2 * Copyright (c) 2000-2005 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
18 #include <linux/stddef.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/vmalloc.h>
24 #include <linux/bio.h>
25 #include <linux/sysctl.h>
26 #include <linux/proc_fs.h>
27 #include <linux/workqueue.h>
28 #include <linux/percpu.h>
29 #include <linux/blkdev.h>
30 #include <linux/hash.h>
31 #include <linux/kthread.h>
32 #include "xfs_linux.h"
34 STATIC kmem_cache_t
*pagebuf_zone
;
35 STATIC kmem_shaker_t pagebuf_shake
;
36 STATIC
int xfsbufd_wakeup(int, gfp_t
);
37 STATIC
void pagebuf_delwri_queue(xfs_buf_t
*, int);
39 STATIC
struct workqueue_struct
*xfslogd_workqueue
;
40 struct workqueue_struct
*xfsdatad_workqueue
;
50 ktrace_enter(pagebuf_trace_buf
,
52 (void *)(unsigned long)pb
->pb_flags
,
53 (void *)(unsigned long)pb
->pb_hold
.counter
,
54 (void *)(unsigned long)pb
->pb_sema
.count
.counter
,
57 (void *)(unsigned long)((pb
->pb_file_offset
>>32) & 0xffffffff),
58 (void *)(unsigned long)(pb
->pb_file_offset
& 0xffffffff),
59 (void *)(unsigned long)pb
->pb_buffer_length
,
60 NULL
, NULL
, NULL
, NULL
, NULL
);
62 ktrace_t
*pagebuf_trace_buf
;
63 #define PAGEBUF_TRACE_SIZE 4096
64 #define PB_TRACE(pb, id, data) \
65 pagebuf_trace(pb, id, (void *)data, (void *)__builtin_return_address(0))
67 #define PB_TRACE(pb, id, data) do { } while (0)
70 #ifdef PAGEBUF_LOCK_TRACKING
71 # define PB_SET_OWNER(pb) ((pb)->pb_last_holder = current->pid)
72 # define PB_CLEAR_OWNER(pb) ((pb)->pb_last_holder = -1)
73 # define PB_GET_OWNER(pb) ((pb)->pb_last_holder)
75 # define PB_SET_OWNER(pb) do { } while (0)
76 # define PB_CLEAR_OWNER(pb) do { } while (0)
77 # define PB_GET_OWNER(pb) do { } while (0)
80 #define pb_to_gfp(flags) \
81 ((((flags) & PBF_READ_AHEAD) ? __GFP_NORETRY : \
82 ((flags) & PBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
84 #define pb_to_km(flags) \
85 (((flags) & PBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
87 #define pagebuf_allocate(flags) \
88 kmem_zone_alloc(pagebuf_zone, pb_to_km(flags))
89 #define pagebuf_deallocate(pb) \
90 kmem_zone_free(pagebuf_zone, (pb));
93 * Page Region interfaces.
95 * For pages in filesystems where the blocksize is smaller than the
96 * pagesize, we use the page->private field (long) to hold a bitmap
97 * of uptodate regions within the page.
99 * Each such region is "bytes per page / bits per long" bytes long.
101 * NBPPR == number-of-bytes-per-page-region
102 * BTOPR == bytes-to-page-region (rounded up)
103 * BTOPRT == bytes-to-page-region-truncated (rounded down)
105 #if (BITS_PER_LONG == 32)
106 #define PRSHIFT (PAGE_CACHE_SHIFT - 5) /* (32 == 1<<5) */
107 #elif (BITS_PER_LONG == 64)
108 #define PRSHIFT (PAGE_CACHE_SHIFT - 6) /* (64 == 1<<6) */
110 #error BITS_PER_LONG must be 32 or 64
112 #define NBPPR (PAGE_CACHE_SIZE/BITS_PER_LONG)
113 #define BTOPR(b) (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
114 #define BTOPRT(b) (((unsigned int)(b) >> PRSHIFT))
124 first
= BTOPR(offset
);
125 final
= BTOPRT(offset
+ length
- 1);
126 first
= min(first
, final
);
129 mask
<<= BITS_PER_LONG
- (final
- first
);
130 mask
>>= BITS_PER_LONG
- (final
);
132 ASSERT(offset
+ length
<= PAGE_CACHE_SIZE
);
133 ASSERT((final
- first
) < BITS_PER_LONG
&& (final
- first
) >= 0);
144 set_page_private(page
,
145 page_private(page
) | page_region_mask(offset
, length
));
146 if (page_private(page
) == ~0UL)
147 SetPageUptodate(page
);
156 unsigned long mask
= page_region_mask(offset
, length
);
158 return (mask
&& (page_private(page
) & mask
) == mask
);
162 * Mapping of multi-page buffers into contiguous virtual space
165 typedef struct a_list
{
170 STATIC a_list_t
*as_free_head
;
171 STATIC
int as_list_len
;
172 STATIC
DEFINE_SPINLOCK(as_lock
);
175 * Try to batch vunmaps because they are costly.
183 aentry
= kmalloc(sizeof(a_list_t
), GFP_ATOMIC
& ~__GFP_HIGH
);
184 if (likely(aentry
)) {
186 aentry
->next
= as_free_head
;
187 aentry
->vm_addr
= addr
;
188 as_free_head
= aentry
;
190 spin_unlock(&as_lock
);
197 purge_addresses(void)
199 a_list_t
*aentry
, *old
;
201 if (as_free_head
== NULL
)
205 aentry
= as_free_head
;
208 spin_unlock(&as_lock
);
210 while ((old
= aentry
) != NULL
) {
211 vunmap(aentry
->vm_addr
);
212 aentry
= aentry
->next
;
218 * Internal pagebuf object manipulation
224 xfs_buftarg_t
*target
,
227 page_buf_flags_t flags
)
230 * We don't want certain flags to appear in pb->pb_flags.
232 flags
&= ~(PBF_LOCK
|PBF_MAPPED
|PBF_DONT_BLOCK
|PBF_READ_AHEAD
);
234 memset(pb
, 0, sizeof(xfs_buf_t
));
235 atomic_set(&pb
->pb_hold
, 1);
236 init_MUTEX_LOCKED(&pb
->pb_iodonesema
);
237 INIT_LIST_HEAD(&pb
->pb_list
);
238 INIT_LIST_HEAD(&pb
->pb_hash_list
);
239 init_MUTEX_LOCKED(&pb
->pb_sema
); /* held, no waiters */
241 pb
->pb_target
= target
;
242 pb
->pb_file_offset
= range_base
;
244 * Set buffer_length and count_desired to the same value initially.
245 * I/O routines should use count_desired, which will be the same in
246 * most cases but may be reset (e.g. XFS recovery).
248 pb
->pb_buffer_length
= pb
->pb_count_desired
= range_length
;
249 pb
->pb_flags
= flags
;
250 pb
->pb_bn
= XFS_BUF_DADDR_NULL
;
251 atomic_set(&pb
->pb_pin_count
, 0);
252 init_waitqueue_head(&pb
->pb_waiters
);
254 XFS_STATS_INC(pb_create
);
255 PB_TRACE(pb
, "initialize", target
);
259 * Allocate a page array capable of holding a specified number
260 * of pages, and point the page buf at it.
266 page_buf_flags_t flags
)
268 /* Make sure that we have a page list */
269 if (pb
->pb_pages
== NULL
) {
270 pb
->pb_offset
= page_buf_poff(pb
->pb_file_offset
);
271 pb
->pb_page_count
= page_count
;
272 if (page_count
<= PB_PAGES
) {
273 pb
->pb_pages
= pb
->pb_page_array
;
275 pb
->pb_pages
= kmem_alloc(sizeof(struct page
*) *
276 page_count
, pb_to_km(flags
));
277 if (pb
->pb_pages
== NULL
)
280 memset(pb
->pb_pages
, 0, sizeof(struct page
*) * page_count
);
286 * Frees pb_pages if it was malloced.
292 if (bp
->pb_pages
!= bp
->pb_page_array
) {
293 kmem_free(bp
->pb_pages
,
294 bp
->pb_page_count
* sizeof(struct page
*));
299 * Releases the specified buffer.
301 * The modification state of any associated pages is left unchanged.
302 * The buffer most not be on any hash - use pagebuf_rele instead for
303 * hashed and refcounted buffers
309 PB_TRACE(bp
, "free", 0);
311 ASSERT(list_empty(&bp
->pb_hash_list
));
313 if (bp
->pb_flags
& _PBF_PAGE_CACHE
) {
316 if ((bp
->pb_flags
& PBF_MAPPED
) && (bp
->pb_page_count
> 1))
317 free_address(bp
->pb_addr
- bp
->pb_offset
);
319 for (i
= 0; i
< bp
->pb_page_count
; i
++)
320 page_cache_release(bp
->pb_pages
[i
]);
321 _pagebuf_free_pages(bp
);
322 } else if (bp
->pb_flags
& _PBF_KMEM_ALLOC
) {
324 * XXX(hch): bp->pb_count_desired might be incorrect (see
325 * pagebuf_associate_memory for details), but fortunately
326 * the Linux version of kmem_free ignores the len argument..
328 kmem_free(bp
->pb_addr
, bp
->pb_count_desired
);
329 _pagebuf_free_pages(bp
);
332 pagebuf_deallocate(bp
);
336 * Finds all pages for buffer in question and builds it's page list.
339 _pagebuf_lookup_pages(
343 struct address_space
*mapping
= bp
->pb_target
->pbr_mapping
;
344 size_t blocksize
= bp
->pb_target
->pbr_bsize
;
345 size_t size
= bp
->pb_count_desired
;
346 size_t nbytes
, offset
;
347 gfp_t gfp_mask
= pb_to_gfp(flags
);
348 unsigned short page_count
, i
;
353 end
= bp
->pb_file_offset
+ bp
->pb_buffer_length
;
354 page_count
= page_buf_btoc(end
) - page_buf_btoct(bp
->pb_file_offset
);
356 error
= _pagebuf_get_pages(bp
, page_count
, flags
);
359 bp
->pb_flags
|= _PBF_PAGE_CACHE
;
361 offset
= bp
->pb_offset
;
362 first
= bp
->pb_file_offset
>> PAGE_CACHE_SHIFT
;
364 for (i
= 0; i
< bp
->pb_page_count
; i
++) {
369 page
= find_or_create_page(mapping
, first
+ i
, gfp_mask
);
370 if (unlikely(page
== NULL
)) {
371 if (flags
& PBF_READ_AHEAD
) {
372 bp
->pb_page_count
= i
;
373 for (i
= 0; i
< bp
->pb_page_count
; i
++)
374 unlock_page(bp
->pb_pages
[i
]);
379 * This could deadlock.
381 * But until all the XFS lowlevel code is revamped to
382 * handle buffer allocation failures we can't do much.
384 if (!(++retries
% 100))
386 "XFS: possible memory allocation "
387 "deadlock in %s (mode:0x%x)\n",
388 __FUNCTION__
, gfp_mask
);
390 XFS_STATS_INC(pb_page_retries
);
391 xfsbufd_wakeup(0, gfp_mask
);
392 blk_congestion_wait(WRITE
, HZ
/50);
396 XFS_STATS_INC(pb_page_found
);
398 nbytes
= min_t(size_t, size
, PAGE_CACHE_SIZE
- offset
);
401 if (!PageUptodate(page
)) {
403 if (blocksize
>= PAGE_CACHE_SIZE
) {
404 if (flags
& PBF_READ
)
406 } else if (!PagePrivate(page
)) {
407 if (test_page_region(page
, offset
, nbytes
))
412 bp
->pb_pages
[i
] = page
;
416 if (!bp
->pb_locked
) {
417 for (i
= 0; i
< bp
->pb_page_count
; i
++)
418 unlock_page(bp
->pb_pages
[i
]);
421 if (page_count
== bp
->pb_page_count
)
422 bp
->pb_flags
|= PBF_DONE
;
424 PB_TRACE(bp
, "lookup_pages", (long)page_count
);
429 * Map buffer into kernel address-space if nessecary.
436 /* A single page buffer is always mappable */
437 if (bp
->pb_page_count
== 1) {
438 bp
->pb_addr
= page_address(bp
->pb_pages
[0]) + bp
->pb_offset
;
439 bp
->pb_flags
|= PBF_MAPPED
;
440 } else if (flags
& PBF_MAPPED
) {
441 if (as_list_len
> 64)
443 bp
->pb_addr
= vmap(bp
->pb_pages
, bp
->pb_page_count
,
444 VM_MAP
, PAGE_KERNEL
);
445 if (unlikely(bp
->pb_addr
== NULL
))
447 bp
->pb_addr
+= bp
->pb_offset
;
448 bp
->pb_flags
|= PBF_MAPPED
;
455 * Finding and Reading Buffers
461 * Looks up, and creates if absent, a lockable buffer for
462 * a given range of an inode. The buffer is returned
463 * locked. If other overlapping buffers exist, they are
464 * released before the new buffer is created and locked,
465 * which may imply that this call will block until those buffers
466 * are unlocked. No I/O is implied by this call.
470 xfs_buftarg_t
*btp
, /* block device target */
471 loff_t ioff
, /* starting offset of range */
472 size_t isize
, /* length of range */
473 page_buf_flags_t flags
, /* PBF_TRYLOCK */
474 xfs_buf_t
*new_pb
)/* newly allocated buffer */
481 range_base
= (ioff
<< BBSHIFT
);
482 range_length
= (isize
<< BBSHIFT
);
484 /* Check for IOs smaller than the sector size / not sector aligned */
485 ASSERT(!(range_length
< (1 << btp
->pbr_sshift
)));
486 ASSERT(!(range_base
& (loff_t
)btp
->pbr_smask
));
488 hash
= &btp
->bt_hash
[hash_long((unsigned long)ioff
, btp
->bt_hashshift
)];
490 spin_lock(&hash
->bh_lock
);
492 list_for_each_entry_safe(pb
, n
, &hash
->bh_list
, pb_hash_list
) {
493 ASSERT(btp
== pb
->pb_target
);
494 if (pb
->pb_file_offset
== range_base
&&
495 pb
->pb_buffer_length
== range_length
) {
497 * If we look at something bring it to the
498 * front of the list for next time.
500 atomic_inc(&pb
->pb_hold
);
501 list_move(&pb
->pb_hash_list
, &hash
->bh_list
);
508 _pagebuf_initialize(new_pb
, btp
, range_base
,
509 range_length
, flags
);
510 new_pb
->pb_hash
= hash
;
511 list_add(&new_pb
->pb_hash_list
, &hash
->bh_list
);
513 XFS_STATS_INC(pb_miss_locked
);
516 spin_unlock(&hash
->bh_lock
);
520 spin_unlock(&hash
->bh_lock
);
522 /* Attempt to get the semaphore without sleeping,
523 * if this does not work then we need to drop the
524 * spinlock and do a hard attempt on the semaphore.
526 if (down_trylock(&pb
->pb_sema
)) {
527 if (!(flags
& PBF_TRYLOCK
)) {
528 /* wait for buffer ownership */
529 PB_TRACE(pb
, "get_lock", 0);
531 XFS_STATS_INC(pb_get_locked_waited
);
533 /* We asked for a trylock and failed, no need
534 * to look at file offset and length here, we
535 * know that this pagebuf at least overlaps our
536 * pagebuf and is locked, therefore our buffer
537 * either does not exist, or is this buffer
541 XFS_STATS_INC(pb_busy_locked
);
549 if (pb
->pb_flags
& PBF_STALE
) {
550 ASSERT((pb
->pb_flags
& _PBF_DELWRI_Q
) == 0);
551 pb
->pb_flags
&= PBF_MAPPED
;
553 PB_TRACE(pb
, "got_lock", 0);
554 XFS_STATS_INC(pb_get_locked
);
559 * xfs_buf_get_flags assembles a buffer covering the specified range.
561 * Storage in memory for all portions of the buffer will be allocated,
562 * although backing storage may not be.
565 xfs_buf_get_flags( /* allocate a buffer */
566 xfs_buftarg_t
*target
,/* target for buffer */
567 loff_t ioff
, /* starting offset of range */
568 size_t isize
, /* length of range */
569 page_buf_flags_t flags
) /* PBF_TRYLOCK */
571 xfs_buf_t
*pb
, *new_pb
;
574 new_pb
= pagebuf_allocate(flags
);
575 if (unlikely(!new_pb
))
578 pb
= _pagebuf_find(target
, ioff
, isize
, flags
, new_pb
);
580 error
= _pagebuf_lookup_pages(pb
, flags
);
584 pagebuf_deallocate(new_pb
);
585 if (unlikely(pb
== NULL
))
589 for (i
= 0; i
< pb
->pb_page_count
; i
++)
590 mark_page_accessed(pb
->pb_pages
[i
]);
592 if (!(pb
->pb_flags
& PBF_MAPPED
)) {
593 error
= _pagebuf_map_pages(pb
, flags
);
594 if (unlikely(error
)) {
595 printk(KERN_WARNING
"%s: failed to map pages\n",
601 XFS_STATS_INC(pb_get
);
604 * Always fill in the block number now, the mapped cases can do
605 * their own overlay of this later.
608 pb
->pb_count_desired
= pb
->pb_buffer_length
;
610 PB_TRACE(pb
, "get", (unsigned long)flags
);
614 if (flags
& (PBF_LOCK
| PBF_TRYLOCK
))
622 xfs_buftarg_t
*target
,
625 page_buf_flags_t flags
)
631 pb
= xfs_buf_get_flags(target
, ioff
, isize
, flags
);
633 if (!XFS_BUF_ISDONE(pb
)) {
634 PB_TRACE(pb
, "read", (unsigned long)flags
);
635 XFS_STATS_INC(pb_get_read
);
636 pagebuf_iostart(pb
, flags
);
637 } else if (flags
& PBF_ASYNC
) {
638 PB_TRACE(pb
, "read_async", (unsigned long)flags
);
640 * Read ahead call which is already satisfied,
645 PB_TRACE(pb
, "read_done", (unsigned long)flags
);
646 /* We do not want read in the flags */
647 pb
->pb_flags
&= ~PBF_READ
;
654 if (flags
& (PBF_LOCK
| PBF_TRYLOCK
))
661 * If we are not low on memory then do the readahead in a deadlock
666 xfs_buftarg_t
*target
,
669 page_buf_flags_t flags
)
671 struct backing_dev_info
*bdi
;
673 bdi
= target
->pbr_mapping
->backing_dev_info
;
674 if (bdi_read_congested(bdi
))
677 flags
|= (PBF_TRYLOCK
|PBF_ASYNC
|PBF_READ_AHEAD
);
678 xfs_buf_read_flags(target
, ioff
, isize
, flags
);
684 xfs_buftarg_t
*target
)
688 pb
= pagebuf_allocate(0);
690 _pagebuf_initialize(pb
, target
, 0, len
, 0);
694 static inline struct page
*
698 if (((unsigned long)addr
< VMALLOC_START
) ||
699 ((unsigned long)addr
>= VMALLOC_END
)) {
700 return virt_to_page(addr
);
702 return vmalloc_to_page(addr
);
707 pagebuf_associate_memory(
719 page_count
= PAGE_CACHE_ALIGN(len
) >> PAGE_CACHE_SHIFT
;
720 offset
= (off_t
) mem
- ((off_t
)mem
& PAGE_CACHE_MASK
);
721 if (offset
&& (len
> PAGE_CACHE_SIZE
))
724 /* Free any previous set of page pointers */
726 _pagebuf_free_pages(pb
);
731 rval
= _pagebuf_get_pages(pb
, page_count
, 0);
735 pb
->pb_offset
= offset
;
736 ptr
= (size_t) mem
& PAGE_CACHE_MASK
;
737 end
= PAGE_CACHE_ALIGN((size_t) mem
+ len
);
739 /* set up first page */
740 pb
->pb_pages
[0] = mem_to_page(mem
);
742 ptr
+= PAGE_CACHE_SIZE
;
743 pb
->pb_page_count
= ++i
;
745 pb
->pb_pages
[i
] = mem_to_page((void *)ptr
);
746 pb
->pb_page_count
= ++i
;
747 ptr
+= PAGE_CACHE_SIZE
;
751 pb
->pb_count_desired
= pb
->pb_buffer_length
= len
;
752 pb
->pb_flags
|= PBF_MAPPED
;
758 pagebuf_get_no_daddr(
760 xfs_buftarg_t
*target
)
762 size_t malloc_len
= len
;
767 bp
= pagebuf_allocate(0);
768 if (unlikely(bp
== NULL
))
770 _pagebuf_initialize(bp
, target
, 0, len
, 0);
773 data
= kmem_alloc(malloc_len
, KM_SLEEP
| KM_MAYFAIL
);
774 if (unlikely(data
== NULL
))
777 /* check whether alignment matches.. */
778 if ((__psunsigned_t
)data
!=
779 ((__psunsigned_t
)data
& ~target
->pbr_smask
)) {
780 /* .. else double the size and try again */
781 kmem_free(data
, malloc_len
);
786 error
= pagebuf_associate_memory(bp
, data
, len
);
789 bp
->pb_flags
|= _PBF_KMEM_ALLOC
;
793 PB_TRACE(bp
, "no_daddr", data
);
796 kmem_free(data
, malloc_len
);
806 * Increment reference count on buffer, to hold the buffer concurrently
807 * with another thread which may release (free) the buffer asynchronously.
809 * Must hold the buffer already to call this function.
815 atomic_inc(&pb
->pb_hold
);
816 PB_TRACE(pb
, "hold", 0);
822 * pagebuf_rele releases a hold on the specified buffer. If the
823 * the hold count is 1, pagebuf_rele calls pagebuf_free.
829 xfs_bufhash_t
*hash
= pb
->pb_hash
;
831 PB_TRACE(pb
, "rele", pb
->pb_relse
);
833 if (atomic_dec_and_lock(&pb
->pb_hold
, &hash
->bh_lock
)) {
835 atomic_inc(&pb
->pb_hold
);
836 spin_unlock(&hash
->bh_lock
);
837 (*(pb
->pb_relse
)) (pb
);
838 } else if (pb
->pb_flags
& PBF_FS_MANAGED
) {
839 spin_unlock(&hash
->bh_lock
);
841 ASSERT(!(pb
->pb_flags
& (PBF_DELWRI
|_PBF_DELWRI_Q
)));
842 list_del_init(&pb
->pb_hash_list
);
843 spin_unlock(&hash
->bh_lock
);
848 * Catch reference count leaks
850 ASSERT(atomic_read(&pb
->pb_hold
) >= 0);
856 * Mutual exclusion on buffers. Locking model:
858 * Buffers associated with inodes for which buffer locking
859 * is not enabled are not protected by semaphores, and are
860 * assumed to be exclusively owned by the caller. There is a
861 * spinlock in the buffer, used by the caller when concurrent
862 * access is possible.
868 * pagebuf_cond_lock locks a buffer object, if it is not already locked.
869 * Note that this in no way
870 * locks the underlying pages, so it is only useful for synchronizing
871 * concurrent use of page buffer objects, not for synchronizing independent
872 * access to the underlying pages.
875 pagebuf_cond_lock( /* lock buffer, if not locked */
876 /* returns -EBUSY if locked) */
881 locked
= down_trylock(&pb
->pb_sema
) == 0;
885 PB_TRACE(pb
, "cond_lock", (long)locked
);
886 return(locked
? 0 : -EBUSY
);
889 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
893 * Return lock value for a pagebuf
899 return(atomic_read(&pb
->pb_sema
.count
));
906 * pagebuf_lock locks a buffer object. Note that this in no way
907 * locks the underlying pages, so it is only useful for synchronizing
908 * concurrent use of page buffer objects, not for synchronizing independent
909 * access to the underlying pages.
915 PB_TRACE(pb
, "lock", 0);
916 if (atomic_read(&pb
->pb_io_remaining
))
917 blk_run_address_space(pb
->pb_target
->pbr_mapping
);
920 PB_TRACE(pb
, "locked", 0);
927 * pagebuf_unlock releases the lock on the buffer object created by
928 * pagebuf_lock or pagebuf_cond_lock (not any pinning of underlying pages
929 * created by pagebuf_pin).
931 * If the buffer is marked delwri but is not queued, do so before we
932 * unlock the buffer as we need to set flags correctly. We also need to
933 * take a reference for the delwri queue because the unlocker is going to
934 * drop their's and they don't know we just queued it.
937 pagebuf_unlock( /* unlock buffer */
938 xfs_buf_t
*pb
) /* buffer to unlock */
940 if ((pb
->pb_flags
& (PBF_DELWRI
|_PBF_DELWRI_Q
)) == PBF_DELWRI
) {
941 atomic_inc(&pb
->pb_hold
);
942 pb
->pb_flags
|= PBF_ASYNC
;
943 pagebuf_delwri_queue(pb
, 0);
948 PB_TRACE(pb
, "unlock", 0);
953 * Pinning Buffer Storage in Memory
959 * pagebuf_pin locks all of the memory represented by a buffer in
960 * memory. Multiple calls to pagebuf_pin and pagebuf_unpin, for
961 * the same or different buffers affecting a given page, will
962 * properly count the number of outstanding "pin" requests. The
963 * buffer may be released after the pagebuf_pin and a different
964 * buffer used when calling pagebuf_unpin, if desired.
965 * pagebuf_pin should be used by the file system when it wants be
966 * assured that no attempt will be made to force the affected
967 * memory to disk. It does not assure that a given logical page
968 * will not be moved to a different physical page.
974 atomic_inc(&pb
->pb_pin_count
);
975 PB_TRACE(pb
, "pin", (long)pb
->pb_pin_count
.counter
);
981 * pagebuf_unpin reverses the locking of memory performed by
982 * pagebuf_pin. Note that both functions affected the logical
983 * pages associated with the buffer, not the buffer itself.
989 if (atomic_dec_and_test(&pb
->pb_pin_count
)) {
990 wake_up_all(&pb
->pb_waiters
);
992 PB_TRACE(pb
, "unpin", (long)pb
->pb_pin_count
.counter
);
999 return atomic_read(&pb
->pb_pin_count
);
1003 * pagebuf_wait_unpin
1005 * pagebuf_wait_unpin waits until all of the memory associated
1006 * with the buffer is not longer locked in memory. It returns
1007 * immediately if none of the affected pages are locked.
1010 _pagebuf_wait_unpin(
1013 DECLARE_WAITQUEUE (wait
, current
);
1015 if (atomic_read(&pb
->pb_pin_count
) == 0)
1018 add_wait_queue(&pb
->pb_waiters
, &wait
);
1020 set_current_state(TASK_UNINTERRUPTIBLE
);
1021 if (atomic_read(&pb
->pb_pin_count
) == 0)
1023 if (atomic_read(&pb
->pb_io_remaining
))
1024 blk_run_address_space(pb
->pb_target
->pbr_mapping
);
1027 remove_wait_queue(&pb
->pb_waiters
, &wait
);
1028 set_current_state(TASK_RUNNING
);
1032 * Buffer Utility Routines
1038 * pagebuf_iodone marks a buffer for which I/O is in progress
1039 * done with respect to that I/O. The pb_iodone routine, if
1040 * present, will be called as a side-effect.
1043 pagebuf_iodone_work(
1046 xfs_buf_t
*bp
= (xfs_buf_t
*)v
;
1049 (*(bp
->pb_iodone
))(bp
);
1050 else if (bp
->pb_flags
& PBF_ASYNC
)
1059 pb
->pb_flags
&= ~(PBF_READ
| PBF_WRITE
);
1060 if (pb
->pb_error
== 0)
1061 pb
->pb_flags
|= PBF_DONE
;
1063 PB_TRACE(pb
, "iodone", pb
->pb_iodone
);
1065 if ((pb
->pb_iodone
) || (pb
->pb_flags
& PBF_ASYNC
)) {
1067 INIT_WORK(&pb
->pb_iodone_work
, pagebuf_iodone_work
, pb
);
1068 queue_work(xfslogd_workqueue
, &pb
->pb_iodone_work
);
1070 pagebuf_iodone_work(pb
);
1073 up(&pb
->pb_iodonesema
);
1080 * pagebuf_ioerror sets the error code for a buffer.
1083 pagebuf_ioerror( /* mark/clear buffer error flag */
1084 xfs_buf_t
*pb
, /* buffer to mark */
1085 int error
) /* error to store (0 if none) */
1087 ASSERT(error
>= 0 && error
<= 0xffff);
1088 pb
->pb_error
= (unsigned short)error
;
1089 PB_TRACE(pb
, "ioerror", (unsigned long)error
);
1095 * pagebuf_iostart initiates I/O on a buffer, based on the flags supplied.
1096 * If necessary, it will arrange for any disk space allocation required,
1097 * and it will break up the request if the block mappings require it.
1098 * The pb_iodone routine in the buffer supplied will only be called
1099 * when all of the subsidiary I/O requests, if any, have been completed.
1100 * pagebuf_iostart calls the pagebuf_ioinitiate routine or
1101 * pagebuf_iorequest, if the former routine is not defined, to start
1102 * the I/O on a given low-level request.
1105 pagebuf_iostart( /* start I/O on a buffer */
1106 xfs_buf_t
*pb
, /* buffer to start */
1107 page_buf_flags_t flags
) /* PBF_LOCK, PBF_ASYNC, PBF_READ, */
1108 /* PBF_WRITE, PBF_DELWRI, */
1109 /* PBF_DONT_BLOCK */
1113 PB_TRACE(pb
, "iostart", (unsigned long)flags
);
1115 if (flags
& PBF_DELWRI
) {
1116 pb
->pb_flags
&= ~(PBF_READ
| PBF_WRITE
| PBF_ASYNC
);
1117 pb
->pb_flags
|= flags
& (PBF_DELWRI
| PBF_ASYNC
);
1118 pagebuf_delwri_queue(pb
, 1);
1122 pb
->pb_flags
&= ~(PBF_READ
| PBF_WRITE
| PBF_ASYNC
| PBF_DELWRI
| \
1123 PBF_READ_AHEAD
| _PBF_RUN_QUEUES
);
1124 pb
->pb_flags
|= flags
& (PBF_READ
| PBF_WRITE
| PBF_ASYNC
| \
1125 PBF_READ_AHEAD
| _PBF_RUN_QUEUES
);
1127 BUG_ON(pb
->pb_bn
== XFS_BUF_DADDR_NULL
);
1129 /* For writes allow an alternate strategy routine to precede
1130 * the actual I/O request (which may not be issued at all in
1131 * a shutdown situation, for example).
1133 status
= (flags
& PBF_WRITE
) ?
1134 pagebuf_iostrategy(pb
) : pagebuf_iorequest(pb
);
1136 /* Wait for I/O if we are not an async request.
1137 * Note: async I/O request completion will release the buffer,
1138 * and that can already be done by this point. So using the
1139 * buffer pointer from here on, after async I/O, is invalid.
1141 if (!status
&& !(flags
& PBF_ASYNC
))
1142 status
= pagebuf_iowait(pb
);
1148 * Helper routine for pagebuf_iorequest
1151 STATIC __inline__
int
1155 ASSERT(pb
->pb_flags
& (PBF_READ
|PBF_WRITE
));
1156 if (pb
->pb_flags
& PBF_READ
)
1157 return pb
->pb_locked
;
1161 STATIC __inline__
void
1166 if (atomic_dec_and_test(&pb
->pb_io_remaining
) == 1) {
1168 pagebuf_iodone(pb
, schedule
);
1175 unsigned int bytes_done
,
1178 xfs_buf_t
*pb
= (xfs_buf_t
*)bio
->bi_private
;
1179 unsigned int blocksize
= pb
->pb_target
->pbr_bsize
;
1180 struct bio_vec
*bvec
= bio
->bi_io_vec
+ bio
->bi_vcnt
- 1;
1185 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1189 struct page
*page
= bvec
->bv_page
;
1191 if (unlikely(pb
->pb_error
)) {
1192 if (pb
->pb_flags
& PBF_READ
)
1193 ClearPageUptodate(page
);
1195 } else if (blocksize
== PAGE_CACHE_SIZE
) {
1196 SetPageUptodate(page
);
1197 } else if (!PagePrivate(page
) &&
1198 (pb
->pb_flags
& _PBF_PAGE_CACHE
)) {
1199 set_page_region(page
, bvec
->bv_offset
, bvec
->bv_len
);
1202 if (--bvec
>= bio
->bi_io_vec
)
1203 prefetchw(&bvec
->bv_page
->flags
);
1205 if (_pagebuf_iolocked(pb
)) {
1208 } while (bvec
>= bio
->bi_io_vec
);
1210 _pagebuf_iodone(pb
, 1);
1219 int i
, rw
, map_i
, total_nr_pages
, nr_pages
;
1221 int offset
= pb
->pb_offset
;
1222 int size
= pb
->pb_count_desired
;
1223 sector_t sector
= pb
->pb_bn
;
1224 unsigned int blocksize
= pb
->pb_target
->pbr_bsize
;
1225 int locking
= _pagebuf_iolocked(pb
);
1227 total_nr_pages
= pb
->pb_page_count
;
1230 if (pb
->pb_flags
& _PBF_RUN_QUEUES
) {
1231 pb
->pb_flags
&= ~_PBF_RUN_QUEUES
;
1232 rw
= (pb
->pb_flags
& PBF_READ
) ? READ_SYNC
: WRITE_SYNC
;
1234 rw
= (pb
->pb_flags
& PBF_READ
) ? READ
: WRITE
;
1237 if (pb
->pb_flags
& PBF_ORDERED
) {
1238 ASSERT(!(pb
->pb_flags
& PBF_READ
));
1242 /* Special code path for reading a sub page size pagebuf in --
1243 * we populate up the whole page, and hence the other metadata
1244 * in the same page. This optimization is only valid when the
1245 * filesystem block size and the page size are equal.
1247 if ((pb
->pb_buffer_length
< PAGE_CACHE_SIZE
) &&
1248 (pb
->pb_flags
& PBF_READ
) && locking
&&
1249 (blocksize
== PAGE_CACHE_SIZE
)) {
1250 bio
= bio_alloc(GFP_NOIO
, 1);
1252 bio
->bi_bdev
= pb
->pb_target
->pbr_bdev
;
1253 bio
->bi_sector
= sector
- (offset
>> BBSHIFT
);
1254 bio
->bi_end_io
= bio_end_io_pagebuf
;
1255 bio
->bi_private
= pb
;
1257 bio_add_page(bio
, pb
->pb_pages
[0], PAGE_CACHE_SIZE
, 0);
1260 atomic_inc(&pb
->pb_io_remaining
);
1265 /* Lock down the pages which we need to for the request */
1266 if (locking
&& (pb
->pb_flags
& PBF_WRITE
) && (pb
->pb_locked
== 0)) {
1267 for (i
= 0; size
; i
++) {
1268 int nbytes
= PAGE_CACHE_SIZE
- offset
;
1269 struct page
*page
= pb
->pb_pages
[i
];
1279 offset
= pb
->pb_offset
;
1280 size
= pb
->pb_count_desired
;
1284 atomic_inc(&pb
->pb_io_remaining
);
1285 nr_pages
= BIO_MAX_SECTORS
>> (PAGE_SHIFT
- BBSHIFT
);
1286 if (nr_pages
> total_nr_pages
)
1287 nr_pages
= total_nr_pages
;
1289 bio
= bio_alloc(GFP_NOIO
, nr_pages
);
1290 bio
->bi_bdev
= pb
->pb_target
->pbr_bdev
;
1291 bio
->bi_sector
= sector
;
1292 bio
->bi_end_io
= bio_end_io_pagebuf
;
1293 bio
->bi_private
= pb
;
1295 for (; size
&& nr_pages
; nr_pages
--, map_i
++) {
1296 int nbytes
= PAGE_CACHE_SIZE
- offset
;
1301 if (bio_add_page(bio
, pb
->pb_pages
[map_i
],
1302 nbytes
, offset
) < nbytes
)
1306 sector
+= nbytes
>> BBSHIFT
;
1312 if (likely(bio
->bi_size
)) {
1313 submit_bio(rw
, bio
);
1318 pagebuf_ioerror(pb
, EIO
);
1323 * pagebuf_iorequest -- the core I/O request routine.
1326 pagebuf_iorequest( /* start real I/O */
1327 xfs_buf_t
*pb
) /* buffer to convey to device */
1329 PB_TRACE(pb
, "iorequest", 0);
1331 if (pb
->pb_flags
& PBF_DELWRI
) {
1332 pagebuf_delwri_queue(pb
, 1);
1336 if (pb
->pb_flags
& PBF_WRITE
) {
1337 _pagebuf_wait_unpin(pb
);
1342 /* Set the count to 1 initially, this will stop an I/O
1343 * completion callout which happens before we have started
1344 * all the I/O from calling pagebuf_iodone too early.
1346 atomic_set(&pb
->pb_io_remaining
, 1);
1347 _pagebuf_ioapply(pb
);
1348 _pagebuf_iodone(pb
, 0);
1357 * pagebuf_iowait waits for I/O to complete on the buffer supplied.
1358 * It returns immediately if no I/O is pending. In any case, it returns
1359 * the error code, if any, or 0 if there is no error.
1365 PB_TRACE(pb
, "iowait", 0);
1366 if (atomic_read(&pb
->pb_io_remaining
))
1367 blk_run_address_space(pb
->pb_target
->pbr_mapping
);
1368 down(&pb
->pb_iodonesema
);
1369 PB_TRACE(pb
, "iowaited", (long)pb
->pb_error
);
1370 return pb
->pb_error
;
1380 offset
+= pb
->pb_offset
;
1382 page
= pb
->pb_pages
[offset
>> PAGE_CACHE_SHIFT
];
1383 return (caddr_t
) page_address(page
) + (offset
& (PAGE_CACHE_SIZE
- 1));
1389 * Move data into or out of a buffer.
1393 xfs_buf_t
*pb
, /* buffer to process */
1394 size_t boff
, /* starting buffer offset */
1395 size_t bsize
, /* length to copy */
1396 caddr_t data
, /* data address */
1397 page_buf_rw_t mode
) /* read/write flag */
1399 size_t bend
, cpoff
, csize
;
1402 bend
= boff
+ bsize
;
1403 while (boff
< bend
) {
1404 page
= pb
->pb_pages
[page_buf_btoct(boff
+ pb
->pb_offset
)];
1405 cpoff
= page_buf_poff(boff
+ pb
->pb_offset
);
1406 csize
= min_t(size_t,
1407 PAGE_CACHE_SIZE
-cpoff
, pb
->pb_count_desired
-boff
);
1409 ASSERT(((csize
+ cpoff
) <= PAGE_CACHE_SIZE
));
1413 memset(page_address(page
) + cpoff
, 0, csize
);
1416 memcpy(data
, page_address(page
) + cpoff
, csize
);
1419 memcpy(page_address(page
) + cpoff
, data
, csize
);
1428 * Handling of buftargs.
1432 * Wait for any bufs with callbacks that have been submitted but
1433 * have not yet returned... walk the hash list for the target.
1440 xfs_bufhash_t
*hash
;
1443 for (i
= 0; i
< (1 << btp
->bt_hashshift
); i
++) {
1444 hash
= &btp
->bt_hash
[i
];
1446 spin_lock(&hash
->bh_lock
);
1447 list_for_each_entry_safe(bp
, n
, &hash
->bh_list
, pb_hash_list
) {
1448 ASSERT(btp
== bp
->pb_target
);
1449 if (!(bp
->pb_flags
& PBF_FS_MANAGED
)) {
1450 spin_unlock(&hash
->bh_lock
);
1452 * Catch superblock reference count leaks
1455 BUG_ON(bp
->pb_bn
== 0);
1460 spin_unlock(&hash
->bh_lock
);
1465 * Allocate buffer hash table for a given target.
1466 * For devices containing metadata (i.e. not the log/realtime devices)
1467 * we need to allocate a much larger hash table.
1476 btp
->bt_hashshift
= external
? 3 : 8; /* 8 or 256 buckets */
1477 btp
->bt_hashmask
= (1 << btp
->bt_hashshift
) - 1;
1478 btp
->bt_hash
= kmem_zalloc((1 << btp
->bt_hashshift
) *
1479 sizeof(xfs_bufhash_t
), KM_SLEEP
);
1480 for (i
= 0; i
< (1 << btp
->bt_hashshift
); i
++) {
1481 spin_lock_init(&btp
->bt_hash
[i
].bh_lock
);
1482 INIT_LIST_HEAD(&btp
->bt_hash
[i
].bh_list
);
1490 kmem_free(btp
->bt_hash
,
1491 (1 << btp
->bt_hashshift
) * sizeof(xfs_bufhash_t
));
1492 btp
->bt_hash
= NULL
;
1500 xfs_flush_buftarg(btp
, 1);
1502 xfs_blkdev_put(btp
->pbr_bdev
);
1503 xfs_free_bufhash(btp
);
1504 iput(btp
->pbr_mapping
->host
);
1505 kmem_free(btp
, sizeof(*btp
));
1509 xfs_setsize_buftarg_flags(
1511 unsigned int blocksize
,
1512 unsigned int sectorsize
,
1515 btp
->pbr_bsize
= blocksize
;
1516 btp
->pbr_sshift
= ffs(sectorsize
) - 1;
1517 btp
->pbr_smask
= sectorsize
- 1;
1519 if (set_blocksize(btp
->pbr_bdev
, sectorsize
)) {
1521 "XFS: Cannot set_blocksize to %u on device %s\n",
1522 sectorsize
, XFS_BUFTARG_NAME(btp
));
1527 (PAGE_CACHE_SIZE
/ BITS_PER_LONG
) > sectorsize
) {
1529 "XFS: %u byte sectors in use on device %s. "
1530 "This is suboptimal; %u or greater is ideal.\n",
1531 sectorsize
, XFS_BUFTARG_NAME(btp
),
1532 (unsigned int)PAGE_CACHE_SIZE
/ BITS_PER_LONG
);
1539 * When allocating the initial buffer target we have not yet
1540 * read in the superblock, so don't know what sized sectors
1541 * are being used is at this early stage. Play safe.
1544 xfs_setsize_buftarg_early(
1546 struct block_device
*bdev
)
1548 return xfs_setsize_buftarg_flags(btp
,
1549 PAGE_CACHE_SIZE
, bdev_hardsect_size(bdev
), 0);
1553 xfs_setsize_buftarg(
1555 unsigned int blocksize
,
1556 unsigned int sectorsize
)
1558 return xfs_setsize_buftarg_flags(btp
, blocksize
, sectorsize
, 1);
1562 xfs_mapping_buftarg(
1564 struct block_device
*bdev
)
1566 struct backing_dev_info
*bdi
;
1567 struct inode
*inode
;
1568 struct address_space
*mapping
;
1569 static struct address_space_operations mapping_aops
= {
1570 .sync_page
= block_sync_page
,
1573 inode
= new_inode(bdev
->bd_inode
->i_sb
);
1576 "XFS: Cannot allocate mapping inode for device %s\n",
1577 XFS_BUFTARG_NAME(btp
));
1580 inode
->i_mode
= S_IFBLK
;
1581 inode
->i_bdev
= bdev
;
1582 inode
->i_rdev
= bdev
->bd_dev
;
1583 bdi
= blk_get_backing_dev_info(bdev
);
1585 bdi
= &default_backing_dev_info
;
1586 mapping
= &inode
->i_data
;
1587 mapping
->a_ops
= &mapping_aops
;
1588 mapping
->backing_dev_info
= bdi
;
1589 mapping_set_gfp_mask(mapping
, GFP_NOFS
);
1590 btp
->pbr_mapping
= mapping
;
1596 struct block_device
*bdev
,
1601 btp
= kmem_zalloc(sizeof(*btp
), KM_SLEEP
);
1603 btp
->pbr_dev
= bdev
->bd_dev
;
1604 btp
->pbr_bdev
= bdev
;
1605 if (xfs_setsize_buftarg_early(btp
, bdev
))
1607 if (xfs_mapping_buftarg(btp
, bdev
))
1609 xfs_alloc_bufhash(btp
, external
);
1613 kmem_free(btp
, sizeof(*btp
));
1619 * Pagebuf delayed write buffer handling
1622 STATIC
LIST_HEAD(pbd_delwrite_queue
);
1623 STATIC
DEFINE_SPINLOCK(pbd_delwrite_lock
);
1626 pagebuf_delwri_queue(
1630 PB_TRACE(pb
, "delwri_q", (long)unlock
);
1631 ASSERT((pb
->pb_flags
& (PBF_DELWRI
|PBF_ASYNC
)) ==
1632 (PBF_DELWRI
|PBF_ASYNC
));
1634 spin_lock(&pbd_delwrite_lock
);
1635 /* If already in the queue, dequeue and place at tail */
1636 if (!list_empty(&pb
->pb_list
)) {
1637 ASSERT(pb
->pb_flags
& _PBF_DELWRI_Q
);
1639 atomic_dec(&pb
->pb_hold
);
1641 list_del(&pb
->pb_list
);
1644 pb
->pb_flags
|= _PBF_DELWRI_Q
;
1645 list_add_tail(&pb
->pb_list
, &pbd_delwrite_queue
);
1646 pb
->pb_queuetime
= jiffies
;
1647 spin_unlock(&pbd_delwrite_lock
);
1654 pagebuf_delwri_dequeue(
1659 spin_lock(&pbd_delwrite_lock
);
1660 if ((pb
->pb_flags
& PBF_DELWRI
) && !list_empty(&pb
->pb_list
)) {
1661 ASSERT(pb
->pb_flags
& _PBF_DELWRI_Q
);
1662 list_del_init(&pb
->pb_list
);
1665 pb
->pb_flags
&= ~(PBF_DELWRI
|_PBF_DELWRI_Q
);
1666 spin_unlock(&pbd_delwrite_lock
);
1671 PB_TRACE(pb
, "delwri_dq", (long)dequeued
);
1675 pagebuf_runall_queues(
1676 struct workqueue_struct
*queue
)
1678 flush_workqueue(queue
);
1681 /* Defines for pagebuf daemon */
1682 STATIC
struct task_struct
*xfsbufd_task
;
1683 STATIC
int xfsbufd_force_flush
;
1684 STATIC
int xfsbufd_force_sleep
;
1691 if (xfsbufd_force_sleep
)
1693 xfsbufd_force_flush
= 1;
1695 wake_up_process(xfsbufd_task
);
1703 struct list_head tmp
;
1705 xfs_buftarg_t
*target
;
1708 current
->flags
|= PF_MEMALLOC
;
1710 INIT_LIST_HEAD(&tmp
);
1712 if (unlikely(freezing(current
))) {
1713 xfsbufd_force_sleep
= 1;
1716 xfsbufd_force_sleep
= 0;
1719 schedule_timeout_interruptible
1720 (xfs_buf_timer_centisecs
* msecs_to_jiffies(10));
1722 age
= xfs_buf_age_centisecs
* msecs_to_jiffies(10);
1723 spin_lock(&pbd_delwrite_lock
);
1724 list_for_each_entry_safe(pb
, n
, &pbd_delwrite_queue
, pb_list
) {
1725 PB_TRACE(pb
, "walkq1", (long)pagebuf_ispin(pb
));
1726 ASSERT(pb
->pb_flags
& PBF_DELWRI
);
1728 if (!pagebuf_ispin(pb
) && !pagebuf_cond_lock(pb
)) {
1729 if (!xfsbufd_force_flush
&&
1730 time_before(jiffies
,
1731 pb
->pb_queuetime
+ age
)) {
1736 pb
->pb_flags
&= ~(PBF_DELWRI
|_PBF_DELWRI_Q
);
1737 pb
->pb_flags
|= PBF_WRITE
;
1738 list_move(&pb
->pb_list
, &tmp
);
1741 spin_unlock(&pbd_delwrite_lock
);
1743 while (!list_empty(&tmp
)) {
1744 pb
= list_entry(tmp
.next
, xfs_buf_t
, pb_list
);
1745 target
= pb
->pb_target
;
1747 list_del_init(&pb
->pb_list
);
1748 pagebuf_iostrategy(pb
);
1750 blk_run_address_space(target
->pbr_mapping
);
1753 if (as_list_len
> 0)
1756 xfsbufd_force_flush
= 0;
1757 } while (!kthread_should_stop());
1763 * Go through all incore buffers, and release buffers if they belong to
1764 * the given device. This is used in filesystem error handling to
1765 * preserve the consistency of its metadata.
1769 xfs_buftarg_t
*target
,
1772 struct list_head tmp
;
1776 pagebuf_runall_queues(xfsdatad_workqueue
);
1777 pagebuf_runall_queues(xfslogd_workqueue
);
1779 INIT_LIST_HEAD(&tmp
);
1780 spin_lock(&pbd_delwrite_lock
);
1781 list_for_each_entry_safe(pb
, n
, &pbd_delwrite_queue
, pb_list
) {
1783 if (pb
->pb_target
!= target
)
1786 ASSERT(pb
->pb_flags
& (PBF_DELWRI
|_PBF_DELWRI_Q
));
1787 PB_TRACE(pb
, "walkq2", (long)pagebuf_ispin(pb
));
1788 if (pagebuf_ispin(pb
)) {
1793 list_move(&pb
->pb_list
, &tmp
);
1795 spin_unlock(&pbd_delwrite_lock
);
1798 * Dropped the delayed write list lock, now walk the temporary list
1800 list_for_each_entry_safe(pb
, n
, &tmp
, pb_list
) {
1802 pb
->pb_flags
&= ~(PBF_DELWRI
|_PBF_DELWRI_Q
);
1803 pb
->pb_flags
|= PBF_WRITE
;
1805 pb
->pb_flags
&= ~PBF_ASYNC
;
1807 list_del_init(&pb
->pb_list
);
1809 pagebuf_iostrategy(pb
);
1813 * Remaining list items must be flushed before returning
1815 while (!list_empty(&tmp
)) {
1816 pb
= list_entry(tmp
.next
, xfs_buf_t
, pb_list
);
1818 list_del_init(&pb
->pb_list
);
1824 blk_run_address_space(target
->pbr_mapping
);
1832 int error
= -ENOMEM
;
1834 #ifdef PAGEBUF_TRACE
1835 pagebuf_trace_buf
= ktrace_alloc(PAGEBUF_TRACE_SIZE
, KM_SLEEP
);
1838 pagebuf_zone
= kmem_zone_init(sizeof(xfs_buf_t
), "xfs_buf");
1840 goto out_free_trace_buf
;
1842 xfslogd_workqueue
= create_workqueue("xfslogd");
1843 if (!xfslogd_workqueue
)
1844 goto out_free_buf_zone
;
1846 xfsdatad_workqueue
= create_workqueue("xfsdatad");
1847 if (!xfsdatad_workqueue
)
1848 goto out_destroy_xfslogd_workqueue
;
1850 xfsbufd_task
= kthread_run(xfsbufd
, NULL
, "xfsbufd");
1851 if (IS_ERR(xfsbufd_task
)) {
1852 error
= PTR_ERR(xfsbufd_task
);
1853 goto out_destroy_xfsdatad_workqueue
;
1856 pagebuf_shake
= kmem_shake_register(xfsbufd_wakeup
);
1858 goto out_stop_xfsbufd
;
1863 kthread_stop(xfsbufd_task
);
1864 out_destroy_xfsdatad_workqueue
:
1865 destroy_workqueue(xfsdatad_workqueue
);
1866 out_destroy_xfslogd_workqueue
:
1867 destroy_workqueue(xfslogd_workqueue
);
1869 kmem_zone_destroy(pagebuf_zone
);
1871 #ifdef PAGEBUF_TRACE
1872 ktrace_free(pagebuf_trace_buf
);
1878 pagebuf_terminate(void)
1880 kmem_shake_deregister(pagebuf_shake
);
1881 kthread_stop(xfsbufd_task
);
1882 destroy_workqueue(xfsdatad_workqueue
);
1883 destroy_workqueue(xfslogd_workqueue
);
1884 kmem_zone_destroy(pagebuf_zone
);
1885 #ifdef PAGEBUF_TRACE
1886 ktrace_free(pagebuf_trace_buf
);