2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
19 #include <linux/stddef.h>
20 #include <linux/errno.h>
21 #include <linux/gfp.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/vmalloc.h>
25 #include <linux/bio.h>
26 #include <linux/sysctl.h>
27 #include <linux/proc_fs.h>
28 #include <linux/workqueue.h>
29 #include <linux/percpu.h>
30 #include <linux/blkdev.h>
31 #include <linux/hash.h>
32 #include <linux/kthread.h>
33 #include <linux/migrate.h>
34 #include <linux/backing-dev.h>
35 #include <linux/freezer.h>
41 #include "xfs_mount.h"
42 #include "xfs_trace.h"
44 static kmem_zone_t
*xfs_buf_zone
;
46 static struct workqueue_struct
*xfslogd_workqueue
;
48 #ifdef XFS_BUF_LOCK_TRACKING
49 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
50 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
51 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
53 # define XB_SET_OWNER(bp) do { } while (0)
54 # define XB_CLEAR_OWNER(bp) do { } while (0)
55 # define XB_GET_OWNER(bp) do { } while (0)
58 #define xb_to_gfp(flags) \
59 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
60 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
62 #define xb_to_km(flags) \
63 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
71 * Return true if the buffer is vmapped.
73 * The XBF_MAPPED flag is set if the buffer should be mapped, but the
74 * code is clever enough to know it doesn't have to map a single page,
75 * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
77 return (bp
->b_flags
& XBF_MAPPED
) && bp
->b_page_count
> 1;
84 return (bp
->b_page_count
* PAGE_SIZE
) - bp
->b_offset
;
88 * xfs_buf_lru_add - add a buffer to the LRU.
90 * The LRU takes a new reference to the buffer so that it will only be freed
91 * once the shrinker takes the buffer off the LRU.
97 struct xfs_buftarg
*btp
= bp
->b_target
;
99 spin_lock(&btp
->bt_lru_lock
);
100 if (list_empty(&bp
->b_lru
)) {
101 atomic_inc(&bp
->b_hold
);
102 list_add_tail(&bp
->b_lru
, &btp
->bt_lru
);
105 spin_unlock(&btp
->bt_lru_lock
);
109 * xfs_buf_lru_del - remove a buffer from the LRU
111 * The unlocked check is safe here because it only occurs when there are not
112 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
113 * to optimise the shrinker removing the buffer from the LRU and calling
114 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
121 struct xfs_buftarg
*btp
= bp
->b_target
;
123 if (list_empty(&bp
->b_lru
))
126 spin_lock(&btp
->bt_lru_lock
);
127 if (!list_empty(&bp
->b_lru
)) {
128 list_del_init(&bp
->b_lru
);
131 spin_unlock(&btp
->bt_lru_lock
);
135 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
136 * b_lru_ref count so that the buffer is freed immediately when the buffer
137 * reference count falls to zero. If the buffer is already on the LRU, we need
138 * to remove the reference that LRU holds on the buffer.
140 * This prevents build-up of stale buffers on the LRU.
146 ASSERT(xfs_buf_islocked(bp
));
148 bp
->b_flags
|= XBF_STALE
;
151 * Clear the delwri status so that a delwri queue walker will not
152 * flush this buffer to disk now that it is stale. The delwri queue has
153 * a reference to the buffer, so this is safe to do.
155 bp
->b_flags
&= ~_XBF_DELWRI_Q
;
157 atomic_set(&(bp
)->b_lru_ref
, 0);
158 if (!list_empty(&bp
->b_lru
)) {
159 struct xfs_buftarg
*btp
= bp
->b_target
;
161 spin_lock(&btp
->bt_lru_lock
);
162 if (!list_empty(&bp
->b_lru
)) {
163 list_del_init(&bp
->b_lru
);
165 atomic_dec(&bp
->b_hold
);
167 spin_unlock(&btp
->bt_lru_lock
);
169 ASSERT(atomic_read(&bp
->b_hold
) >= 1);
174 struct xfs_buftarg
*target
,
175 xfs_off_t range_base
,
177 xfs_buf_flags_t flags
)
181 bp
= kmem_zone_alloc(xfs_buf_zone
, xb_to_km(flags
));
186 * We don't want certain flags to appear in b_flags.
188 flags
&= ~(XBF_LOCK
|XBF_MAPPED
|XBF_DONT_BLOCK
|XBF_READ_AHEAD
);
190 memset(bp
, 0, sizeof(xfs_buf_t
));
191 atomic_set(&bp
->b_hold
, 1);
192 atomic_set(&bp
->b_lru_ref
, 1);
193 init_completion(&bp
->b_iowait
);
194 INIT_LIST_HEAD(&bp
->b_lru
);
195 INIT_LIST_HEAD(&bp
->b_list
);
196 RB_CLEAR_NODE(&bp
->b_rbnode
);
197 sema_init(&bp
->b_sema
, 0); /* held, no waiters */
199 bp
->b_target
= target
;
200 bp
->b_file_offset
= range_base
;
202 * Set buffer_length and count_desired to the same value initially.
203 * I/O routines should use count_desired, which will be the same in
204 * most cases but may be reset (e.g. XFS recovery).
206 bp
->b_buffer_length
= bp
->b_count_desired
= range_length
;
208 bp
->b_bn
= XFS_BUF_DADDR_NULL
;
209 atomic_set(&bp
->b_pin_count
, 0);
210 init_waitqueue_head(&bp
->b_waiters
);
212 XFS_STATS_INC(xb_create
);
213 trace_xfs_buf_init(bp
, _RET_IP_
);
219 * Allocate a page array capable of holding a specified number
220 * of pages, and point the page buf at it.
226 xfs_buf_flags_t flags
)
228 /* Make sure that we have a page list */
229 if (bp
->b_pages
== NULL
) {
230 bp
->b_page_count
= page_count
;
231 if (page_count
<= XB_PAGES
) {
232 bp
->b_pages
= bp
->b_page_array
;
234 bp
->b_pages
= kmem_alloc(sizeof(struct page
*) *
235 page_count
, xb_to_km(flags
));
236 if (bp
->b_pages
== NULL
)
239 memset(bp
->b_pages
, 0, sizeof(struct page
*) * page_count
);
245 * Frees b_pages if it was allocated.
251 if (bp
->b_pages
!= bp
->b_page_array
) {
252 kmem_free(bp
->b_pages
);
258 * Releases the specified buffer.
260 * The modification state of any associated pages is left unchanged.
261 * The buffer most not be on any hash - use xfs_buf_rele instead for
262 * hashed and refcounted buffers
268 trace_xfs_buf_free(bp
, _RET_IP_
);
270 ASSERT(list_empty(&bp
->b_lru
));
272 if (bp
->b_flags
& _XBF_PAGES
) {
275 if (xfs_buf_is_vmapped(bp
))
276 vm_unmap_ram(bp
->b_addr
- bp
->b_offset
,
279 for (i
= 0; i
< bp
->b_page_count
; i
++) {
280 struct page
*page
= bp
->b_pages
[i
];
284 } else if (bp
->b_flags
& _XBF_KMEM
)
285 kmem_free(bp
->b_addr
);
286 _xfs_buf_free_pages(bp
);
287 kmem_zone_free(xfs_buf_zone
, bp
);
291 * Allocates all the pages for buffer in question and builds it's page list.
294 xfs_buf_allocate_memory(
298 size_t size
= bp
->b_count_desired
;
299 size_t nbytes
, offset
;
300 gfp_t gfp_mask
= xb_to_gfp(flags
);
301 unsigned short page_count
, i
;
306 * for buffers that are contained within a single page, just allocate
307 * the memory from the heap - there's no need for the complexity of
308 * page arrays to keep allocation down to order 0.
310 if (bp
->b_buffer_length
< PAGE_SIZE
) {
311 bp
->b_addr
= kmem_alloc(bp
->b_buffer_length
, xb_to_km(flags
));
313 /* low memory - use alloc_page loop instead */
317 if (((unsigned long)(bp
->b_addr
+ bp
->b_buffer_length
- 1) &
319 ((unsigned long)bp
->b_addr
& PAGE_MASK
)) {
320 /* b_addr spans two pages - use alloc_page instead */
321 kmem_free(bp
->b_addr
);
325 bp
->b_offset
= offset_in_page(bp
->b_addr
);
326 bp
->b_pages
= bp
->b_page_array
;
327 bp
->b_pages
[0] = virt_to_page(bp
->b_addr
);
328 bp
->b_page_count
= 1;
329 bp
->b_flags
|= XBF_MAPPED
| _XBF_KMEM
;
334 end
= bp
->b_file_offset
+ bp
->b_buffer_length
;
335 page_count
= xfs_buf_btoc(end
) - xfs_buf_btoct(bp
->b_file_offset
);
336 error
= _xfs_buf_get_pages(bp
, page_count
, flags
);
340 offset
= bp
->b_offset
;
341 bp
->b_flags
|= _XBF_PAGES
;
343 for (i
= 0; i
< bp
->b_page_count
; i
++) {
347 page
= alloc_page(gfp_mask
);
348 if (unlikely(page
== NULL
)) {
349 if (flags
& XBF_READ_AHEAD
) {
350 bp
->b_page_count
= i
;
356 * This could deadlock.
358 * But until all the XFS lowlevel code is revamped to
359 * handle buffer allocation failures we can't do much.
361 if (!(++retries
% 100))
363 "possible memory allocation deadlock in %s (mode:0x%x)",
366 XFS_STATS_INC(xb_page_retries
);
367 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
371 XFS_STATS_INC(xb_page_found
);
373 nbytes
= min_t(size_t, size
, PAGE_SIZE
- offset
);
375 bp
->b_pages
[i
] = page
;
381 for (i
= 0; i
< bp
->b_page_count
; i
++)
382 __free_page(bp
->b_pages
[i
]);
387 * Map buffer into kernel address-space if necessary.
394 ASSERT(bp
->b_flags
& _XBF_PAGES
);
395 if (bp
->b_page_count
== 1) {
396 /* A single page buffer is always mappable */
397 bp
->b_addr
= page_address(bp
->b_pages
[0]) + bp
->b_offset
;
398 bp
->b_flags
|= XBF_MAPPED
;
399 } else if (flags
& XBF_MAPPED
) {
403 bp
->b_addr
= vm_map_ram(bp
->b_pages
, bp
->b_page_count
,
408 } while (retried
++ <= 1);
412 bp
->b_addr
+= bp
->b_offset
;
413 bp
->b_flags
|= XBF_MAPPED
;
420 * Finding and Reading Buffers
424 * Look up, and creates if absent, a lockable buffer for
425 * a given range of an inode. The buffer is returned
426 * locked. No I/O is implied by this call.
430 xfs_buftarg_t
*btp
, /* block device target */
431 xfs_off_t ioff
, /* starting offset of range */
432 size_t isize
, /* length of range */
433 xfs_buf_flags_t flags
,
436 xfs_off_t range_base
;
438 struct xfs_perag
*pag
;
439 struct rb_node
**rbp
;
440 struct rb_node
*parent
;
443 range_base
= (ioff
<< BBSHIFT
);
444 range_length
= (isize
<< BBSHIFT
);
446 /* Check for IOs smaller than the sector size / not sector aligned */
447 ASSERT(!(range_length
< (1 << btp
->bt_sshift
)));
448 ASSERT(!(range_base
& (xfs_off_t
)btp
->bt_smask
));
451 pag
= xfs_perag_get(btp
->bt_mount
,
452 xfs_daddr_to_agno(btp
->bt_mount
, ioff
));
455 spin_lock(&pag
->pag_buf_lock
);
456 rbp
= &pag
->pag_buf_tree
.rb_node
;
461 bp
= rb_entry(parent
, struct xfs_buf
, b_rbnode
);
463 if (range_base
< bp
->b_file_offset
)
464 rbp
= &(*rbp
)->rb_left
;
465 else if (range_base
> bp
->b_file_offset
)
466 rbp
= &(*rbp
)->rb_right
;
469 * found a block offset match. If the range doesn't
470 * match, the only way this is allowed is if the buffer
471 * in the cache is stale and the transaction that made
472 * it stale has not yet committed. i.e. we are
473 * reallocating a busy extent. Skip this buffer and
474 * continue searching to the right for an exact match.
476 if (bp
->b_buffer_length
!= range_length
) {
477 ASSERT(bp
->b_flags
& XBF_STALE
);
478 rbp
= &(*rbp
)->rb_right
;
481 atomic_inc(&bp
->b_hold
);
488 rb_link_node(&new_bp
->b_rbnode
, parent
, rbp
);
489 rb_insert_color(&new_bp
->b_rbnode
, &pag
->pag_buf_tree
);
490 /* the buffer keeps the perag reference until it is freed */
492 spin_unlock(&pag
->pag_buf_lock
);
494 XFS_STATS_INC(xb_miss_locked
);
495 spin_unlock(&pag
->pag_buf_lock
);
501 spin_unlock(&pag
->pag_buf_lock
);
504 if (!xfs_buf_trylock(bp
)) {
505 if (flags
& XBF_TRYLOCK
) {
507 XFS_STATS_INC(xb_busy_locked
);
511 XFS_STATS_INC(xb_get_locked_waited
);
515 * if the buffer is stale, clear all the external state associated with
516 * it. We need to keep flags such as how we allocated the buffer memory
519 if (bp
->b_flags
& XBF_STALE
) {
520 ASSERT((bp
->b_flags
& _XBF_DELWRI_Q
) == 0);
521 bp
->b_flags
&= XBF_MAPPED
| _XBF_KMEM
| _XBF_PAGES
;
524 trace_xfs_buf_find(bp
, flags
, _RET_IP_
);
525 XFS_STATS_INC(xb_get_locked
);
530 * Assembles a buffer covering the specified range. The code is optimised for
531 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
532 * more hits than misses.
536 xfs_buftarg_t
*target
,/* target for buffer */
537 xfs_off_t ioff
, /* starting offset of range */
538 size_t isize
, /* length of range */
539 xfs_buf_flags_t flags
)
542 struct xfs_buf
*new_bp
;
545 bp
= _xfs_buf_find(target
, ioff
, isize
, flags
, NULL
);
549 new_bp
= xfs_buf_alloc(target
, ioff
<< BBSHIFT
, isize
<< BBSHIFT
,
551 if (unlikely(!new_bp
))
554 error
= xfs_buf_allocate_memory(new_bp
, flags
);
556 kmem_zone_free(xfs_buf_zone
, new_bp
);
560 bp
= _xfs_buf_find(target
, ioff
, isize
, flags
, new_bp
);
562 xfs_buf_free(new_bp
);
567 xfs_buf_free(new_bp
);
570 * Now we have a workable buffer, fill in the block number so
571 * that we can do IO on it.
574 bp
->b_count_desired
= bp
->b_buffer_length
;
577 if (!(bp
->b_flags
& XBF_MAPPED
)) {
578 error
= _xfs_buf_map_pages(bp
, flags
);
579 if (unlikely(error
)) {
580 xfs_warn(target
->bt_mount
,
581 "%s: failed to map pages\n", __func__
);
586 XFS_STATS_INC(xb_get
);
587 trace_xfs_buf_get(bp
, flags
, _RET_IP_
);
591 if (flags
& (XBF_LOCK
| XBF_TRYLOCK
))
600 xfs_buf_flags_t flags
)
602 ASSERT(!(flags
& XBF_WRITE
));
603 ASSERT(bp
->b_bn
!= XFS_BUF_DADDR_NULL
);
605 bp
->b_flags
&= ~(XBF_WRITE
| XBF_ASYNC
| XBF_READ_AHEAD
);
606 bp
->b_flags
|= flags
& (XBF_READ
| XBF_ASYNC
| XBF_READ_AHEAD
);
608 xfs_buf_iorequest(bp
);
609 if (flags
& XBF_ASYNC
)
611 return xfs_buf_iowait(bp
);
616 xfs_buftarg_t
*target
,
619 xfs_buf_flags_t flags
)
625 bp
= xfs_buf_get(target
, ioff
, isize
, flags
);
627 trace_xfs_buf_read(bp
, flags
, _RET_IP_
);
629 if (!XFS_BUF_ISDONE(bp
)) {
630 XFS_STATS_INC(xb_get_read
);
631 _xfs_buf_read(bp
, flags
);
632 } else if (flags
& XBF_ASYNC
) {
634 * Read ahead call which is already satisfied,
639 /* We do not want read in the flags */
640 bp
->b_flags
&= ~XBF_READ
;
647 if (flags
& (XBF_LOCK
| XBF_TRYLOCK
))
654 * If we are not low on memory then do the readahead in a deadlock
659 xfs_buftarg_t
*target
,
663 if (bdi_read_congested(target
->bt_bdi
))
666 xfs_buf_read(target
, ioff
, isize
,
667 XBF_TRYLOCK
|XBF_ASYNC
|XBF_READ_AHEAD
|XBF_DONT_BLOCK
);
671 * Read an uncached buffer from disk. Allocates and returns a locked
672 * buffer containing the disk contents or nothing.
675 xfs_buf_read_uncached(
676 struct xfs_mount
*mp
,
677 struct xfs_buftarg
*target
,
685 bp
= xfs_buf_get_uncached(target
, length
, flags
);
689 /* set up the buffer for a read IO */
690 XFS_BUF_SET_ADDR(bp
, daddr
);
694 error
= xfs_buf_iowait(bp
);
703 * Return a buffer allocated as an empty buffer and associated to external
704 * memory via xfs_buf_associate_memory() back to it's empty state.
712 _xfs_buf_free_pages(bp
);
715 bp
->b_page_count
= 0;
717 bp
->b_file_offset
= 0;
718 bp
->b_buffer_length
= bp
->b_count_desired
= len
;
719 bp
->b_bn
= XFS_BUF_DADDR_NULL
;
720 bp
->b_flags
&= ~XBF_MAPPED
;
723 static inline struct page
*
727 if ((!is_vmalloc_addr(addr
))) {
728 return virt_to_page(addr
);
730 return vmalloc_to_page(addr
);
735 xfs_buf_associate_memory(
742 unsigned long pageaddr
;
743 unsigned long offset
;
747 pageaddr
= (unsigned long)mem
& PAGE_MASK
;
748 offset
= (unsigned long)mem
- pageaddr
;
749 buflen
= PAGE_ALIGN(len
+ offset
);
750 page_count
= buflen
>> PAGE_SHIFT
;
752 /* Free any previous set of page pointers */
754 _xfs_buf_free_pages(bp
);
759 rval
= _xfs_buf_get_pages(bp
, page_count
, XBF_DONT_BLOCK
);
763 bp
->b_offset
= offset
;
765 for (i
= 0; i
< bp
->b_page_count
; i
++) {
766 bp
->b_pages
[i
] = mem_to_page((void *)pageaddr
);
767 pageaddr
+= PAGE_SIZE
;
770 bp
->b_count_desired
= len
;
771 bp
->b_buffer_length
= buflen
;
772 bp
->b_flags
|= XBF_MAPPED
;
778 xfs_buf_get_uncached(
779 struct xfs_buftarg
*target
,
783 unsigned long page_count
= PAGE_ALIGN(len
) >> PAGE_SHIFT
;
787 bp
= xfs_buf_alloc(target
, 0, len
, 0);
788 if (unlikely(bp
== NULL
))
791 error
= _xfs_buf_get_pages(bp
, page_count
, 0);
795 for (i
= 0; i
< page_count
; i
++) {
796 bp
->b_pages
[i
] = alloc_page(xb_to_gfp(flags
));
800 bp
->b_flags
|= _XBF_PAGES
;
802 error
= _xfs_buf_map_pages(bp
, XBF_MAPPED
);
803 if (unlikely(error
)) {
804 xfs_warn(target
->bt_mount
,
805 "%s: failed to map pages\n", __func__
);
809 trace_xfs_buf_get_uncached(bp
, _RET_IP_
);
814 __free_page(bp
->b_pages
[i
]);
815 _xfs_buf_free_pages(bp
);
817 kmem_zone_free(xfs_buf_zone
, bp
);
823 * Increment reference count on buffer, to hold the buffer concurrently
824 * with another thread which may release (free) the buffer asynchronously.
825 * Must hold the buffer already to call this function.
831 trace_xfs_buf_hold(bp
, _RET_IP_
);
832 atomic_inc(&bp
->b_hold
);
836 * Releases a hold on the specified buffer. If the
837 * the hold count is 1, calls xfs_buf_free.
843 struct xfs_perag
*pag
= bp
->b_pag
;
845 trace_xfs_buf_rele(bp
, _RET_IP_
);
848 ASSERT(list_empty(&bp
->b_lru
));
849 ASSERT(RB_EMPTY_NODE(&bp
->b_rbnode
));
850 if (atomic_dec_and_test(&bp
->b_hold
))
855 ASSERT(!RB_EMPTY_NODE(&bp
->b_rbnode
));
857 ASSERT(atomic_read(&bp
->b_hold
) > 0);
858 if (atomic_dec_and_lock(&bp
->b_hold
, &pag
->pag_buf_lock
)) {
859 if (!(bp
->b_flags
& XBF_STALE
) &&
860 atomic_read(&bp
->b_lru_ref
)) {
862 spin_unlock(&pag
->pag_buf_lock
);
865 ASSERT(!(bp
->b_flags
& _XBF_DELWRI_Q
));
866 rb_erase(&bp
->b_rbnode
, &pag
->pag_buf_tree
);
867 spin_unlock(&pag
->pag_buf_lock
);
876 * Lock a buffer object, if it is not already locked.
878 * If we come across a stale, pinned, locked buffer, we know that we are
879 * being asked to lock a buffer that has been reallocated. Because it is
880 * pinned, we know that the log has not been pushed to disk and hence it
881 * will still be locked. Rather than continuing to have trylock attempts
882 * fail until someone else pushes the log, push it ourselves before
883 * returning. This means that the xfsaild will not get stuck trying
884 * to push on stale inode buffers.
892 locked
= down_trylock(&bp
->b_sema
) == 0;
895 else if (atomic_read(&bp
->b_pin_count
) && (bp
->b_flags
& XBF_STALE
))
896 xfs_log_force(bp
->b_target
->bt_mount
, 0);
898 trace_xfs_buf_trylock(bp
, _RET_IP_
);
903 * Lock a buffer object.
905 * If we come across a stale, pinned, locked buffer, we know that we
906 * are being asked to lock a buffer that has been reallocated. Because
907 * it is pinned, we know that the log has not been pushed to disk and
908 * hence it will still be locked. Rather than sleeping until someone
909 * else pushes the log, push it ourselves before trying to get the lock.
915 trace_xfs_buf_lock(bp
, _RET_IP_
);
917 if (atomic_read(&bp
->b_pin_count
) && (bp
->b_flags
& XBF_STALE
))
918 xfs_log_force(bp
->b_target
->bt_mount
, 0);
922 trace_xfs_buf_lock_done(bp
, _RET_IP_
);
932 trace_xfs_buf_unlock(bp
, _RET_IP_
);
939 DECLARE_WAITQUEUE (wait
, current
);
941 if (atomic_read(&bp
->b_pin_count
) == 0)
944 add_wait_queue(&bp
->b_waiters
, &wait
);
946 set_current_state(TASK_UNINTERRUPTIBLE
);
947 if (atomic_read(&bp
->b_pin_count
) == 0)
951 remove_wait_queue(&bp
->b_waiters
, &wait
);
952 set_current_state(TASK_RUNNING
);
956 * Buffer Utility Routines
961 struct work_struct
*work
)
964 container_of(work
, xfs_buf_t
, b_iodone_work
);
967 (*(bp
->b_iodone
))(bp
);
968 else if (bp
->b_flags
& XBF_ASYNC
)
977 trace_xfs_buf_iodone(bp
, _RET_IP_
);
979 bp
->b_flags
&= ~(XBF_READ
| XBF_WRITE
| XBF_READ_AHEAD
);
980 if (bp
->b_error
== 0)
981 bp
->b_flags
|= XBF_DONE
;
983 if ((bp
->b_iodone
) || (bp
->b_flags
& XBF_ASYNC
)) {
985 INIT_WORK(&bp
->b_iodone_work
, xfs_buf_iodone_work
);
986 queue_work(xfslogd_workqueue
, &bp
->b_iodone_work
);
988 xfs_buf_iodone_work(&bp
->b_iodone_work
);
991 complete(&bp
->b_iowait
);
1000 ASSERT(error
>= 0 && error
<= 0xffff);
1001 bp
->b_error
= (unsigned short)error
;
1002 trace_xfs_buf_ioerror(bp
, error
, _RET_IP_
);
1006 xfs_buf_ioerror_alert(
1010 xfs_alert(bp
->b_target
->bt_mount
,
1011 "metadata I/O error: block 0x%llx (\"%s\") error %d buf count %zd",
1012 (__uint64_t
)XFS_BUF_ADDR(bp
), func
,
1013 bp
->b_error
, XFS_BUF_COUNT(bp
));
1022 ASSERT(xfs_buf_islocked(bp
));
1024 bp
->b_flags
|= XBF_WRITE
;
1025 bp
->b_flags
&= ~(XBF_ASYNC
| XBF_READ
| _XBF_DELWRI_Q
);
1029 error
= xfs_buf_iowait(bp
);
1031 xfs_force_shutdown(bp
->b_target
->bt_mount
,
1032 SHUTDOWN_META_IO_ERROR
);
1038 * Called when we want to stop a buffer from getting written or read.
1039 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1040 * so that the proper iodone callbacks get called.
1046 #ifdef XFSERRORDEBUG
1047 ASSERT(XFS_BUF_ISREAD(bp
) || bp
->b_iodone
);
1051 * No need to wait until the buffer is unpinned, we aren't flushing it.
1053 xfs_buf_ioerror(bp
, EIO
);
1056 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1062 xfs_buf_ioend(bp
, 0);
1068 * Same as xfs_bioerror, except that we are releasing the buffer
1069 * here ourselves, and avoiding the xfs_buf_ioend call.
1070 * This is meant for userdata errors; metadata bufs come with
1071 * iodone functions attached, so that we can track down errors.
1077 int64_t fl
= bp
->b_flags
;
1079 * No need to wait until the buffer is unpinned.
1080 * We aren't flushing it.
1082 * chunkhold expects B_DONE to be set, whether
1083 * we actually finish the I/O or not. We don't want to
1084 * change that interface.
1089 bp
->b_iodone
= NULL
;
1090 if (!(fl
& XBF_ASYNC
)) {
1092 * Mark b_error and B_ERROR _both_.
1093 * Lot's of chunkcache code assumes that.
1094 * There's no reason to mark error for
1097 xfs_buf_ioerror(bp
, EIO
);
1098 complete(&bp
->b_iowait
);
1108 * All xfs metadata buffers except log state machine buffers
1109 * get this attached as their b_bdstrat callback function.
1110 * This is so that we can catch a buffer
1111 * after prematurely unpinning it to forcibly shutdown the filesystem.
1117 if (XFS_FORCED_SHUTDOWN(bp
->b_target
->bt_mount
)) {
1118 trace_xfs_bdstrat_shut(bp
, _RET_IP_
);
1120 * Metadata write that didn't get logged but
1121 * written delayed anyway. These aren't associated
1122 * with a transaction, and can be ignored.
1124 if (!bp
->b_iodone
&& !XFS_BUF_ISREAD(bp
))
1125 return xfs_bioerror_relse(bp
);
1127 return xfs_bioerror(bp
);
1130 xfs_buf_iorequest(bp
);
1135 * Wrapper around bdstrat so that we can stop data from going to disk in case
1136 * we are shutting down the filesystem. Typically user data goes thru this
1137 * path; one of the exceptions is the superblock.
1141 struct xfs_mount
*mp
,
1144 if (XFS_FORCED_SHUTDOWN(mp
)) {
1145 trace_xfs_bdstrat_shut(bp
, _RET_IP_
);
1146 xfs_bioerror_relse(bp
);
1150 xfs_buf_iorequest(bp
);
1158 if (atomic_dec_and_test(&bp
->b_io_remaining
) == 1)
1159 xfs_buf_ioend(bp
, schedule
);
1167 xfs_buf_t
*bp
= (xfs_buf_t
*)bio
->bi_private
;
1169 xfs_buf_ioerror(bp
, -error
);
1171 if (!error
&& xfs_buf_is_vmapped(bp
) && (bp
->b_flags
& XBF_READ
))
1172 invalidate_kernel_vmap_range(bp
->b_addr
, xfs_buf_vmap_len(bp
));
1174 _xfs_buf_ioend(bp
, 1);
1182 int rw
, map_i
, total_nr_pages
, nr_pages
;
1184 int offset
= bp
->b_offset
;
1185 int size
= bp
->b_count_desired
;
1186 sector_t sector
= bp
->b_bn
;
1188 total_nr_pages
= bp
->b_page_count
;
1191 if (bp
->b_flags
& XBF_WRITE
) {
1192 if (bp
->b_flags
& XBF_SYNCIO
)
1196 if (bp
->b_flags
& XBF_FUA
)
1198 if (bp
->b_flags
& XBF_FLUSH
)
1200 } else if (bp
->b_flags
& XBF_READ_AHEAD
) {
1206 /* we only use the buffer cache for meta-data */
1210 atomic_inc(&bp
->b_io_remaining
);
1211 nr_pages
= BIO_MAX_SECTORS
>> (PAGE_SHIFT
- BBSHIFT
);
1212 if (nr_pages
> total_nr_pages
)
1213 nr_pages
= total_nr_pages
;
1215 bio
= bio_alloc(GFP_NOIO
, nr_pages
);
1216 bio
->bi_bdev
= bp
->b_target
->bt_bdev
;
1217 bio
->bi_sector
= sector
;
1218 bio
->bi_end_io
= xfs_buf_bio_end_io
;
1219 bio
->bi_private
= bp
;
1222 for (; size
&& nr_pages
; nr_pages
--, map_i
++) {
1223 int rbytes
, nbytes
= PAGE_SIZE
- offset
;
1228 rbytes
= bio_add_page(bio
, bp
->b_pages
[map_i
], nbytes
, offset
);
1229 if (rbytes
< nbytes
)
1233 sector
+= nbytes
>> BBSHIFT
;
1238 if (likely(bio
->bi_size
)) {
1239 if (xfs_buf_is_vmapped(bp
)) {
1240 flush_kernel_vmap_range(bp
->b_addr
,
1241 xfs_buf_vmap_len(bp
));
1243 submit_bio(rw
, bio
);
1247 xfs_buf_ioerror(bp
, EIO
);
1256 trace_xfs_buf_iorequest(bp
, _RET_IP_
);
1258 ASSERT(!(bp
->b_flags
& _XBF_DELWRI_Q
));
1260 if (bp
->b_flags
& XBF_WRITE
)
1261 xfs_buf_wait_unpin(bp
);
1264 /* Set the count to 1 initially, this will stop an I/O
1265 * completion callout which happens before we have started
1266 * all the I/O from calling xfs_buf_ioend too early.
1268 atomic_set(&bp
->b_io_remaining
, 1);
1269 _xfs_buf_ioapply(bp
);
1270 _xfs_buf_ioend(bp
, 0);
1276 * Waits for I/O to complete on the buffer supplied. It returns immediately if
1277 * no I/O is pending or there is already a pending error on the buffer. It
1278 * returns the I/O error code, if any, or 0 if there was no error.
1284 trace_xfs_buf_iowait(bp
, _RET_IP_
);
1287 wait_for_completion(&bp
->b_iowait
);
1289 trace_xfs_buf_iowait_done(bp
, _RET_IP_
);
1300 if (bp
->b_flags
& XBF_MAPPED
)
1301 return bp
->b_addr
+ offset
;
1303 offset
+= bp
->b_offset
;
1304 page
= bp
->b_pages
[offset
>> PAGE_SHIFT
];
1305 return (xfs_caddr_t
)page_address(page
) + (offset
& (PAGE_SIZE
-1));
1309 * Move data into or out of a buffer.
1313 xfs_buf_t
*bp
, /* buffer to process */
1314 size_t boff
, /* starting buffer offset */
1315 size_t bsize
, /* length to copy */
1316 void *data
, /* data address */
1317 xfs_buf_rw_t mode
) /* read/write/zero flag */
1319 size_t bend
, cpoff
, csize
;
1322 bend
= boff
+ bsize
;
1323 while (boff
< bend
) {
1324 page
= bp
->b_pages
[xfs_buf_btoct(boff
+ bp
->b_offset
)];
1325 cpoff
= xfs_buf_poff(boff
+ bp
->b_offset
);
1326 csize
= min_t(size_t,
1327 PAGE_SIZE
-cpoff
, bp
->b_count_desired
-boff
);
1329 ASSERT(((csize
+ cpoff
) <= PAGE_SIZE
));
1333 memset(page_address(page
) + cpoff
, 0, csize
);
1336 memcpy(data
, page_address(page
) + cpoff
, csize
);
1339 memcpy(page_address(page
) + cpoff
, data
, csize
);
1348 * Handling of buffer targets (buftargs).
1352 * Wait for any bufs with callbacks that have been submitted but have not yet
1353 * returned. These buffers will have an elevated hold count, so wait on those
1354 * while freeing all the buffers only held by the LRU.
1358 struct xfs_buftarg
*btp
)
1363 spin_lock(&btp
->bt_lru_lock
);
1364 while (!list_empty(&btp
->bt_lru
)) {
1365 bp
= list_first_entry(&btp
->bt_lru
, struct xfs_buf
, b_lru
);
1366 if (atomic_read(&bp
->b_hold
) > 1) {
1367 spin_unlock(&btp
->bt_lru_lock
);
1372 * clear the LRU reference count so the buffer doesn't get
1373 * ignored in xfs_buf_rele().
1375 atomic_set(&bp
->b_lru_ref
, 0);
1376 spin_unlock(&btp
->bt_lru_lock
);
1378 spin_lock(&btp
->bt_lru_lock
);
1380 spin_unlock(&btp
->bt_lru_lock
);
1385 struct shrinker
*shrink
,
1386 struct shrink_control
*sc
)
1388 struct xfs_buftarg
*btp
= container_of(shrink
,
1389 struct xfs_buftarg
, bt_shrinker
);
1391 int nr_to_scan
= sc
->nr_to_scan
;
1395 return btp
->bt_lru_nr
;
1397 spin_lock(&btp
->bt_lru_lock
);
1398 while (!list_empty(&btp
->bt_lru
)) {
1399 if (nr_to_scan
-- <= 0)
1402 bp
= list_first_entry(&btp
->bt_lru
, struct xfs_buf
, b_lru
);
1405 * Decrement the b_lru_ref count unless the value is already
1406 * zero. If the value is already zero, we need to reclaim the
1407 * buffer, otherwise it gets another trip through the LRU.
1409 if (!atomic_add_unless(&bp
->b_lru_ref
, -1, 0)) {
1410 list_move_tail(&bp
->b_lru
, &btp
->bt_lru
);
1415 * remove the buffer from the LRU now to avoid needing another
1416 * lock round trip inside xfs_buf_rele().
1418 list_move(&bp
->b_lru
, &dispose
);
1421 spin_unlock(&btp
->bt_lru_lock
);
1423 while (!list_empty(&dispose
)) {
1424 bp
= list_first_entry(&dispose
, struct xfs_buf
, b_lru
);
1425 list_del_init(&bp
->b_lru
);
1429 return btp
->bt_lru_nr
;
1434 struct xfs_mount
*mp
,
1435 struct xfs_buftarg
*btp
)
1437 unregister_shrinker(&btp
->bt_shrinker
);
1439 if (mp
->m_flags
& XFS_MOUNT_BARRIER
)
1440 xfs_blkdev_issue_flush(btp
);
1446 xfs_setsize_buftarg_flags(
1448 unsigned int blocksize
,
1449 unsigned int sectorsize
,
1452 btp
->bt_bsize
= blocksize
;
1453 btp
->bt_sshift
= ffs(sectorsize
) - 1;
1454 btp
->bt_smask
= sectorsize
- 1;
1456 if (set_blocksize(btp
->bt_bdev
, sectorsize
)) {
1457 char name
[BDEVNAME_SIZE
];
1459 bdevname(btp
->bt_bdev
, name
);
1461 xfs_warn(btp
->bt_mount
,
1462 "Cannot set_blocksize to %u on device %s\n",
1471 * When allocating the initial buffer target we have not yet
1472 * read in the superblock, so don't know what sized sectors
1473 * are being used is at this early stage. Play safe.
1476 xfs_setsize_buftarg_early(
1478 struct block_device
*bdev
)
1480 return xfs_setsize_buftarg_flags(btp
,
1481 PAGE_SIZE
, bdev_logical_block_size(bdev
), 0);
1485 xfs_setsize_buftarg(
1487 unsigned int blocksize
,
1488 unsigned int sectorsize
)
1490 return xfs_setsize_buftarg_flags(btp
, blocksize
, sectorsize
, 1);
1495 struct xfs_mount
*mp
,
1496 struct block_device
*bdev
,
1502 btp
= kmem_zalloc(sizeof(*btp
), KM_SLEEP
);
1505 btp
->bt_dev
= bdev
->bd_dev
;
1506 btp
->bt_bdev
= bdev
;
1507 btp
->bt_bdi
= blk_get_backing_dev_info(bdev
);
1511 INIT_LIST_HEAD(&btp
->bt_lru
);
1512 spin_lock_init(&btp
->bt_lru_lock
);
1513 if (xfs_setsize_buftarg_early(btp
, bdev
))
1515 btp
->bt_shrinker
.shrink
= xfs_buftarg_shrink
;
1516 btp
->bt_shrinker
.seeks
= DEFAULT_SEEKS
;
1517 register_shrinker(&btp
->bt_shrinker
);
1526 * Add a buffer to the delayed write list.
1528 * This queues a buffer for writeout if it hasn't already been. Note that
1529 * neither this routine nor the buffer list submission functions perform
1530 * any internal synchronization. It is expected that the lists are thread-local
1533 * Returns true if we queued up the buffer, or false if it already had
1534 * been on the buffer list.
1537 xfs_buf_delwri_queue(
1539 struct list_head
*list
)
1541 ASSERT(xfs_buf_islocked(bp
));
1542 ASSERT(!(bp
->b_flags
& XBF_READ
));
1545 * If the buffer is already marked delwri it already is queued up
1546 * by someone else for imediate writeout. Just ignore it in that
1549 if (bp
->b_flags
& _XBF_DELWRI_Q
) {
1550 trace_xfs_buf_delwri_queued(bp
, _RET_IP_
);
1554 trace_xfs_buf_delwri_queue(bp
, _RET_IP_
);
1557 * If a buffer gets written out synchronously or marked stale while it
1558 * is on a delwri list we lazily remove it. To do this, the other party
1559 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1560 * It remains referenced and on the list. In a rare corner case it
1561 * might get readded to a delwri list after the synchronous writeout, in
1562 * which case we need just need to re-add the flag here.
1564 bp
->b_flags
|= _XBF_DELWRI_Q
;
1565 if (list_empty(&bp
->b_list
)) {
1566 atomic_inc(&bp
->b_hold
);
1567 list_add_tail(&bp
->b_list
, list
);
1574 * Compare function is more complex than it needs to be because
1575 * the return value is only 32 bits and we are doing comparisons
1581 struct list_head
*a
,
1582 struct list_head
*b
)
1584 struct xfs_buf
*ap
= container_of(a
, struct xfs_buf
, b_list
);
1585 struct xfs_buf
*bp
= container_of(b
, struct xfs_buf
, b_list
);
1588 diff
= ap
->b_bn
- bp
->b_bn
;
1597 __xfs_buf_delwri_submit(
1598 struct list_head
*buffer_list
,
1599 struct list_head
*io_list
,
1602 struct blk_plug plug
;
1603 struct xfs_buf
*bp
, *n
;
1606 list_for_each_entry_safe(bp
, n
, buffer_list
, b_list
) {
1608 if (xfs_buf_ispinned(bp
)) {
1612 if (!xfs_buf_trylock(bp
))
1619 * Someone else might have written the buffer synchronously or
1620 * marked it stale in the meantime. In that case only the
1621 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1622 * reference and remove it from the list here.
1624 if (!(bp
->b_flags
& _XBF_DELWRI_Q
)) {
1625 list_del_init(&bp
->b_list
);
1630 list_move_tail(&bp
->b_list
, io_list
);
1631 trace_xfs_buf_delwri_split(bp
, _RET_IP_
);
1634 list_sort(NULL
, io_list
, xfs_buf_cmp
);
1636 blk_start_plug(&plug
);
1637 list_for_each_entry_safe(bp
, n
, io_list
, b_list
) {
1638 bp
->b_flags
&= ~(_XBF_DELWRI_Q
| XBF_ASYNC
);
1639 bp
->b_flags
|= XBF_WRITE
;
1642 bp
->b_flags
|= XBF_ASYNC
;
1643 list_del_init(&bp
->b_list
);
1647 blk_finish_plug(&plug
);
1653 * Write out a buffer list asynchronously.
1655 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1656 * out and not wait for I/O completion on any of the buffers. This interface
1657 * is only safely useable for callers that can track I/O completion by higher
1658 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1662 xfs_buf_delwri_submit_nowait(
1663 struct list_head
*buffer_list
)
1665 LIST_HEAD (io_list
);
1666 return __xfs_buf_delwri_submit(buffer_list
, &io_list
, false);
1670 * Write out a buffer list synchronously.
1672 * This will take the @buffer_list, write all buffers out and wait for I/O
1673 * completion on all of the buffers. @buffer_list is consumed by the function,
1674 * so callers must have some other way of tracking buffers if they require such
1678 xfs_buf_delwri_submit(
1679 struct list_head
*buffer_list
)
1681 LIST_HEAD (io_list
);
1682 int error
= 0, error2
;
1685 __xfs_buf_delwri_submit(buffer_list
, &io_list
, true);
1687 /* Wait for IO to complete. */
1688 while (!list_empty(&io_list
)) {
1689 bp
= list_first_entry(&io_list
, struct xfs_buf
, b_list
);
1691 list_del_init(&bp
->b_list
);
1692 error2
= xfs_buf_iowait(bp
);
1704 xfs_buf_zone
= kmem_zone_init_flags(sizeof(xfs_buf_t
), "xfs_buf",
1705 KM_ZONE_HWALIGN
, NULL
);
1709 xfslogd_workqueue
= alloc_workqueue("xfslogd",
1710 WQ_MEM_RECLAIM
| WQ_HIGHPRI
, 1);
1711 if (!xfslogd_workqueue
)
1712 goto out_free_buf_zone
;
1717 kmem_zone_destroy(xfs_buf_zone
);
1723 xfs_buf_terminate(void)
1725 destroy_workqueue(xfslogd_workqueue
);
1726 kmem_zone_destroy(xfs_buf_zone
);