcibly shutdown the filesystem.
1121 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1122 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1124 * Metadata write that didn't get logged but
1125 * written delayed anyway. These aren't associated
1126 * with a transaction, and can be ignored.
1128 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1129 return xfs_bioerror_relse(bp);
1131 return xfs_bioerror(bp);
1134 xfs_buf_iorequest(bp);
1139 * Wrapper around bdstrat so that we can stop data from going to disk in case
1140 * we are shutting down the filesystem. Typically user data goes thru this
1141 * path; one of the exceptions is the superblock.
1145 struct xfs_mount *mp,
1148 if (XFS_FORCED_SHUTDOWN(mp)) {
1149 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1150 xfs_bioerror_relse(bp);
1154 xfs_buf_iorequest(bp);
1162 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1163 xfs_buf_ioend(bp, schedule);
1171 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1173 xfs_buf_ioerror(bp, -error);
1175 if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1176 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1178 _xfs_buf_ioend(bp, 1);
1186 int rw, map_i, total_nr_pages, nr_pages;
1188 int offset = bp->b_offset;
1189 int size = bp->b_count_desired;
1190 sector_t sector = bp->b_bn;
1192 total_nr_pages = bp->b_page_count;
1195 if (bp->b_flags & XBF_WRITE) {
1196 if (bp->b_flags & XBF_SYNCIO)
1200 if (bp->b_flags & XBF_FUA)
1202 if (bp->b_flags & XBF_FLUSH)
1204 } else if (bp->b_flags & XBF_READ_AHEAD) {
1210 /* we only use the buffer cache for meta-data */
1214 atomic_inc(&bp->b_io_remaining);
1215 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1216 if (nr_pages > total_nr_pages)
1217 nr_pages = total_nr_pages;
1219 bio = bio_alloc(GFP_NOIO, nr_pages);
1220 bio->bi_bdev = bp->b_target->bt_bdev;
1221 bio->bi_sector = sector;
1222 bio->bi_end_io = xfs_buf_bio_end_io;
1223 bio->bi_private = bp;
1226 for (; size && nr_pages; nr_pages--, map_i++) {
1227 int rbytes, nbytes = PAGE_SIZE - offset;
1232 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1233 if (rbytes < nbytes)
1237 sector += nbytes >> BBSHIFT;
1242 if (likely(bio->bi_size)) {
1243 if (xfs_buf_is_vmapped(bp)) {
1244 flush_kernel_vmap_range(bp->b_addr,
1245 xfs_buf_vmap_len(bp));
1247 submit_bio(rw, bio);
1251 xfs_buf_ioerror(bp, EIO);
1260 trace_xfs_buf_iorequest(bp, _RET_IP_);
1262 ASSERT(!(bp->b_flags & XBF_DELWRI));
1264 if (bp->b_flags & XBF_WRITE)
1265 xfs_buf_wait_unpin(bp);
1268 /* Set the count to 1 initially, this will stop an I/O
1269 * completion callout which happens before we have started
1270 * all the I/O from calling xfs_buf_ioend too early.
1272 atomic_set(&bp->b_io_remaining, 1);
1273 _xfs_buf_ioapply(bp);
1274 _xfs_buf_ioend(bp, 0);
1281 * Waits for I/O to complete on the buffer supplied.
1282 * It returns immediately if no I/O is pending.
1283 * It returns the I/O error code, if any, or 0 if there was no error.
1289 trace_xfs_buf_iowait(bp, _RET_IP_);
1291 wait_for_completion(&bp->b_iowait);
1293 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1304 if (bp->b_flags & XBF_MAPPED)
1305 return bp->b_addr + offset;
1307 offset += bp->b_offset;
1308 page = bp->b_pages[offset >> PAGE_SHIFT];
1309 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1313 * Move data into or out of a buffer.
1317 xfs_buf_t *bp, /* buffer to process */
1318 size_t boff, /* starting buffer offset */
1319 size_t bsize, /* length to copy */
1320 void *data, /* data address */
1321 xfs_buf_rw_t mode) /* read/write/zero flag */
1323 size_t bend, cpoff, csize;
1326 bend = boff + bsize;
1327 while (boff < bend) {
1328 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1329 cpoff = xfs_buf_poff(boff + bp->b_offset);
1330 csize = min_t(size_t,
1331 PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1333 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1337 memset(page_address(page) + cpoff, 0, csize);
1340 memcpy(data, page_address(page) + cpoff, csize);
1343 memcpy(page_address(page) + cpoff, data, csize);
1352 * Handling of buffer targets (buftargs).
1356 * Wait for any bufs with callbacks that have been submitted but have not yet
1357 * returned. These buffers will have an elevated hold count, so wait on those
1358 * while freeing all the buffers only held by the LRU.
1362 struct xfs_buftarg *btp)
1367 spin_lock(&btp->bt_lru_lock);
1368 while (!list_empty(&btp->bt_lru)) {
1369 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1370 if (atomic_read(&bp->b_hold) > 1) {
1371 spin_unlock(&btp->bt_lru_lock);
1376 * clear the LRU reference count so the bufer doesn't get
1377 * ignored in xfs_buf_rele().
1379 atomic_set(&bp->b_lru_ref, 0);
1380 spin_unlock(&btp->bt_lru_lock);
1382 spin_lock(&btp->bt_lru_lock);
1384 spin_unlock(&btp->bt_lru_lock);
1389 struct shrinker *shrink,
1390 struct shrink_control *sc)
1392 struct xfs_buftarg *btp = container_of(shrink,
1393 struct xfs_buftarg, bt_shrinker);
1395 int nr_to_scan = sc->nr_to_scan;
1399 return btp->bt_lru_nr;
1401 spin_lock(&btp->bt_lru_lock);
1402 while (!list_empty(&btp->bt_lru)) {
1403 if (nr_to_scan-- <= 0)
1406 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1409 * Decrement the b_lru_ref count unless the value is already
1410 * zero. If the value is already zero, we need to reclaim the
1411 * buffer, otherwise it gets another trip through the LRU.
1413 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1414 list_move_tail(&bp->b_lru, &btp->bt_lru);
1419 * remove the buffer from the LRU now to avoid needing another
1420 * lock round trip inside xfs_buf_rele().
1422 list_move(&bp->b_lru, &dispose);
1425 spin_unlock(&btp->bt_lru_lock);
1427 while (!list_empty(&dispose)) {
1428 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1429 list_del_init(&bp->b_lru);
1433 return btp->bt_lru_nr;
1438 struct xfs_mount *mp,
1439 struct xfs_buftarg *btp)
1441 unregister_shrinker(&btp->bt_shrinker);
1443 xfs_flush_buftarg(btp, 1);
1444 if (mp->m_flags & XFS_MOUNT_BARRIER)
1445 xfs_blkdev_issue_flush(btp);
1447 kthread_stop(btp->bt_task);
1452 xfs_setsize_buftarg_flags(
1454 unsigned int blocksize,
1455 unsigned int sectorsize,
1458 btp->bt_bsize = blocksize;
1459 btp->bt_sshift = ffs(sectorsize) - 1;
1460 btp->bt_smask = sectorsize - 1;
1462 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1463 xfs_warn(btp->bt_mount,
1464 "Cannot set_blocksize to %u on device %s\n",
1465 sectorsize, xfs_buf_target_name(btp));
1473 * When allocating the initial buffer target we have not yet
1474 * read in the superblock, so don't know what sized sectors
1475 * are being used is at this early stage. Play safe.
1478 xfs_setsize_buftarg_early(
1480 struct block_device *bdev)
1482 return xfs_setsize_buftarg_flags(btp,
1483 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1487 xfs_setsize_buftarg(
1489 unsigned int blocksize,
1490 unsigned int sectorsize)
1492 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1496 xfs_alloc_delwri_queue(
1500 INIT_LIST_HEAD(&btp->bt_delwri_queue);
1501 spin_lock_init(&btp->bt_delwri_lock);
1503 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1504 if (IS_ERR(btp->bt_task))
1505 return PTR_ERR(btp->bt_task);
1511 struct xfs_mount *mp,
1512 struct block_device *bdev,
1518 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1521 btp->bt_dev = bdev->bd_dev;
1522 btp->bt_bdev = bdev;
1523 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1527 INIT_LIST_HEAD(&btp->bt_lru);
1528 spin_lock_init(&btp->bt_lru_lock);
1529 if (xfs_setsize_buftarg_early(btp, bdev))
1531 if (xfs_alloc_delwri_queue(btp, fsname))
1533 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1534 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1535 register_shrinker(&btp->bt_shrinker);
1545 * Delayed write buffer handling
1548 xfs_buf_delwri_queue(
1551 struct xfs_buftarg *btp = bp->b_target;
1553 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1555 ASSERT(!(bp->b_flags & XBF_READ));
1557 spin_lock(&btp->bt_delwri_lock);
1558 if (!list_empty(&bp->b_list)) {
1559 /* if already in the queue, move it to the tail */
1560 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1561 list_move_tail(&bp->b_list, &btp->bt_delwri_queue);
1563 /* start xfsbufd as it is about to have something to do */
1564 if (list_empty(&btp->bt_delwri_queue))
1565 wake_up_process(bp->b_target->bt_task);
1567 atomic_inc(&bp->b_hold);
1568 bp->b_flags |= XBF_DELWRI | _XBF_DELWRI_Q | XBF_ASYNC;
1569 list_add_tail(&bp->b_list, &btp->bt_delwri_queue);
1571 bp->b_queuetime = jiffies;
1572 spin_unlock(&btp->bt_delwri_lock);
1576 xfs_buf_delwri_dequeue(
1581 spin_lock(&bp->b_target->bt_delwri_lock);
1582 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1583 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1584 list_del_init(&bp->b_list);
1587 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1588 spin_unlock(&bp->b_target->bt_delwri_lock);
1593 trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1597 * If a delwri buffer needs to be pushed before it has aged out, then promote
1598 * it to the head of the delwri queue so that it will be flushed on the next
1599 * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1600 * than the age currently needed to flush the buffer. Hence the next time the
1601 * xfsbufd sees it is guaranteed to be considered old enough to flush.
1604 xfs_buf_delwri_promote(
1607 struct xfs_buftarg *btp = bp->b_target;
1608 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1610 ASSERT(bp->b_flags & XBF_DELWRI);
1611 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1614 * Check the buffer age before locking the delayed write queue as we
1615 * don't need to promote buffers that are already past the flush age.
1617 if (bp->b_queuetime < jiffies - age)
1619 bp->b_queuetime = jiffies - age;
1620 spin_lock(&btp->bt_delwri_lock);
1621 list_move(&bp->b_list, &btp->bt_delwri_queue);
1622 spin_unlock(&btp->bt_delwri_lock);
1626 xfs_buf_runall_queues(
1627 struct workqueue_struct *queue)
1629 flush_workqueue(queue);
1633 * Move as many buffers as specified to the supplied list
1634 * idicating if we skipped any buffers to prevent deadlocks.
1637 xfs_buf_delwri_split(
1638 xfs_buftarg_t *target,
1639 struct list_head *list,
1646 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1647 INIT_LIST_HEAD(list);
1648 spin_lock(&target->bt_delwri_lock);
1649 list_for_each_entry_safe(bp, n, &target->bt_delwri_queue, b_list) {
1650 ASSERT(bp->b_flags & XBF_DELWRI);
1652 if (!xfs_buf_ispinned(bp) && xfs_buf_trylock(bp)) {
1654 time_before(jiffies, bp->b_queuetime + age)) {
1659 bp->b_flags &= ~(XBF_DELWRI | _XBF_DELWRI_Q);
1660 bp->b_flags |= XBF_WRITE;
1661 list_move_tail(&bp->b_list, list);
1662 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1667 spin_unlock(&target->bt_delwri_lock);
1672 * Compare function is more complex than it needs to be because
1673 * the return value is only 32 bits and we are doing comparisons
1679 struct list_head *a,
1680 struct list_head *b)
1682 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1683 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1686 diff = ap->b_bn - bp->b_bn;
1698 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1700 current->flags |= PF_MEMALLOC;
1705 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1706 long tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1707 struct list_head tmp;
1708 struct blk_plug plug;
1710 if (unlikely(freezing(current))) {
1711 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1714 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1717 /* sleep for a long time if there is nothing to do. */
1718 if (list_empty(&target->bt_delwri_queue))
1719 tout = MAX_SCHEDULE_TIMEOUT;
1720 schedule_timeout_interruptible(tout);
1722 xfs_buf_delwri_split(target, &tmp, age);
1723 list_sort(NULL, &tmp, xfs_buf_cmp);
1725 blk_start_plug(&plug);
1726 while (!list_empty(&tmp)) {
1728 bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1729 list_del_init(&bp->b_list);
1732 blk_finish_plug(&plug);
1733 } while (!kthread_should_stop());
1739 * Go through all incore buffers, and release buffers if they belong to
1740 * the given device. This is used in filesystem error handling to
1741 * preserve the consistency of its metadata.
1745 xfs_buftarg_t *target,
1750 LIST_HEAD(tmp_list);
1751 LIST_HEAD(wait_list);
1752 struct blk_plug plug;
1754 xfs_buf_runall_queues(xfsconvertd_workqueue);
1755 xfs_buf_runall_queues(xfsdatad_workqueue);
1756 xfs_buf_runall_queues(xfslogd_workqueue);
1758 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1759 pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1762 * Dropped the delayed write list lock, now walk the temporary list.
1763 * All I/O is issued async and then if we need to wait for completion
1764 * we do that after issuing all the IO.
1766 list_sort(NULL, &tmp_list, xfs_buf_cmp);
1768 blk_start_plug(&plug);
1769 while (!list_empty(&tmp_list)) {
1770 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1771 ASSERT(target == bp->b_target);
1772 list_del_init(&bp->b_list);
1774 bp->b_flags &= ~XBF_ASYNC;
1775 list_add(&bp->b_list, &wait_list);
1779 blk_finish_plug(&plug);
1782 /* Wait for IO to complete. */
1783 while (!list_empty(&wait_list)) {
1784 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1786 list_del_init(&bp->b_list);
1798 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1799 KM_ZONE_HWALIGN, NULL);
1803 xfslogd_workqueue = alloc_workqueue("xfslogd",
1804 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1805 if (!xfslogd_workqueue)
1806 goto out_free_buf_zone;
1808 xfsdatad_workqueue = alloc_workqueue("xfsdatad", WQ_MEM_RECLAIM, 1);
1809 if (!xfsdatad_workqueue)
1810 goto out_destroy_xfslogd_workqueue;
1812 xfsconvertd_workqueue = alloc_workqueue("xfsconvertd",
1814 if (!xfsconvertd_workqueue)
1815 goto out_destroy_xfsdatad_workqueue;
1819 out_destroy_xfsdatad_workqueue:
1820 destroy_workqueue(xfsdatad_workqueue);
1821 out_destroy_xfslogd_workqueue:
1822 destroy_workqueue(xfslogd_workqueue);
1824 kmem_zone_destroy(xfs_buf_zone);
1830 xfs_buf_terminate(void)
1832 destroy_workqueue(xfsconvertd_workqueue);
1833 destroy_workqueue(xfsdatad_workqueue);
1834 destroy_workqueue(xfslogd_workqueue);
1835 kmem_zone_destroy(xfs_buf_zone);