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 rb_link_node(&new_bp
->b_rbnode
, parent
, rbp
);
481 rb_insert_color(&new_bp
->b_rbnode
, &pag
->pag_buf_tree
);
482 /* the buffer keeps the perag reference until it is freed */
484 spin_unlock(&pag
->pag_buf_lock
);
486 XFS_STATS_INC(xb_miss_locked
);
487 spin_unlock(&pag
->pag_buf_lock
);
493 spin_unlock(&pag
->pag_buf_lock
);
496 if (!xfs_buf_trylock(bp
)) {
497 if (flags
& XBF_TRYLOCK
) {
499 XFS_STATS_INC(xb_busy_locked
);
503 XFS_STATS_INC(xb_get_locked_waited
);
507 * if the buffer is stale, clear all the external state associated with
508 * it. We need to keep flags such as how we allocated the buffer memory
511 if (bp
->b_flags
& XBF_STALE
) {
512 ASSERT((bp
->b_flags
& _XBF_DELWRI_Q
) == 0);
513 bp
->b_flags
&= XBF_MAPPED
| _XBF_KMEM
| _XBF_PAGES
;
516 trace_xfs_buf_find(bp
, flags
, _RET_IP_
);
517 XFS_STATS_INC(xb_get_locked
);
522 * Assembles a buffer covering the specified range. The code is optimised for
523 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
524 * more hits than misses.
528 xfs_buftarg_t
*target
,/* target for buffer */
529 xfs_off_t ioff
, /* starting offset of range */
530 size_t isize
, /* length of range */
531 xfs_buf_flags_t flags
)
534 struct xfs_buf
*new_bp
;
537 bp
= _xfs_buf_find(target
, ioff
, isize
, flags
, NULL
);
541 new_bp
= xfs_buf_allocate(flags
);
542 if (unlikely(!new_bp
))
545 _xfs_buf_initialize(new_bp
, target
,
546 ioff
<< BBSHIFT
, isize
<< BBSHIFT
, flags
);
548 bp
= _xfs_buf_find(target
, ioff
, isize
, flags
, new_bp
);
550 xfs_buf_deallocate(new_bp
);
555 error
= xfs_buf_allocate_memory(bp
, flags
);
559 xfs_buf_deallocate(new_bp
);
562 * Now we have a workable buffer, fill in the block number so
563 * that we can do IO on it.
566 bp
->b_count_desired
= bp
->b_buffer_length
;
569 if (!(bp
->b_flags
& XBF_MAPPED
)) {
570 error
= _xfs_buf_map_pages(bp
, flags
);
571 if (unlikely(error
)) {
572 xfs_warn(target
->bt_mount
,
573 "%s: failed to map pages\n", __func__
);
578 XFS_STATS_INC(xb_get
);
579 trace_xfs_buf_get(bp
, flags
, _RET_IP_
);
583 if (flags
& (XBF_LOCK
| XBF_TRYLOCK
))
592 xfs_buf_flags_t flags
)
596 ASSERT(!(flags
& (XBF_DELWRI
|XBF_WRITE
)));
597 ASSERT(bp
->b_bn
!= XFS_BUF_DADDR_NULL
);
599 bp
->b_flags
&= ~(XBF_WRITE
| XBF_ASYNC
| XBF_DELWRI
| XBF_READ_AHEAD
);
600 bp
->b_flags
|= flags
& (XBF_READ
| XBF_ASYNC
| XBF_READ_AHEAD
);
602 status
= xfs_buf_iorequest(bp
);
603 if (status
|| bp
->b_error
|| (flags
& XBF_ASYNC
))
605 return xfs_buf_iowait(bp
);
610 xfs_buftarg_t
*target
,
613 xfs_buf_flags_t flags
)
619 bp
= xfs_buf_get(target
, ioff
, isize
, flags
);
621 trace_xfs_buf_read(bp
, flags
, _RET_IP_
);
623 if (!XFS_BUF_ISDONE(bp
)) {
624 XFS_STATS_INC(xb_get_read
);
625 _xfs_buf_read(bp
, flags
);
626 } else if (flags
& XBF_ASYNC
) {
628 * Read ahead call which is already satisfied,
633 /* We do not want read in the flags */
634 bp
->b_flags
&= ~XBF_READ
;
641 if (flags
& (XBF_LOCK
| XBF_TRYLOCK
))
648 * If we are not low on memory then do the readahead in a deadlock
653 xfs_buftarg_t
*target
,
657 if (bdi_read_congested(target
->bt_bdi
))
660 xfs_buf_read(target
, ioff
, isize
,
661 XBF_TRYLOCK
|XBF_ASYNC
|XBF_READ_AHEAD
|XBF_DONT_BLOCK
);
665 * Read an uncached buffer from disk. Allocates and returns a locked
666 * buffer containing the disk contents or nothing.
669 xfs_buf_read_uncached(
670 struct xfs_mount
*mp
,
671 struct xfs_buftarg
*target
,
679 bp
= xfs_buf_get_uncached(target
, length
, flags
);
683 /* set up the buffer for a read IO */
684 XFS_BUF_SET_ADDR(bp
, daddr
);
688 error
= xfs_buf_iowait(bp
);
689 if (error
|| bp
->b_error
) {
699 xfs_buftarg_t
*target
)
703 bp
= xfs_buf_allocate(0);
705 _xfs_buf_initialize(bp
, target
, 0, len
, 0);
710 * Return a buffer allocated as an empty buffer and associated to external
711 * memory via xfs_buf_associate_memory() back to it's empty state.
719 _xfs_buf_free_pages(bp
);
722 bp
->b_page_count
= 0;
724 bp
->b_file_offset
= 0;
725 bp
->b_buffer_length
= bp
->b_count_desired
= len
;
726 bp
->b_bn
= XFS_BUF_DADDR_NULL
;
727 bp
->b_flags
&= ~XBF_MAPPED
;
730 static inline struct page
*
734 if ((!is_vmalloc_addr(addr
))) {
735 return virt_to_page(addr
);
737 return vmalloc_to_page(addr
);
742 xfs_buf_associate_memory(
749 unsigned long pageaddr
;
750 unsigned long offset
;
754 pageaddr
= (unsigned long)mem
& PAGE_MASK
;
755 offset
= (unsigned long)mem
- pageaddr
;
756 buflen
= PAGE_ALIGN(len
+ offset
);
757 page_count
= buflen
>> PAGE_SHIFT
;
759 /* Free any previous set of page pointers */
761 _xfs_buf_free_pages(bp
);
766 rval
= _xfs_buf_get_pages(bp
, page_count
, XBF_DONT_BLOCK
);
770 bp
->b_offset
= offset
;
772 for (i
= 0; i
< bp
->b_page_count
; i
++) {
773 bp
->b_pages
[i
] = mem_to_page((void *)pageaddr
);
774 pageaddr
+= PAGE_SIZE
;
777 bp
->b_count_desired
= len
;
778 bp
->b_buffer_length
= buflen
;
779 bp
->b_flags
|= XBF_MAPPED
;
785 xfs_buf_get_uncached(
786 struct xfs_buftarg
*target
,
790 unsigned long page_count
= PAGE_ALIGN(len
) >> PAGE_SHIFT
;
794 bp
= xfs_buf_allocate(0);
795 if (unlikely(bp
== NULL
))
797 _xfs_buf_initialize(bp
, target
, 0, len
, 0);
799 error
= _xfs_buf_get_pages(bp
, page_count
, 0);
803 for (i
= 0; i
< page_count
; i
++) {
804 bp
->b_pages
[i
] = alloc_page(xb_to_gfp(flags
));
808 bp
->b_flags
|= _XBF_PAGES
;
810 error
= _xfs_buf_map_pages(bp
, XBF_MAPPED
);
811 if (unlikely(error
)) {
812 xfs_warn(target
->bt_mount
,
813 "%s: failed to map pages\n", __func__
);
817 trace_xfs_buf_get_uncached(bp
, _RET_IP_
);
822 __free_page(bp
->b_pages
[i
]);
823 _xfs_buf_free_pages(bp
);
825 xfs_buf_deallocate(bp
);
831 * Increment reference count on buffer, to hold the buffer concurrently
832 * with another thread which may release (free) the buffer asynchronously.
833 * Must hold the buffer already to call this function.
839 trace_xfs_buf_hold(bp
, _RET_IP_
);
840 atomic_inc(&bp
->b_hold
);
844 * Releases a hold on the specified buffer. If the
845 * the hold count is 1, calls xfs_buf_free.
851 struct xfs_perag
*pag
= bp
->b_pag
;
853 trace_xfs_buf_rele(bp
, _RET_IP_
);
856 ASSERT(list_empty(&bp
->b_lru
));
857 ASSERT(RB_EMPTY_NODE(&bp
->b_rbnode
));
858 if (atomic_dec_and_test(&bp
->b_hold
))
863 ASSERT(!RB_EMPTY_NODE(&bp
->b_rbnode
));
865 ASSERT(atomic_read(&bp
->b_hold
) > 0);
866 if (atomic_dec_and_lock(&bp
->b_hold
, &pag
->pag_buf_lock
)) {
867 if (!(bp
->b_flags
& XBF_STALE
) &&
868 atomic_read(&bp
->b_lru_ref
)) {
870 spin_unlock(&pag
->pag_buf_lock
);
873 ASSERT(!(bp
->b_flags
& (XBF_DELWRI
|_XBF_DELWRI_Q
)));
874 rb_erase(&bp
->b_rbnode
, &pag
->pag_buf_tree
);
875 spin_unlock(&pag
->pag_buf_lock
);
884 * Lock a buffer object, if it is not already locked.
886 * If we come across a stale, pinned, locked buffer, we know that we are
887 * being asked to lock a buffer that has been reallocated. Because it is
888 * pinned, we know that the log has not been pushed to disk and hence it
889 * will still be locked. Rather than continuing to have trylock attempts
890 * fail until someone else pushes the log, push it ourselves before
891 * returning. This means that the xfsaild will not get stuck trying
892 * to push on stale inode buffers.
900 locked
= down_trylock(&bp
->b_sema
) == 0;
903 else if (atomic_read(&bp
->b_pin_count
) && (bp
->b_flags
& XBF_STALE
))
904 xfs_log_force(bp
->b_target
->bt_mount
, 0);
906 trace_xfs_buf_trylock(bp
, _RET_IP_
);
911 * Lock a buffer object.
913 * If we come across a stale, pinned, locked buffer, we know that we
914 * are being asked to lock a buffer that has been reallocated. Because
915 * it is pinned, we know that the log has not been pushed to disk and
916 * hence it will still be locked. Rather than sleeping until someone
917 * else pushes the log, push it ourselves before trying to get the lock.
923 trace_xfs_buf_lock(bp
, _RET_IP_
);
925 if (atomic_read(&bp
->b_pin_count
) && (bp
->b_flags
& XBF_STALE
))
926 xfs_log_force(bp
->b_target
->bt_mount
, 0);
930 trace_xfs_buf_lock_done(bp
, _RET_IP_
);
934 * Releases the lock on the buffer object.
935 * If the buffer is marked delwri but is not queued, do so before we
936 * unlock the buffer as we need to set flags correctly. We also need to
937 * take a reference for the delwri queue because the unlocker is going to
938 * drop their's and they don't know we just queued it.
947 trace_xfs_buf_unlock(bp
, _RET_IP_
);
954 DECLARE_WAITQUEUE (wait
, current
);
956 if (atomic_read(&bp
->b_pin_count
) == 0)
959 add_wait_queue(&bp
->b_waiters
, &wait
);
961 set_current_state(TASK_UNINTERRUPTIBLE
);
962 if (atomic_read(&bp
->b_pin_count
) == 0)
966 remove_wait_queue(&bp
->b_waiters
, &wait
);
967 set_current_state(TASK_RUNNING
);
971 * Buffer Utility Routines
976 struct work_struct
*work
)
979 container_of(work
, xfs_buf_t
, b_iodone_work
);
982 (*(bp
->b_iodone
))(bp
);
983 else if (bp
->b_flags
& XBF_ASYNC
)
992 trace_xfs_buf_iodone(bp
, _RET_IP_
);
994 bp
->b_flags
&= ~(XBF_READ
| XBF_WRITE
| XBF_READ_AHEAD
);
995 if (bp
->b_error
== 0)
996 bp
->b_flags
|= XBF_DONE
;
998 if ((bp
->b_iodone
) || (bp
->b_flags
& XBF_ASYNC
)) {
1000 INIT_WORK(&bp
->b_iodone_work
, xfs_buf_iodone_work
);
1001 queue_work(xfslogd_workqueue
, &bp
->b_iodone_work
);
1003 xfs_buf_iodone_work(&bp
->b_iodone_work
);
1006 complete(&bp
->b_iowait
);
1015 ASSERT(error
>= 0 && error
<= 0xffff);
1016 bp
->b_error
= (unsigned short)error
;
1017 trace_xfs_buf_ioerror(bp
, error
, _RET_IP_
);
1026 bp
->b_flags
|= XBF_WRITE
;
1027 bp
->b_flags
&= ~(XBF_ASYNC
| XBF_READ
);
1029 xfs_buf_delwri_dequeue(bp
);
1032 error
= xfs_buf_iowait(bp
);
1034 xfs_force_shutdown(bp
->b_target
->bt_mount
,
1035 SHUTDOWN_META_IO_ERROR
);
1041 * Called when we want to stop a buffer from getting written or read.
1042 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1043 * so that the proper iodone callbacks get called.
1049 #ifdef XFSERRORDEBUG
1050 ASSERT(XFS_BUF_ISREAD(bp
) || bp
->b_iodone
);
1054 * No need to wait until the buffer is unpinned, we aren't flushing it.
1056 xfs_buf_ioerror(bp
, EIO
);
1059 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1062 xfs_buf_delwri_dequeue(bp
);
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.
1091 xfs_buf_delwri_dequeue(bp
);
1094 bp
->b_iodone
= NULL
;
1095 if (!(fl
& XBF_ASYNC
)) {
1097 * Mark b_error and B_ERROR _both_.
1098 * Lot's of chunkcache code assumes that.
1099 * There's no reason to mark error for
1102 xfs_buf_ioerror(bp
, EIO
);
1103 XFS_BUF_FINISH_IOWAIT(bp
);
1113 * All xfs metadata buffers except log state machine buffers
1114 * get this attached as their b_bdstrat callback function.
1115 * This is so that we can catch a buffer
1116 * after prematurely unpinning it to forcibly shutdown the filesystem.
1122 if (XFS_FORCED_SHUTDOWN(bp
->b_target
->bt_mount
)) {
1123 trace_xfs_bdstrat_shut(bp
, _RET_IP_
);
1125 * Metadata write that didn't get logged but
1126 * written delayed anyway. These aren't associated
1127 * with a transaction, and can be ignored.
1129 if (!bp
->b_iodone
&& !XFS_BUF_ISREAD(bp
))
1130 return xfs_bioerror_relse(bp
);
1132 return xfs_bioerror(bp
);
1135 xfs_buf_iorequest(bp
);
1140 * Wrapper around bdstrat so that we can stop data from going to disk in case
1141 * we are shutting down the filesystem. Typically user data goes thru this
1142 * path; one of the exceptions is the superblock.
1146 struct xfs_mount
*mp
,
1149 if (XFS_FORCED_SHUTDOWN(mp
)) {
1150 trace_xfs_bdstrat_shut(bp
, _RET_IP_
);
1151 xfs_bioerror_relse(bp
);
1155 xfs_buf_iorequest(bp
);
1163 if (atomic_dec_and_test(&bp
->b_io_remaining
) == 1)
1164 xfs_buf_ioend(bp
, schedule
);
1172 xfs_buf_t
*bp
= (xfs_buf_t
*)bio
->bi_private
;
1174 xfs_buf_ioerror(bp
, -error
);
1176 if (!error
&& xfs_buf_is_vmapped(bp
) && (bp
->b_flags
& XBF_READ
))
1177 invalidate_kernel_vmap_range(bp
->b_addr
, xfs_buf_vmap_len(bp
));
1179 _xfs_buf_ioend(bp
, 1);
1187 int rw
, map_i
, total_nr_pages
, nr_pages
;
1189 int offset
= bp
->b_offset
;
1190 int size
= bp
->b_count_desired
;
1191 sector_t sector
= bp
->b_bn
;
1193 total_nr_pages
= bp
->b_page_count
;
1196 if (bp
->b_flags
& XBF_WRITE
) {
1197 if (bp
->b_flags
& XBF_SYNCIO
)
1201 if (bp
->b_flags
& XBF_FUA
)
1203 if (bp
->b_flags
& XBF_FLUSH
)
1205 } else if (bp
->b_flags
& XBF_READ_AHEAD
) {
1211 /* we only use the buffer cache for meta-data */
1215 atomic_inc(&bp
->b_io_remaining
);
1216 nr_pages
= BIO_MAX_SECTORS
>> (PAGE_SHIFT
- BBSHIFT
);
1217 if (nr_pages
> total_nr_pages
)
1218 nr_pages
= total_nr_pages
;
1220 bio
= bio_alloc(GFP_NOIO
, nr_pages
);
1221 bio
->bi_bdev
= bp
->b_target
->bt_bdev
;
1222 bio
->bi_sector
= sector
;
1223 bio
->bi_end_io
= xfs_buf_bio_end_io
;
1224 bio
->bi_private
= bp
;
1227 for (; size
&& nr_pages
; nr_pages
--, map_i
++) {
1228 int rbytes
, nbytes
= PAGE_SIZE
- offset
;
1233 rbytes
= bio_add_page(bio
, bp
->b_pages
[map_i
], nbytes
, offset
);
1234 if (rbytes
< nbytes
)
1238 sector
+= nbytes
>> BBSHIFT
;
1243 if (likely(bio
->bi_size
)) {
1244 if (xfs_buf_is_vmapped(bp
)) {
1245 flush_kernel_vmap_range(bp
->b_addr
,
1246 xfs_buf_vmap_len(bp
));
1248 submit_bio(rw
, bio
);
1252 xfs_buf_ioerror(bp
, EIO
);
1261 trace_xfs_buf_iorequest(bp
, _RET_IP_
);
1263 ASSERT(!(bp
->b_flags
& XBF_DELWRI
));
1265 if (bp
->b_flags
& XBF_WRITE
)
1266 xfs_buf_wait_unpin(bp
);
1269 /* Set the count to 1 initially, this will stop an I/O
1270 * completion callout which happens before we have started
1271 * all the I/O from calling xfs_buf_ioend too early.
1273 atomic_set(&bp
->b_io_remaining
, 1);
1274 _xfs_buf_ioapply(bp
);
1275 _xfs_buf_ioend(bp
, 0);
1282 * Waits for I/O to complete on the buffer supplied.
1283 * It returns immediately if no I/O is pending.
1284 * It returns the I/O error code, if any, or 0 if there was no error.
1290 trace_xfs_buf_iowait(bp
, _RET_IP_
);
1292 wait_for_completion(&bp
->b_iowait
);
1294 trace_xfs_buf_iowait_done(bp
, _RET_IP_
);
1305 if (bp
->b_flags
& XBF_MAPPED
)
1306 return bp
->b_addr
+ offset
;
1308 offset
+= bp
->b_offset
;
1309 page
= bp
->b_pages
[offset
>> PAGE_SHIFT
];
1310 return (xfs_caddr_t
)page_address(page
) + (offset
& (PAGE_SIZE
-1));
1314 * Move data into or out of a buffer.
1318 xfs_buf_t
*bp
, /* buffer to process */
1319 size_t boff
, /* starting buffer offset */
1320 size_t bsize
, /* length to copy */
1321 void *data
, /* data address */
1322 xfs_buf_rw_t mode
) /* read/write/zero flag */
1324 size_t bend
, cpoff
, csize
;
1327 bend
= boff
+ bsize
;
1328 while (boff
< bend
) {
1329 page
= bp
->b_pages
[xfs_buf_btoct(boff
+ bp
->b_offset
)];
1330 cpoff
= xfs_buf_poff(boff
+ bp
->b_offset
);
1331 csize
= min_t(size_t,
1332 PAGE_SIZE
-cpoff
, bp
->b_count_desired
-boff
);
1334 ASSERT(((csize
+ cpoff
) <= PAGE_SIZE
));
1338 memset(page_address(page
) + cpoff
, 0, csize
);
1341 memcpy(data
, page_address(page
) + cpoff
, csize
);
1344 memcpy(page_address(page
) + cpoff
, data
, csize
);
1353 * Handling of buffer targets (buftargs).
1357 * Wait for any bufs with callbacks that have been submitted but have not yet
1358 * returned. These buffers will have an elevated hold count, so wait on those
1359 * while freeing all the buffers only held by the LRU.
1363 struct xfs_buftarg
*btp
)
1368 spin_lock(&btp
->bt_lru_lock
);
1369 while (!list_empty(&btp
->bt_lru
)) {
1370 bp
= list_first_entry(&btp
->bt_lru
, struct xfs_buf
, b_lru
);
1371 if (atomic_read(&bp
->b_hold
) > 1) {
1372 spin_unlock(&btp
->bt_lru_lock
);
1377 * clear the LRU reference count so the bufer doesn't get
1378 * ignored in xfs_buf_rele().
1380 atomic_set(&bp
->b_lru_ref
, 0);
1381 spin_unlock(&btp
->bt_lru_lock
);
1383 spin_lock(&btp
->bt_lru_lock
);
1385 spin_unlock(&btp
->bt_lru_lock
);
1390 struct shrinker
*shrink
,
1391 struct shrink_control
*sc
)
1393 struct xfs_buftarg
*btp
= container_of(shrink
,
1394 struct xfs_buftarg
, bt_shrinker
);
1396 int nr_to_scan
= sc
->nr_to_scan
;
1400 return btp
->bt_lru_nr
;
1402 spin_lock(&btp
->bt_lru_lock
);
1403 while (!list_empty(&btp
->bt_lru
)) {
1404 if (nr_to_scan
-- <= 0)
1407 bp
= list_first_entry(&btp
->bt_lru
, struct xfs_buf
, b_lru
);
1410 * Decrement the b_lru_ref count unless the value is already
1411 * zero. If the value is already zero, we need to reclaim the
1412 * buffer, otherwise it gets another trip through the LRU.
1414 if (!atomic_add_unless(&bp
->b_lru_ref
, -1, 0)) {
1415 list_move_tail(&bp
->b_lru
, &btp
->bt_lru
);
1420 * remove the buffer from the LRU now to avoid needing another
1421 * lock round trip inside xfs_buf_rele().
1423 list_move(&bp
->b_lru
, &dispose
);
1426 spin_unlock(&btp
->bt_lru_lock
);
1428 while (!list_empty(&dispose
)) {
1429 bp
= list_first_entry(&dispose
, struct xfs_buf
, b_lru
);
1430 list_del_init(&bp
->b_lru
);
1434 return btp
->bt_lru_nr
;
1439 struct xfs_mount
*mp
,
1440 struct xfs_buftarg
*btp
)
1442 unregister_shrinker(&btp
->bt_shrinker
);
1444 xfs_flush_buftarg(btp
, 1);
1445 if (mp
->m_flags
& XFS_MOUNT_BARRIER
)
1446 xfs_blkdev_issue_flush(btp
);
1448 kthread_stop(btp
->bt_task
);
1453 xfs_setsize_buftarg_flags(
1455 unsigned int blocksize
,
1456 unsigned int sectorsize
,
1459 btp
->bt_bsize
= blocksize
;
1460 btp
->bt_sshift
= ffs(sectorsize
) - 1;
1461 btp
->bt_smask
= sectorsize
- 1;
1463 if (set_blocksize(btp
->bt_bdev
, sectorsize
)) {
1464 xfs_warn(btp
->bt_mount
,
1465 "Cannot set_blocksize to %u on device %s\n",
1466 sectorsize
, xfs_buf_target_name(btp
));
1474 * When allocating the initial buffer target we have not yet
1475 * read in the superblock, so don't know what sized sectors
1476 * are being used is at this early stage. Play safe.
1479 xfs_setsize_buftarg_early(
1481 struct block_device
*bdev
)
1483 return xfs_setsize_buftarg_flags(btp
,
1484 PAGE_SIZE
, bdev_logical_block_size(bdev
), 0);
1488 xfs_setsize_buftarg(
1490 unsigned int blocksize
,
1491 unsigned int sectorsize
)
1493 return xfs_setsize_buftarg_flags(btp
, blocksize
, sectorsize
, 1);
1497 xfs_alloc_delwri_queue(
1501 INIT_LIST_HEAD(&btp
->bt_delwri_queue
);
1502 spin_lock_init(&btp
->bt_delwri_lock
);
1504 btp
->bt_task
= kthread_run(xfsbufd
, btp
, "xfsbufd/%s", fsname
);
1505 if (IS_ERR(btp
->bt_task
))
1506 return PTR_ERR(btp
->bt_task
);
1512 struct xfs_mount
*mp
,
1513 struct block_device
*bdev
,
1519 btp
= kmem_zalloc(sizeof(*btp
), KM_SLEEP
);
1522 btp
->bt_dev
= bdev
->bd_dev
;
1523 btp
->bt_bdev
= bdev
;
1524 btp
->bt_bdi
= blk_get_backing_dev_info(bdev
);
1528 INIT_LIST_HEAD(&btp
->bt_lru
);
1529 spin_lock_init(&btp
->bt_lru_lock
);
1530 if (xfs_setsize_buftarg_early(btp
, bdev
))
1532 if (xfs_alloc_delwri_queue(btp
, fsname
))
1534 btp
->bt_shrinker
.shrink
= xfs_buftarg_shrink
;
1535 btp
->bt_shrinker
.seeks
= DEFAULT_SEEKS
;
1536 register_shrinker(&btp
->bt_shrinker
);
1546 * Delayed write buffer handling
1549 xfs_buf_delwri_queue(
1552 struct xfs_buftarg
*btp
= bp
->b_target
;
1554 trace_xfs_buf_delwri_queue(bp
, _RET_IP_
);
1556 ASSERT(!(bp
->b_flags
& XBF_READ
));
1558 spin_lock(&btp
->bt_delwri_lock
);
1559 if (!list_empty(&bp
->b_list
)) {
1560 /* if already in the queue, move it to the tail */
1561 ASSERT(bp
->b_flags
& _XBF_DELWRI_Q
);
1562 list_move_tail(&bp
->b_list
, &btp
->bt_delwri_queue
);
1564 /* start xfsbufd as it is about to have something to do */
1565 if (list_empty(&btp
->bt_delwri_queue
))
1566 wake_up_process(bp
->b_target
->bt_task
);
1568 atomic_inc(&bp
->b_hold
);
1569 bp
->b_flags
|= XBF_DELWRI
| _XBF_DELWRI_Q
| XBF_ASYNC
;
1570 list_add_tail(&bp
->b_list
, &btp
->bt_delwri_queue
);
1572 bp
->b_queuetime
= jiffies
;
1573 spin_unlock(&btp
->bt_delwri_lock
);
1577 xfs_buf_delwri_dequeue(
1582 spin_lock(&bp
->b_target
->bt_delwri_lock
);
1583 if ((bp
->b_flags
& XBF_DELWRI
) && !list_empty(&bp
->b_list
)) {
1584 ASSERT(bp
->b_flags
& _XBF_DELWRI_Q
);
1585 list_del_init(&bp
->b_list
);
1588 bp
->b_flags
&= ~(XBF_DELWRI
|_XBF_DELWRI_Q
);
1589 spin_unlock(&bp
->b_target
->bt_delwri_lock
);
1594 trace_xfs_buf_delwri_dequeue(bp
, _RET_IP_
);
1598 * If a delwri buffer needs to be pushed before it has aged out, then promote
1599 * it to the head of the delwri queue so that it will be flushed on the next
1600 * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1601 * than the age currently needed to flush the buffer. Hence the next time the
1602 * xfsbufd sees it is guaranteed to be considered old enough to flush.
1605 xfs_buf_delwri_promote(
1608 struct xfs_buftarg
*btp
= bp
->b_target
;
1609 long age
= xfs_buf_age_centisecs
* msecs_to_jiffies(10) + 1;
1611 ASSERT(bp
->b_flags
& XBF_DELWRI
);
1612 ASSERT(bp
->b_flags
& _XBF_DELWRI_Q
);
1615 * Check the buffer age before locking the delayed write queue as we
1616 * don't need to promote buffers that are already past the flush age.
1618 if (bp
->b_queuetime
< jiffies
- age
)
1620 bp
->b_queuetime
= jiffies
- age
;
1621 spin_lock(&btp
->bt_delwri_lock
);
1622 list_move(&bp
->b_list
, &btp
->bt_delwri_queue
);
1623 spin_unlock(&btp
->bt_delwri_lock
);
1627 xfs_buf_runall_queues(
1628 struct workqueue_struct
*queue
)
1630 flush_workqueue(queue
);
1634 * Move as many buffers as specified to the supplied list
1635 * idicating if we skipped any buffers to prevent deadlocks.
1638 xfs_buf_delwri_split(
1639 xfs_buftarg_t
*target
,
1640 struct list_head
*list
,
1647 force
= test_and_clear_bit(XBT_FORCE_FLUSH
, &target
->bt_flags
);
1648 INIT_LIST_HEAD(list
);
1649 spin_lock(&target
->bt_delwri_lock
);
1650 list_for_each_entry_safe(bp
, n
, &target
->bt_delwri_queue
, b_list
) {
1651 ASSERT(bp
->b_flags
& XBF_DELWRI
);
1653 if (!xfs_buf_ispinned(bp
) && xfs_buf_trylock(bp
)) {
1655 time_before(jiffies
, bp
->b_queuetime
+ age
)) {
1660 bp
->b_flags
&= ~(XBF_DELWRI
| _XBF_DELWRI_Q
);
1661 bp
->b_flags
|= XBF_WRITE
;
1662 list_move_tail(&bp
->b_list
, list
);
1663 trace_xfs_buf_delwri_split(bp
, _RET_IP_
);
1668 spin_unlock(&target
->bt_delwri_lock
);
1673 * Compare function is more complex than it needs to be because
1674 * the return value is only 32 bits and we are doing comparisons
1680 struct list_head
*a
,
1681 struct list_head
*b
)
1683 struct xfs_buf
*ap
= container_of(a
, struct xfs_buf
, b_list
);
1684 struct xfs_buf
*bp
= container_of(b
, struct xfs_buf
, b_list
);
1687 diff
= ap
->b_bn
- bp
->b_bn
;
1699 xfs_buftarg_t
*target
= (xfs_buftarg_t
*)data
;
1701 current
->flags
|= PF_MEMALLOC
;
1706 long age
= xfs_buf_age_centisecs
* msecs_to_jiffies(10);
1707 long tout
= xfs_buf_timer_centisecs
* msecs_to_jiffies(10);
1708 struct list_head tmp
;
1709 struct blk_plug plug
;
1711 if (unlikely(freezing(current
))) {
1712 set_bit(XBT_FORCE_SLEEP
, &target
->bt_flags
);
1715 clear_bit(XBT_FORCE_SLEEP
, &target
->bt_flags
);
1718 /* sleep for a long time if there is nothing to do. */
1719 if (list_empty(&target
->bt_delwri_queue
))
1720 tout
= MAX_SCHEDULE_TIMEOUT
;
1721 schedule_timeout_interruptible(tout
);
1723 xfs_buf_delwri_split(target
, &tmp
, age
);
1724 list_sort(NULL
, &tmp
, xfs_buf_cmp
);
1726 blk_start_plug(&plug
);
1727 while (!list_empty(&tmp
)) {
1729 bp
= list_first_entry(&tmp
, struct xfs_buf
, b_list
);
1730 list_del_init(&bp
->b_list
);
1733 blk_finish_plug(&plug
);
1734 } while (!kthread_should_stop());
1740 * Go through all incore buffers, and release buffers if they belong to
1741 * the given device. This is used in filesystem error handling to
1742 * preserve the consistency of its metadata.
1746 xfs_buftarg_t
*target
,
1751 LIST_HEAD(tmp_list
);
1752 LIST_HEAD(wait_list
);
1753 struct blk_plug plug
;
1755 xfs_buf_runall_queues(xfsconvertd_workqueue
);
1756 xfs_buf_runall_queues(xfsdatad_workqueue
);
1757 xfs_buf_runall_queues(xfslogd_workqueue
);
1759 set_bit(XBT_FORCE_FLUSH
, &target
->bt_flags
);
1760 pincount
= xfs_buf_delwri_split(target
, &tmp_list
, 0);
1763 * Dropped the delayed write list lock, now walk the temporary list.
1764 * All I/O is issued async and then if we need to wait for completion
1765 * we do that after issuing all the IO.
1767 list_sort(NULL
, &tmp_list
, xfs_buf_cmp
);
1769 blk_start_plug(&plug
);
1770 while (!list_empty(&tmp_list
)) {
1771 bp
= list_first_entry(&tmp_list
, struct xfs_buf
, b_list
);
1772 ASSERT(target
== bp
->b_target
);
1773 list_del_init(&bp
->b_list
);
1775 bp
->b_flags
&= ~XBF_ASYNC
;
1776 list_add(&bp
->b_list
, &wait_list
);
1780 blk_finish_plug(&plug
);
1783 /* Wait for IO to complete. */
1784 while (!list_empty(&wait_list
)) {
1785 bp
= list_first_entry(&wait_list
, struct xfs_buf
, b_list
);
1787 list_del_init(&bp
->b_list
);
1799 xfs_buf_zone
= kmem_zone_init_flags(sizeof(xfs_buf_t
), "xfs_buf",
1800 KM_ZONE_HWALIGN
, NULL
);
1804 xfslogd_workqueue
= alloc_workqueue("xfslogd",
1805 WQ_MEM_RECLAIM
| WQ_HIGHPRI
, 1);
1806 if (!xfslogd_workqueue
)
1807 goto out_free_buf_zone
;
1809 xfsdatad_workqueue
= alloc_workqueue("xfsdatad", WQ_MEM_RECLAIM
, 1);
1810 if (!xfsdatad_workqueue
)
1811 goto out_destroy_xfslogd_workqueue
;
1813 xfsconvertd_workqueue
= alloc_workqueue("xfsconvertd",
1815 if (!xfsconvertd_workqueue
)
1816 goto out_destroy_xfsdatad_workqueue
;
1820 out_destroy_xfsdatad_workqueue
:
1821 destroy_workqueue(xfsdatad_workqueue
);
1822 out_destroy_xfslogd_workqueue
:
1823 destroy_workqueue(xfslogd_workqueue
);
1825 kmem_zone_destroy(xfs_buf_zone
);
1831 xfs_buf_terminate(void)
1833 destroy_workqueue(xfsconvertd_workqueue
);
1834 destroy_workqueue(xfsdatad_workqueue
);
1835 destroy_workqueue(xfslogd_workqueue
);
1836 kmem_zone_destroy(xfs_buf_zone
);