2 * Copyright (c) 2000-2005 Silicon Graphics, Inc. All Rights Reserved.
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms of version 2 of the GNU General Public License as
6 * published by the Free Software Foundation.
8 * This program is distributed in the hope that it would be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
12 * Further, this software is distributed without any warranty that it is
13 * free of the rightful claim of any third person regarding infringement
14 * or the like. Any license provided herein, whether implied or
15 * otherwise, applies only to this software file. Patent licenses, if
16 * any, provided herein do not apply to combinations of this program with
17 * other software, or any other product whatsoever.
19 * You should have received a copy of the GNU General Public License along
20 * with this program; if not, write the Free Software Foundation, Inc., 59
21 * Temple Place - Suite 330, Boston MA 02111-1307, USA.
23 * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
24 * Mountain View, CA 94043, or:
28 * For further information regarding this notice, see:
30 * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
34 * The xfs_buf.c code provides an abstract buffer cache model on top
35 * of the Linux page cache. Cached metadata blocks for a file system
36 * are hashed to the inode for the block device. xfs_buf.c assembles
37 * buffers (xfs_buf_t) on demand to aggregate such cached pages for I/O.
39 * Written by Steve Lord, Jim Mostek, Russell Cattelan
40 * and Rajagopal Ananthanarayanan ("ananth") at SGI.
44 #include <linux/stddef.h>
45 #include <linux/errno.h>
46 #include <linux/slab.h>
47 #include <linux/pagemap.h>
48 #include <linux/init.h>
49 #include <linux/vmalloc.h>
50 #include <linux/bio.h>
51 #include <linux/sysctl.h>
52 #include <linux/proc_fs.h>
53 #include <linux/workqueue.h>
54 #include <linux/percpu.h>
55 #include <linux/blkdev.h>
56 #include <linux/hash.h>
57 #include <linux/kthread.h>
59 #include "xfs_linux.h"
65 STATIC kmem_cache_t
*pagebuf_zone
;
66 STATIC kmem_shaker_t pagebuf_shake
;
67 STATIC
int xfsbufd_wakeup(int, gfp_t
);
68 STATIC
void pagebuf_delwri_queue(xfs_buf_t
*, int);
70 STATIC
struct workqueue_struct
*xfslogd_workqueue
;
71 struct workqueue_struct
*xfsdatad_workqueue
;
85 ktrace_enter(pagebuf_trace_buf
,
87 (void *)(unsigned long)pb
->pb_flags
,
88 (void *)(unsigned long)pb
->pb_hold
.counter
,
89 (void *)(unsigned long)pb
->pb_sema
.count
.counter
,
92 (void *)(unsigned long)((pb
->pb_file_offset
>>32) & 0xffffffff),
93 (void *)(unsigned long)(pb
->pb_file_offset
& 0xffffffff),
94 (void *)(unsigned long)pb
->pb_buffer_length
,
95 NULL
, NULL
, NULL
, NULL
, NULL
);
97 ktrace_t
*pagebuf_trace_buf
;
98 #define PAGEBUF_TRACE_SIZE 4096
99 #define PB_TRACE(pb, id, data) \
100 pagebuf_trace(pb, id, (void *)data, (void *)__builtin_return_address(0))
102 #define PB_TRACE(pb, id, data) do { } while (0)
105 #ifdef PAGEBUF_LOCK_TRACKING
106 # define PB_SET_OWNER(pb) ((pb)->pb_last_holder = current->pid)
107 # define PB_CLEAR_OWNER(pb) ((pb)->pb_last_holder = -1)
108 # define PB_GET_OWNER(pb) ((pb)->pb_last_holder)
110 # define PB_SET_OWNER(pb) do { } while (0)
111 # define PB_CLEAR_OWNER(pb) do { } while (0)
112 # define PB_GET_OWNER(pb) do { } while (0)
116 * Pagebuf allocation / freeing.
119 #define pb_to_gfp(flags) \
120 ((((flags) & PBF_READ_AHEAD) ? __GFP_NORETRY : \
121 ((flags) & PBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
123 #define pb_to_km(flags) \
124 (((flags) & PBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
127 #define pagebuf_allocate(flags) \
128 kmem_zone_alloc(pagebuf_zone, pb_to_km(flags))
129 #define pagebuf_deallocate(pb) \
130 kmem_zone_free(pagebuf_zone, (pb));
133 * Page Region interfaces.
135 * For pages in filesystems where the blocksize is smaller than the
136 * pagesize, we use the page->private field (long) to hold a bitmap
137 * of uptodate regions within the page.
139 * Each such region is "bytes per page / bits per long" bytes long.
141 * NBPPR == number-of-bytes-per-page-region
142 * BTOPR == bytes-to-page-region (rounded up)
143 * BTOPRT == bytes-to-page-region-truncated (rounded down)
145 #if (BITS_PER_LONG == 32)
146 #define PRSHIFT (PAGE_CACHE_SHIFT - 5) /* (32 == 1<<5) */
147 #elif (BITS_PER_LONG == 64)
148 #define PRSHIFT (PAGE_CACHE_SHIFT - 6) /* (64 == 1<<6) */
150 #error BITS_PER_LONG must be 32 or 64
152 #define NBPPR (PAGE_CACHE_SIZE/BITS_PER_LONG)
153 #define BTOPR(b) (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
154 #define BTOPRT(b) (((unsigned int)(b) >> PRSHIFT))
164 first
= BTOPR(offset
);
165 final
= BTOPRT(offset
+ length
- 1);
166 first
= min(first
, final
);
169 mask
<<= BITS_PER_LONG
- (final
- first
);
170 mask
>>= BITS_PER_LONG
- (final
);
172 ASSERT(offset
+ length
<= PAGE_CACHE_SIZE
);
173 ASSERT((final
- first
) < BITS_PER_LONG
&& (final
- first
) >= 0);
184 page
->private |= page_region_mask(offset
, length
);
185 if (page
->private == ~0UL)
186 SetPageUptodate(page
);
195 unsigned long mask
= page_region_mask(offset
, length
);
197 return (mask
&& (page
->private & mask
) == mask
);
201 * Mapping of multi-page buffers into contiguous virtual space
204 typedef struct a_list
{
209 STATIC a_list_t
*as_free_head
;
210 STATIC
int as_list_len
;
211 STATIC
DEFINE_SPINLOCK(as_lock
);
214 * Try to batch vunmaps because they are costly.
222 aentry
= kmalloc(sizeof(a_list_t
), GFP_ATOMIC
& ~__GFP_HIGH
);
223 if (likely(aentry
)) {
225 aentry
->next
= as_free_head
;
226 aentry
->vm_addr
= addr
;
227 as_free_head
= aentry
;
229 spin_unlock(&as_lock
);
236 purge_addresses(void)
238 a_list_t
*aentry
, *old
;
240 if (as_free_head
== NULL
)
244 aentry
= as_free_head
;
247 spin_unlock(&as_lock
);
249 while ((old
= aentry
) != NULL
) {
250 vunmap(aentry
->vm_addr
);
251 aentry
= aentry
->next
;
257 * Internal pagebuf object manipulation
263 xfs_buftarg_t
*target
,
266 page_buf_flags_t flags
)
269 * We don't want certain flags to appear in pb->pb_flags.
271 flags
&= ~(PBF_LOCK
|PBF_MAPPED
|PBF_DONT_BLOCK
|PBF_READ_AHEAD
);
273 memset(pb
, 0, sizeof(xfs_buf_t
));
274 atomic_set(&pb
->pb_hold
, 1);
275 init_MUTEX_LOCKED(&pb
->pb_iodonesema
);
276 INIT_LIST_HEAD(&pb
->pb_list
);
277 INIT_LIST_HEAD(&pb
->pb_hash_list
);
278 init_MUTEX_LOCKED(&pb
->pb_sema
); /* held, no waiters */
280 pb
->pb_target
= target
;
281 pb
->pb_file_offset
= range_base
;
283 * Set buffer_length and count_desired to the same value initially.
284 * I/O routines should use count_desired, which will be the same in
285 * most cases but may be reset (e.g. XFS recovery).
287 pb
->pb_buffer_length
= pb
->pb_count_desired
= range_length
;
288 pb
->pb_flags
= flags
| PBF_NONE
;
289 pb
->pb_bn
= XFS_BUF_DADDR_NULL
;
290 atomic_set(&pb
->pb_pin_count
, 0);
291 init_waitqueue_head(&pb
->pb_waiters
);
293 XFS_STATS_INC(pb_create
);
294 PB_TRACE(pb
, "initialize", target
);
298 * Allocate a page array capable of holding a specified number
299 * of pages, and point the page buf at it.
305 page_buf_flags_t flags
)
307 /* Make sure that we have a page list */
308 if (pb
->pb_pages
== NULL
) {
309 pb
->pb_offset
= page_buf_poff(pb
->pb_file_offset
);
310 pb
->pb_page_count
= page_count
;
311 if (page_count
<= PB_PAGES
) {
312 pb
->pb_pages
= pb
->pb_page_array
;
314 pb
->pb_pages
= kmem_alloc(sizeof(struct page
*) *
315 page_count
, pb_to_km(flags
));
316 if (pb
->pb_pages
== NULL
)
319 memset(pb
->pb_pages
, 0, sizeof(struct page
*) * page_count
);
325 * Frees pb_pages if it was malloced.
331 if (bp
->pb_pages
!= bp
->pb_page_array
) {
332 kmem_free(bp
->pb_pages
,
333 bp
->pb_page_count
* sizeof(struct page
*));
338 * Releases the specified buffer.
340 * The modification state of any associated pages is left unchanged.
341 * The buffer most not be on any hash - use pagebuf_rele instead for
342 * hashed and refcounted buffers
348 PB_TRACE(bp
, "free", 0);
350 ASSERT(list_empty(&bp
->pb_hash_list
));
352 if (bp
->pb_flags
& _PBF_PAGE_CACHE
) {
355 if ((bp
->pb_flags
& PBF_MAPPED
) && (bp
->pb_page_count
> 1))
356 free_address(bp
->pb_addr
- bp
->pb_offset
);
358 for (i
= 0; i
< bp
->pb_page_count
; i
++)
359 page_cache_release(bp
->pb_pages
[i
]);
360 _pagebuf_free_pages(bp
);
361 } else if (bp
->pb_flags
& _PBF_KMEM_ALLOC
) {
363 * XXX(hch): bp->pb_count_desired might be incorrect (see
364 * pagebuf_associate_memory for details), but fortunately
365 * the Linux version of kmem_free ignores the len argument..
367 kmem_free(bp
->pb_addr
, bp
->pb_count_desired
);
368 _pagebuf_free_pages(bp
);
371 pagebuf_deallocate(bp
);
375 * Finds all pages for buffer in question and builds it's page list.
378 _pagebuf_lookup_pages(
382 struct address_space
*mapping
= bp
->pb_target
->pbr_mapping
;
383 size_t blocksize
= bp
->pb_target
->pbr_bsize
;
384 size_t size
= bp
->pb_count_desired
;
385 size_t nbytes
, offset
;
386 gfp_t gfp_mask
= pb_to_gfp(flags
);
387 unsigned short page_count
, i
;
392 end
= bp
->pb_file_offset
+ bp
->pb_buffer_length
;
393 page_count
= page_buf_btoc(end
) - page_buf_btoct(bp
->pb_file_offset
);
395 error
= _pagebuf_get_pages(bp
, page_count
, flags
);
398 bp
->pb_flags
|= _PBF_PAGE_CACHE
;
400 offset
= bp
->pb_offset
;
401 first
= bp
->pb_file_offset
>> PAGE_CACHE_SHIFT
;
403 for (i
= 0; i
< bp
->pb_page_count
; i
++) {
408 page
= find_or_create_page(mapping
, first
+ i
, gfp_mask
);
409 if (unlikely(page
== NULL
)) {
410 if (flags
& PBF_READ_AHEAD
) {
411 bp
->pb_page_count
= i
;
412 for (i
= 0; i
< bp
->pb_page_count
; i
++)
413 unlock_page(bp
->pb_pages
[i
]);
418 * This could deadlock.
420 * But until all the XFS lowlevel code is revamped to
421 * handle buffer allocation failures we can't do much.
423 if (!(++retries
% 100))
425 "XFS: possible memory allocation "
426 "deadlock in %s (mode:0x%x)\n",
427 __FUNCTION__
, gfp_mask
);
429 XFS_STATS_INC(pb_page_retries
);
430 xfsbufd_wakeup(0, gfp_mask
);
431 blk_congestion_wait(WRITE
, HZ
/50);
435 XFS_STATS_INC(pb_page_found
);
437 nbytes
= min_t(size_t, size
, PAGE_CACHE_SIZE
- offset
);
440 if (!PageUptodate(page
)) {
442 if (blocksize
>= PAGE_CACHE_SIZE
) {
443 if (flags
& PBF_READ
)
445 } else if (!PagePrivate(page
)) {
446 if (test_page_region(page
, offset
, nbytes
))
451 bp
->pb_pages
[i
] = page
;
455 if (!bp
->pb_locked
) {
456 for (i
= 0; i
< bp
->pb_page_count
; i
++)
457 unlock_page(bp
->pb_pages
[i
]);
461 /* if we have any uptodate pages, mark that in the buffer */
462 bp
->pb_flags
&= ~PBF_NONE
;
464 /* if some pages aren't uptodate, mark that in the buffer */
465 if (page_count
!= bp
->pb_page_count
)
466 bp
->pb_flags
|= PBF_PARTIAL
;
469 PB_TRACE(bp
, "lookup_pages", (long)page_count
);
474 * Map buffer into kernel address-space if nessecary.
481 /* A single page buffer is always mappable */
482 if (bp
->pb_page_count
== 1) {
483 bp
->pb_addr
= page_address(bp
->pb_pages
[0]) + bp
->pb_offset
;
484 bp
->pb_flags
|= PBF_MAPPED
;
485 } else if (flags
& PBF_MAPPED
) {
486 if (as_list_len
> 64)
488 bp
->pb_addr
= vmap(bp
->pb_pages
, bp
->pb_page_count
,
489 VM_MAP
, PAGE_KERNEL
);
490 if (unlikely(bp
->pb_addr
== NULL
))
492 bp
->pb_addr
+= bp
->pb_offset
;
493 bp
->pb_flags
|= PBF_MAPPED
;
500 * Finding and Reading Buffers
506 * Looks up, and creates if absent, a lockable buffer for
507 * a given range of an inode. The buffer is returned
508 * locked. If other overlapping buffers exist, they are
509 * released before the new buffer is created and locked,
510 * which may imply that this call will block until those buffers
511 * are unlocked. No I/O is implied by this call.
515 xfs_buftarg_t
*btp
, /* block device target */
516 loff_t ioff
, /* starting offset of range */
517 size_t isize
, /* length of range */
518 page_buf_flags_t flags
, /* PBF_TRYLOCK */
519 xfs_buf_t
*new_pb
)/* newly allocated buffer */
526 range_base
= (ioff
<< BBSHIFT
);
527 range_length
= (isize
<< BBSHIFT
);
529 /* Check for IOs smaller than the sector size / not sector aligned */
530 ASSERT(!(range_length
< (1 << btp
->pbr_sshift
)));
531 ASSERT(!(range_base
& (loff_t
)btp
->pbr_smask
));
533 hash
= &btp
->bt_hash
[hash_long((unsigned long)ioff
, btp
->bt_hashshift
)];
535 spin_lock(&hash
->bh_lock
);
537 list_for_each_entry_safe(pb
, n
, &hash
->bh_list
, pb_hash_list
) {
538 ASSERT(btp
== pb
->pb_target
);
539 if (pb
->pb_file_offset
== range_base
&&
540 pb
->pb_buffer_length
== range_length
) {
542 * If we look at something bring it to the
543 * front of the list for next time.
545 atomic_inc(&pb
->pb_hold
);
546 list_move(&pb
->pb_hash_list
, &hash
->bh_list
);
553 _pagebuf_initialize(new_pb
, btp
, range_base
,
554 range_length
, flags
);
555 new_pb
->pb_hash
= hash
;
556 list_add(&new_pb
->pb_hash_list
, &hash
->bh_list
);
558 XFS_STATS_INC(pb_miss_locked
);
561 spin_unlock(&hash
->bh_lock
);
565 spin_unlock(&hash
->bh_lock
);
567 /* Attempt to get the semaphore without sleeping,
568 * if this does not work then we need to drop the
569 * spinlock and do a hard attempt on the semaphore.
571 if (down_trylock(&pb
->pb_sema
)) {
572 if (!(flags
& PBF_TRYLOCK
)) {
573 /* wait for buffer ownership */
574 PB_TRACE(pb
, "get_lock", 0);
576 XFS_STATS_INC(pb_get_locked_waited
);
578 /* We asked for a trylock and failed, no need
579 * to look at file offset and length here, we
580 * know that this pagebuf at least overlaps our
581 * pagebuf and is locked, therefore our buffer
582 * either does not exist, or is this buffer
586 XFS_STATS_INC(pb_busy_locked
);
594 if (pb
->pb_flags
& PBF_STALE
) {
595 ASSERT((pb
->pb_flags
& _PBF_DELWRI_Q
) == 0);
596 pb
->pb_flags
&= PBF_MAPPED
;
598 PB_TRACE(pb
, "got_lock", 0);
599 XFS_STATS_INC(pb_get_locked
);
604 * xfs_buf_get_flags assembles a buffer covering the specified range.
606 * Storage in memory for all portions of the buffer will be allocated,
607 * although backing storage may not be.
610 xfs_buf_get_flags( /* allocate a buffer */
611 xfs_buftarg_t
*target
,/* target for buffer */
612 loff_t ioff
, /* starting offset of range */
613 size_t isize
, /* length of range */
614 page_buf_flags_t flags
) /* PBF_TRYLOCK */
616 xfs_buf_t
*pb
, *new_pb
;
619 new_pb
= pagebuf_allocate(flags
);
620 if (unlikely(!new_pb
))
623 pb
= _pagebuf_find(target
, ioff
, isize
, flags
, new_pb
);
625 error
= _pagebuf_lookup_pages(pb
, flags
);
629 pagebuf_deallocate(new_pb
);
630 if (unlikely(pb
== NULL
))
634 for (i
= 0; i
< pb
->pb_page_count
; i
++)
635 mark_page_accessed(pb
->pb_pages
[i
]);
637 if (!(pb
->pb_flags
& PBF_MAPPED
)) {
638 error
= _pagebuf_map_pages(pb
, flags
);
639 if (unlikely(error
)) {
640 printk(KERN_WARNING
"%s: failed to map pages\n",
646 XFS_STATS_INC(pb_get
);
649 * Always fill in the block number now, the mapped cases can do
650 * their own overlay of this later.
653 pb
->pb_count_desired
= pb
->pb_buffer_length
;
655 PB_TRACE(pb
, "get", (unsigned long)flags
);
659 if (flags
& (PBF_LOCK
| PBF_TRYLOCK
))
667 xfs_buftarg_t
*target
,
670 page_buf_flags_t flags
)
676 pb
= xfs_buf_get_flags(target
, ioff
, isize
, flags
);
678 if (PBF_NOT_DONE(pb
)) {
679 PB_TRACE(pb
, "read", (unsigned long)flags
);
680 XFS_STATS_INC(pb_get_read
);
681 pagebuf_iostart(pb
, flags
);
682 } else if (flags
& PBF_ASYNC
) {
683 PB_TRACE(pb
, "read_async", (unsigned long)flags
);
685 * Read ahead call which is already satisfied,
690 PB_TRACE(pb
, "read_done", (unsigned long)flags
);
691 /* We do not want read in the flags */
692 pb
->pb_flags
&= ~PBF_READ
;
699 if (flags
& (PBF_LOCK
| PBF_TRYLOCK
))
706 * If we are not low on memory then do the readahead in a deadlock
711 xfs_buftarg_t
*target
,
714 page_buf_flags_t flags
)
716 struct backing_dev_info
*bdi
;
718 bdi
= target
->pbr_mapping
->backing_dev_info
;
719 if (bdi_read_congested(bdi
))
722 flags
|= (PBF_TRYLOCK
|PBF_ASYNC
|PBF_READ_AHEAD
);
723 xfs_buf_read_flags(target
, ioff
, isize
, flags
);
729 xfs_buftarg_t
*target
)
733 pb
= pagebuf_allocate(0);
735 _pagebuf_initialize(pb
, target
, 0, len
, 0);
739 static inline struct page
*
743 if (((unsigned long)addr
< VMALLOC_START
) ||
744 ((unsigned long)addr
>= VMALLOC_END
)) {
745 return virt_to_page(addr
);
747 return vmalloc_to_page(addr
);
752 pagebuf_associate_memory(
764 page_count
= PAGE_CACHE_ALIGN(len
) >> PAGE_CACHE_SHIFT
;
765 offset
= (off_t
) mem
- ((off_t
)mem
& PAGE_CACHE_MASK
);
766 if (offset
&& (len
> PAGE_CACHE_SIZE
))
769 /* Free any previous set of page pointers */
771 _pagebuf_free_pages(pb
);
776 rval
= _pagebuf_get_pages(pb
, page_count
, 0);
780 pb
->pb_offset
= offset
;
781 ptr
= (size_t) mem
& PAGE_CACHE_MASK
;
782 end
= PAGE_CACHE_ALIGN((size_t) mem
+ len
);
784 /* set up first page */
785 pb
->pb_pages
[0] = mem_to_page(mem
);
787 ptr
+= PAGE_CACHE_SIZE
;
788 pb
->pb_page_count
= ++i
;
790 pb
->pb_pages
[i
] = mem_to_page((void *)ptr
);
791 pb
->pb_page_count
= ++i
;
792 ptr
+= PAGE_CACHE_SIZE
;
796 pb
->pb_count_desired
= pb
->pb_buffer_length
= len
;
797 pb
->pb_flags
|= PBF_MAPPED
;
803 pagebuf_get_no_daddr(
805 xfs_buftarg_t
*target
)
807 size_t malloc_len
= len
;
812 bp
= pagebuf_allocate(0);
813 if (unlikely(bp
== NULL
))
815 _pagebuf_initialize(bp
, target
, 0, len
, PBF_FORCEIO
);
818 data
= kmem_alloc(malloc_len
, KM_SLEEP
| KM_MAYFAIL
);
819 if (unlikely(data
== NULL
))
822 /* check whether alignment matches.. */
823 if ((__psunsigned_t
)data
!=
824 ((__psunsigned_t
)data
& ~target
->pbr_smask
)) {
825 /* .. else double the size and try again */
826 kmem_free(data
, malloc_len
);
831 error
= pagebuf_associate_memory(bp
, data
, len
);
834 bp
->pb_flags
|= _PBF_KMEM_ALLOC
;
838 PB_TRACE(bp
, "no_daddr", data
);
841 kmem_free(data
, malloc_len
);
851 * Increment reference count on buffer, to hold the buffer concurrently
852 * with another thread which may release (free) the buffer asynchronously.
854 * Must hold the buffer already to call this function.
860 atomic_inc(&pb
->pb_hold
);
861 PB_TRACE(pb
, "hold", 0);
867 * pagebuf_rele releases a hold on the specified buffer. If the
868 * the hold count is 1, pagebuf_rele calls pagebuf_free.
874 xfs_bufhash_t
*hash
= pb
->pb_hash
;
876 PB_TRACE(pb
, "rele", pb
->pb_relse
);
879 * pagebuf_lookup buffers are not hashed, not delayed write,
880 * and don't have their own release routines. Special case.
882 if (unlikely(!hash
)) {
883 ASSERT(!pb
->pb_relse
);
884 if (atomic_dec_and_test(&pb
->pb_hold
))
889 if (atomic_dec_and_lock(&pb
->pb_hold
, &hash
->bh_lock
)) {
893 atomic_inc(&pb
->pb_hold
);
894 spin_unlock(&hash
->bh_lock
);
895 (*(pb
->pb_relse
)) (pb
);
896 spin_lock(&hash
->bh_lock
);
900 if (pb
->pb_flags
& PBF_FS_MANAGED
) {
905 ASSERT((pb
->pb_flags
& (PBF_DELWRI
|_PBF_DELWRI_Q
)) == 0);
906 list_del_init(&pb
->pb_hash_list
);
907 spin_unlock(&hash
->bh_lock
);
910 spin_unlock(&hash
->bh_lock
);
914 * Catch reference count leaks
916 ASSERT(atomic_read(&pb
->pb_hold
) >= 0);
922 * Mutual exclusion on buffers. Locking model:
924 * Buffers associated with inodes for which buffer locking
925 * is not enabled are not protected by semaphores, and are
926 * assumed to be exclusively owned by the caller. There is a
927 * spinlock in the buffer, used by the caller when concurrent
928 * access is possible.
934 * pagebuf_cond_lock locks a buffer object, if it is not already locked.
935 * Note that this in no way
936 * locks the underlying pages, so it is only useful for synchronizing
937 * concurrent use of page buffer objects, not for synchronizing independent
938 * access to the underlying pages.
941 pagebuf_cond_lock( /* lock buffer, if not locked */
942 /* returns -EBUSY if locked) */
947 locked
= down_trylock(&pb
->pb_sema
) == 0;
951 PB_TRACE(pb
, "cond_lock", (long)locked
);
952 return(locked
? 0 : -EBUSY
);
955 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
959 * Return lock value for a pagebuf
965 return(atomic_read(&pb
->pb_sema
.count
));
972 * pagebuf_lock locks a buffer object. Note that this in no way
973 * locks the underlying pages, so it is only useful for synchronizing
974 * concurrent use of page buffer objects, not for synchronizing independent
975 * access to the underlying pages.
981 PB_TRACE(pb
, "lock", 0);
982 if (atomic_read(&pb
->pb_io_remaining
))
983 blk_run_address_space(pb
->pb_target
->pbr_mapping
);
986 PB_TRACE(pb
, "locked", 0);
993 * pagebuf_unlock releases the lock on the buffer object created by
994 * pagebuf_lock or pagebuf_cond_lock (not any pinning of underlying pages
995 * created by pagebuf_pin).
997 * If the buffer is marked delwri but is not queued, do so before we
998 * unlock the buffer as we need to set flags correctly. We also need to
999 * take a reference for the delwri queue because the unlocker is going to
1000 * drop their's and they don't know we just queued it.
1003 pagebuf_unlock( /* unlock buffer */
1004 xfs_buf_t
*pb
) /* buffer to unlock */
1006 if ((pb
->pb_flags
& (PBF_DELWRI
|_PBF_DELWRI_Q
)) == PBF_DELWRI
) {
1007 atomic_inc(&pb
->pb_hold
);
1008 pb
->pb_flags
|= PBF_ASYNC
;
1009 pagebuf_delwri_queue(pb
, 0);
1014 PB_TRACE(pb
, "unlock", 0);
1019 * Pinning Buffer Storage in Memory
1025 * pagebuf_pin locks all of the memory represented by a buffer in
1026 * memory. Multiple calls to pagebuf_pin and pagebuf_unpin, for
1027 * the same or different buffers affecting a given page, will
1028 * properly count the number of outstanding "pin" requests. The
1029 * buffer may be released after the pagebuf_pin and a different
1030 * buffer used when calling pagebuf_unpin, if desired.
1031 * pagebuf_pin should be used by the file system when it wants be
1032 * assured that no attempt will be made to force the affected
1033 * memory to disk. It does not assure that a given logical page
1034 * will not be moved to a different physical page.
1040 atomic_inc(&pb
->pb_pin_count
);
1041 PB_TRACE(pb
, "pin", (long)pb
->pb_pin_count
.counter
);
1047 * pagebuf_unpin reverses the locking of memory performed by
1048 * pagebuf_pin. Note that both functions affected the logical
1049 * pages associated with the buffer, not the buffer itself.
1055 if (atomic_dec_and_test(&pb
->pb_pin_count
)) {
1056 wake_up_all(&pb
->pb_waiters
);
1058 PB_TRACE(pb
, "unpin", (long)pb
->pb_pin_count
.counter
);
1065 return atomic_read(&pb
->pb_pin_count
);
1069 * pagebuf_wait_unpin
1071 * pagebuf_wait_unpin waits until all of the memory associated
1072 * with the buffer is not longer locked in memory. It returns
1073 * immediately if none of the affected pages are locked.
1076 _pagebuf_wait_unpin(
1079 DECLARE_WAITQUEUE (wait
, current
);
1081 if (atomic_read(&pb
->pb_pin_count
) == 0)
1084 add_wait_queue(&pb
->pb_waiters
, &wait
);
1086 set_current_state(TASK_UNINTERRUPTIBLE
);
1087 if (atomic_read(&pb
->pb_pin_count
) == 0)
1089 if (atomic_read(&pb
->pb_io_remaining
))
1090 blk_run_address_space(pb
->pb_target
->pbr_mapping
);
1093 remove_wait_queue(&pb
->pb_waiters
, &wait
);
1094 set_current_state(TASK_RUNNING
);
1098 * Buffer Utility Routines
1104 * pagebuf_iodone marks a buffer for which I/O is in progress
1105 * done with respect to that I/O. The pb_iodone routine, if
1106 * present, will be called as a side-effect.
1109 pagebuf_iodone_work(
1112 xfs_buf_t
*bp
= (xfs_buf_t
*)v
;
1115 (*(bp
->pb_iodone
))(bp
);
1116 else if (bp
->pb_flags
& PBF_ASYNC
)
1126 pb
->pb_flags
&= ~(PBF_READ
| PBF_WRITE
);
1127 if (pb
->pb_error
== 0) {
1128 pb
->pb_flags
&= ~(PBF_PARTIAL
| PBF_NONE
);
1131 PB_TRACE(pb
, "iodone", pb
->pb_iodone
);
1133 if ((pb
->pb_iodone
) || (pb
->pb_flags
& PBF_ASYNC
)) {
1135 INIT_WORK(&pb
->pb_iodone_work
, pagebuf_iodone_work
, pb
);
1136 queue_work(dataio
? xfsdatad_workqueue
:
1137 xfslogd_workqueue
, &pb
->pb_iodone_work
);
1139 pagebuf_iodone_work(pb
);
1142 up(&pb
->pb_iodonesema
);
1149 * pagebuf_ioerror sets the error code for a buffer.
1152 pagebuf_ioerror( /* mark/clear buffer error flag */
1153 xfs_buf_t
*pb
, /* buffer to mark */
1154 int error
) /* error to store (0 if none) */
1156 ASSERT(error
>= 0 && error
<= 0xffff);
1157 pb
->pb_error
= (unsigned short)error
;
1158 PB_TRACE(pb
, "ioerror", (unsigned long)error
);
1164 * pagebuf_iostart initiates I/O on a buffer, based on the flags supplied.
1165 * If necessary, it will arrange for any disk space allocation required,
1166 * and it will break up the request if the block mappings require it.
1167 * The pb_iodone routine in the buffer supplied will only be called
1168 * when all of the subsidiary I/O requests, if any, have been completed.
1169 * pagebuf_iostart calls the pagebuf_ioinitiate routine or
1170 * pagebuf_iorequest, if the former routine is not defined, to start
1171 * the I/O on a given low-level request.
1174 pagebuf_iostart( /* start I/O on a buffer */
1175 xfs_buf_t
*pb
, /* buffer to start */
1176 page_buf_flags_t flags
) /* PBF_LOCK, PBF_ASYNC, PBF_READ, */
1177 /* PBF_WRITE, PBF_DELWRI, */
1178 /* PBF_DONT_BLOCK */
1182 PB_TRACE(pb
, "iostart", (unsigned long)flags
);
1184 if (flags
& PBF_DELWRI
) {
1185 pb
->pb_flags
&= ~(PBF_READ
| PBF_WRITE
| PBF_ASYNC
);
1186 pb
->pb_flags
|= flags
& (PBF_DELWRI
| PBF_ASYNC
);
1187 pagebuf_delwri_queue(pb
, 1);
1191 pb
->pb_flags
&= ~(PBF_READ
| PBF_WRITE
| PBF_ASYNC
| PBF_DELWRI
| \
1192 PBF_READ_AHEAD
| _PBF_RUN_QUEUES
);
1193 pb
->pb_flags
|= flags
& (PBF_READ
| PBF_WRITE
| PBF_ASYNC
| \
1194 PBF_READ_AHEAD
| _PBF_RUN_QUEUES
);
1196 BUG_ON(pb
->pb_bn
== XFS_BUF_DADDR_NULL
);
1198 /* For writes allow an alternate strategy routine to precede
1199 * the actual I/O request (which may not be issued at all in
1200 * a shutdown situation, for example).
1202 status
= (flags
& PBF_WRITE
) ?
1203 pagebuf_iostrategy(pb
) : pagebuf_iorequest(pb
);
1205 /* Wait for I/O if we are not an async request.
1206 * Note: async I/O request completion will release the buffer,
1207 * and that can already be done by this point. So using the
1208 * buffer pointer from here on, after async I/O, is invalid.
1210 if (!status
&& !(flags
& PBF_ASYNC
))
1211 status
= pagebuf_iowait(pb
);
1217 * Helper routine for pagebuf_iorequest
1220 STATIC __inline__
int
1224 ASSERT(pb
->pb_flags
& (PBF_READ
|PBF_WRITE
));
1225 if (pb
->pb_flags
& PBF_READ
)
1226 return pb
->pb_locked
;
1230 STATIC __inline__
void
1235 if (atomic_dec_and_test(&pb
->pb_io_remaining
) == 1) {
1237 pagebuf_iodone(pb
, (pb
->pb_flags
& PBF_FS_DATAIOD
), schedule
);
1244 unsigned int bytes_done
,
1247 xfs_buf_t
*pb
= (xfs_buf_t
*)bio
->bi_private
;
1248 unsigned int blocksize
= pb
->pb_target
->pbr_bsize
;
1249 struct bio_vec
*bvec
= bio
->bi_io_vec
+ bio
->bi_vcnt
- 1;
1254 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1258 struct page
*page
= bvec
->bv_page
;
1260 if (unlikely(pb
->pb_error
)) {
1261 if (pb
->pb_flags
& PBF_READ
)
1262 ClearPageUptodate(page
);
1264 } else if (blocksize
== PAGE_CACHE_SIZE
) {
1265 SetPageUptodate(page
);
1266 } else if (!PagePrivate(page
) &&
1267 (pb
->pb_flags
& _PBF_PAGE_CACHE
)) {
1268 set_page_region(page
, bvec
->bv_offset
, bvec
->bv_len
);
1271 if (--bvec
>= bio
->bi_io_vec
)
1272 prefetchw(&bvec
->bv_page
->flags
);
1274 if (_pagebuf_iolocked(pb
)) {
1277 } while (bvec
>= bio
->bi_io_vec
);
1279 _pagebuf_iodone(pb
, 1);
1288 int i
, rw
, map_i
, total_nr_pages
, nr_pages
;
1290 int offset
= pb
->pb_offset
;
1291 int size
= pb
->pb_count_desired
;
1292 sector_t sector
= pb
->pb_bn
;
1293 unsigned int blocksize
= pb
->pb_target
->pbr_bsize
;
1294 int locking
= _pagebuf_iolocked(pb
);
1296 total_nr_pages
= pb
->pb_page_count
;
1299 if (pb
->pb_flags
& _PBF_RUN_QUEUES
) {
1300 pb
->pb_flags
&= ~_PBF_RUN_QUEUES
;
1301 rw
= (pb
->pb_flags
& PBF_READ
) ? READ_SYNC
: WRITE_SYNC
;
1303 rw
= (pb
->pb_flags
& PBF_READ
) ? READ
: WRITE
;
1306 /* Special code path for reading a sub page size pagebuf in --
1307 * we populate up the whole page, and hence the other metadata
1308 * in the same page. This optimization is only valid when the
1309 * filesystem block size and the page size are equal.
1311 if ((pb
->pb_buffer_length
< PAGE_CACHE_SIZE
) &&
1312 (pb
->pb_flags
& PBF_READ
) && locking
&&
1313 (blocksize
== PAGE_CACHE_SIZE
)) {
1314 bio
= bio_alloc(GFP_NOIO
, 1);
1316 bio
->bi_bdev
= pb
->pb_target
->pbr_bdev
;
1317 bio
->bi_sector
= sector
- (offset
>> BBSHIFT
);
1318 bio
->bi_end_io
= bio_end_io_pagebuf
;
1319 bio
->bi_private
= pb
;
1321 bio_add_page(bio
, pb
->pb_pages
[0], PAGE_CACHE_SIZE
, 0);
1324 atomic_inc(&pb
->pb_io_remaining
);
1329 /* Lock down the pages which we need to for the request */
1330 if (locking
&& (pb
->pb_flags
& PBF_WRITE
) && (pb
->pb_locked
== 0)) {
1331 for (i
= 0; size
; i
++) {
1332 int nbytes
= PAGE_CACHE_SIZE
- offset
;
1333 struct page
*page
= pb
->pb_pages
[i
];
1343 offset
= pb
->pb_offset
;
1344 size
= pb
->pb_count_desired
;
1348 atomic_inc(&pb
->pb_io_remaining
);
1349 nr_pages
= BIO_MAX_SECTORS
>> (PAGE_SHIFT
- BBSHIFT
);
1350 if (nr_pages
> total_nr_pages
)
1351 nr_pages
= total_nr_pages
;
1353 bio
= bio_alloc(GFP_NOIO
, nr_pages
);
1354 bio
->bi_bdev
= pb
->pb_target
->pbr_bdev
;
1355 bio
->bi_sector
= sector
;
1356 bio
->bi_end_io
= bio_end_io_pagebuf
;
1357 bio
->bi_private
= pb
;
1359 for (; size
&& nr_pages
; nr_pages
--, map_i
++) {
1360 int nbytes
= PAGE_CACHE_SIZE
- offset
;
1365 if (bio_add_page(bio
, pb
->pb_pages
[map_i
],
1366 nbytes
, offset
) < nbytes
)
1370 sector
+= nbytes
>> BBSHIFT
;
1376 if (likely(bio
->bi_size
)) {
1377 submit_bio(rw
, bio
);
1382 pagebuf_ioerror(pb
, EIO
);
1387 * pagebuf_iorequest -- the core I/O request routine.
1390 pagebuf_iorequest( /* start real I/O */
1391 xfs_buf_t
*pb
) /* buffer to convey to device */
1393 PB_TRACE(pb
, "iorequest", 0);
1395 if (pb
->pb_flags
& PBF_DELWRI
) {
1396 pagebuf_delwri_queue(pb
, 1);
1400 if (pb
->pb_flags
& PBF_WRITE
) {
1401 _pagebuf_wait_unpin(pb
);
1406 /* Set the count to 1 initially, this will stop an I/O
1407 * completion callout which happens before we have started
1408 * all the I/O from calling pagebuf_iodone too early.
1410 atomic_set(&pb
->pb_io_remaining
, 1);
1411 _pagebuf_ioapply(pb
);
1412 _pagebuf_iodone(pb
, 0);
1421 * pagebuf_iowait waits for I/O to complete on the buffer supplied.
1422 * It returns immediately if no I/O is pending. In any case, it returns
1423 * the error code, if any, or 0 if there is no error.
1429 PB_TRACE(pb
, "iowait", 0);
1430 if (atomic_read(&pb
->pb_io_remaining
))
1431 blk_run_address_space(pb
->pb_target
->pbr_mapping
);
1432 down(&pb
->pb_iodonesema
);
1433 PB_TRACE(pb
, "iowaited", (long)pb
->pb_error
);
1434 return pb
->pb_error
;
1444 offset
+= pb
->pb_offset
;
1446 page
= pb
->pb_pages
[offset
>> PAGE_CACHE_SHIFT
];
1447 return (caddr_t
) page_address(page
) + (offset
& (PAGE_CACHE_SIZE
- 1));
1453 * Move data into or out of a buffer.
1457 xfs_buf_t
*pb
, /* buffer to process */
1458 size_t boff
, /* starting buffer offset */
1459 size_t bsize
, /* length to copy */
1460 caddr_t data
, /* data address */
1461 page_buf_rw_t mode
) /* read/write flag */
1463 size_t bend
, cpoff
, csize
;
1466 bend
= boff
+ bsize
;
1467 while (boff
< bend
) {
1468 page
= pb
->pb_pages
[page_buf_btoct(boff
+ pb
->pb_offset
)];
1469 cpoff
= page_buf_poff(boff
+ pb
->pb_offset
);
1470 csize
= min_t(size_t,
1471 PAGE_CACHE_SIZE
-cpoff
, pb
->pb_count_desired
-boff
);
1473 ASSERT(((csize
+ cpoff
) <= PAGE_CACHE_SIZE
));
1477 memset(page_address(page
) + cpoff
, 0, csize
);
1480 memcpy(data
, page_address(page
) + cpoff
, csize
);
1483 memcpy(page_address(page
) + cpoff
, data
, csize
);
1492 * Handling of buftargs.
1496 * Wait for any bufs with callbacks that have been submitted but
1497 * have not yet returned... walk the hash list for the target.
1504 xfs_bufhash_t
*hash
;
1507 for (i
= 0; i
< (1 << btp
->bt_hashshift
); i
++) {
1508 hash
= &btp
->bt_hash
[i
];
1510 spin_lock(&hash
->bh_lock
);
1511 list_for_each_entry_safe(bp
, n
, &hash
->bh_list
, pb_hash_list
) {
1512 ASSERT(btp
== bp
->pb_target
);
1513 if (!(bp
->pb_flags
& PBF_FS_MANAGED
)) {
1514 spin_unlock(&hash
->bh_lock
);
1516 * Catch superblock reference count leaks
1519 BUG_ON(bp
->pb_bn
== 0);
1524 spin_unlock(&hash
->bh_lock
);
1529 * Allocate buffer hash table for a given target.
1530 * For devices containing metadata (i.e. not the log/realtime devices)
1531 * we need to allocate a much larger hash table.
1540 btp
->bt_hashshift
= external
? 3 : 8; /* 8 or 256 buckets */
1541 btp
->bt_hashmask
= (1 << btp
->bt_hashshift
) - 1;
1542 btp
->bt_hash
= kmem_zalloc((1 << btp
->bt_hashshift
) *
1543 sizeof(xfs_bufhash_t
), KM_SLEEP
);
1544 for (i
= 0; i
< (1 << btp
->bt_hashshift
); i
++) {
1545 spin_lock_init(&btp
->bt_hash
[i
].bh_lock
);
1546 INIT_LIST_HEAD(&btp
->bt_hash
[i
].bh_list
);
1554 kmem_free(btp
->bt_hash
,
1555 (1 << btp
->bt_hashshift
) * sizeof(xfs_bufhash_t
));
1556 btp
->bt_hash
= NULL
;
1564 xfs_flush_buftarg(btp
, 1);
1566 xfs_blkdev_put(btp
->pbr_bdev
);
1567 xfs_free_bufhash(btp
);
1568 iput(btp
->pbr_mapping
->host
);
1569 kmem_free(btp
, sizeof(*btp
));
1573 xfs_setsize_buftarg_flags(
1575 unsigned int blocksize
,
1576 unsigned int sectorsize
,
1579 btp
->pbr_bsize
= blocksize
;
1580 btp
->pbr_sshift
= ffs(sectorsize
) - 1;
1581 btp
->pbr_smask
= sectorsize
- 1;
1583 if (set_blocksize(btp
->pbr_bdev
, sectorsize
)) {
1585 "XFS: Cannot set_blocksize to %u on device %s\n",
1586 sectorsize
, XFS_BUFTARG_NAME(btp
));
1591 (PAGE_CACHE_SIZE
/ BITS_PER_LONG
) > sectorsize
) {
1593 "XFS: %u byte sectors in use on device %s. "
1594 "This is suboptimal; %u or greater is ideal.\n",
1595 sectorsize
, XFS_BUFTARG_NAME(btp
),
1596 (unsigned int)PAGE_CACHE_SIZE
/ BITS_PER_LONG
);
1603 * When allocating the initial buffer target we have not yet
1604 * read in the superblock, so don't know what sized sectors
1605 * are being used is at this early stage. Play safe.
1608 xfs_setsize_buftarg_early(
1610 struct block_device
*bdev
)
1612 return xfs_setsize_buftarg_flags(btp
,
1613 PAGE_CACHE_SIZE
, bdev_hardsect_size(bdev
), 0);
1617 xfs_setsize_buftarg(
1619 unsigned int blocksize
,
1620 unsigned int sectorsize
)
1622 return xfs_setsize_buftarg_flags(btp
, blocksize
, sectorsize
, 1);
1626 xfs_mapping_buftarg(
1628 struct block_device
*bdev
)
1630 struct backing_dev_info
*bdi
;
1631 struct inode
*inode
;
1632 struct address_space
*mapping
;
1633 static struct address_space_operations mapping_aops
= {
1634 .sync_page
= block_sync_page
,
1637 inode
= new_inode(bdev
->bd_inode
->i_sb
);
1640 "XFS: Cannot allocate mapping inode for device %s\n",
1641 XFS_BUFTARG_NAME(btp
));
1644 inode
->i_mode
= S_IFBLK
;
1645 inode
->i_bdev
= bdev
;
1646 inode
->i_rdev
= bdev
->bd_dev
;
1647 bdi
= blk_get_backing_dev_info(bdev
);
1649 bdi
= &default_backing_dev_info
;
1650 mapping
= &inode
->i_data
;
1651 mapping
->a_ops
= &mapping_aops
;
1652 mapping
->backing_dev_info
= bdi
;
1653 mapping_set_gfp_mask(mapping
, GFP_NOFS
);
1654 btp
->pbr_mapping
= mapping
;
1660 struct block_device
*bdev
,
1665 btp
= kmem_zalloc(sizeof(*btp
), KM_SLEEP
);
1667 btp
->pbr_dev
= bdev
->bd_dev
;
1668 btp
->pbr_bdev
= bdev
;
1669 if (xfs_setsize_buftarg_early(btp
, bdev
))
1671 if (xfs_mapping_buftarg(btp
, bdev
))
1673 xfs_alloc_bufhash(btp
, external
);
1677 kmem_free(btp
, sizeof(*btp
));
1683 * Pagebuf delayed write buffer handling
1686 STATIC
LIST_HEAD(pbd_delwrite_queue
);
1687 STATIC
DEFINE_SPINLOCK(pbd_delwrite_lock
);
1690 pagebuf_delwri_queue(
1694 PB_TRACE(pb
, "delwri_q", (long)unlock
);
1695 ASSERT((pb
->pb_flags
& (PBF_DELWRI
|PBF_ASYNC
)) ==
1696 (PBF_DELWRI
|PBF_ASYNC
));
1698 spin_lock(&pbd_delwrite_lock
);
1699 /* If already in the queue, dequeue and place at tail */
1700 if (!list_empty(&pb
->pb_list
)) {
1701 ASSERT(pb
->pb_flags
& _PBF_DELWRI_Q
);
1703 atomic_dec(&pb
->pb_hold
);
1705 list_del(&pb
->pb_list
);
1708 pb
->pb_flags
|= _PBF_DELWRI_Q
;
1709 list_add_tail(&pb
->pb_list
, &pbd_delwrite_queue
);
1710 pb
->pb_queuetime
= jiffies
;
1711 spin_unlock(&pbd_delwrite_lock
);
1718 pagebuf_delwri_dequeue(
1723 spin_lock(&pbd_delwrite_lock
);
1724 if ((pb
->pb_flags
& PBF_DELWRI
) && !list_empty(&pb
->pb_list
)) {
1725 ASSERT(pb
->pb_flags
& _PBF_DELWRI_Q
);
1726 list_del_init(&pb
->pb_list
);
1729 pb
->pb_flags
&= ~(PBF_DELWRI
|_PBF_DELWRI_Q
);
1730 spin_unlock(&pbd_delwrite_lock
);
1735 PB_TRACE(pb
, "delwri_dq", (long)dequeued
);
1739 pagebuf_runall_queues(
1740 struct workqueue_struct
*queue
)
1742 flush_workqueue(queue
);
1745 /* Defines for pagebuf daemon */
1746 STATIC
struct task_struct
*xfsbufd_task
;
1747 STATIC
int xfsbufd_force_flush
;
1748 STATIC
int xfsbufd_force_sleep
;
1755 if (xfsbufd_force_sleep
)
1757 xfsbufd_force_flush
= 1;
1759 wake_up_process(xfsbufd_task
);
1767 struct list_head tmp
;
1769 xfs_buftarg_t
*target
;
1772 current
->flags
|= PF_MEMALLOC
;
1774 INIT_LIST_HEAD(&tmp
);
1776 if (unlikely(freezing(current
))) {
1777 xfsbufd_force_sleep
= 1;
1780 xfsbufd_force_sleep
= 0;
1783 schedule_timeout_interruptible
1784 (xfs_buf_timer_centisecs
* msecs_to_jiffies(10));
1786 age
= xfs_buf_age_centisecs
* msecs_to_jiffies(10);
1787 spin_lock(&pbd_delwrite_lock
);
1788 list_for_each_entry_safe(pb
, n
, &pbd_delwrite_queue
, pb_list
) {
1789 PB_TRACE(pb
, "walkq1", (long)pagebuf_ispin(pb
));
1790 ASSERT(pb
->pb_flags
& PBF_DELWRI
);
1792 if (!pagebuf_ispin(pb
) && !pagebuf_cond_lock(pb
)) {
1793 if (!xfsbufd_force_flush
&&
1794 time_before(jiffies
,
1795 pb
->pb_queuetime
+ age
)) {
1800 pb
->pb_flags
&= ~(PBF_DELWRI
|_PBF_DELWRI_Q
);
1801 pb
->pb_flags
|= PBF_WRITE
;
1802 list_move(&pb
->pb_list
, &tmp
);
1805 spin_unlock(&pbd_delwrite_lock
);
1807 while (!list_empty(&tmp
)) {
1808 pb
= list_entry(tmp
.next
, xfs_buf_t
, pb_list
);
1809 target
= pb
->pb_target
;
1811 list_del_init(&pb
->pb_list
);
1812 pagebuf_iostrategy(pb
);
1814 blk_run_address_space(target
->pbr_mapping
);
1817 if (as_list_len
> 0)
1820 xfsbufd_force_flush
= 0;
1821 } while (!kthread_should_stop());
1827 * Go through all incore buffers, and release buffers if they belong to
1828 * the given device. This is used in filesystem error handling to
1829 * preserve the consistency of its metadata.
1833 xfs_buftarg_t
*target
,
1836 struct list_head tmp
;
1840 pagebuf_runall_queues(xfsdatad_workqueue
);
1841 pagebuf_runall_queues(xfslogd_workqueue
);
1843 INIT_LIST_HEAD(&tmp
);
1844 spin_lock(&pbd_delwrite_lock
);
1845 list_for_each_entry_safe(pb
, n
, &pbd_delwrite_queue
, pb_list
) {
1847 if (pb
->pb_target
!= target
)
1850 ASSERT(pb
->pb_flags
& (PBF_DELWRI
|_PBF_DELWRI_Q
));
1851 PB_TRACE(pb
, "walkq2", (long)pagebuf_ispin(pb
));
1852 if (pagebuf_ispin(pb
)) {
1857 list_move(&pb
->pb_list
, &tmp
);
1859 spin_unlock(&pbd_delwrite_lock
);
1862 * Dropped the delayed write list lock, now walk the temporary list
1864 list_for_each_entry_safe(pb
, n
, &tmp
, pb_list
) {
1866 pb
->pb_flags
&= ~(PBF_DELWRI
|_PBF_DELWRI_Q
);
1867 pb
->pb_flags
|= PBF_WRITE
;
1869 pb
->pb_flags
&= ~PBF_ASYNC
;
1871 list_del_init(&pb
->pb_list
);
1873 pagebuf_iostrategy(pb
);
1877 * Remaining list items must be flushed before returning
1879 while (!list_empty(&tmp
)) {
1880 pb
= list_entry(tmp
.next
, xfs_buf_t
, pb_list
);
1882 list_del_init(&pb
->pb_list
);
1888 blk_run_address_space(target
->pbr_mapping
);
1894 xfs_buf_daemons_start(void)
1896 int error
= -ENOMEM
;
1898 xfslogd_workqueue
= create_workqueue("xfslogd");
1899 if (!xfslogd_workqueue
)
1902 xfsdatad_workqueue
= create_workqueue("xfsdatad");
1903 if (!xfsdatad_workqueue
)
1904 goto out_destroy_xfslogd_workqueue
;
1906 xfsbufd_task
= kthread_run(xfsbufd
, NULL
, "xfsbufd");
1907 if (IS_ERR(xfsbufd_task
)) {
1908 error
= PTR_ERR(xfsbufd_task
);
1909 goto out_destroy_xfsdatad_workqueue
;
1913 out_destroy_xfsdatad_workqueue
:
1914 destroy_workqueue(xfsdatad_workqueue
);
1915 out_destroy_xfslogd_workqueue
:
1916 destroy_workqueue(xfslogd_workqueue
);
1922 * Note: do not mark as __exit, it is called from pagebuf_terminate.
1925 xfs_buf_daemons_stop(void)
1927 kthread_stop(xfsbufd_task
);
1928 destroy_workqueue(xfslogd_workqueue
);
1929 destroy_workqueue(xfsdatad_workqueue
);
1933 * Initialization and Termination
1939 int error
= -ENOMEM
;
1941 pagebuf_zone
= kmem_zone_init(sizeof(xfs_buf_t
), "xfs_buf");
1945 #ifdef PAGEBUF_TRACE
1946 pagebuf_trace_buf
= ktrace_alloc(PAGEBUF_TRACE_SIZE
, KM_SLEEP
);
1949 error
= xfs_buf_daemons_start();
1951 goto out_free_buf_zone
;
1953 pagebuf_shake
= kmem_shake_register(xfsbufd_wakeup
);
1954 if (!pagebuf_shake
) {
1956 goto out_stop_daemons
;
1962 xfs_buf_daemons_stop();
1964 #ifdef PAGEBUF_TRACE
1965 ktrace_free(pagebuf_trace_buf
);
1967 kmem_zone_destroy(pagebuf_zone
);
1974 * pagebuf_terminate.
1976 * Note: do not mark as __exit, this is also called from the __init code.
1979 pagebuf_terminate(void)
1981 xfs_buf_daemons_stop();
1983 #ifdef PAGEBUF_TRACE
1984 ktrace_free(pagebuf_trace_buf
);
1987 kmem_zone_destroy(pagebuf_zone
);
1988 kmem_shake_deregister(pagebuf_shake
);