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>
58 #include "xfs_linux.h"
64 STATIC kmem_cache_t
*pagebuf_zone
;
65 STATIC kmem_shaker_t pagebuf_shake
;
66 STATIC
int xfsbufd_wakeup(int, unsigned int);
67 STATIC
void pagebuf_delwri_queue(xfs_buf_t
*, int);
69 STATIC
struct workqueue_struct
*xfslogd_workqueue
;
70 struct workqueue_struct
*xfsdatad_workqueue
;
84 ktrace_enter(pagebuf_trace_buf
,
86 (void *)(unsigned long)pb
->pb_flags
,
87 (void *)(unsigned long)pb
->pb_hold
.counter
,
88 (void *)(unsigned long)pb
->pb_sema
.count
.counter
,
91 (void *)(unsigned long)((pb
->pb_file_offset
>>32) & 0xffffffff),
92 (void *)(unsigned long)(pb
->pb_file_offset
& 0xffffffff),
93 (void *)(unsigned long)pb
->pb_buffer_length
,
94 NULL
, NULL
, NULL
, NULL
, NULL
);
96 ktrace_t
*pagebuf_trace_buf
;
97 #define PAGEBUF_TRACE_SIZE 4096
98 #define PB_TRACE(pb, id, data) \
99 pagebuf_trace(pb, id, (void *)data, (void *)__builtin_return_address(0))
101 #define PB_TRACE(pb, id, data) do { } while (0)
104 #ifdef PAGEBUF_LOCK_TRACKING
105 # define PB_SET_OWNER(pb) ((pb)->pb_last_holder = current->pid)
106 # define PB_CLEAR_OWNER(pb) ((pb)->pb_last_holder = -1)
107 # define PB_GET_OWNER(pb) ((pb)->pb_last_holder)
109 # define PB_SET_OWNER(pb) do { } while (0)
110 # define PB_CLEAR_OWNER(pb) do { } while (0)
111 # define PB_GET_OWNER(pb) do { } while (0)
115 * Pagebuf allocation / freeing.
118 #define pb_to_gfp(flags) \
119 ((((flags) & PBF_READ_AHEAD) ? __GFP_NORETRY : \
120 ((flags) & PBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
122 #define pb_to_km(flags) \
123 (((flags) & PBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
126 #define pagebuf_allocate(flags) \
127 kmem_zone_alloc(pagebuf_zone, pb_to_km(flags))
128 #define pagebuf_deallocate(pb) \
129 kmem_zone_free(pagebuf_zone, (pb));
132 * Page Region interfaces.
134 * For pages in filesystems where the blocksize is smaller than the
135 * pagesize, we use the page->private field (long) to hold a bitmap
136 * of uptodate regions within the page.
138 * Each such region is "bytes per page / bits per long" bytes long.
140 * NBPPR == number-of-bytes-per-page-region
141 * BTOPR == bytes-to-page-region (rounded up)
142 * BTOPRT == bytes-to-page-region-truncated (rounded down)
144 #if (BITS_PER_LONG == 32)
145 #define PRSHIFT (PAGE_CACHE_SHIFT - 5) /* (32 == 1<<5) */
146 #elif (BITS_PER_LONG == 64)
147 #define PRSHIFT (PAGE_CACHE_SHIFT - 6) /* (64 == 1<<6) */
149 #error BITS_PER_LONG must be 32 or 64
151 #define NBPPR (PAGE_CACHE_SIZE/BITS_PER_LONG)
152 #define BTOPR(b) (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
153 #define BTOPRT(b) (((unsigned int)(b) >> PRSHIFT))
163 first
= BTOPR(offset
);
164 final
= BTOPRT(offset
+ length
- 1);
165 first
= min(first
, final
);
168 mask
<<= BITS_PER_LONG
- (final
- first
);
169 mask
>>= BITS_PER_LONG
- (final
);
171 ASSERT(offset
+ length
<= PAGE_CACHE_SIZE
);
172 ASSERT((final
- first
) < BITS_PER_LONG
&& (final
- first
) >= 0);
183 page
->private |= page_region_mask(offset
, length
);
184 if (page
->private == ~0UL)
185 SetPageUptodate(page
);
194 unsigned long mask
= page_region_mask(offset
, length
);
196 return (mask
&& (page
->private & mask
) == mask
);
200 * Mapping of multi-page buffers into contiguous virtual space
203 typedef struct a_list
{
208 STATIC a_list_t
*as_free_head
;
209 STATIC
int as_list_len
;
210 STATIC
DEFINE_SPINLOCK(as_lock
);
213 * Try to batch vunmaps because they are costly.
221 aentry
= kmalloc(sizeof(a_list_t
), GFP_ATOMIC
& ~__GFP_HIGH
);
222 if (likely(aentry
)) {
224 aentry
->next
= as_free_head
;
225 aentry
->vm_addr
= addr
;
226 as_free_head
= aentry
;
228 spin_unlock(&as_lock
);
235 purge_addresses(void)
237 a_list_t
*aentry
, *old
;
239 if (as_free_head
== NULL
)
243 aentry
= as_free_head
;
246 spin_unlock(&as_lock
);
248 while ((old
= aentry
) != NULL
) {
249 vunmap(aentry
->vm_addr
);
250 aentry
= aentry
->next
;
256 * Internal pagebuf object manipulation
262 xfs_buftarg_t
*target
,
265 page_buf_flags_t flags
)
268 * We don't want certain flags to appear in pb->pb_flags.
270 flags
&= ~(PBF_LOCK
|PBF_MAPPED
|PBF_DONT_BLOCK
|PBF_READ_AHEAD
);
272 memset(pb
, 0, sizeof(xfs_buf_t
));
273 atomic_set(&pb
->pb_hold
, 1);
274 init_MUTEX_LOCKED(&pb
->pb_iodonesema
);
275 INIT_LIST_HEAD(&pb
->pb_list
);
276 INIT_LIST_HEAD(&pb
->pb_hash_list
);
277 init_MUTEX_LOCKED(&pb
->pb_sema
); /* held, no waiters */
279 pb
->pb_target
= target
;
280 pb
->pb_file_offset
= range_base
;
282 * Set buffer_length and count_desired to the same value initially.
283 * I/O routines should use count_desired, which will be the same in
284 * most cases but may be reset (e.g. XFS recovery).
286 pb
->pb_buffer_length
= pb
->pb_count_desired
= range_length
;
287 pb
->pb_flags
= flags
| PBF_NONE
;
288 pb
->pb_bn
= XFS_BUF_DADDR_NULL
;
289 atomic_set(&pb
->pb_pin_count
, 0);
290 init_waitqueue_head(&pb
->pb_waiters
);
292 XFS_STATS_INC(pb_create
);
293 PB_TRACE(pb
, "initialize", target
);
297 * Allocate a page array capable of holding a specified number
298 * of pages, and point the page buf at it.
304 page_buf_flags_t flags
)
306 /* Make sure that we have a page list */
307 if (pb
->pb_pages
== NULL
) {
308 pb
->pb_offset
= page_buf_poff(pb
->pb_file_offset
);
309 pb
->pb_page_count
= page_count
;
310 if (page_count
<= PB_PAGES
) {
311 pb
->pb_pages
= pb
->pb_page_array
;
313 pb
->pb_pages
= kmem_alloc(sizeof(struct page
*) *
314 page_count
, pb_to_km(flags
));
315 if (pb
->pb_pages
== NULL
)
318 memset(pb
->pb_pages
, 0, sizeof(struct page
*) * page_count
);
324 * Frees pb_pages if it was malloced.
330 if (bp
->pb_pages
!= bp
->pb_page_array
) {
331 kmem_free(bp
->pb_pages
,
332 bp
->pb_page_count
* sizeof(struct page
*));
337 * Releases the specified buffer.
339 * The modification state of any associated pages is left unchanged.
340 * The buffer most not be on any hash - use pagebuf_rele instead for
341 * hashed and refcounted buffers
347 PB_TRACE(bp
, "free", 0);
349 ASSERT(list_empty(&bp
->pb_hash_list
));
351 if (bp
->pb_flags
& _PBF_PAGE_CACHE
) {
354 if ((bp
->pb_flags
& PBF_MAPPED
) && (bp
->pb_page_count
> 1))
355 free_address(bp
->pb_addr
- bp
->pb_offset
);
357 for (i
= 0; i
< bp
->pb_page_count
; i
++)
358 page_cache_release(bp
->pb_pages
[i
]);
359 _pagebuf_free_pages(bp
);
360 } else if (bp
->pb_flags
& _PBF_KMEM_ALLOC
) {
362 * XXX(hch): bp->pb_count_desired might be incorrect (see
363 * pagebuf_associate_memory for details), but fortunately
364 * the Linux version of kmem_free ignores the len argument..
366 kmem_free(bp
->pb_addr
, bp
->pb_count_desired
);
367 _pagebuf_free_pages(bp
);
370 pagebuf_deallocate(bp
);
374 * Finds all pages for buffer in question and builds it's page list.
377 _pagebuf_lookup_pages(
381 struct address_space
*mapping
= bp
->pb_target
->pbr_mapping
;
382 size_t blocksize
= bp
->pb_target
->pbr_bsize
;
383 size_t size
= bp
->pb_count_desired
;
384 size_t nbytes
, offset
;
385 int gfp_mask
= pb_to_gfp(flags
);
386 unsigned short page_count
, i
;
391 end
= bp
->pb_file_offset
+ bp
->pb_buffer_length
;
392 page_count
= page_buf_btoc(end
) - page_buf_btoct(bp
->pb_file_offset
);
394 error
= _pagebuf_get_pages(bp
, page_count
, flags
);
397 bp
->pb_flags
|= _PBF_PAGE_CACHE
;
399 offset
= bp
->pb_offset
;
400 first
= bp
->pb_file_offset
>> PAGE_CACHE_SHIFT
;
402 for (i
= 0; i
< bp
->pb_page_count
; i
++) {
407 page
= find_or_create_page(mapping
, first
+ i
, gfp_mask
);
408 if (unlikely(page
== NULL
)) {
409 if (flags
& PBF_READ_AHEAD
) {
410 bp
->pb_page_count
= i
;
411 for (i
= 0; i
< bp
->pb_page_count
; i
++)
412 unlock_page(bp
->pb_pages
[i
]);
417 * This could deadlock.
419 * But until all the XFS lowlevel code is revamped to
420 * handle buffer allocation failures we can't do much.
422 if (!(++retries
% 100))
424 "XFS: possible memory allocation "
425 "deadlock in %s (mode:0x%x)\n",
426 __FUNCTION__
, gfp_mask
);
428 XFS_STATS_INC(pb_page_retries
);
429 xfsbufd_wakeup(0, gfp_mask
);
430 blk_congestion_wait(WRITE
, HZ
/50);
434 XFS_STATS_INC(pb_page_found
);
436 nbytes
= min_t(size_t, size
, PAGE_CACHE_SIZE
- offset
);
439 if (!PageUptodate(page
)) {
441 if (blocksize
>= PAGE_CACHE_SIZE
) {
442 if (flags
& PBF_READ
)
444 } else if (!PagePrivate(page
)) {
445 if (test_page_region(page
, offset
, nbytes
))
450 bp
->pb_pages
[i
] = page
;
454 if (!bp
->pb_locked
) {
455 for (i
= 0; i
< bp
->pb_page_count
; i
++)
456 unlock_page(bp
->pb_pages
[i
]);
460 /* if we have any uptodate pages, mark that in the buffer */
461 bp
->pb_flags
&= ~PBF_NONE
;
463 /* if some pages aren't uptodate, mark that in the buffer */
464 if (page_count
!= bp
->pb_page_count
)
465 bp
->pb_flags
|= PBF_PARTIAL
;
468 PB_TRACE(bp
, "lookup_pages", (long)page_count
);
473 * Map buffer into kernel address-space if nessecary.
480 /* A single page buffer is always mappable */
481 if (bp
->pb_page_count
== 1) {
482 bp
->pb_addr
= page_address(bp
->pb_pages
[0]) + bp
->pb_offset
;
483 bp
->pb_flags
|= PBF_MAPPED
;
484 } else if (flags
& PBF_MAPPED
) {
485 if (as_list_len
> 64)
487 bp
->pb_addr
= vmap(bp
->pb_pages
, bp
->pb_page_count
,
488 VM_MAP
, PAGE_KERNEL
);
489 if (unlikely(bp
->pb_addr
== NULL
))
491 bp
->pb_addr
+= bp
->pb_offset
;
492 bp
->pb_flags
|= PBF_MAPPED
;
499 * Finding and Reading Buffers
505 * Looks up, and creates if absent, a lockable buffer for
506 * a given range of an inode. The buffer is returned
507 * locked. If other overlapping buffers exist, they are
508 * released before the new buffer is created and locked,
509 * which may imply that this call will block until those buffers
510 * are unlocked. No I/O is implied by this call.
514 xfs_buftarg_t
*btp
, /* block device target */
515 loff_t ioff
, /* starting offset of range */
516 size_t isize
, /* length of range */
517 page_buf_flags_t flags
, /* PBF_TRYLOCK */
518 xfs_buf_t
*new_pb
)/* newly allocated buffer */
525 range_base
= (ioff
<< BBSHIFT
);
526 range_length
= (isize
<< BBSHIFT
);
528 /* Check for IOs smaller than the sector size / not sector aligned */
529 ASSERT(!(range_length
< (1 << btp
->pbr_sshift
)));
530 ASSERT(!(range_base
& (loff_t
)btp
->pbr_smask
));
532 hash
= &btp
->bt_hash
[hash_long((unsigned long)ioff
, btp
->bt_hashshift
)];
534 spin_lock(&hash
->bh_lock
);
536 list_for_each_entry_safe(pb
, n
, &hash
->bh_list
, pb_hash_list
) {
537 ASSERT(btp
== pb
->pb_target
);
538 if (pb
->pb_file_offset
== range_base
&&
539 pb
->pb_buffer_length
== range_length
) {
541 * If we look at something bring it to the
542 * front of the list for next time.
544 atomic_inc(&pb
->pb_hold
);
545 list_move(&pb
->pb_hash_list
, &hash
->bh_list
);
552 _pagebuf_initialize(new_pb
, btp
, range_base
,
553 range_length
, flags
);
554 new_pb
->pb_hash
= hash
;
555 list_add(&new_pb
->pb_hash_list
, &hash
->bh_list
);
557 XFS_STATS_INC(pb_miss_locked
);
560 spin_unlock(&hash
->bh_lock
);
564 spin_unlock(&hash
->bh_lock
);
566 /* Attempt to get the semaphore without sleeping,
567 * if this does not work then we need to drop the
568 * spinlock and do a hard attempt on the semaphore.
570 if (down_trylock(&pb
->pb_sema
)) {
571 if (!(flags
& PBF_TRYLOCK
)) {
572 /* wait for buffer ownership */
573 PB_TRACE(pb
, "get_lock", 0);
575 XFS_STATS_INC(pb_get_locked_waited
);
577 /* We asked for a trylock and failed, no need
578 * to look at file offset and length here, we
579 * know that this pagebuf at least overlaps our
580 * pagebuf and is locked, therefore our buffer
581 * either does not exist, or is this buffer
585 XFS_STATS_INC(pb_busy_locked
);
593 if (pb
->pb_flags
& PBF_STALE
) {
594 ASSERT((pb
->pb_flags
& _PBF_DELWRI_Q
) == 0);
595 pb
->pb_flags
&= PBF_MAPPED
;
597 PB_TRACE(pb
, "got_lock", 0);
598 XFS_STATS_INC(pb_get_locked
);
603 * xfs_buf_get_flags assembles a buffer covering the specified range.
605 * Storage in memory for all portions of the buffer will be allocated,
606 * although backing storage may not be.
609 xfs_buf_get_flags( /* allocate a buffer */
610 xfs_buftarg_t
*target
,/* target for buffer */
611 loff_t ioff
, /* starting offset of range */
612 size_t isize
, /* length of range */
613 page_buf_flags_t flags
) /* PBF_TRYLOCK */
615 xfs_buf_t
*pb
, *new_pb
;
618 new_pb
= pagebuf_allocate(flags
);
619 if (unlikely(!new_pb
))
622 pb
= _pagebuf_find(target
, ioff
, isize
, flags
, new_pb
);
624 error
= _pagebuf_lookup_pages(pb
, flags
);
628 pagebuf_deallocate(new_pb
);
629 if (unlikely(pb
== NULL
))
633 for (i
= 0; i
< pb
->pb_page_count
; i
++)
634 mark_page_accessed(pb
->pb_pages
[i
]);
636 if (!(pb
->pb_flags
& PBF_MAPPED
)) {
637 error
= _pagebuf_map_pages(pb
, flags
);
638 if (unlikely(error
)) {
639 printk(KERN_WARNING
"%s: failed to map pages\n",
645 XFS_STATS_INC(pb_get
);
648 * Always fill in the block number now, the mapped cases can do
649 * their own overlay of this later.
652 pb
->pb_count_desired
= pb
->pb_buffer_length
;
654 PB_TRACE(pb
, "get", (unsigned long)flags
);
658 if (flags
& (PBF_LOCK
| PBF_TRYLOCK
))
666 xfs_buftarg_t
*target
,
669 page_buf_flags_t flags
)
675 pb
= xfs_buf_get_flags(target
, ioff
, isize
, flags
);
677 if (PBF_NOT_DONE(pb
)) {
678 PB_TRACE(pb
, "read", (unsigned long)flags
);
679 XFS_STATS_INC(pb_get_read
);
680 pagebuf_iostart(pb
, flags
);
681 } else if (flags
& PBF_ASYNC
) {
682 PB_TRACE(pb
, "read_async", (unsigned long)flags
);
684 * Read ahead call which is already satisfied,
689 PB_TRACE(pb
, "read_done", (unsigned long)flags
);
690 /* We do not want read in the flags */
691 pb
->pb_flags
&= ~PBF_READ
;
698 if (flags
& (PBF_LOCK
| PBF_TRYLOCK
))
705 * If we are not low on memory then do the readahead in a deadlock
710 xfs_buftarg_t
*target
,
713 page_buf_flags_t flags
)
715 struct backing_dev_info
*bdi
;
717 bdi
= target
->pbr_mapping
->backing_dev_info
;
718 if (bdi_read_congested(bdi
))
721 flags
|= (PBF_TRYLOCK
|PBF_ASYNC
|PBF_READ_AHEAD
);
722 xfs_buf_read_flags(target
, ioff
, isize
, flags
);
728 xfs_buftarg_t
*target
)
732 pb
= pagebuf_allocate(0);
734 _pagebuf_initialize(pb
, target
, 0, len
, 0);
738 static inline struct page
*
742 if (((unsigned long)addr
< VMALLOC_START
) ||
743 ((unsigned long)addr
>= VMALLOC_END
)) {
744 return virt_to_page(addr
);
746 return vmalloc_to_page(addr
);
751 pagebuf_associate_memory(
763 page_count
= PAGE_CACHE_ALIGN(len
) >> PAGE_CACHE_SHIFT
;
764 offset
= (off_t
) mem
- ((off_t
)mem
& PAGE_CACHE_MASK
);
765 if (offset
&& (len
> PAGE_CACHE_SIZE
))
768 /* Free any previous set of page pointers */
770 _pagebuf_free_pages(pb
);
775 rval
= _pagebuf_get_pages(pb
, page_count
, 0);
779 pb
->pb_offset
= offset
;
780 ptr
= (size_t) mem
& PAGE_CACHE_MASK
;
781 end
= PAGE_CACHE_ALIGN((size_t) mem
+ len
);
783 /* set up first page */
784 pb
->pb_pages
[0] = mem_to_page(mem
);
786 ptr
+= PAGE_CACHE_SIZE
;
787 pb
->pb_page_count
= ++i
;
789 pb
->pb_pages
[i
] = mem_to_page((void *)ptr
);
790 pb
->pb_page_count
= ++i
;
791 ptr
+= PAGE_CACHE_SIZE
;
795 pb
->pb_count_desired
= pb
->pb_buffer_length
= len
;
796 pb
->pb_flags
|= PBF_MAPPED
;
802 pagebuf_get_no_daddr(
804 xfs_buftarg_t
*target
)
806 size_t malloc_len
= len
;
811 bp
= pagebuf_allocate(0);
812 if (unlikely(bp
== NULL
))
814 _pagebuf_initialize(bp
, target
, 0, len
, PBF_FORCEIO
);
817 data
= kmem_alloc(malloc_len
, KM_SLEEP
| KM_MAYFAIL
);
818 if (unlikely(data
== NULL
))
821 /* check whether alignment matches.. */
822 if ((__psunsigned_t
)data
!=
823 ((__psunsigned_t
)data
& ~target
->pbr_smask
)) {
824 /* .. else double the size and try again */
825 kmem_free(data
, malloc_len
);
830 error
= pagebuf_associate_memory(bp
, data
, len
);
833 bp
->pb_flags
|= _PBF_KMEM_ALLOC
;
837 PB_TRACE(bp
, "no_daddr", data
);
840 kmem_free(data
, malloc_len
);
850 * Increment reference count on buffer, to hold the buffer concurrently
851 * with another thread which may release (free) the buffer asynchronously.
853 * Must hold the buffer already to call this function.
859 atomic_inc(&pb
->pb_hold
);
860 PB_TRACE(pb
, "hold", 0);
866 * pagebuf_rele releases a hold on the specified buffer. If the
867 * the hold count is 1, pagebuf_rele calls pagebuf_free.
873 xfs_bufhash_t
*hash
= pb
->pb_hash
;
875 PB_TRACE(pb
, "rele", pb
->pb_relse
);
878 * pagebuf_lookup buffers are not hashed, not delayed write,
879 * and don't have their own release routines. Special case.
881 if (unlikely(!hash
)) {
882 ASSERT(!pb
->pb_relse
);
883 if (atomic_dec_and_test(&pb
->pb_hold
))
888 if (atomic_dec_and_lock(&pb
->pb_hold
, &hash
->bh_lock
)) {
892 atomic_inc(&pb
->pb_hold
);
893 spin_unlock(&hash
->bh_lock
);
894 (*(pb
->pb_relse
)) (pb
);
895 spin_lock(&hash
->bh_lock
);
899 if (pb
->pb_flags
& PBF_FS_MANAGED
) {
904 ASSERT((pb
->pb_flags
& (PBF_DELWRI
|_PBF_DELWRI_Q
)) == 0);
905 list_del_init(&pb
->pb_hash_list
);
906 spin_unlock(&hash
->bh_lock
);
909 spin_unlock(&hash
->bh_lock
);
913 * Catch reference count leaks
915 ASSERT(atomic_read(&pb
->pb_hold
) >= 0);
921 * Mutual exclusion on buffers. Locking model:
923 * Buffers associated with inodes for which buffer locking
924 * is not enabled are not protected by semaphores, and are
925 * assumed to be exclusively owned by the caller. There is a
926 * spinlock in the buffer, used by the caller when concurrent
927 * access is possible.
933 * pagebuf_cond_lock locks a buffer object, if it is not already locked.
934 * Note that this in no way
935 * locks the underlying pages, so it is only useful for synchronizing
936 * concurrent use of page buffer objects, not for synchronizing independent
937 * access to the underlying pages.
940 pagebuf_cond_lock( /* lock buffer, if not locked */
941 /* returns -EBUSY if locked) */
946 locked
= down_trylock(&pb
->pb_sema
) == 0;
950 PB_TRACE(pb
, "cond_lock", (long)locked
);
951 return(locked
? 0 : -EBUSY
);
954 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
958 * Return lock value for a pagebuf
964 return(atomic_read(&pb
->pb_sema
.count
));
971 * pagebuf_lock locks a buffer object. Note that this in no way
972 * locks the underlying pages, so it is only useful for synchronizing
973 * concurrent use of page buffer objects, not for synchronizing independent
974 * access to the underlying pages.
980 PB_TRACE(pb
, "lock", 0);
981 if (atomic_read(&pb
->pb_io_remaining
))
982 blk_run_address_space(pb
->pb_target
->pbr_mapping
);
985 PB_TRACE(pb
, "locked", 0);
992 * pagebuf_unlock releases the lock on the buffer object created by
993 * pagebuf_lock or pagebuf_cond_lock (not any pinning of underlying pages
994 * created by pagebuf_pin).
996 * If the buffer is marked delwri but is not queued, do so before we
997 * unlock the buffer as we need to set flags correctly. We also need to
998 * take a reference for the delwri queue because the unlocker is going to
999 * drop their's and they don't know we just queued it.
1002 pagebuf_unlock( /* unlock buffer */
1003 xfs_buf_t
*pb
) /* buffer to unlock */
1005 if ((pb
->pb_flags
& (PBF_DELWRI
|_PBF_DELWRI_Q
)) == PBF_DELWRI
) {
1006 atomic_inc(&pb
->pb_hold
);
1007 pb
->pb_flags
|= PBF_ASYNC
;
1008 pagebuf_delwri_queue(pb
, 0);
1013 PB_TRACE(pb
, "unlock", 0);
1018 * Pinning Buffer Storage in Memory
1024 * pagebuf_pin locks all of the memory represented by a buffer in
1025 * memory. Multiple calls to pagebuf_pin and pagebuf_unpin, for
1026 * the same or different buffers affecting a given page, will
1027 * properly count the number of outstanding "pin" requests. The
1028 * buffer may be released after the pagebuf_pin and a different
1029 * buffer used when calling pagebuf_unpin, if desired.
1030 * pagebuf_pin should be used by the file system when it wants be
1031 * assured that no attempt will be made to force the affected
1032 * memory to disk. It does not assure that a given logical page
1033 * will not be moved to a different physical page.
1039 atomic_inc(&pb
->pb_pin_count
);
1040 PB_TRACE(pb
, "pin", (long)pb
->pb_pin_count
.counter
);
1046 * pagebuf_unpin reverses the locking of memory performed by
1047 * pagebuf_pin. Note that both functions affected the logical
1048 * pages associated with the buffer, not the buffer itself.
1054 if (atomic_dec_and_test(&pb
->pb_pin_count
)) {
1055 wake_up_all(&pb
->pb_waiters
);
1057 PB_TRACE(pb
, "unpin", (long)pb
->pb_pin_count
.counter
);
1064 return atomic_read(&pb
->pb_pin_count
);
1068 * pagebuf_wait_unpin
1070 * pagebuf_wait_unpin waits until all of the memory associated
1071 * with the buffer is not longer locked in memory. It returns
1072 * immediately if none of the affected pages are locked.
1075 _pagebuf_wait_unpin(
1078 DECLARE_WAITQUEUE (wait
, current
);
1080 if (atomic_read(&pb
->pb_pin_count
) == 0)
1083 add_wait_queue(&pb
->pb_waiters
, &wait
);
1085 set_current_state(TASK_UNINTERRUPTIBLE
);
1086 if (atomic_read(&pb
->pb_pin_count
) == 0)
1088 if (atomic_read(&pb
->pb_io_remaining
))
1089 blk_run_address_space(pb
->pb_target
->pbr_mapping
);
1092 remove_wait_queue(&pb
->pb_waiters
, &wait
);
1093 set_current_state(TASK_RUNNING
);
1097 * Buffer Utility Routines
1103 * pagebuf_iodone marks a buffer for which I/O is in progress
1104 * done with respect to that I/O. The pb_iodone routine, if
1105 * present, will be called as a side-effect.
1108 pagebuf_iodone_work(
1111 xfs_buf_t
*bp
= (xfs_buf_t
*)v
;
1114 (*(bp
->pb_iodone
))(bp
);
1115 else if (bp
->pb_flags
& PBF_ASYNC
)
1125 pb
->pb_flags
&= ~(PBF_READ
| PBF_WRITE
);
1126 if (pb
->pb_error
== 0) {
1127 pb
->pb_flags
&= ~(PBF_PARTIAL
| PBF_NONE
);
1130 PB_TRACE(pb
, "iodone", pb
->pb_iodone
);
1132 if ((pb
->pb_iodone
) || (pb
->pb_flags
& PBF_ASYNC
)) {
1134 INIT_WORK(&pb
->pb_iodone_work
, pagebuf_iodone_work
, pb
);
1135 queue_work(dataio
? xfsdatad_workqueue
:
1136 xfslogd_workqueue
, &pb
->pb_iodone_work
);
1138 pagebuf_iodone_work(pb
);
1141 up(&pb
->pb_iodonesema
);
1148 * pagebuf_ioerror sets the error code for a buffer.
1151 pagebuf_ioerror( /* mark/clear buffer error flag */
1152 xfs_buf_t
*pb
, /* buffer to mark */
1153 int error
) /* error to store (0 if none) */
1155 ASSERT(error
>= 0 && error
<= 0xffff);
1156 pb
->pb_error
= (unsigned short)error
;
1157 PB_TRACE(pb
, "ioerror", (unsigned long)error
);
1163 * pagebuf_iostart initiates I/O on a buffer, based on the flags supplied.
1164 * If necessary, it will arrange for any disk space allocation required,
1165 * and it will break up the request if the block mappings require it.
1166 * The pb_iodone routine in the buffer supplied will only be called
1167 * when all of the subsidiary I/O requests, if any, have been completed.
1168 * pagebuf_iostart calls the pagebuf_ioinitiate routine or
1169 * pagebuf_iorequest, if the former routine is not defined, to start
1170 * the I/O on a given low-level request.
1173 pagebuf_iostart( /* start I/O on a buffer */
1174 xfs_buf_t
*pb
, /* buffer to start */
1175 page_buf_flags_t flags
) /* PBF_LOCK, PBF_ASYNC, PBF_READ, */
1176 /* PBF_WRITE, PBF_DELWRI, */
1177 /* PBF_DONT_BLOCK */
1181 PB_TRACE(pb
, "iostart", (unsigned long)flags
);
1183 if (flags
& PBF_DELWRI
) {
1184 pb
->pb_flags
&= ~(PBF_READ
| PBF_WRITE
| PBF_ASYNC
);
1185 pb
->pb_flags
|= flags
& (PBF_DELWRI
| PBF_ASYNC
);
1186 pagebuf_delwri_queue(pb
, 1);
1190 pb
->pb_flags
&= ~(PBF_READ
| PBF_WRITE
| PBF_ASYNC
| PBF_DELWRI
| \
1191 PBF_READ_AHEAD
| _PBF_RUN_QUEUES
);
1192 pb
->pb_flags
|= flags
& (PBF_READ
| PBF_WRITE
| PBF_ASYNC
| \
1193 PBF_READ_AHEAD
| _PBF_RUN_QUEUES
);
1195 BUG_ON(pb
->pb_bn
== XFS_BUF_DADDR_NULL
);
1197 /* For writes allow an alternate strategy routine to precede
1198 * the actual I/O request (which may not be issued at all in
1199 * a shutdown situation, for example).
1201 status
= (flags
& PBF_WRITE
) ?
1202 pagebuf_iostrategy(pb
) : pagebuf_iorequest(pb
);
1204 /* Wait for I/O if we are not an async request.
1205 * Note: async I/O request completion will release the buffer,
1206 * and that can already be done by this point. So using the
1207 * buffer pointer from here on, after async I/O, is invalid.
1209 if (!status
&& !(flags
& PBF_ASYNC
))
1210 status
= pagebuf_iowait(pb
);
1216 * Helper routine for pagebuf_iorequest
1219 STATIC __inline__
int
1223 ASSERT(pb
->pb_flags
& (PBF_READ
|PBF_WRITE
));
1224 if (pb
->pb_flags
& PBF_READ
)
1225 return pb
->pb_locked
;
1229 STATIC __inline__
void
1234 if (atomic_dec_and_test(&pb
->pb_io_remaining
) == 1) {
1236 pagebuf_iodone(pb
, (pb
->pb_flags
& PBF_FS_DATAIOD
), schedule
);
1243 unsigned int bytes_done
,
1246 xfs_buf_t
*pb
= (xfs_buf_t
*)bio
->bi_private
;
1247 unsigned int blocksize
= pb
->pb_target
->pbr_bsize
;
1248 struct bio_vec
*bvec
= bio
->bi_io_vec
+ bio
->bi_vcnt
- 1;
1253 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1257 struct page
*page
= bvec
->bv_page
;
1259 if (unlikely(pb
->pb_error
)) {
1260 if (pb
->pb_flags
& PBF_READ
)
1261 ClearPageUptodate(page
);
1263 } else if (blocksize
== PAGE_CACHE_SIZE
) {
1264 SetPageUptodate(page
);
1265 } else if (!PagePrivate(page
) &&
1266 (pb
->pb_flags
& _PBF_PAGE_CACHE
)) {
1267 set_page_region(page
, bvec
->bv_offset
, bvec
->bv_len
);
1270 if (--bvec
>= bio
->bi_io_vec
)
1271 prefetchw(&bvec
->bv_page
->flags
);
1273 if (_pagebuf_iolocked(pb
)) {
1276 } while (bvec
>= bio
->bi_io_vec
);
1278 _pagebuf_iodone(pb
, 1);
1287 int i
, rw
, map_i
, total_nr_pages
, nr_pages
;
1289 int offset
= pb
->pb_offset
;
1290 int size
= pb
->pb_count_desired
;
1291 sector_t sector
= pb
->pb_bn
;
1292 unsigned int blocksize
= pb
->pb_target
->pbr_bsize
;
1293 int locking
= _pagebuf_iolocked(pb
);
1295 total_nr_pages
= pb
->pb_page_count
;
1298 if (pb
->pb_flags
& _PBF_RUN_QUEUES
) {
1299 pb
->pb_flags
&= ~_PBF_RUN_QUEUES
;
1300 rw
= (pb
->pb_flags
& PBF_READ
) ? READ_SYNC
: WRITE_SYNC
;
1302 rw
= (pb
->pb_flags
& PBF_READ
) ? READ
: WRITE
;
1305 /* Special code path for reading a sub page size pagebuf in --
1306 * we populate up the whole page, and hence the other metadata
1307 * in the same page. This optimization is only valid when the
1308 * filesystem block size and the page size are equal.
1310 if ((pb
->pb_buffer_length
< PAGE_CACHE_SIZE
) &&
1311 (pb
->pb_flags
& PBF_READ
) && locking
&&
1312 (blocksize
== PAGE_CACHE_SIZE
)) {
1313 bio
= bio_alloc(GFP_NOIO
, 1);
1315 bio
->bi_bdev
= pb
->pb_target
->pbr_bdev
;
1316 bio
->bi_sector
= sector
- (offset
>> BBSHIFT
);
1317 bio
->bi_end_io
= bio_end_io_pagebuf
;
1318 bio
->bi_private
= pb
;
1320 bio_add_page(bio
, pb
->pb_pages
[0], PAGE_CACHE_SIZE
, 0);
1323 atomic_inc(&pb
->pb_io_remaining
);
1328 /* Lock down the pages which we need to for the request */
1329 if (locking
&& (pb
->pb_flags
& PBF_WRITE
) && (pb
->pb_locked
== 0)) {
1330 for (i
= 0; size
; i
++) {
1331 int nbytes
= PAGE_CACHE_SIZE
- offset
;
1332 struct page
*page
= pb
->pb_pages
[i
];
1342 offset
= pb
->pb_offset
;
1343 size
= pb
->pb_count_desired
;
1347 atomic_inc(&pb
->pb_io_remaining
);
1348 nr_pages
= BIO_MAX_SECTORS
>> (PAGE_SHIFT
- BBSHIFT
);
1349 if (nr_pages
> total_nr_pages
)
1350 nr_pages
= total_nr_pages
;
1352 bio
= bio_alloc(GFP_NOIO
, nr_pages
);
1353 bio
->bi_bdev
= pb
->pb_target
->pbr_bdev
;
1354 bio
->bi_sector
= sector
;
1355 bio
->bi_end_io
= bio_end_io_pagebuf
;
1356 bio
->bi_private
= pb
;
1358 for (; size
&& nr_pages
; nr_pages
--, map_i
++) {
1359 int nbytes
= PAGE_CACHE_SIZE
- offset
;
1364 if (bio_add_page(bio
, pb
->pb_pages
[map_i
],
1365 nbytes
, offset
) < nbytes
)
1369 sector
+= nbytes
>> BBSHIFT
;
1375 if (likely(bio
->bi_size
)) {
1376 submit_bio(rw
, bio
);
1381 pagebuf_ioerror(pb
, EIO
);
1386 * pagebuf_iorequest -- the core I/O request routine.
1389 pagebuf_iorequest( /* start real I/O */
1390 xfs_buf_t
*pb
) /* buffer to convey to device */
1392 PB_TRACE(pb
, "iorequest", 0);
1394 if (pb
->pb_flags
& PBF_DELWRI
) {
1395 pagebuf_delwri_queue(pb
, 1);
1399 if (pb
->pb_flags
& PBF_WRITE
) {
1400 _pagebuf_wait_unpin(pb
);
1405 /* Set the count to 1 initially, this will stop an I/O
1406 * completion callout which happens before we have started
1407 * all the I/O from calling pagebuf_iodone too early.
1409 atomic_set(&pb
->pb_io_remaining
, 1);
1410 _pagebuf_ioapply(pb
);
1411 _pagebuf_iodone(pb
, 0);
1420 * pagebuf_iowait waits for I/O to complete on the buffer supplied.
1421 * It returns immediately if no I/O is pending. In any case, it returns
1422 * the error code, if any, or 0 if there is no error.
1428 PB_TRACE(pb
, "iowait", 0);
1429 if (atomic_read(&pb
->pb_io_remaining
))
1430 blk_run_address_space(pb
->pb_target
->pbr_mapping
);
1431 down(&pb
->pb_iodonesema
);
1432 PB_TRACE(pb
, "iowaited", (long)pb
->pb_error
);
1433 return pb
->pb_error
;
1443 offset
+= pb
->pb_offset
;
1445 page
= pb
->pb_pages
[offset
>> PAGE_CACHE_SHIFT
];
1446 return (caddr_t
) page_address(page
) + (offset
& (PAGE_CACHE_SIZE
- 1));
1452 * Move data into or out of a buffer.
1456 xfs_buf_t
*pb
, /* buffer to process */
1457 size_t boff
, /* starting buffer offset */
1458 size_t bsize
, /* length to copy */
1459 caddr_t data
, /* data address */
1460 page_buf_rw_t mode
) /* read/write flag */
1462 size_t bend
, cpoff
, csize
;
1465 bend
= boff
+ bsize
;
1466 while (boff
< bend
) {
1467 page
= pb
->pb_pages
[page_buf_btoct(boff
+ pb
->pb_offset
)];
1468 cpoff
= page_buf_poff(boff
+ pb
->pb_offset
);
1469 csize
= min_t(size_t,
1470 PAGE_CACHE_SIZE
-cpoff
, pb
->pb_count_desired
-boff
);
1472 ASSERT(((csize
+ cpoff
) <= PAGE_CACHE_SIZE
));
1476 memset(page_address(page
) + cpoff
, 0, csize
);
1479 memcpy(data
, page_address(page
) + cpoff
, csize
);
1482 memcpy(page_address(page
) + cpoff
, data
, csize
);
1491 * Handling of buftargs.
1495 * Wait for any bufs with callbacks that have been submitted but
1496 * have not yet returned... walk the hash list for the target.
1503 xfs_bufhash_t
*hash
;
1506 for (i
= 0; i
< (1 << btp
->bt_hashshift
); i
++) {
1507 hash
= &btp
->bt_hash
[i
];
1509 spin_lock(&hash
->bh_lock
);
1510 list_for_each_entry_safe(bp
, n
, &hash
->bh_list
, pb_hash_list
) {
1511 ASSERT(btp
== bp
->pb_target
);
1512 if (!(bp
->pb_flags
& PBF_FS_MANAGED
)) {
1513 spin_unlock(&hash
->bh_lock
);
1515 * Catch superblock reference count leaks
1518 BUG_ON(bp
->pb_bn
== 0);
1523 spin_unlock(&hash
->bh_lock
);
1528 * Allocate buffer hash table for a given target.
1529 * For devices containing metadata (i.e. not the log/realtime devices)
1530 * we need to allocate a much larger hash table.
1539 btp
->bt_hashshift
= external
? 3 : 8; /* 8 or 256 buckets */
1540 btp
->bt_hashmask
= (1 << btp
->bt_hashshift
) - 1;
1541 btp
->bt_hash
= kmem_zalloc((1 << btp
->bt_hashshift
) *
1542 sizeof(xfs_bufhash_t
), KM_SLEEP
);
1543 for (i
= 0; i
< (1 << btp
->bt_hashshift
); i
++) {
1544 spin_lock_init(&btp
->bt_hash
[i
].bh_lock
);
1545 INIT_LIST_HEAD(&btp
->bt_hash
[i
].bh_list
);
1553 kmem_free(btp
->bt_hash
,
1554 (1 << btp
->bt_hashshift
) * sizeof(xfs_bufhash_t
));
1555 btp
->bt_hash
= NULL
;
1563 xfs_flush_buftarg(btp
, 1);
1565 xfs_blkdev_put(btp
->pbr_bdev
);
1566 xfs_free_bufhash(btp
);
1567 iput(btp
->pbr_mapping
->host
);
1568 kmem_free(btp
, sizeof(*btp
));
1572 xfs_setsize_buftarg_flags(
1574 unsigned int blocksize
,
1575 unsigned int sectorsize
,
1578 btp
->pbr_bsize
= blocksize
;
1579 btp
->pbr_sshift
= ffs(sectorsize
) - 1;
1580 btp
->pbr_smask
= sectorsize
- 1;
1582 if (set_blocksize(btp
->pbr_bdev
, sectorsize
)) {
1584 "XFS: Cannot set_blocksize to %u on device %s\n",
1585 sectorsize
, XFS_BUFTARG_NAME(btp
));
1590 (PAGE_CACHE_SIZE
/ BITS_PER_LONG
) > sectorsize
) {
1592 "XFS: %u byte sectors in use on device %s. "
1593 "This is suboptimal; %u or greater is ideal.\n",
1594 sectorsize
, XFS_BUFTARG_NAME(btp
),
1595 (unsigned int)PAGE_CACHE_SIZE
/ BITS_PER_LONG
);
1602 * When allocating the initial buffer target we have not yet
1603 * read in the superblock, so don't know what sized sectors
1604 * are being used is at this early stage. Play safe.
1607 xfs_setsize_buftarg_early(
1609 struct block_device
*bdev
)
1611 return xfs_setsize_buftarg_flags(btp
,
1612 PAGE_CACHE_SIZE
, bdev_hardsect_size(bdev
), 0);
1616 xfs_setsize_buftarg(
1618 unsigned int blocksize
,
1619 unsigned int sectorsize
)
1621 return xfs_setsize_buftarg_flags(btp
, blocksize
, sectorsize
, 1);
1625 xfs_mapping_buftarg(
1627 struct block_device
*bdev
)
1629 struct backing_dev_info
*bdi
;
1630 struct inode
*inode
;
1631 struct address_space
*mapping
;
1632 static struct address_space_operations mapping_aops
= {
1633 .sync_page
= block_sync_page
,
1636 inode
= new_inode(bdev
->bd_inode
->i_sb
);
1639 "XFS: Cannot allocate mapping inode for device %s\n",
1640 XFS_BUFTARG_NAME(btp
));
1643 inode
->i_mode
= S_IFBLK
;
1644 inode
->i_bdev
= bdev
;
1645 inode
->i_rdev
= bdev
->bd_dev
;
1646 bdi
= blk_get_backing_dev_info(bdev
);
1648 bdi
= &default_backing_dev_info
;
1649 mapping
= &inode
->i_data
;
1650 mapping
->a_ops
= &mapping_aops
;
1651 mapping
->backing_dev_info
= bdi
;
1652 mapping_set_gfp_mask(mapping
, GFP_NOFS
);
1653 btp
->pbr_mapping
= mapping
;
1659 struct block_device
*bdev
,
1664 btp
= kmem_zalloc(sizeof(*btp
), KM_SLEEP
);
1666 btp
->pbr_dev
= bdev
->bd_dev
;
1667 btp
->pbr_bdev
= bdev
;
1668 if (xfs_setsize_buftarg_early(btp
, bdev
))
1670 if (xfs_mapping_buftarg(btp
, bdev
))
1672 xfs_alloc_bufhash(btp
, external
);
1676 kmem_free(btp
, sizeof(*btp
));
1682 * Pagebuf delayed write buffer handling
1685 STATIC
LIST_HEAD(pbd_delwrite_queue
);
1686 STATIC
DEFINE_SPINLOCK(pbd_delwrite_lock
);
1689 pagebuf_delwri_queue(
1693 PB_TRACE(pb
, "delwri_q", (long)unlock
);
1694 ASSERT((pb
->pb_flags
& (PBF_DELWRI
|PBF_ASYNC
)) ==
1695 (PBF_DELWRI
|PBF_ASYNC
));
1697 spin_lock(&pbd_delwrite_lock
);
1698 /* If already in the queue, dequeue and place at tail */
1699 if (!list_empty(&pb
->pb_list
)) {
1700 ASSERT(pb
->pb_flags
& _PBF_DELWRI_Q
);
1702 atomic_dec(&pb
->pb_hold
);
1704 list_del(&pb
->pb_list
);
1707 pb
->pb_flags
|= _PBF_DELWRI_Q
;
1708 list_add_tail(&pb
->pb_list
, &pbd_delwrite_queue
);
1709 pb
->pb_queuetime
= jiffies
;
1710 spin_unlock(&pbd_delwrite_lock
);
1717 pagebuf_delwri_dequeue(
1722 spin_lock(&pbd_delwrite_lock
);
1723 if ((pb
->pb_flags
& PBF_DELWRI
) && !list_empty(&pb
->pb_list
)) {
1724 ASSERT(pb
->pb_flags
& _PBF_DELWRI_Q
);
1725 list_del_init(&pb
->pb_list
);
1728 pb
->pb_flags
&= ~(PBF_DELWRI
|_PBF_DELWRI_Q
);
1729 spin_unlock(&pbd_delwrite_lock
);
1734 PB_TRACE(pb
, "delwri_dq", (long)dequeued
);
1738 pagebuf_runall_queues(
1739 struct workqueue_struct
*queue
)
1741 flush_workqueue(queue
);
1744 /* Defines for pagebuf daemon */
1745 STATIC
DECLARE_COMPLETION(xfsbufd_done
);
1746 STATIC
struct task_struct
*xfsbufd_task
;
1747 STATIC
int xfsbufd_active
;
1748 STATIC
int xfsbufd_force_flush
;
1749 STATIC
int xfsbufd_force_sleep
;
1756 if (xfsbufd_force_sleep
)
1758 xfsbufd_force_flush
= 1;
1760 wake_up_process(xfsbufd_task
);
1768 struct list_head tmp
;
1770 xfs_buftarg_t
*target
;
1773 /* Set up the thread */
1774 daemonize("xfsbufd");
1775 current
->flags
|= PF_MEMALLOC
;
1777 xfsbufd_task
= current
;
1781 INIT_LIST_HEAD(&tmp
);
1783 if (unlikely(freezing(current
))) {
1784 xfsbufd_force_sleep
= 1;
1787 xfsbufd_force_sleep
= 0;
1790 set_current_state(TASK_INTERRUPTIBLE
);
1791 schedule_timeout((xfs_buf_timer_centisecs
* HZ
) / 100);
1793 age
= (xfs_buf_age_centisecs
* HZ
) / 100;
1794 spin_lock(&pbd_delwrite_lock
);
1795 list_for_each_entry_safe(pb
, n
, &pbd_delwrite_queue
, pb_list
) {
1796 PB_TRACE(pb
, "walkq1", (long)pagebuf_ispin(pb
));
1797 ASSERT(pb
->pb_flags
& PBF_DELWRI
);
1799 if (!pagebuf_ispin(pb
) && !pagebuf_cond_lock(pb
)) {
1800 if (!xfsbufd_force_flush
&&
1801 time_before(jiffies
,
1802 pb
->pb_queuetime
+ age
)) {
1807 pb
->pb_flags
&= ~(PBF_DELWRI
|_PBF_DELWRI_Q
);
1808 pb
->pb_flags
|= PBF_WRITE
;
1809 list_move(&pb
->pb_list
, &tmp
);
1812 spin_unlock(&pbd_delwrite_lock
);
1814 while (!list_empty(&tmp
)) {
1815 pb
= list_entry(tmp
.next
, xfs_buf_t
, pb_list
);
1816 target
= pb
->pb_target
;
1818 list_del_init(&pb
->pb_list
);
1819 pagebuf_iostrategy(pb
);
1821 blk_run_address_space(target
->pbr_mapping
);
1824 if (as_list_len
> 0)
1827 xfsbufd_force_flush
= 0;
1828 } while (xfsbufd_active
);
1830 complete_and_exit(&xfsbufd_done
, 0);
1834 * Go through all incore buffers, and release buffers if they belong to
1835 * the given device. This is used in filesystem error handling to
1836 * preserve the consistency of its metadata.
1840 xfs_buftarg_t
*target
,
1843 struct list_head tmp
;
1847 pagebuf_runall_queues(xfsdatad_workqueue
);
1848 pagebuf_runall_queues(xfslogd_workqueue
);
1850 INIT_LIST_HEAD(&tmp
);
1851 spin_lock(&pbd_delwrite_lock
);
1852 list_for_each_entry_safe(pb
, n
, &pbd_delwrite_queue
, pb_list
) {
1854 if (pb
->pb_target
!= target
)
1857 ASSERT(pb
->pb_flags
& (PBF_DELWRI
|_PBF_DELWRI_Q
));
1858 PB_TRACE(pb
, "walkq2", (long)pagebuf_ispin(pb
));
1859 if (pagebuf_ispin(pb
)) {
1864 list_move(&pb
->pb_list
, &tmp
);
1866 spin_unlock(&pbd_delwrite_lock
);
1869 * Dropped the delayed write list lock, now walk the temporary list
1871 list_for_each_entry_safe(pb
, n
, &tmp
, pb_list
) {
1873 pb
->pb_flags
&= ~(PBF_DELWRI
|_PBF_DELWRI_Q
);
1874 pb
->pb_flags
|= PBF_WRITE
;
1876 pb
->pb_flags
&= ~PBF_ASYNC
;
1878 list_del_init(&pb
->pb_list
);
1880 pagebuf_iostrategy(pb
);
1884 * Remaining list items must be flushed before returning
1886 while (!list_empty(&tmp
)) {
1887 pb
= list_entry(tmp
.next
, xfs_buf_t
, pb_list
);
1889 list_del_init(&pb
->pb_list
);
1895 blk_run_address_space(target
->pbr_mapping
);
1901 xfs_buf_daemons_start(void)
1903 int error
= -ENOMEM
;
1905 xfslogd_workqueue
= create_workqueue("xfslogd");
1906 if (!xfslogd_workqueue
)
1909 xfsdatad_workqueue
= create_workqueue("xfsdatad");
1910 if (!xfsdatad_workqueue
)
1911 goto out_destroy_xfslogd_workqueue
;
1913 error
= kernel_thread(xfsbufd
, NULL
, CLONE_FS
|CLONE_FILES
);
1915 goto out_destroy_xfsdatad_workqueue
;
1918 out_destroy_xfsdatad_workqueue
:
1919 destroy_workqueue(xfsdatad_workqueue
);
1920 out_destroy_xfslogd_workqueue
:
1921 destroy_workqueue(xfslogd_workqueue
);
1927 * Note: do not mark as __exit, it is called from pagebuf_terminate.
1930 xfs_buf_daemons_stop(void)
1934 wait_for_completion(&xfsbufd_done
);
1936 destroy_workqueue(xfslogd_workqueue
);
1937 destroy_workqueue(xfsdatad_workqueue
);
1941 * Initialization and Termination
1947 int error
= -ENOMEM
;
1949 pagebuf_zone
= kmem_zone_init(sizeof(xfs_buf_t
), "xfs_buf");
1953 #ifdef PAGEBUF_TRACE
1954 pagebuf_trace_buf
= ktrace_alloc(PAGEBUF_TRACE_SIZE
, KM_SLEEP
);
1957 error
= xfs_buf_daemons_start();
1959 goto out_free_buf_zone
;
1961 pagebuf_shake
= kmem_shake_register(xfsbufd_wakeup
);
1962 if (!pagebuf_shake
) {
1964 goto out_stop_daemons
;
1970 xfs_buf_daemons_stop();
1972 #ifdef PAGEBUF_TRACE
1973 ktrace_free(pagebuf_trace_buf
);
1975 kmem_zone_destroy(pagebuf_zone
);
1982 * pagebuf_terminate.
1984 * Note: do not mark as __exit, this is also called from the __init code.
1987 pagebuf_terminate(void)
1989 xfs_buf_daemons_stop();
1991 #ifdef PAGEBUF_TRACE
1992 ktrace_free(pagebuf_trace_buf
);
1995 kmem_zone_destroy(pagebuf_zone
);
1996 kmem_shake_deregister(pagebuf_shake
);