2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
19 #include <linux/stddef.h>
20 #include <linux/errno.h>
21 #include <linux/gfp.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/vmalloc.h>
25 #include <linux/bio.h>
26 #include <linux/sysctl.h>
27 #include <linux/proc_fs.h>
28 #include <linux/workqueue.h>
29 #include <linux/percpu.h>
30 #include <linux/blkdev.h>
31 #include <linux/hash.h>
32 #include <linux/kthread.h>
33 #include <linux/migrate.h>
34 #include <linux/backing-dev.h>
35 #include <linux/freezer.h>
40 #include "xfs_mount.h"
41 #include "xfs_trace.h"
43 static kmem_zone_t
*xfs_buf_zone
;
45 static struct workqueue_struct
*xfslogd_workqueue
;
47 #ifdef XFS_BUF_LOCK_TRACKING
48 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
49 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
50 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
52 # define XB_SET_OWNER(bp) do { } while (0)
53 # define XB_CLEAR_OWNER(bp) do { } while (0)
54 # define XB_GET_OWNER(bp) do { } while (0)
57 #define xb_to_gfp(flags) \
58 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN)
66 * Return true if the buffer is vmapped.
68 * b_addr is null if the buffer is not mapped, but the code is clever
69 * enough to know it doesn't have to map a single page, so the check has
70 * to be both for b_addr and bp->b_page_count > 1.
72 return bp
->b_addr
&& bp
->b_page_count
> 1;
79 return (bp
->b_page_count
* PAGE_SIZE
) - bp
->b_offset
;
83 * xfs_buf_lru_add - add a buffer to the LRU.
85 * The LRU takes a new reference to the buffer so that it will only be freed
86 * once the shrinker takes the buffer off the LRU.
92 struct xfs_buftarg
*btp
= bp
->b_target
;
94 spin_lock(&btp
->bt_lru_lock
);
95 if (list_empty(&bp
->b_lru
)) {
96 atomic_inc(&bp
->b_hold
);
97 list_add_tail(&bp
->b_lru
, &btp
->bt_lru
);
100 spin_unlock(&btp
->bt_lru_lock
);
104 * xfs_buf_lru_del - remove a buffer from the LRU
106 * The unlocked check is safe here because it only occurs when there are not
107 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
108 * to optimise the shrinker removing the buffer from the LRU and calling
109 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
116 struct xfs_buftarg
*btp
= bp
->b_target
;
118 if (list_empty(&bp
->b_lru
))
121 spin_lock(&btp
->bt_lru_lock
);
122 if (!list_empty(&bp
->b_lru
)) {
123 list_del_init(&bp
->b_lru
);
126 spin_unlock(&btp
->bt_lru_lock
);
130 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
131 * b_lru_ref count so that the buffer is freed immediately when the buffer
132 * reference count falls to zero. If the buffer is already on the LRU, we need
133 * to remove the reference that LRU holds on the buffer.
135 * This prevents build-up of stale buffers on the LRU.
141 ASSERT(xfs_buf_islocked(bp
));
143 bp
->b_flags
|= XBF_STALE
;
146 * Clear the delwri status so that a delwri queue walker will not
147 * flush this buffer to disk now that it is stale. The delwri queue has
148 * a reference to the buffer, so this is safe to do.
150 bp
->b_flags
&= ~_XBF_DELWRI_Q
;
152 atomic_set(&(bp
)->b_lru_ref
, 0);
153 if (!list_empty(&bp
->b_lru
)) {
154 struct xfs_buftarg
*btp
= bp
->b_target
;
156 spin_lock(&btp
->bt_lru_lock
);
157 if (!list_empty(&bp
->b_lru
)) {
158 list_del_init(&bp
->b_lru
);
160 atomic_dec(&bp
->b_hold
);
162 spin_unlock(&btp
->bt_lru_lock
);
164 ASSERT(atomic_read(&bp
->b_hold
) >= 1);
169 struct xfs_buftarg
*target
,
172 xfs_buf_flags_t flags
)
176 bp
= kmem_zone_zalloc(xfs_buf_zone
, KM_NOFS
);
181 * We don't want certain flags to appear in b_flags unless they are
182 * specifically set by later operations on the buffer.
184 flags
&= ~(XBF_UNMAPPED
| XBF_TRYLOCK
| XBF_ASYNC
| XBF_READ_AHEAD
);
186 atomic_set(&bp
->b_hold
, 1);
187 atomic_set(&bp
->b_lru_ref
, 1);
188 init_completion(&bp
->b_iowait
);
189 INIT_LIST_HEAD(&bp
->b_lru
);
190 INIT_LIST_HEAD(&bp
->b_list
);
191 RB_CLEAR_NODE(&bp
->b_rbnode
);
192 sema_init(&bp
->b_sema
, 0); /* held, no waiters */
194 bp
->b_target
= target
;
197 * Set length and io_length to the same value initially.
198 * I/O routines should use io_length, which will be the same in
199 * most cases but may be reset (e.g. XFS recovery).
201 bp
->b_length
= numblks
;
202 bp
->b_io_length
= numblks
;
206 * We do not set the block number here in the buffer because we have not
207 * finished initialising the buffer. We insert the buffer into the cache
208 * in this state, so this ensures that we are unable to do IO on a
209 * buffer that hasn't been fully initialised.
211 bp
->b_bn
= XFS_BUF_DADDR_NULL
;
212 atomic_set(&bp
->b_pin_count
, 0);
213 init_waitqueue_head(&bp
->b_waiters
);
215 XFS_STATS_INC(xb_create
);
216 trace_xfs_buf_init(bp
, _RET_IP_
);
222 * Allocate a page array capable of holding a specified number
223 * of pages, and point the page buf at it.
229 xfs_buf_flags_t flags
)
231 /* Make sure that we have a page list */
232 if (bp
->b_pages
== NULL
) {
233 bp
->b_page_count
= page_count
;
234 if (page_count
<= XB_PAGES
) {
235 bp
->b_pages
= bp
->b_page_array
;
237 bp
->b_pages
= kmem_alloc(sizeof(struct page
*) *
238 page_count
, KM_NOFS
);
239 if (bp
->b_pages
== NULL
)
242 memset(bp
->b_pages
, 0, sizeof(struct page
*) * page_count
);
248 * Frees b_pages if it was allocated.
254 if (bp
->b_pages
!= bp
->b_page_array
) {
255 kmem_free(bp
->b_pages
);
261 * Releases the specified buffer.
263 * The modification state of any associated pages is left unchanged.
264 * The buffer most not be on any hash - use xfs_buf_rele instead for
265 * hashed and refcounted buffers
271 trace_xfs_buf_free(bp
, _RET_IP_
);
273 ASSERT(list_empty(&bp
->b_lru
));
275 if (bp
->b_flags
& _XBF_PAGES
) {
278 if (xfs_buf_is_vmapped(bp
))
279 vm_unmap_ram(bp
->b_addr
- bp
->b_offset
,
282 for (i
= 0; i
< bp
->b_page_count
; i
++) {
283 struct page
*page
= bp
->b_pages
[i
];
287 } else if (bp
->b_flags
& _XBF_KMEM
)
288 kmem_free(bp
->b_addr
);
289 _xfs_buf_free_pages(bp
);
290 kmem_zone_free(xfs_buf_zone
, bp
);
294 * Allocates all the pages for buffer in question and builds it's page list.
297 xfs_buf_allocate_memory(
302 size_t nbytes
, offset
;
303 gfp_t gfp_mask
= xb_to_gfp(flags
);
304 unsigned short page_count
, i
;
305 xfs_off_t start
, end
;
309 * for buffers that are contained within a single page, just allocate
310 * the memory from the heap - there's no need for the complexity of
311 * page arrays to keep allocation down to order 0.
313 size
= BBTOB(bp
->b_length
);
314 if (size
< PAGE_SIZE
) {
315 bp
->b_addr
= kmem_alloc(size
, KM_NOFS
);
317 /* low memory - use alloc_page loop instead */
321 if (((unsigned long)(bp
->b_addr
+ size
- 1) & PAGE_MASK
) !=
322 ((unsigned long)bp
->b_addr
& PAGE_MASK
)) {
323 /* b_addr spans two pages - use alloc_page instead */
324 kmem_free(bp
->b_addr
);
328 bp
->b_offset
= offset_in_page(bp
->b_addr
);
329 bp
->b_pages
= bp
->b_page_array
;
330 bp
->b_pages
[0] = virt_to_page(bp
->b_addr
);
331 bp
->b_page_count
= 1;
332 bp
->b_flags
|= _XBF_KMEM
;
337 start
= BBTOB(bp
->b_bn
) >> PAGE_SHIFT
;
338 end
= (BBTOB(bp
->b_bn
+ bp
->b_length
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
339 page_count
= end
- start
;
340 error
= _xfs_buf_get_pages(bp
, page_count
, flags
);
344 offset
= bp
->b_offset
;
345 bp
->b_flags
|= _XBF_PAGES
;
347 for (i
= 0; i
< bp
->b_page_count
; i
++) {
351 page
= alloc_page(gfp_mask
);
352 if (unlikely(page
== NULL
)) {
353 if (flags
& XBF_READ_AHEAD
) {
354 bp
->b_page_count
= i
;
360 * This could deadlock.
362 * But until all the XFS lowlevel code is revamped to
363 * handle buffer allocation failures we can't do much.
365 if (!(++retries
% 100))
367 "possible memory allocation deadlock in %s (mode:0x%x)",
370 XFS_STATS_INC(xb_page_retries
);
371 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
375 XFS_STATS_INC(xb_page_found
);
377 nbytes
= min_t(size_t, size
, PAGE_SIZE
- offset
);
379 bp
->b_pages
[i
] = page
;
385 for (i
= 0; i
< bp
->b_page_count
; i
++)
386 __free_page(bp
->b_pages
[i
]);
391 * Map buffer into kernel address-space if necessary.
398 ASSERT(bp
->b_flags
& _XBF_PAGES
);
399 if (bp
->b_page_count
== 1) {
400 /* A single page buffer is always mappable */
401 bp
->b_addr
= page_address(bp
->b_pages
[0]) + bp
->b_offset
;
402 } else if (flags
& XBF_UNMAPPED
) {
408 bp
->b_addr
= vm_map_ram(bp
->b_pages
, bp
->b_page_count
,
413 } while (retried
++ <= 1);
417 bp
->b_addr
+= bp
->b_offset
;
424 * Finding and Reading Buffers
428 * Look up, and creates if absent, a lockable buffer for
429 * a given range of an inode. The buffer is returned
430 * locked. No I/O is implied by this call.
434 struct xfs_buftarg
*btp
,
437 xfs_buf_flags_t flags
,
441 struct xfs_perag
*pag
;
442 struct rb_node
**rbp
;
443 struct rb_node
*parent
;
446 numbytes
= BBTOB(numblks
);
448 /* Check for IOs smaller than the sector size / not sector aligned */
449 ASSERT(!(numbytes
< (1 << btp
->bt_sshift
)));
450 ASSERT(!(BBTOB(blkno
) & (xfs_off_t
)btp
->bt_smask
));
453 pag
= xfs_perag_get(btp
->bt_mount
,
454 xfs_daddr_to_agno(btp
->bt_mount
, blkno
));
457 spin_lock(&pag
->pag_buf_lock
);
458 rbp
= &pag
->pag_buf_tree
.rb_node
;
463 bp
= rb_entry(parent
, struct xfs_buf
, b_rbnode
);
465 if (blkno
< bp
->b_bn
)
466 rbp
= &(*rbp
)->rb_left
;
467 else if (blkno
> bp
->b_bn
)
468 rbp
= &(*rbp
)->rb_right
;
471 * found a block number match. If the range doesn't
472 * match, the only way this is allowed is if the buffer
473 * in the cache is stale and the transaction that made
474 * it stale has not yet committed. i.e. we are
475 * reallocating a busy extent. Skip this buffer and
476 * continue searching to the right for an exact match.
478 if (bp
->b_length
!= numblks
) {
479 ASSERT(bp
->b_flags
& XBF_STALE
);
480 rbp
= &(*rbp
)->rb_right
;
483 atomic_inc(&bp
->b_hold
);
490 rb_link_node(&new_bp
->b_rbnode
, parent
, rbp
);
491 rb_insert_color(&new_bp
->b_rbnode
, &pag
->pag_buf_tree
);
492 /* the buffer keeps the perag reference until it is freed */
494 spin_unlock(&pag
->pag_buf_lock
);
496 XFS_STATS_INC(xb_miss_locked
);
497 spin_unlock(&pag
->pag_buf_lock
);
503 spin_unlock(&pag
->pag_buf_lock
);
506 if (!xfs_buf_trylock(bp
)) {
507 if (flags
& XBF_TRYLOCK
) {
509 XFS_STATS_INC(xb_busy_locked
);
513 XFS_STATS_INC(xb_get_locked_waited
);
517 * if the buffer is stale, clear all the external state associated with
518 * it. We need to keep flags such as how we allocated the buffer memory
521 if (bp
->b_flags
& XBF_STALE
) {
522 ASSERT((bp
->b_flags
& _XBF_DELWRI_Q
) == 0);
523 bp
->b_flags
&= _XBF_KMEM
| _XBF_PAGES
;
526 trace_xfs_buf_find(bp
, flags
, _RET_IP_
);
527 XFS_STATS_INC(xb_get_locked
);
532 * Assembles a buffer covering the specified range. The code is optimised for
533 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
534 * more hits than misses.
538 xfs_buftarg_t
*target
,
541 xfs_buf_flags_t flags
)
544 struct xfs_buf
*new_bp
;
547 bp
= _xfs_buf_find(target
, blkno
, numblks
, flags
, NULL
);
551 new_bp
= xfs_buf_alloc(target
, blkno
, numblks
, flags
);
552 if (unlikely(!new_bp
))
555 error
= xfs_buf_allocate_memory(new_bp
, flags
);
557 kmem_zone_free(xfs_buf_zone
, new_bp
);
561 bp
= _xfs_buf_find(target
, blkno
, numblks
, flags
, new_bp
);
563 xfs_buf_free(new_bp
);
568 xfs_buf_free(new_bp
);
571 * Now we have a workable buffer, fill in the block number so
572 * that we can do IO on it.
575 bp
->b_io_length
= bp
->b_length
;
579 error
= _xfs_buf_map_pages(bp
, flags
);
580 if (unlikely(error
)) {
581 xfs_warn(target
->bt_mount
,
582 "%s: failed to map pages\n", __func__
);
588 XFS_STATS_INC(xb_get
);
589 trace_xfs_buf_get(bp
, flags
, _RET_IP_
);
596 xfs_buf_flags_t flags
)
598 ASSERT(!(flags
& XBF_WRITE
));
599 ASSERT(bp
->b_bn
!= XFS_BUF_DADDR_NULL
);
601 bp
->b_flags
&= ~(XBF_WRITE
| XBF_ASYNC
| XBF_READ_AHEAD
);
602 bp
->b_flags
|= flags
& (XBF_READ
| XBF_ASYNC
| XBF_READ_AHEAD
);
604 xfs_buf_iorequest(bp
);
605 if (flags
& XBF_ASYNC
)
607 return xfs_buf_iowait(bp
);
612 xfs_buftarg_t
*target
,
615 xfs_buf_flags_t flags
)
621 bp
= xfs_buf_get(target
, blkno
, numblks
, flags
);
623 trace_xfs_buf_read(bp
, flags
, _RET_IP_
);
625 if (!XFS_BUF_ISDONE(bp
)) {
626 XFS_STATS_INC(xb_get_read
);
627 _xfs_buf_read(bp
, flags
);
628 } else if (flags
& XBF_ASYNC
) {
630 * Read ahead call which is already satisfied,
636 /* We do not want read in the flags */
637 bp
->b_flags
&= ~XBF_READ
;
645 * If we are not low on memory then do the readahead in a deadlock
650 xfs_buftarg_t
*target
,
654 if (bdi_read_congested(target
->bt_bdi
))
657 xfs_buf_read(target
, blkno
, numblks
,
658 XBF_TRYLOCK
|XBF_ASYNC
|XBF_READ_AHEAD
);
662 * Read an uncached buffer from disk. Allocates and returns a locked
663 * buffer containing the disk contents or nothing.
666 xfs_buf_read_uncached(
667 struct xfs_buftarg
*target
,
675 bp
= xfs_buf_get_uncached(target
, numblks
, flags
);
679 /* set up the buffer for a read IO */
680 XFS_BUF_SET_ADDR(bp
, daddr
);
683 xfsbdstrat(target
->bt_mount
, bp
);
684 error
= xfs_buf_iowait(bp
);
693 * Return a buffer allocated as an empty buffer and associated to external
694 * memory via xfs_buf_associate_memory() back to it's empty state.
702 _xfs_buf_free_pages(bp
);
705 bp
->b_page_count
= 0;
707 bp
->b_length
= numblks
;
708 bp
->b_io_length
= numblks
;
709 bp
->b_bn
= XFS_BUF_DADDR_NULL
;
712 static inline struct page
*
716 if ((!is_vmalloc_addr(addr
))) {
717 return virt_to_page(addr
);
719 return vmalloc_to_page(addr
);
724 xfs_buf_associate_memory(
731 unsigned long pageaddr
;
732 unsigned long offset
;
736 pageaddr
= (unsigned long)mem
& PAGE_MASK
;
737 offset
= (unsigned long)mem
- pageaddr
;
738 buflen
= PAGE_ALIGN(len
+ offset
);
739 page_count
= buflen
>> PAGE_SHIFT
;
741 /* Free any previous set of page pointers */
743 _xfs_buf_free_pages(bp
);
748 rval
= _xfs_buf_get_pages(bp
, page_count
, 0);
752 bp
->b_offset
= offset
;
754 for (i
= 0; i
< bp
->b_page_count
; i
++) {
755 bp
->b_pages
[i
] = mem_to_page((void *)pageaddr
);
756 pageaddr
+= PAGE_SIZE
;
759 bp
->b_io_length
= BTOBB(len
);
760 bp
->b_length
= BTOBB(buflen
);
766 xfs_buf_get_uncached(
767 struct xfs_buftarg
*target
,
771 unsigned long page_count
;
775 bp
= xfs_buf_alloc(target
, 0, numblks
, 0);
776 if (unlikely(bp
== NULL
))
779 page_count
= PAGE_ALIGN(numblks
<< BBSHIFT
) >> PAGE_SHIFT
;
780 error
= _xfs_buf_get_pages(bp
, page_count
, 0);
784 for (i
= 0; i
< page_count
; i
++) {
785 bp
->b_pages
[i
] = alloc_page(xb_to_gfp(flags
));
789 bp
->b_flags
|= _XBF_PAGES
;
791 error
= _xfs_buf_map_pages(bp
, 0);
792 if (unlikely(error
)) {
793 xfs_warn(target
->bt_mount
,
794 "%s: failed to map pages\n", __func__
);
798 trace_xfs_buf_get_uncached(bp
, _RET_IP_
);
803 __free_page(bp
->b_pages
[i
]);
804 _xfs_buf_free_pages(bp
);
806 kmem_zone_free(xfs_buf_zone
, bp
);
812 * Increment reference count on buffer, to hold the buffer concurrently
813 * with another thread which may release (free) the buffer asynchronously.
814 * Must hold the buffer already to call this function.
820 trace_xfs_buf_hold(bp
, _RET_IP_
);
821 atomic_inc(&bp
->b_hold
);
825 * Releases a hold on the specified buffer. If the
826 * the hold count is 1, calls xfs_buf_free.
832 struct xfs_perag
*pag
= bp
->b_pag
;
834 trace_xfs_buf_rele(bp
, _RET_IP_
);
837 ASSERT(list_empty(&bp
->b_lru
));
838 ASSERT(RB_EMPTY_NODE(&bp
->b_rbnode
));
839 if (atomic_dec_and_test(&bp
->b_hold
))
844 ASSERT(!RB_EMPTY_NODE(&bp
->b_rbnode
));
846 ASSERT(atomic_read(&bp
->b_hold
) > 0);
847 if (atomic_dec_and_lock(&bp
->b_hold
, &pag
->pag_buf_lock
)) {
848 if (!(bp
->b_flags
& XBF_STALE
) &&
849 atomic_read(&bp
->b_lru_ref
)) {
851 spin_unlock(&pag
->pag_buf_lock
);
854 ASSERT(!(bp
->b_flags
& _XBF_DELWRI_Q
));
855 rb_erase(&bp
->b_rbnode
, &pag
->pag_buf_tree
);
856 spin_unlock(&pag
->pag_buf_lock
);
865 * Lock a buffer object, if it is not already locked.
867 * If we come across a stale, pinned, locked buffer, we know that we are
868 * being asked to lock a buffer that has been reallocated. Because it is
869 * pinned, we know that the log has not been pushed to disk and hence it
870 * will still be locked. Rather than continuing to have trylock attempts
871 * fail until someone else pushes the log, push it ourselves before
872 * returning. This means that the xfsaild will not get stuck trying
873 * to push on stale inode buffers.
881 locked
= down_trylock(&bp
->b_sema
) == 0;
884 else if (atomic_read(&bp
->b_pin_count
) && (bp
->b_flags
& XBF_STALE
))
885 xfs_log_force(bp
->b_target
->bt_mount
, 0);
887 trace_xfs_buf_trylock(bp
, _RET_IP_
);
892 * Lock a buffer object.
894 * If we come across a stale, pinned, locked buffer, we know that we
895 * are being asked to lock a buffer that has been reallocated. Because
896 * it is pinned, we know that the log has not been pushed to disk and
897 * hence it will still be locked. Rather than sleeping until someone
898 * else pushes the log, push it ourselves before trying to get the lock.
904 trace_xfs_buf_lock(bp
, _RET_IP_
);
906 if (atomic_read(&bp
->b_pin_count
) && (bp
->b_flags
& XBF_STALE
))
907 xfs_log_force(bp
->b_target
->bt_mount
, 0);
911 trace_xfs_buf_lock_done(bp
, _RET_IP_
);
921 trace_xfs_buf_unlock(bp
, _RET_IP_
);
928 DECLARE_WAITQUEUE (wait
, current
);
930 if (atomic_read(&bp
->b_pin_count
) == 0)
933 add_wait_queue(&bp
->b_waiters
, &wait
);
935 set_current_state(TASK_UNINTERRUPTIBLE
);
936 if (atomic_read(&bp
->b_pin_count
) == 0)
940 remove_wait_queue(&bp
->b_waiters
, &wait
);
941 set_current_state(TASK_RUNNING
);
945 * Buffer Utility Routines
950 struct work_struct
*work
)
953 container_of(work
, xfs_buf_t
, b_iodone_work
);
956 (*(bp
->b_iodone
))(bp
);
957 else if (bp
->b_flags
& XBF_ASYNC
)
966 trace_xfs_buf_iodone(bp
, _RET_IP_
);
968 bp
->b_flags
&= ~(XBF_READ
| XBF_WRITE
| XBF_READ_AHEAD
);
969 if (bp
->b_error
== 0)
970 bp
->b_flags
|= XBF_DONE
;
972 if ((bp
->b_iodone
) || (bp
->b_flags
& XBF_ASYNC
)) {
974 INIT_WORK(&bp
->b_iodone_work
, xfs_buf_iodone_work
);
975 queue_work(xfslogd_workqueue
, &bp
->b_iodone_work
);
977 xfs_buf_iodone_work(&bp
->b_iodone_work
);
980 complete(&bp
->b_iowait
);
989 ASSERT(error
>= 0 && error
<= 0xffff);
990 bp
->b_error
= (unsigned short)error
;
991 trace_xfs_buf_ioerror(bp
, error
, _RET_IP_
);
995 xfs_buf_ioerror_alert(
999 xfs_alert(bp
->b_target
->bt_mount
,
1000 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1001 (__uint64_t
)XFS_BUF_ADDR(bp
), func
, bp
->b_error
, bp
->b_length
);
1010 ASSERT(xfs_buf_islocked(bp
));
1012 bp
->b_flags
|= XBF_WRITE
;
1013 bp
->b_flags
&= ~(XBF_ASYNC
| XBF_READ
| _XBF_DELWRI_Q
);
1017 error
= xfs_buf_iowait(bp
);
1019 xfs_force_shutdown(bp
->b_target
->bt_mount
,
1020 SHUTDOWN_META_IO_ERROR
);
1026 * Called when we want to stop a buffer from getting written or read.
1027 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1028 * so that the proper iodone callbacks get called.
1034 #ifdef XFSERRORDEBUG
1035 ASSERT(XFS_BUF_ISREAD(bp
) || bp
->b_iodone
);
1039 * No need to wait until the buffer is unpinned, we aren't flushing it.
1041 xfs_buf_ioerror(bp
, EIO
);
1044 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1050 xfs_buf_ioend(bp
, 0);
1056 * Same as xfs_bioerror, except that we are releasing the buffer
1057 * here ourselves, and avoiding the xfs_buf_ioend call.
1058 * This is meant for userdata errors; metadata bufs come with
1059 * iodone functions attached, so that we can track down errors.
1065 int64_t fl
= bp
->b_flags
;
1067 * No need to wait until the buffer is unpinned.
1068 * We aren't flushing it.
1070 * chunkhold expects B_DONE to be set, whether
1071 * we actually finish the I/O or not. We don't want to
1072 * change that interface.
1077 bp
->b_iodone
= NULL
;
1078 if (!(fl
& XBF_ASYNC
)) {
1080 * Mark b_error and B_ERROR _both_.
1081 * Lot's of chunkcache code assumes that.
1082 * There's no reason to mark error for
1085 xfs_buf_ioerror(bp
, EIO
);
1086 complete(&bp
->b_iowait
);
1096 * All xfs metadata buffers except log state machine buffers
1097 * get this attached as their b_bdstrat callback function.
1098 * This is so that we can catch a buffer
1099 * after prematurely unpinning it to forcibly shutdown the filesystem.
1105 if (XFS_FORCED_SHUTDOWN(bp
->b_target
->bt_mount
)) {
1106 trace_xfs_bdstrat_shut(bp
, _RET_IP_
);
1108 * Metadata write that didn't get logged but
1109 * written delayed anyway. These aren't associated
1110 * with a transaction, and can be ignored.
1112 if (!bp
->b_iodone
&& !XFS_BUF_ISREAD(bp
))
1113 return xfs_bioerror_relse(bp
);
1115 return xfs_bioerror(bp
);
1118 xfs_buf_iorequest(bp
);
1123 * Wrapper around bdstrat so that we can stop data from going to disk in case
1124 * we are shutting down the filesystem. Typically user data goes thru this
1125 * path; one of the exceptions is the superblock.
1129 struct xfs_mount
*mp
,
1132 if (XFS_FORCED_SHUTDOWN(mp
)) {
1133 trace_xfs_bdstrat_shut(bp
, _RET_IP_
);
1134 xfs_bioerror_relse(bp
);
1138 xfs_buf_iorequest(bp
);
1146 if (atomic_dec_and_test(&bp
->b_io_remaining
) == 1)
1147 xfs_buf_ioend(bp
, schedule
);
1155 xfs_buf_t
*bp
= (xfs_buf_t
*)bio
->bi_private
;
1157 xfs_buf_ioerror(bp
, -error
);
1159 if (!error
&& xfs_buf_is_vmapped(bp
) && (bp
->b_flags
& XBF_READ
))
1160 invalidate_kernel_vmap_range(bp
->b_addr
, xfs_buf_vmap_len(bp
));
1162 _xfs_buf_ioend(bp
, 1);
1170 int rw
, map_i
, total_nr_pages
, nr_pages
;
1172 int offset
= bp
->b_offset
;
1173 int size
= BBTOB(bp
->b_io_length
);
1174 sector_t sector
= bp
->b_bn
;
1176 total_nr_pages
= bp
->b_page_count
;
1179 if (bp
->b_flags
& XBF_WRITE
) {
1180 if (bp
->b_flags
& XBF_SYNCIO
)
1184 if (bp
->b_flags
& XBF_FUA
)
1186 if (bp
->b_flags
& XBF_FLUSH
)
1188 } else if (bp
->b_flags
& XBF_READ_AHEAD
) {
1194 /* we only use the buffer cache for meta-data */
1198 atomic_inc(&bp
->b_io_remaining
);
1199 nr_pages
= BIO_MAX_SECTORS
>> (PAGE_SHIFT
- BBSHIFT
);
1200 if (nr_pages
> total_nr_pages
)
1201 nr_pages
= total_nr_pages
;
1203 bio
= bio_alloc(GFP_NOIO
, nr_pages
);
1204 bio
->bi_bdev
= bp
->b_target
->bt_bdev
;
1205 bio
->bi_sector
= sector
;
1206 bio
->bi_end_io
= xfs_buf_bio_end_io
;
1207 bio
->bi_private
= bp
;
1210 for (; size
&& nr_pages
; nr_pages
--, map_i
++) {
1211 int rbytes
, nbytes
= PAGE_SIZE
- offset
;
1216 rbytes
= bio_add_page(bio
, bp
->b_pages
[map_i
], nbytes
, offset
);
1217 if (rbytes
< nbytes
)
1221 sector
+= BTOBB(nbytes
);
1226 if (likely(bio
->bi_size
)) {
1227 if (xfs_buf_is_vmapped(bp
)) {
1228 flush_kernel_vmap_range(bp
->b_addr
,
1229 xfs_buf_vmap_len(bp
));
1231 submit_bio(rw
, bio
);
1235 xfs_buf_ioerror(bp
, EIO
);
1244 trace_xfs_buf_iorequest(bp
, _RET_IP_
);
1246 ASSERT(!(bp
->b_flags
& _XBF_DELWRI_Q
));
1248 if (bp
->b_flags
& XBF_WRITE
)
1249 xfs_buf_wait_unpin(bp
);
1252 /* Set the count to 1 initially, this will stop an I/O
1253 * completion callout which happens before we have started
1254 * all the I/O from calling xfs_buf_ioend too early.
1256 atomic_set(&bp
->b_io_remaining
, 1);
1257 _xfs_buf_ioapply(bp
);
1258 _xfs_buf_ioend(bp
, 0);
1264 * Waits for I/O to complete on the buffer supplied. It returns immediately if
1265 * no I/O is pending or there is already a pending error on the buffer. It
1266 * returns the I/O error code, if any, or 0 if there was no error.
1272 trace_xfs_buf_iowait(bp
, _RET_IP_
);
1275 wait_for_completion(&bp
->b_iowait
);
1277 trace_xfs_buf_iowait_done(bp
, _RET_IP_
);
1289 return bp
->b_addr
+ offset
;
1291 offset
+= bp
->b_offset
;
1292 page
= bp
->b_pages
[offset
>> PAGE_SHIFT
];
1293 return (xfs_caddr_t
)page_address(page
) + (offset
& (PAGE_SIZE
-1));
1297 * Move data into or out of a buffer.
1301 xfs_buf_t
*bp
, /* buffer to process */
1302 size_t boff
, /* starting buffer offset */
1303 size_t bsize
, /* length to copy */
1304 void *data
, /* data address */
1305 xfs_buf_rw_t mode
) /* read/write/zero flag */
1309 bend
= boff
+ bsize
;
1310 while (boff
< bend
) {
1312 int page_index
, page_offset
, csize
;
1314 page_index
= (boff
+ bp
->b_offset
) >> PAGE_SHIFT
;
1315 page_offset
= (boff
+ bp
->b_offset
) & ~PAGE_MASK
;
1316 page
= bp
->b_pages
[page_index
];
1317 csize
= min_t(size_t, PAGE_SIZE
- page_offset
,
1318 BBTOB(bp
->b_io_length
) - boff
);
1320 ASSERT((csize
+ page_offset
) <= PAGE_SIZE
);
1324 memset(page_address(page
) + page_offset
, 0, csize
);
1327 memcpy(data
, page_address(page
) + page_offset
, csize
);
1330 memcpy(page_address(page
) + page_offset
, data
, csize
);
1339 * Handling of buffer targets (buftargs).
1343 * Wait for any bufs with callbacks that have been submitted but have not yet
1344 * returned. These buffers will have an elevated hold count, so wait on those
1345 * while freeing all the buffers only held by the LRU.
1349 struct xfs_buftarg
*btp
)
1354 spin_lock(&btp
->bt_lru_lock
);
1355 while (!list_empty(&btp
->bt_lru
)) {
1356 bp
= list_first_entry(&btp
->bt_lru
, struct xfs_buf
, b_lru
);
1357 if (atomic_read(&bp
->b_hold
) > 1) {
1358 spin_unlock(&btp
->bt_lru_lock
);
1363 * clear the LRU reference count so the buffer doesn't get
1364 * ignored in xfs_buf_rele().
1366 atomic_set(&bp
->b_lru_ref
, 0);
1367 spin_unlock(&btp
->bt_lru_lock
);
1369 spin_lock(&btp
->bt_lru_lock
);
1371 spin_unlock(&btp
->bt_lru_lock
);
1376 struct shrinker
*shrink
,
1377 struct shrink_control
*sc
)
1379 struct xfs_buftarg
*btp
= container_of(shrink
,
1380 struct xfs_buftarg
, bt_shrinker
);
1382 int nr_to_scan
= sc
->nr_to_scan
;
1386 return btp
->bt_lru_nr
;
1388 spin_lock(&btp
->bt_lru_lock
);
1389 while (!list_empty(&btp
->bt_lru
)) {
1390 if (nr_to_scan
-- <= 0)
1393 bp
= list_first_entry(&btp
->bt_lru
, struct xfs_buf
, b_lru
);
1396 * Decrement the b_lru_ref count unless the value is already
1397 * zero. If the value is already zero, we need to reclaim the
1398 * buffer, otherwise it gets another trip through the LRU.
1400 if (!atomic_add_unless(&bp
->b_lru_ref
, -1, 0)) {
1401 list_move_tail(&bp
->b_lru
, &btp
->bt_lru
);
1406 * remove the buffer from the LRU now to avoid needing another
1407 * lock round trip inside xfs_buf_rele().
1409 list_move(&bp
->b_lru
, &dispose
);
1412 spin_unlock(&btp
->bt_lru_lock
);
1414 while (!list_empty(&dispose
)) {
1415 bp
= list_first_entry(&dispose
, struct xfs_buf
, b_lru
);
1416 list_del_init(&bp
->b_lru
);
1420 return btp
->bt_lru_nr
;
1425 struct xfs_mount
*mp
,
1426 struct xfs_buftarg
*btp
)
1428 unregister_shrinker(&btp
->bt_shrinker
);
1430 if (mp
->m_flags
& XFS_MOUNT_BARRIER
)
1431 xfs_blkdev_issue_flush(btp
);
1437 xfs_setsize_buftarg_flags(
1439 unsigned int blocksize
,
1440 unsigned int sectorsize
,
1443 btp
->bt_bsize
= blocksize
;
1444 btp
->bt_sshift
= ffs(sectorsize
) - 1;
1445 btp
->bt_smask
= sectorsize
- 1;
1447 if (set_blocksize(btp
->bt_bdev
, sectorsize
)) {
1448 char name
[BDEVNAME_SIZE
];
1450 bdevname(btp
->bt_bdev
, name
);
1452 xfs_warn(btp
->bt_mount
,
1453 "Cannot set_blocksize to %u on device %s\n",
1462 * When allocating the initial buffer target we have not yet
1463 * read in the superblock, so don't know what sized sectors
1464 * are being used is at this early stage. Play safe.
1467 xfs_setsize_buftarg_early(
1469 struct block_device
*bdev
)
1471 return xfs_setsize_buftarg_flags(btp
,
1472 PAGE_SIZE
, bdev_logical_block_size(bdev
), 0);
1476 xfs_setsize_buftarg(
1478 unsigned int blocksize
,
1479 unsigned int sectorsize
)
1481 return xfs_setsize_buftarg_flags(btp
, blocksize
, sectorsize
, 1);
1486 struct xfs_mount
*mp
,
1487 struct block_device
*bdev
,
1493 btp
= kmem_zalloc(sizeof(*btp
), KM_SLEEP
);
1496 btp
->bt_dev
= bdev
->bd_dev
;
1497 btp
->bt_bdev
= bdev
;
1498 btp
->bt_bdi
= blk_get_backing_dev_info(bdev
);
1502 INIT_LIST_HEAD(&btp
->bt_lru
);
1503 spin_lock_init(&btp
->bt_lru_lock
);
1504 if (xfs_setsize_buftarg_early(btp
, bdev
))
1506 btp
->bt_shrinker
.shrink
= xfs_buftarg_shrink
;
1507 btp
->bt_shrinker
.seeks
= DEFAULT_SEEKS
;
1508 register_shrinker(&btp
->bt_shrinker
);
1517 * Add a buffer to the delayed write list.
1519 * This queues a buffer for writeout if it hasn't already been. Note that
1520 * neither this routine nor the buffer list submission functions perform
1521 * any internal synchronization. It is expected that the lists are thread-local
1524 * Returns true if we queued up the buffer, or false if it already had
1525 * been on the buffer list.
1528 xfs_buf_delwri_queue(
1530 struct list_head
*list
)
1532 ASSERT(xfs_buf_islocked(bp
));
1533 ASSERT(!(bp
->b_flags
& XBF_READ
));
1536 * If the buffer is already marked delwri it already is queued up
1537 * by someone else for imediate writeout. Just ignore it in that
1540 if (bp
->b_flags
& _XBF_DELWRI_Q
) {
1541 trace_xfs_buf_delwri_queued(bp
, _RET_IP_
);
1545 trace_xfs_buf_delwri_queue(bp
, _RET_IP_
);
1548 * If a buffer gets written out synchronously or marked stale while it
1549 * is on a delwri list we lazily remove it. To do this, the other party
1550 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1551 * It remains referenced and on the list. In a rare corner case it
1552 * might get readded to a delwri list after the synchronous writeout, in
1553 * which case we need just need to re-add the flag here.
1555 bp
->b_flags
|= _XBF_DELWRI_Q
;
1556 if (list_empty(&bp
->b_list
)) {
1557 atomic_inc(&bp
->b_hold
);
1558 list_add_tail(&bp
->b_list
, list
);
1565 * Compare function is more complex than it needs to be because
1566 * the return value is only 32 bits and we are doing comparisons
1572 struct list_head
*a
,
1573 struct list_head
*b
)
1575 struct xfs_buf
*ap
= container_of(a
, struct xfs_buf
, b_list
);
1576 struct xfs_buf
*bp
= container_of(b
, struct xfs_buf
, b_list
);
1579 diff
= ap
->b_bn
- bp
->b_bn
;
1588 __xfs_buf_delwri_submit(
1589 struct list_head
*buffer_list
,
1590 struct list_head
*io_list
,
1593 struct blk_plug plug
;
1594 struct xfs_buf
*bp
, *n
;
1597 list_for_each_entry_safe(bp
, n
, buffer_list
, b_list
) {
1599 if (xfs_buf_ispinned(bp
)) {
1603 if (!xfs_buf_trylock(bp
))
1610 * Someone else might have written the buffer synchronously or
1611 * marked it stale in the meantime. In that case only the
1612 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1613 * reference and remove it from the list here.
1615 if (!(bp
->b_flags
& _XBF_DELWRI_Q
)) {
1616 list_del_init(&bp
->b_list
);
1621 list_move_tail(&bp
->b_list
, io_list
);
1622 trace_xfs_buf_delwri_split(bp
, _RET_IP_
);
1625 list_sort(NULL
, io_list
, xfs_buf_cmp
);
1627 blk_start_plug(&plug
);
1628 list_for_each_entry_safe(bp
, n
, io_list
, b_list
) {
1629 bp
->b_flags
&= ~(_XBF_DELWRI_Q
| XBF_ASYNC
);
1630 bp
->b_flags
|= XBF_WRITE
;
1633 bp
->b_flags
|= XBF_ASYNC
;
1634 list_del_init(&bp
->b_list
);
1638 blk_finish_plug(&plug
);
1644 * Write out a buffer list asynchronously.
1646 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1647 * out and not wait for I/O completion on any of the buffers. This interface
1648 * is only safely useable for callers that can track I/O completion by higher
1649 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1653 xfs_buf_delwri_submit_nowait(
1654 struct list_head
*buffer_list
)
1656 LIST_HEAD (io_list
);
1657 return __xfs_buf_delwri_submit(buffer_list
, &io_list
, false);
1661 * Write out a buffer list synchronously.
1663 * This will take the @buffer_list, write all buffers out and wait for I/O
1664 * completion on all of the buffers. @buffer_list is consumed by the function,
1665 * so callers must have some other way of tracking buffers if they require such
1669 xfs_buf_delwri_submit(
1670 struct list_head
*buffer_list
)
1672 LIST_HEAD (io_list
);
1673 int error
= 0, error2
;
1676 __xfs_buf_delwri_submit(buffer_list
, &io_list
, true);
1678 /* Wait for IO to complete. */
1679 while (!list_empty(&io_list
)) {
1680 bp
= list_first_entry(&io_list
, struct xfs_buf
, b_list
);
1682 list_del_init(&bp
->b_list
);
1683 error2
= xfs_buf_iowait(bp
);
1695 xfs_buf_zone
= kmem_zone_init_flags(sizeof(xfs_buf_t
), "xfs_buf",
1696 KM_ZONE_HWALIGN
, NULL
);
1700 xfslogd_workqueue
= alloc_workqueue("xfslogd",
1701 WQ_MEM_RECLAIM
| WQ_HIGHPRI
, 1);
1702 if (!xfslogd_workqueue
)
1703 goto out_free_buf_zone
;
1708 kmem_zone_destroy(xfs_buf_zone
);
1714 xfs_buf_terminate(void)
1716 destroy_workqueue(xfslogd_workqueue
);
1717 kmem_zone_destroy(xfs_buf_zone
);