4 * Copyright (C) 1991, 1992 Linus Torvalds
8 * 'buffer.c' implements the buffer-cache functions. Race-conditions have
9 * been avoided by NEVER letting an interrupt change a buffer (except for the
10 * data, of course), but instead letting the caller do it.
13 /* Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95 */
15 /* Removed a lot of unnecessary code and simplified things now that
16 * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
19 /* Speed up hash, lru, and free list operations. Use gfp() for allocating
20 * hash table, use SLAB cache for buffer heads. -DaveM
23 /* Added 32k buffer block sizes - these are required older ARM systems.
27 #include <linux/malloc.h>
28 #include <linux/locks.h>
29 #include <linux/errno.h>
30 #include <linux/swap.h>
31 #include <linux/swapctl.h>
32 #include <linux/smp_lock.h>
33 #include <linux/vmalloc.h>
34 #include <linux/blkdev.h>
35 #include <linux/sysrq.h>
36 #include <linux/file.h>
37 #include <linux/init.h>
38 #include <linux/quotaops.h>
40 #include <asm/uaccess.h>
42 #include <asm/bitops.h>
45 static char buffersize_index
[65] =
46 {-1, 0, 1, -1, 2, -1, -1, -1, 3, -1, -1, -1, -1, -1, -1, -1,
47 4, -1, -1, -1, -1, -1, -1, -1, -1,-1, -1, -1, -1, -1, -1, -1,
48 5, -1, -1, -1, -1, -1, -1, -1, -1,-1, -1, -1, -1, -1, -1, -1,
49 -1, -1, -1, -1, -1, -1, -1, -1, -1,-1, -1, -1, -1, -1, -1, -1,
52 #define BUFSIZE_INDEX(X) ((int) buffersize_index[(X)>>9])
53 #define MAX_BUF_PER_PAGE (PAGE_SIZE / 512)
54 #define NR_RESERVED (2*MAX_BUF_PER_PAGE)
55 #define MAX_UNUSED_BUFFERS NR_RESERVED+20 /* don't ever have more than this
56 number of unused buffer heads */
61 static unsigned long bh_hash_mask
= 0;
63 static int grow_buffers(int size
);
65 static struct buffer_head
** hash_table
;
66 static struct buffer_head
* lru_list
[NR_LIST
] = {NULL
, };
67 static struct buffer_head
* free_list
[NR_SIZES
] = {NULL
, };
69 static kmem_cache_t
*bh_cachep
;
71 static struct buffer_head
* unused_list
= NULL
;
72 static struct buffer_head
* reuse_list
= NULL
;
73 static struct wait_queue
* buffer_wait
= NULL
;
75 static int nr_buffers
= 0;
76 static int nr_buffers_type
[NR_LIST
] = {0,};
77 static int nr_buffer_heads
= 0;
78 static int nr_unused_buffer_heads
= 0;
79 static int nr_hashed_buffers
= 0;
81 /* This is used by some architectures to estimate available memory. */
84 /* Here is the parameter block for the bdflush process. If you add or
85 * remove any of the parameters, make sure to update kernel/sysctl.c.
90 /* The dummy values in this structure are left in there for compatibility
91 * with old programs that play with the /proc entries.
95 int nfract
; /* Percentage of buffer cache dirty to
97 int ndirty
; /* Maximum number of dirty blocks to write out per
99 int nrefill
; /* Number of clean buffers to try to obtain
100 each time we call refill */
101 int nref_dirt
; /* Dirty buffer threshold for activating bdflush
102 when trying to refill buffers. */
103 int interval
; /* Interval (seconds) between spontaneous
105 int age_buffer
; /* Time for normal buffer to age before
107 int age_super
; /* Time for superblock to age before we
109 int dummy2
; /* unused */
110 int dummy3
; /* unused */
112 unsigned int data
[N_PARAM
];
113 } bdf_prm
= {{40, 500, 64, 256, 5, 30*HZ
, 5*HZ
, 1884, 2}};
115 /* These are the min and max parameter values that we will allow to be assigned */
116 int bdflush_min
[N_PARAM
] = { 0, 10, 5, 25, 1, 1*HZ
, 1*HZ
, 1, 1};
117 int bdflush_max
[N_PARAM
] = {100,5000, 2000, 2000,100, 600*HZ
, 600*HZ
, 2047, 5};
119 void wakeup_bdflush(int);
122 * Rewrote the wait-routines to use the "new" wait-queue functionality,
123 * and getting rid of the cli-sti pairs. The wait-queue routines still
124 * need cli-sti, but now it's just a couple of 386 instructions or so.
126 * Note that the real wait_on_buffer() is an inline function that checks
127 * if 'b_wait' is set before calling this, so that the queues aren't set
130 void __wait_on_buffer(struct buffer_head
* bh
)
132 struct task_struct
*tsk
= current
;
133 struct wait_queue wait
;
137 add_wait_queue(&bh
->b_wait
, &wait
);
139 tsk
->state
= TASK_UNINTERRUPTIBLE
;
140 run_task_queue(&tq_disk
);
141 if (buffer_locked(bh
)) {
145 tsk
->state
= TASK_RUNNING
;
146 remove_wait_queue(&bh
->b_wait
, &wait
);
150 /* Call sync_buffers with wait!=0 to ensure that the call does not
151 * return until all buffer writes have completed. Sync() may return
152 * before the writes have finished; fsync() may not.
155 /* Godamity-damn. Some buffers (bitmaps for filesystems)
156 * spontaneously dirty themselves without ever brelse being called.
157 * We will ultimately want to put these in a separate list, but for
158 * now we search all of the lists for dirty buffers.
160 static int sync_buffers(kdev_t dev
, int wait
)
162 int i
, retry
, pass
= 0, err
= 0;
163 struct buffer_head
* bh
, *next
;
165 /* One pass for no-wait, three for wait:
166 * 0) write out all dirty, unlocked buffers;
167 * 1) write out all dirty buffers, waiting if locked;
168 * 2) wait for completion by waiting for all buffers to unlock.
173 /* We search all lists as a failsafe mechanism, not because we expect
174 * there to be dirty buffers on any of the other lists.
176 bh
= lru_list
[BUF_DIRTY
];
179 for (i
= nr_buffers_type
[BUF_DIRTY
]*2 ; i
-- > 0 ; bh
= next
) {
180 if (bh
->b_list
!= BUF_DIRTY
)
182 next
= bh
->b_next_free
;
183 if (!lru_list
[BUF_DIRTY
])
185 if (dev
&& bh
->b_dev
!= dev
)
187 if (buffer_locked(bh
)) {
188 /* Buffer is locked; skip it unless wait is
189 * requested AND pass > 0.
191 if (!wait
|| !pass
) {
199 /* If an unlocked buffer is not uptodate, there has
200 * been an IO error. Skip it.
202 if (wait
&& buffer_req(bh
) && !buffer_locked(bh
) &&
203 !buffer_dirty(bh
) && !buffer_uptodate(bh
)) {
208 /* Don't write clean buffers. Don't write ANY buffers
211 if (!buffer_dirty(bh
) || pass
>= 2)
214 /* Don't bother about locked buffers.
216 * XXX We checked if it was locked above and there is no
217 * XXX way we could have slept in between. -DaveM
219 if (buffer_locked(bh
))
224 ll_rw_block(WRITE
, 1, &bh
);
231 bh
= lru_list
[BUF_LOCKED
];
234 for (i
= nr_buffers_type
[BUF_LOCKED
]*2 ; i
-- > 0 ; bh
= next
) {
235 if (bh
->b_list
!= BUF_LOCKED
)
237 next
= bh
->b_next_free
;
238 if (!lru_list
[BUF_LOCKED
])
240 if (dev
&& bh
->b_dev
!= dev
)
242 if (buffer_locked(bh
)) {
243 /* Buffer is locked; skip it unless wait is
244 * requested AND pass > 0.
246 if (!wait
|| !pass
) {
255 /* If we are waiting for the sync to succeed, and if any dirty
256 * blocks were written, then repeat; on the second pass, only
257 * wait for buffers being written (do not pass to write any
258 * more buffers on the second pass).
260 } while (wait
&& retry
&& ++pass
<=2);
264 void sync_dev(kdev_t dev
)
266 sync_buffers(dev
, 0);
269 sync_buffers(dev
, 0);
272 * FIXME(eric) we need to sync the physical devices here.
273 * This is because some (scsi) controllers have huge amounts of
274 * cache onboard (hundreds of Mb), and we need to instruct
275 * them to commit all of the dirty memory to disk, and we should
276 * not return until this has happened.
278 * This would need to get implemented by going through the assorted
279 * layers so that each block major number can be synced, and this
280 * would call down into the upper and mid-layer scsi.
284 int fsync_dev(kdev_t dev
)
286 sync_buffers(dev
, 0);
290 return sync_buffers(dev
, 1);
293 asmlinkage
int sys_sync(void)
302 * filp may be NULL if called via the msync of a vma.
305 int file_fsync(struct file
*filp
, struct dentry
*dentry
)
307 struct inode
* inode
= dentry
->d_inode
;
308 struct super_block
* sb
;
311 /* sync the inode to buffers */
312 write_inode_now(inode
);
314 /* sync the superblock to buffers */
317 if (sb
->s_op
&& sb
->s_op
->write_super
)
318 sb
->s_op
->write_super(sb
);
320 /* .. finally sync the buffers to disk */
322 return sync_buffers(dev
, 1);
325 asmlinkage
int sys_fsync(unsigned int fd
)
328 struct dentry
* dentry
;
329 struct inode
* inode
;
338 dentry
= file
->f_dentry
;
342 inode
= dentry
->d_inode
;
347 if (!file
->f_op
|| !file
->f_op
->fsync
)
350 /* We need to protect against concurrent writers.. */
352 err
= file
->f_op
->fsync(file
, dentry
);
362 asmlinkage
int sys_fdatasync(unsigned int fd
)
365 struct dentry
* dentry
;
366 struct inode
* inode
;
375 dentry
= file
->f_dentry
;
379 inode
= dentry
->d_inode
;
384 if (!file
->f_op
|| !file
->f_op
->fsync
)
387 /* this needs further work, at the moment it is identical to fsync() */
389 err
= file
->f_op
->fsync(file
, dentry
);
399 void invalidate_buffers(kdev_t dev
)
403 struct buffer_head
* bh
;
405 for(nlist
= 0; nlist
< NR_LIST
; nlist
++) {
406 bh
= lru_list
[nlist
];
407 for (i
= nr_buffers_type
[nlist
]*2 ; --i
> 0 ; bh
= bh
->b_next_free
) {
408 if (bh
->b_dev
!= dev
)
411 if (bh
->b_dev
!= dev
)
416 clear_bit(BH_Protected
, &bh
->b_state
);
417 clear_bit(BH_Uptodate
, &bh
->b_state
);
418 clear_bit(BH_Dirty
, &bh
->b_state
);
419 clear_bit(BH_Req
, &bh
->b_state
);
424 #define _hashfn(dev,block) (((unsigned)(HASHDEV(dev)^block)) & bh_hash_mask)
425 #define hash(dev,block) hash_table[_hashfn(dev,block)]
427 static inline void remove_from_hash_queue(struct buffer_head
* bh
)
429 struct buffer_head
**pprev
= bh
->b_pprev
;
431 struct buffer_head
* next
= bh
->b_next
;
433 next
->b_pprev
= pprev
;
442 static inline void remove_from_lru_list(struct buffer_head
* bh
)
444 if (!(bh
->b_prev_free
) || !(bh
->b_next_free
))
445 panic("VFS: LRU block list corrupted");
446 if (bh
->b_dev
== B_FREE
)
447 panic("LRU list corrupted");
448 bh
->b_prev_free
->b_next_free
= bh
->b_next_free
;
449 bh
->b_next_free
->b_prev_free
= bh
->b_prev_free
;
451 if (lru_list
[bh
->b_list
] == bh
)
452 lru_list
[bh
->b_list
] = bh
->b_next_free
;
453 if (lru_list
[bh
->b_list
] == bh
)
454 lru_list
[bh
->b_list
] = NULL
;
455 bh
->b_next_free
= bh
->b_prev_free
= NULL
;
458 static inline void remove_from_free_list(struct buffer_head
* bh
)
460 int isize
= BUFSIZE_INDEX(bh
->b_size
);
461 if (!(bh
->b_prev_free
) || !(bh
->b_next_free
))
462 panic("VFS: Free block list corrupted");
463 if(bh
->b_dev
!= B_FREE
)
464 panic("Free list corrupted");
465 if(!free_list
[isize
])
466 panic("Free list empty");
467 if(bh
->b_next_free
== bh
)
468 free_list
[isize
] = NULL
;
470 bh
->b_prev_free
->b_next_free
= bh
->b_next_free
;
471 bh
->b_next_free
->b_prev_free
= bh
->b_prev_free
;
472 if (free_list
[isize
] == bh
)
473 free_list
[isize
] = bh
->b_next_free
;
475 bh
->b_next_free
= bh
->b_prev_free
= NULL
;
478 static void remove_from_queues(struct buffer_head
* bh
)
480 if(bh
->b_dev
== B_FREE
) {
481 remove_from_free_list(bh
); /* Free list entries should not be
485 nr_buffers_type
[bh
->b_list
]--;
486 remove_from_hash_queue(bh
);
487 remove_from_lru_list(bh
);
490 static inline void put_last_free(struct buffer_head
* bh
)
493 struct buffer_head
**bhp
= &free_list
[BUFSIZE_INDEX(bh
->b_size
)];
495 bh
->b_dev
= B_FREE
; /* So it is obvious we are on the free list. */
497 /* Add to back of free list. */
500 bh
->b_prev_free
= bh
;
503 bh
->b_next_free
= *bhp
;
504 bh
->b_prev_free
= (*bhp
)->b_prev_free
;
505 (*bhp
)->b_prev_free
->b_next_free
= bh
;
506 (*bhp
)->b_prev_free
= bh
;
510 static void insert_into_queues(struct buffer_head
* bh
)
512 /* put at end of free list */
513 if(bh
->b_dev
== B_FREE
) {
516 struct buffer_head
**bhp
= &lru_list
[bh
->b_list
];
520 bh
->b_prev_free
= bh
;
524 panic("VFS: buffer LRU pointers corrupted");
526 bh
->b_next_free
= *bhp
;
527 bh
->b_prev_free
= (*bhp
)->b_prev_free
;
528 (*bhp
)->b_prev_free
->b_next_free
= bh
;
529 (*bhp
)->b_prev_free
= bh
;
531 nr_buffers_type
[bh
->b_list
]++;
533 /* Put the buffer in new hash-queue if it has a device. */
537 struct buffer_head
**bhp
= &hash(bh
->b_dev
, bh
->b_blocknr
);
538 struct buffer_head
*next
= *bhp
;
542 next
->b_pprev
= &bh
->b_next
;
551 struct buffer_head
* find_buffer(kdev_t dev
, int block
, int size
)
553 struct buffer_head
* next
;
555 next
= hash(dev
,block
);
557 struct buffer_head
*tmp
= next
;
561 if (tmp
->b_blocknr
!= block
|| tmp
->b_size
!= size
|| tmp
->b_dev
!= dev
)
570 * Why like this, I hear you say... The reason is race-conditions.
571 * As we don't lock buffers (unless we are reading them, that is),
572 * something might happen to it while we sleep (ie a read-error
573 * will force it bad). This shouldn't really happen currently, but
576 struct buffer_head
* get_hash_table(kdev_t dev
, int block
, int size
)
578 struct buffer_head
* bh
;
579 bh
= find_buffer(dev
,block
,size
);
585 unsigned int get_hardblocksize(kdev_t dev
)
588 * Get the hard sector size for the given device. If we don't know
589 * what it is, return 0.
591 if (hardsect_size
[MAJOR(dev
)] != NULL
) {
592 int blksize
= hardsect_size
[MAJOR(dev
)][MINOR(dev
)];
598 * We don't know what the hardware sector size for this device is.
599 * Return 0 indicating that we don't know.
604 void set_blocksize(kdev_t dev
, int size
)
606 extern int *blksize_size
[];
608 struct buffer_head
* bh
, *bhnext
;
610 if (!blksize_size
[MAJOR(dev
)])
613 /* Size must be a power of two, and between 512 and PAGE_SIZE */
614 if (size
> PAGE_SIZE
|| size
< 512 || (size
& (size
-1)))
615 panic("Invalid blocksize passed to set_blocksize");
617 if (blksize_size
[MAJOR(dev
)][MINOR(dev
)] == 0 && size
== BLOCK_SIZE
) {
618 blksize_size
[MAJOR(dev
)][MINOR(dev
)] = size
;
621 if (blksize_size
[MAJOR(dev
)][MINOR(dev
)] == size
)
623 sync_buffers(dev
, 2);
624 blksize_size
[MAJOR(dev
)][MINOR(dev
)] = size
;
626 /* We need to be quite careful how we do this - we are moving entries
627 * around on the free list, and we can get in a loop if we are not careful.
629 for(nlist
= 0; nlist
< NR_LIST
; nlist
++) {
630 bh
= lru_list
[nlist
];
631 for (i
= nr_buffers_type
[nlist
]*2 ; --i
> 0 ; bh
= bhnext
) {
635 bhnext
= bh
->b_next_free
;
636 if (bh
->b_dev
!= dev
)
638 if (bh
->b_size
== size
)
643 if (bh
->b_dev
== dev
&& bh
->b_size
!= size
) {
644 clear_bit(BH_Dirty
, &bh
->b_state
);
645 clear_bit(BH_Uptodate
, &bh
->b_state
);
646 clear_bit(BH_Req
, &bh
->b_state
);
649 remove_from_hash_queue(bh
);
655 * We used to try various strange things. Let's not.
657 static void refill_freelist(int size
)
659 if (!grow_buffers(size
)) {
661 current
->policy
|= SCHED_YIELD
;
666 void init_buffer(struct buffer_head
*bh
, kdev_t dev
, int block
,
667 bh_end_io_t
*handler
, void *dev_id
)
670 bh
->b_list
= BUF_CLEAN
;
673 bh
->b_blocknr
= block
;
674 bh
->b_end_io
= handler
;
675 bh
->b_dev_id
= dev_id
;
678 static void end_buffer_io_sync(struct buffer_head
*bh
, int uptodate
)
680 mark_buffer_uptodate(bh
, uptodate
);
685 * Ok, this is getblk, and it isn't very clear, again to hinder
686 * race-conditions. Most of the code is seldom used, (ie repeating),
687 * so it should be much more efficient than it looks.
689 * The algorithm is changed: hopefully better, and an elusive bug removed.
691 * 14.02.92: changed it to sync dirty buffers a bit: better performance
692 * when the filesystem starts to get full of dirty blocks (I hope).
694 struct buffer_head
* getblk(kdev_t dev
, int block
, int size
)
696 struct buffer_head
* bh
;
700 bh
= get_hash_table(dev
, block
, size
);
702 if (!buffer_dirty(bh
)) {
708 isize
= BUFSIZE_INDEX(size
);
710 bh
= free_list
[isize
];
713 remove_from_free_list(bh
);
715 /* OK, FINALLY we know that this buffer is the only one of its kind,
716 * and that it's unused (b_count=0), unlocked, and clean.
718 init_buffer(bh
, dev
, block
, end_buffer_io_sync
, NULL
);
720 insert_into_queues(bh
);
724 * If we block while refilling the free list, somebody may
725 * create the buffer first ... search the hashes again.
728 refill_freelist(size
);
729 if (!find_buffer(dev
,block
,size
))
734 void set_writetime(struct buffer_head
* buf
, int flag
)
738 if (buffer_dirty(buf
)) {
739 /* Move buffer to dirty list if jiffies is clear. */
740 newtime
= jiffies
+ (flag
? bdf_prm
.b_un
.age_super
:
741 bdf_prm
.b_un
.age_buffer
);
742 if(!buf
->b_flushtime
|| buf
->b_flushtime
> newtime
)
743 buf
->b_flushtime
= newtime
;
745 buf
->b_flushtime
= 0;
751 * Put a buffer into the appropriate list, without side-effects.
753 static inline void file_buffer(struct buffer_head
*bh
, int list
)
755 remove_from_queues(bh
);
757 insert_into_queues(bh
);
761 * A buffer may need to be moved from one buffer list to another
762 * (e.g. in case it is not shared any more). Handle this.
764 void refile_buffer(struct buffer_head
* buf
)
768 if(buf
->b_dev
== B_FREE
) {
769 printk("Attempt to refile free buffer\n");
772 if (buffer_dirty(buf
))
774 else if (buffer_locked(buf
))
775 dispose
= BUF_LOCKED
;
778 if(dispose
!= buf
->b_list
) {
779 file_buffer(buf
, dispose
);
780 if(dispose
== BUF_DIRTY
) {
781 int too_many
= (nr_buffers
* bdf_prm
.b_un
.nfract
/100);
783 /* This buffer is dirty, maybe we need to start flushing.
784 * If too high a percentage of the buffers are dirty...
786 if (nr_buffers_type
[BUF_DIRTY
] > too_many
)
789 /* If this is a loop device, and
790 * more than half of the buffers are dirty...
791 * (Prevents no-free-buffers deadlock with loop device.)
793 if (MAJOR(buf
->b_dev
) == LOOP_MAJOR
&&
794 nr_buffers_type
[BUF_DIRTY
]*2>nr_buffers
)
801 * Release a buffer head
803 void __brelse(struct buffer_head
* buf
)
805 /* If dirty, mark the time this buffer should be written back. */
806 set_writetime(buf
, 0);
814 printk("VFS: brelse: Trying to free free buffer\n");
818 * bforget() is like brelse(), except it puts the buffer on the
819 * free list if it can.. We can NOT free the buffer if:
820 * - there are other users of it
821 * - it is locked and thus can have active IO
823 void __bforget(struct buffer_head
* buf
)
825 if (buf
->b_count
!= 1 || buffer_locked(buf
)) {
831 remove_from_queues(buf
);
836 * bread() reads a specified block and returns the buffer that contains
837 * it. It returns NULL if the block was unreadable.
839 struct buffer_head
* bread(kdev_t dev
, int block
, int size
)
841 struct buffer_head
* bh
;
843 bh
= getblk(dev
, block
, size
);
844 if (buffer_uptodate(bh
))
846 ll_rw_block(READ
, 1, &bh
);
848 if (buffer_uptodate(bh
))
855 * Ok, breada can be used as bread, but additionally to mark other
856 * blocks for reading as well. End the argument list with a negative
862 struct buffer_head
* breada(kdev_t dev
, int block
, int bufsize
,
863 unsigned int pos
, unsigned int filesize
)
865 struct buffer_head
* bhlist
[NBUF
];
867 struct buffer_head
* bh
;
877 bh
= getblk(dev
, block
, bufsize
);
878 index
= BUFSIZE_INDEX(bh
->b_size
);
880 if (buffer_uptodate(bh
))
882 else ll_rw_block(READ
, 1, &bh
);
884 blocks
= (filesize
- pos
) >> (9+index
);
886 if (blocks
< (read_ahead
[MAJOR(dev
)] >> index
))
887 blocks
= read_ahead
[MAJOR(dev
)] >> index
;
891 /* if (blocks) printk("breada (new) %d blocks\n",blocks); */
896 for(i
=1; i
<blocks
; i
++) {
897 bh
= getblk(dev
,block
+i
,bufsize
);
898 if (buffer_uptodate(bh
)) {
902 else bhlist
[j
++] = bh
;
905 /* Request the read for these buffers, and then release them. */
907 ll_rw_block(READA
, (j
-1), bhlist
+1);
911 /* Wait for this buffer, and then continue on. */
914 if (buffer_uptodate(bh
))
921 * Note: the caller should wake up the buffer_wait list if needed.
923 static void put_unused_buffer_head(struct buffer_head
* bh
)
925 if (nr_unused_buffer_heads
>= MAX_UNUSED_BUFFERS
) {
927 kmem_cache_free(bh_cachep
, bh
);
931 memset(bh
,0,sizeof(*bh
));
932 nr_unused_buffer_heads
++;
933 bh
->b_next_free
= unused_list
;
938 * We can't put completed temporary IO buffer_heads directly onto the
939 * unused_list when they become unlocked, since the device driver
940 * end_request routines still expect access to the buffer_head's
941 * fields after the final unlock. So, the device driver puts them on
942 * the reuse_list instead once IO completes, and we recover these to
943 * the unused_list here.
945 * Note that we don't do a wakeup here, but return a flag indicating
946 * whether we got any buffer heads. A task ready to sleep can check
947 * the returned value, and any tasks already sleeping will have been
948 * awakened when the buffer heads were added to the reuse list.
950 static inline int recover_reusable_buffer_heads(void)
952 struct buffer_head
*head
= xchg(&reuse_list
, NULL
);
957 struct buffer_head
*bh
= head
;
958 head
= head
->b_next_free
;
959 put_unused_buffer_head(bh
);
967 * Reserve NR_RESERVED buffer heads for async IO requests to avoid
968 * no-buffer-head deadlock. Return NULL on failure; waiting for
969 * buffer heads is now handled in create_buffers().
971 static struct buffer_head
* get_unused_buffer_head(int async
)
973 struct buffer_head
* bh
;
975 recover_reusable_buffer_heads();
976 if (nr_unused_buffer_heads
> NR_RESERVED
) {
978 unused_list
= bh
->b_next_free
;
979 nr_unused_buffer_heads
--;
983 /* This is critical. We can't swap out pages to get
984 * more buffer heads, because the swap-out may need
985 * more buffer-heads itself. Thus SLAB_BUFFER.
987 if((bh
= kmem_cache_alloc(bh_cachep
, SLAB_BUFFER
)) != NULL
) {
988 memset(bh
, 0, sizeof(*bh
));
994 * If we need an async buffer, use the reserved buffer heads.
996 if (async
&& unused_list
) {
998 unused_list
= bh
->b_next_free
;
999 nr_unused_buffer_heads
--;
1005 * (Pending further analysis ...)
1006 * Ordinary (non-async) requests can use a different memory priority
1007 * to free up pages. Any swapping thus generated will use async
1011 (bh
= kmem_cache_alloc(bh_cachep
, SLAB_KERNEL
)) != NULL
) {
1012 memset(bh
, 0, sizeof(*bh
));
1022 * Create the appropriate buffers when given a page for data area and
1023 * the size of each buffer.. Use the bh->b_this_page linked list to
1024 * follow the buffers created. Return NULL if unable to create more
1026 * The async flag is used to differentiate async IO (paging, swapping)
1027 * from ordinary buffer allocations, and only async requests are allowed
1028 * to sleep waiting for buffer heads.
1030 static struct buffer_head
* create_buffers(unsigned long page
,
1031 unsigned long size
, int async
)
1033 struct wait_queue wait
= { current
, NULL
};
1034 struct buffer_head
*bh
, *head
;
1040 while ((offset
-= size
) >= 0) {
1041 bh
= get_unused_buffer_head(async
);
1045 bh
->b_dev
= B_FREE
; /* Flag as unused */
1046 bh
->b_this_page
= head
;
1050 bh
->b_next_free
= NULL
;
1054 bh
->b_data
= (char *) (page
+offset
);
1059 * In case anything failed, we just free everything we got.
1065 head
= head
->b_this_page
;
1066 put_unused_buffer_head(bh
);
1069 /* Wake up any waiters ... */
1070 wake_up(&buffer_wait
);
1074 * Return failure for non-async IO requests. Async IO requests
1075 * are not allowed to fail, so we have to wait until buffer heads
1076 * become available. But we don't want tasks sleeping with
1077 * partially complete buffers, so all were released above.
1082 /* We're _really_ low on memory. Now we just
1083 * wait for old buffer heads to become free due to
1084 * finishing IO. Since this is an async request and
1085 * the reserve list is empty, we're sure there are
1086 * async buffer heads in use.
1088 run_task_queue(&tq_disk
);
1091 * Set our state for sleeping, then check again for buffer heads.
1092 * This ensures we won't miss a wake_up from an interrupt.
1094 add_wait_queue(&buffer_wait
, &wait
);
1095 current
->state
= TASK_UNINTERRUPTIBLE
;
1096 if (!recover_reusable_buffer_heads())
1098 remove_wait_queue(&buffer_wait
, &wait
);
1099 current
->state
= TASK_RUNNING
;
1103 /* Run the hooks that have to be done when a page I/O has completed. */
1104 static inline void after_unlock_page (struct page
* page
)
1106 if (test_and_clear_bit(PG_decr_after
, &page
->flags
)) {
1107 atomic_dec(&nr_async_pages
);
1109 printk ("DebugVM: Finished IO on page %p, nr_async_pages %d\n",
1110 (char *) page_address(page
),
1111 atomic_read(&nr_async_pages
));
1114 if (test_and_clear_bit(PG_swap_unlock_after
, &page
->flags
))
1115 swap_after_unlock_page(page
->offset
);
1116 if (test_and_clear_bit(PG_free_after
, &page
->flags
))
1121 * Free all temporary buffers belonging to a page.
1122 * This needs to be called with interrupts disabled.
1124 static inline void free_async_buffers (struct buffer_head
* bh
)
1126 struct buffer_head
*tmp
, *tail
;
1129 * Link all the buffers into the b_next_free list,
1130 * so we only have to do one xchg() operation ...
1133 while ((tmp
= tail
->b_this_page
) != bh
) {
1134 tail
->b_next_free
= tmp
;
1138 /* Update the reuse list */
1139 tail
->b_next_free
= xchg(&reuse_list
, NULL
);
1142 /* Wake up any waiters ... */
1143 wake_up(&buffer_wait
);
1146 static void end_buffer_io_async(struct buffer_head
* bh
, int uptodate
)
1148 unsigned long flags
;
1149 struct buffer_head
*tmp
;
1152 mark_buffer_uptodate(bh
, uptodate
);
1155 /* This is a temporary buffer used for page I/O. */
1156 page
= mem_map
+ MAP_NR(bh
->b_data
);
1157 if (!PageLocked(page
))
1159 if (bh
->b_count
!= 1)
1162 if (!test_bit(BH_Uptodate
, &bh
->b_state
))
1163 set_bit(PG_error
, &page
->flags
);
1166 * Be _very_ careful from here on. Bad things can happen if
1167 * two buffer heads end IO at almost the same time and both
1168 * decide that the page is now completely done.
1170 * Async buffer_heads are here only as labels for IO, and get
1171 * thrown away once the IO for this page is complete. IO is
1172 * deemed complete once all buffers have been visited
1173 * (b_count==0) and are now unlocked. We must make sure that
1174 * only the _last_ buffer that decrements its count is the one
1175 * that free's the page..
1184 tmp
= tmp
->b_this_page
;
1185 } while (tmp
!= bh
);
1187 /* OK, the async IO on this page is complete. */
1188 free_async_buffers(bh
);
1189 restore_flags(flags
);
1190 clear_bit(PG_locked
, &page
->flags
);
1191 wake_up(&page
->wait
);
1192 after_unlock_page(page
);
1196 restore_flags(flags
);
1200 printk ("Whoops: end_buffer_io_async: async io complete on unlocked page\n");
1204 printk ("Whoops: end_buffer_io_async: b_count != 1 on async io.\n");
1209 * Start I/O on a page.
1210 * This function expects the page to be locked and may return before I/O is complete.
1211 * You then have to check page->locked, page->uptodate, and maybe wait on page->wait.
1213 int brw_page(int rw
, struct page
*page
, kdev_t dev
, int b
[], int size
, int bmap
)
1215 struct buffer_head
*bh
, *prev
, *next
, *arr
[MAX_BUF_PER_PAGE
];
1218 if (!PageLocked(page
))
1219 panic("brw_page: page not locked for I/O");
1220 clear_bit(PG_uptodate
, &page
->flags
);
1221 clear_bit(PG_error
, &page
->flags
);
1223 * Allocate async buffer heads pointing to this page, just for I/O.
1224 * They do _not_ show up in the buffer hash table!
1225 * They are _not_ registered in page->buffers either!
1227 bh
= create_buffers(page_address(page
), size
, 1);
1229 /* WSH: exit here leaves page->count incremented */
1230 clear_bit(PG_locked
, &page
->flags
);
1231 wake_up(&page
->wait
);
1237 struct buffer_head
* tmp
;
1240 init_buffer(next
, dev
, block
, end_buffer_io_async
, NULL
);
1241 set_bit(BH_Uptodate
, &next
->b_state
);
1244 * When we use bmap, we define block zero to represent
1245 * a hole. ll_rw_page, however, may legitimately
1246 * access block zero, and we need to distinguish the
1249 if (bmap
&& !block
) {
1250 memset(next
->b_data
, 0, size
);
1254 tmp
= get_hash_table(dev
, block
, size
);
1256 if (!buffer_uptodate(tmp
)) {
1258 ll_rw_block(READ
, 1, &tmp
);
1259 wait_on_buffer(tmp
);
1262 memcpy(next
->b_data
, tmp
->b_data
, size
);
1264 memcpy(tmp
->b_data
, next
->b_data
, size
);
1265 mark_buffer_dirty(tmp
, 0);
1272 clear_bit(BH_Uptodate
, &next
->b_state
);
1274 set_bit(BH_Dirty
, &next
->b_state
);
1276 } while (prev
= next
, (next
= next
->b_this_page
) != NULL
);
1277 prev
->b_this_page
= bh
;
1280 ll_rw_block(rw
, nr
, arr
);
1281 /* The rest of the work is done in mark_buffer_uptodate()
1282 * and unlock_buffer(). */
1284 unsigned long flags
;
1285 clear_bit(PG_locked
, &page
->flags
);
1286 set_bit(PG_uptodate
, &page
->flags
);
1287 wake_up(&page
->wait
);
1290 free_async_buffers(bh
);
1291 restore_flags(flags
);
1292 after_unlock_page(page
);
1299 * This is called by end_request() when I/O has completed.
1301 void mark_buffer_uptodate(struct buffer_head
* bh
, int on
)
1304 struct buffer_head
*tmp
= bh
;
1305 set_bit(BH_Uptodate
, &bh
->b_state
);
1306 /* If a page has buffers and all these buffers are uptodate,
1307 * then the page is uptodate. */
1309 if (!test_bit(BH_Uptodate
, &tmp
->b_state
))
1311 tmp
=tmp
->b_this_page
;
1312 } while (tmp
&& tmp
!= bh
);
1313 set_bit(PG_uptodate
, &mem_map
[MAP_NR(bh
->b_data
)].flags
);
1316 clear_bit(BH_Uptodate
, &bh
->b_state
);
1320 * Generic "readpage" function for block devices that have the normal
1321 * bmap functionality. This is most of the block device filesystems.
1322 * Reads the page asynchronously --- the unlock_buffer() and
1323 * mark_buffer_uptodate() functions propagate buffer state into the
1324 * page struct once IO has completed.
1326 int generic_readpage(struct file
* file
, struct page
* page
)
1328 struct dentry
*dentry
= file
->f_dentry
;
1329 struct inode
*inode
= dentry
->d_inode
;
1330 unsigned long block
;
1331 int *p
, nr
[PAGE_SIZE
/512];
1334 atomic_inc(&page
->count
);
1335 set_bit(PG_locked
, &page
->flags
);
1336 set_bit(PG_free_after
, &page
->flags
);
1338 i
= PAGE_SIZE
>> inode
->i_sb
->s_blocksize_bits
;
1339 block
= page
->offset
>> inode
->i_sb
->s_blocksize_bits
;
1342 *p
= inode
->i_op
->bmap(inode
, block
);
1349 brw_page(READ
, page
, inode
->i_dev
, nr
, inode
->i_sb
->s_blocksize
, 1);
1354 * Try to increase the number of buffers available: the size argument
1355 * is used to determine what kind of buffers we want.
1357 static int grow_buffers(int size
)
1360 struct buffer_head
*bh
, *tmp
;
1361 struct buffer_head
* insert_point
;
1364 if ((size
& 511) || (size
> PAGE_SIZE
)) {
1365 printk("VFS: grow_buffers: size = %d\n",size
);
1369 if (!(page
= __get_free_page(GFP_BUFFER
)))
1371 bh
= create_buffers(page
, size
, 0);
1377 isize
= BUFSIZE_INDEX(size
);
1378 insert_point
= free_list
[isize
];
1383 tmp
->b_next_free
= insert_point
->b_next_free
;
1384 tmp
->b_prev_free
= insert_point
;
1385 insert_point
->b_next_free
->b_prev_free
= tmp
;
1386 insert_point
->b_next_free
= tmp
;
1388 tmp
->b_prev_free
= tmp
;
1389 tmp
->b_next_free
= tmp
;
1393 if (tmp
->b_this_page
)
1394 tmp
= tmp
->b_this_page
;
1398 tmp
->b_this_page
= bh
;
1399 free_list
[isize
] = bh
;
1400 mem_map
[MAP_NR(page
)].buffers
= bh
;
1401 buffermem
+= PAGE_SIZE
;
1406 * Can the buffer be thrown out?
1408 #define BUFFER_BUSY_BITS ((1<<BH_Dirty) | (1<<BH_Lock) | (1<<BH_Protected))
1409 #define buffer_busy(bh) ((bh)->b_count || ((bh)->b_state & BUFFER_BUSY_BITS))
1412 * try_to_free_buffers() checks if all the buffers on this particular page
1413 * are unused, and free's the page if so.
1415 * Wake up bdflush() if this fails - if we're running low on memory due
1416 * to dirty buffers, we need to flush them out as quickly as possible.
1418 int try_to_free_buffers(struct page
* page_map
)
1420 struct buffer_head
* tmp
, * bh
= page_map
->buffers
;
1424 struct buffer_head
* p
= tmp
;
1426 tmp
= tmp
->b_this_page
;
1427 if (!buffer_busy(p
))
1432 } while (tmp
!= bh
);
1436 struct buffer_head
* p
= tmp
;
1437 tmp
= tmp
->b_this_page
;
1439 remove_from_queues(p
);
1440 put_unused_buffer_head(p
);
1441 } while (tmp
!= bh
);
1443 /* Wake up anyone waiting for buffer heads */
1444 wake_up(&buffer_wait
);
1446 /* And free the page */
1447 buffermem
-= PAGE_SIZE
;
1448 page_map
->buffers
= NULL
;
1449 __free_page(page_map
);
1453 /* ================== Debugging =================== */
1455 void show_buffers(void)
1457 struct buffer_head
* bh
;
1458 int found
= 0, locked
= 0, dirty
= 0, used
= 0, lastused
= 0;
1461 static char *buf_types
[NR_LIST
] = {"CLEAN","LOCKED","DIRTY"};
1463 printk("Buffer memory: %6dkB\n",buffermem
>>10);
1464 printk("Buffer heads: %6d\n",nr_buffer_heads
);
1465 printk("Buffer blocks: %6d\n",nr_buffers
);
1466 printk("Buffer hashed: %6d\n",nr_hashed_buffers
);
1468 for(nlist
= 0; nlist
< NR_LIST
; nlist
++) {
1469 found
= locked
= dirty
= used
= lastused
= protected = 0;
1470 bh
= lru_list
[nlist
];
1475 if (buffer_locked(bh
))
1477 if (buffer_protected(bh
))
1479 if (buffer_dirty(bh
))
1482 used
++, lastused
= found
;
1483 bh
= bh
->b_next_free
;
1484 } while (bh
!= lru_list
[nlist
]);
1485 printk("%8s: %d buffers, %d used (last=%d), "
1486 "%d locked, %d protected, %d dirty\n",
1487 buf_types
[nlist
], found
, used
, lastused
,
1488 locked
, protected, dirty
);
1493 /* ===================== Init ======================= */
1496 * allocate the hash table and init the free list
1497 * Use gfp() for the hash table to decrease TLB misses, use
1498 * SLAB cache for buffer heads.
1500 void __init
buffer_init(unsigned long memory_size
)
1503 unsigned int nr_hash
;
1505 /* we need to guess at the right sort of size for a buffer cache.
1506 the heuristic from working with large databases and getting
1507 fsync times (ext2) manageable, is the following */
1510 for (order
= 5; (1UL << order
) < memory_size
; order
++);
1512 /* try to allocate something until we get it or we're asking
1513 for something that is really too small */
1516 nr_hash
= (1UL << order
) * PAGE_SIZE
/
1517 sizeof(struct buffer_head
*);
1518 hash_table
= (struct buffer_head
**)
1519 __get_free_pages(GFP_ATOMIC
, order
);
1520 } while (hash_table
== NULL
&& --order
> 4);
1523 panic("Failed to allocate buffer hash table\n");
1524 memset(hash_table
, 0, nr_hash
* sizeof(struct buffer_head
*));
1525 bh_hash_mask
= nr_hash
-1;
1527 bh_cachep
= kmem_cache_create("buffer_head",
1528 sizeof(struct buffer_head
),
1530 SLAB_HWCACHE_ALIGN
, NULL
, NULL
);
1532 panic("Cannot create buffer head SLAB cache\n");
1534 * Allocate the reserved buffer heads.
1536 while (nr_buffer_heads
< NR_RESERVED
) {
1537 struct buffer_head
* bh
;
1539 bh
= kmem_cache_alloc(bh_cachep
, SLAB_ATOMIC
);
1542 put_unused_buffer_head(bh
);
1546 lru_list
[BUF_CLEAN
] = 0;
1547 grow_buffers(BLOCK_SIZE
);
1551 /* ====================== bdflush support =================== */
1553 /* This is a simple kernel daemon, whose job it is to provide a dynamic
1554 * response to dirty buffers. Once this process is activated, we write back
1555 * a limited number of buffers to the disks and then go back to sleep again.
1557 static struct wait_queue
* bdflush_done
= NULL
;
1558 struct task_struct
*bdflush_tsk
= 0;
1560 void wakeup_bdflush(int wait
)
1562 if (current
== bdflush_tsk
)
1564 wake_up_process(bdflush_tsk
);
1566 run_task_queue(&tq_disk
);
1567 sleep_on(&bdflush_done
);
1573 * Here we attempt to write back old buffers.
1574 * To prevent deadlocks for a loop device:
1575 * 1) Do non-blocking writes to loop (avoids deadlock with running
1576 * out of request blocks).
1577 * 2) But do a blocking write if the only dirty buffers are loop buffers
1578 * (otherwise we go into an infinite busy-loop).
1579 * 3) Quit writing loop blocks if a freelist went low (avoids deadlock
1580 * with running out of free buffers for loop's "real" device).
1583 static inline void sync_old_buffers(void)
1587 int wrta_cmd
= WRITEA
;
1589 int ncount
= 0, nwritten
= 0;
1591 struct buffer_head
* bh
, *next
;
1594 bh
= lru_list
[BUF_CLEAN
];
1596 for(i
= nr_buffers_type
[BUF_CLEAN
]; --i
> 0; bh
= next
) {
1597 next
= bh
->b_next_free
;
1599 /* Dirty/locked buffer on clean list? Refile it */
1600 if (buffer_locked(bh
) || buffer_dirty(bh
)) {
1607 bh
= lru_list
[BUF_LOCKED
];
1609 for(i
= nr_buffers_type
[BUF_LOCKED
]; --i
> 0; bh
= next
) {
1610 next
= bh
->b_next_free
;
1612 /* Unlocked buffer on locked list? Refile it */
1613 if (!buffer_locked(bh
))
1618 bh
= lru_list
[BUF_DIRTY
];
1620 for (i
= nr_buffers_type
[BUF_DIRTY
];
1621 i
-- > 0 && ndirty
< bdf_prm
.b_un
.ndirty
;
1623 /* We may have stalled while waiting for
1625 if(bh
->b_list
!= BUF_DIRTY
)
1627 next
= bh
->b_next_free
;
1628 if(!lru_list
[BUF_DIRTY
]) {
1629 printk("Dirty list empty %d\n", i
);
1633 /* Clean buffer on dirty list? Refile it */
1634 if (!buffer_dirty(bh
)) {
1639 if (buffer_locked(bh
))
1641 /* Should we write back buffers that are
1642 shared or not?? Currently dirty buffers
1643 are not shared, so it does not matter */
1647 bh
->b_flushtime
= 0;
1648 if (MAJOR(bh
->b_dev
) == LOOP_MAJOR
) {
1649 ll_rw_block(wrta_cmd
,1, &bh
);
1651 if (buffer_dirty(bh
))
1655 ll_rw_block(WRITE
, 1, &bh
);
1659 /* If we didn't write anything, but there are still
1660 * dirty buffers, then make the next write to a
1661 * loop device to be a blocking write.
1662 * This lets us block--which we _must_ do! */
1664 && nr_buffers_type
[BUF_DIRTY
] > 0 && wrta_cmd
!= WRITE
) {
1670 if (ncount
) printk("sync_old_buffers: %d dirty buffers not on dirty list\n", ncount
);
1671 printk("wrote %d/%d buffers...", nwritten
, ndirty
);
1673 run_task_queue(&tq_disk
);
1677 /* This is the interface to bdflush. As we get more sophisticated, we can
1678 * pass tuning parameters to this "process", to adjust how it behaves.
1679 * We would want to verify each parameter, however, to make sure that it
1682 asmlinkage
int sys_bdflush(int func
, long data
)
1684 int i
, error
= -EPERM
;
1687 if (!capable(CAP_SYS_ADMIN
))
1691 /* Func 1 used to call sync_old_buffers; a user space
1692 daemon would call it periodically. This is no
1693 longer necessary. Returning -EPERM here makes the
1694 daemon silently exit. */
1697 /* Basically func 1 means read param 1, 2 means write param 1, etc */
1701 if (i
< 0 || i
>= N_PARAM
)
1703 if((func
& 1) == 0) {
1704 error
= put_user(bdf_prm
.data
[i
], (int*)data
);
1707 if (data
< bdflush_min
[i
] || data
> bdflush_max
[i
])
1709 bdf_prm
.data
[i
] = data
;
1714 /* Having func 0 used to launch the actual bdflush and then never
1715 * return (unless explicitly killed). We return zero here to
1716 * remain semi-compatible with present update(8) programs.
1724 /* This is the actual bdflush daemon itself. It used to be started
1725 * from the syscall above, but now we launch it ourselves internally
1726 * with kernel_thread(...) directly after the first thread in
1727 * init/main.c. Every so often, or when woken up by another task that
1728 * needs memory, we call sync_old_buffers to partially clear the dirty list.
1731 int bdflush(void * unused
)
1733 long remaining
= HZ
* bdf_prm
.b_un
.interval
;
1734 struct task_struct
*tsk
= current
;
1737 * We have a bare-bones task_struct, and really should fill
1738 * in a few more things so "top" and /proc/2/{exe,root,cwd}
1739 * display semi-sane things. Not real crucial though...
1744 tsk
->dumpable
= 0; /* inhibit ptrace() */
1745 strcpy(tsk
->comm
, "kflushd");
1746 sigfillset(&tsk
->blocked
);
1750 * As a kernel thread we want to tamper with system buffers
1751 * and other internals and thus be subject to the SMP locking
1752 * rules. (On a uniprocessor box this does nothing).
1757 tsk
->state
= TASK_INTERRUPTIBLE
;
1758 remaining
= schedule_timeout(remaining
);
1761 printk("bdflush() activated...");
1763 CHECK_EMERGENCY_SYNC
1765 if (remaining
== 0) {
1767 * Also try to flush inodes and supers, since
1768 * otherwise there would be no way of ensuring
1769 * that these quantities ever get written
1770 * back. Ideally, we would have a timestamp
1771 * on the inodes and superblocks so that we
1772 * could write back only the old ones.
1776 remaining
= HZ
* bdf_prm
.b_un
.interval
;
1779 /* Keep flushing till there aren't very many dirty buffers */
1782 } while(nr_buffers_type
[BUF_DIRTY
] > nr_buffers
* bdf_prm
.b_un
.nfract
/100);
1784 wake_up(&bdflush_done
);
1786 printk("sleeping again.\n");