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 /* Thread it... -DaveM */
29 #include <linux/sched.h>
31 #include <linux/malloc.h>
32 #include <linux/locks.h>
33 #include <linux/errno.h>
34 #include <linux/swap.h>
35 #include <linux/swapctl.h>
36 #include <linux/smp_lock.h>
37 #include <linux/vmalloc.h>
38 #include <linux/blkdev.h>
39 #include <linux/sysrq.h>
40 #include <linux/file.h>
41 #include <linux/init.h>
42 #include <linux/quotaops.h>
44 #include <asm/uaccess.h>
46 #include <asm/bitops.h>
47 #include <asm/mmu_context.h>
50 static char buffersize_index
[65] =
51 {-1, 0, 1, -1, 2, -1, -1, -1, 3, -1, -1, -1, -1, -1, -1, -1,
52 4, -1, -1, -1, -1, -1, -1, -1, -1,-1, -1, -1, -1, -1, -1, -1,
53 5, -1, -1, -1, -1, -1, -1, -1, -1,-1, -1, -1, -1, -1, -1, -1,
54 -1, -1, -1, -1, -1, -1, -1, -1, -1,-1, -1, -1, -1, -1, -1, -1,
57 #define BUFSIZE_INDEX(X) ((int) buffersize_index[(X)>>9])
58 #define MAX_BUF_PER_PAGE (PAGE_SIZE / 512)
59 #define NR_RESERVED (2*MAX_BUF_PER_PAGE)
60 #define MAX_UNUSED_BUFFERS NR_RESERVED+20 /* don't ever have more than this
61 number of unused buffer heads */
63 /* Anti-deadlock ordering:
64 * lru_list_lock > hash_table_lock > free_list_lock > unused_list_lock
70 static unsigned int bh_hash_mask
= 0;
71 static unsigned int bh_hash_shift
= 0;
72 static struct buffer_head
**hash_table
;
73 static rwlock_t hash_table_lock
= RW_LOCK_UNLOCKED
;
75 static struct buffer_head
*lru_list
[NR_LIST
];
76 static spinlock_t lru_list_lock
= SPIN_LOCK_UNLOCKED
;
77 static int nr_buffers_type
[NR_LIST
] = {0,};
79 static struct buffer_head
* unused_list
= NULL
;
80 static int nr_unused_buffer_heads
= 0;
81 static spinlock_t unused_list_lock
= SPIN_LOCK_UNLOCKED
;
82 static DECLARE_WAIT_QUEUE_HEAD(buffer_wait
);
85 struct buffer_head
*list
;
88 static struct bh_free_head free_list
[NR_SIZES
];
90 static kmem_cache_t
*bh_cachep
;
92 static int grow_buffers(int size
);
94 /* This is used by some architectures to estimate available memory. */
95 atomic_t buffermem
= ATOMIC_INIT(0);
97 /* Here is the parameter block for the bdflush process. If you add or
98 * remove any of the parameters, make sure to update kernel/sysctl.c.
103 /* The dummy values in this structure are left in there for compatibility
104 * with old programs that play with the /proc entries.
106 union bdflush_param
{
108 int nfract
; /* Percentage of buffer cache dirty to
110 int ndirty
; /* Maximum number of dirty blocks to write out per
112 int nrefill
; /* Number of clean buffers to try to obtain
113 each time we call refill */
114 int nref_dirt
; /* Dirty buffer threshold for activating bdflush
115 when trying to refill buffers. */
116 int dummy1
; /* unused */
117 int age_buffer
; /* Time for normal buffer to age before we flush it */
118 int age_super
; /* Time for superblock to age before we flush it */
119 int dummy2
; /* unused */
120 int dummy3
; /* unused */
122 unsigned int data
[N_PARAM
];
123 } bdf_prm
= {{40, 500, 64, 256, 15, 30*HZ
, 5*HZ
, 1884, 2}};
125 /* These are the min and max parameter values that we will allow to be assigned */
126 int bdflush_min
[N_PARAM
] = { 0, 10, 5, 25, 0, 1*HZ
, 1*HZ
, 1, 1};
127 int bdflush_max
[N_PARAM
] = {100,50000, 20000, 20000,1000, 6000*HZ
, 6000*HZ
, 2047, 5};
129 void wakeup_bdflush(int);
132 * Rewrote the wait-routines to use the "new" wait-queue functionality,
133 * and getting rid of the cli-sti pairs. The wait-queue routines still
134 * need cli-sti, but now it's just a couple of 386 instructions or so.
136 * Note that the real wait_on_buffer() is an inline function that checks
137 * if 'b_wait' is set before calling this, so that the queues aren't set
140 void __wait_on_buffer(struct buffer_head
* bh
)
142 struct task_struct
*tsk
= current
;
143 DECLARE_WAITQUEUE(wait
, tsk
);
145 atomic_inc(&bh
->b_count
);
146 add_wait_queue(&bh
->b_wait
, &wait
);
148 tsk
->state
= TASK_UNINTERRUPTIBLE
;
149 run_task_queue(&tq_disk
);
150 if (buffer_locked(bh
)) {
154 tsk
->state
= TASK_RUNNING
;
155 remove_wait_queue(&bh
->b_wait
, &wait
);
156 atomic_dec(&bh
->b_count
);
159 /* Call sync_buffers with wait!=0 to ensure that the call does not
160 * return until all buffer writes have completed. Sync() may return
161 * before the writes have finished; fsync() may not.
164 /* Godamity-damn. Some buffers (bitmaps for filesystems)
165 * spontaneously dirty themselves without ever brelse being called.
166 * We will ultimately want to put these in a separate list, but for
167 * now we search all of the lists for dirty buffers.
169 static int sync_buffers(kdev_t dev
, int wait
)
171 int i
, retry
, pass
= 0, err
= 0;
172 struct buffer_head
* bh
, *next
;
174 /* One pass for no-wait, three for wait:
175 * 0) write out all dirty, unlocked buffers;
176 * 1) write out all dirty buffers, waiting if locked;
177 * 2) wait for completion by waiting for all buffers to unlock.
182 /* We search all lists as a failsafe mechanism, not because we expect
183 * there to be dirty buffers on any of the other lists.
186 spin_lock(&lru_list_lock
);
187 bh
= lru_list
[BUF_DIRTY
];
191 for (i
= nr_buffers_type
[BUF_DIRTY
]*2 ; i
-- > 0 ; bh
= next
) {
192 next
= bh
->b_next_free
;
194 if (!lru_list
[BUF_DIRTY
])
196 if (dev
&& bh
->b_dev
!= dev
)
198 if (buffer_locked(bh
)) {
199 /* Buffer is locked; skip it unless wait is
200 * requested AND pass > 0.
202 if (!wait
|| !pass
) {
206 atomic_inc(&bh
->b_count
);
207 spin_unlock(&lru_list_lock
);
209 atomic_dec(&bh
->b_count
);
213 /* If an unlocked buffer is not uptodate, there has
214 * been an IO error. Skip it.
216 if (wait
&& buffer_req(bh
) && !buffer_locked(bh
) &&
217 !buffer_dirty(bh
) && !buffer_uptodate(bh
)) {
222 /* Don't write clean buffers. Don't write ANY buffers
225 if (!buffer_dirty(bh
) || pass
>= 2)
228 atomic_inc(&bh
->b_count
);
230 spin_unlock(&lru_list_lock
);
231 ll_rw_block(WRITE
, 1, &bh
);
232 atomic_dec(&bh
->b_count
);
238 bh
= lru_list
[BUF_LOCKED
];
240 spin_unlock(&lru_list_lock
);
243 for (i
= nr_buffers_type
[BUF_LOCKED
]*2 ; i
-- > 0 ; bh
= next
) {
244 next
= bh
->b_next_free
;
246 if (!lru_list
[BUF_LOCKED
])
248 if (dev
&& bh
->b_dev
!= dev
)
250 if (buffer_locked(bh
)) {
251 /* Buffer is locked; skip it unless wait is
252 * requested AND pass > 0.
254 if (!wait
|| !pass
) {
258 atomic_inc(&bh
->b_count
);
259 spin_unlock(&lru_list_lock
);
261 spin_lock(&lru_list_lock
);
262 atomic_dec(&bh
->b_count
);
266 spin_unlock(&lru_list_lock
);
268 /* If we are waiting for the sync to succeed, and if any dirty
269 * blocks were written, then repeat; on the second pass, only
270 * wait for buffers being written (do not pass to write any
271 * more buffers on the second pass).
273 } while (wait
&& retry
&& ++pass
<=2);
277 void sync_dev(kdev_t dev
)
279 sync_buffers(dev
, 0);
282 sync_buffers(dev
, 0);
285 * FIXME(eric) we need to sync the physical devices here.
286 * This is because some (scsi) controllers have huge amounts of
287 * cache onboard (hundreds of Mb), and we need to instruct
288 * them to commit all of the dirty memory to disk, and we should
289 * not return until this has happened.
291 * This would need to get implemented by going through the assorted
292 * layers so that each block major number can be synced, and this
293 * would call down into the upper and mid-layer scsi.
297 int fsync_dev(kdev_t dev
)
299 sync_buffers(dev
, 0);
307 return sync_buffers(dev
, 1);
310 asmlinkage
int sys_sync(void)
317 * filp may be NULL if called via the msync of a vma.
320 int file_fsync(struct file
*filp
, struct dentry
*dentry
)
322 struct inode
* inode
= dentry
->d_inode
;
323 struct super_block
* sb
;
326 /* sync the inode to buffers */
327 write_inode_now(inode
);
329 /* sync the superblock to buffers */
332 if (sb
->s_op
&& sb
->s_op
->write_super
)
333 sb
->s_op
->write_super(sb
);
335 /* .. finally sync the buffers to disk */
337 return sync_buffers(dev
, 1);
340 asmlinkage
int sys_fsync(unsigned int fd
)
343 struct dentry
* dentry
;
344 struct inode
* inode
;
353 dentry
= file
->f_dentry
;
357 inode
= dentry
->d_inode
;
362 if (!file
->f_op
|| !file
->f_op
->fsync
)
365 /* We need to protect against concurrent writers.. */
367 err
= file
->f_op
->fsync(file
, dentry
);
377 asmlinkage
int sys_fdatasync(unsigned int fd
)
380 struct dentry
* dentry
;
381 struct inode
* inode
;
390 dentry
= file
->f_dentry
;
394 inode
= dentry
->d_inode
;
399 if (!file
->f_op
|| !file
->f_op
->fsync
)
402 /* this needs further work, at the moment it is identical to fsync() */
404 err
= file
->f_op
->fsync(file
, dentry
);
414 void invalidate_buffers(kdev_t dev
)
418 spin_lock(&lru_list_lock
);
419 for(nlist
= 0; nlist
< NR_LIST
; nlist
++) {
420 struct buffer_head
* bh
;
423 bh
= lru_list
[nlist
];
426 for (i
= nr_buffers_type
[nlist
]*2 ; --i
> 0 ; bh
= bh
->b_next_free
) {
427 if (bh
->b_dev
!= dev
)
429 if (buffer_locked(bh
)) {
430 atomic_inc(&bh
->b_count
);
431 spin_unlock(&lru_list_lock
);
433 spin_lock(&lru_list_lock
);
434 atomic_dec(&bh
->b_count
);
437 if (atomic_read(&bh
->b_count
))
440 clear_bit(BH_Protected
, &bh
->b_state
);
441 clear_bit(BH_Uptodate
, &bh
->b_state
);
442 clear_bit(BH_Dirty
, &bh
->b_state
);
443 clear_bit(BH_Req
, &bh
->b_state
);
446 spin_unlock(&lru_list_lock
);
449 /* After several hours of tedious analysis, the following hash
450 * function won. Do not mess with it... -DaveM
452 #define _hashfn(dev,block) \
453 ((((dev)<<(bh_hash_shift - 6)) ^ ((dev)<<(bh_hash_shift - 9))) ^ \
454 (((block)<<(bh_hash_shift - 6)) ^ ((block) >> 13) ^ ((block) << (bh_hash_shift - 12))))
455 #define hash(dev,block) hash_table[(_hashfn(dev,block) & bh_hash_mask)]
457 static __inline__
void __hash_link(struct buffer_head
*bh
, struct buffer_head
**head
)
459 if ((bh
->b_next
= *head
) != NULL
)
460 bh
->b_next
->b_pprev
= &bh
->b_next
;
465 static __inline__
void __hash_unlink(struct buffer_head
*bh
)
468 bh
->b_next
->b_pprev
= bh
->b_pprev
;
469 *(bh
->b_pprev
) = bh
->b_next
;
473 static void __insert_into_lru_list(struct buffer_head
* bh
, int blist
)
475 struct buffer_head
**bhp
= &lru_list
[blist
];
479 bh
->b_prev_free
= bh
;
481 bh
->b_next_free
= *bhp
;
482 bh
->b_prev_free
= (*bhp
)->b_prev_free
;
483 (*bhp
)->b_prev_free
->b_next_free
= bh
;
484 (*bhp
)->b_prev_free
= bh
;
485 nr_buffers_type
[blist
]++;
488 static void __remove_from_lru_list(struct buffer_head
* bh
, int blist
)
490 if (bh
->b_prev_free
|| bh
->b_next_free
) {
491 bh
->b_prev_free
->b_next_free
= bh
->b_next_free
;
492 bh
->b_next_free
->b_prev_free
= bh
->b_prev_free
;
493 if (lru_list
[blist
] == bh
)
494 lru_list
[blist
] = bh
->b_next_free
;
495 if (lru_list
[blist
] == bh
)
496 lru_list
[blist
] = NULL
;
497 bh
->b_next_free
= bh
->b_prev_free
= NULL
;
498 nr_buffers_type
[blist
]--;
502 static void __remove_from_free_list(struct buffer_head
* bh
, int index
)
504 if(bh
->b_next_free
== bh
)
505 free_list
[index
].list
= NULL
;
507 bh
->b_prev_free
->b_next_free
= bh
->b_next_free
;
508 bh
->b_next_free
->b_prev_free
= bh
->b_prev_free
;
509 if (free_list
[index
].list
== bh
)
510 free_list
[index
].list
= bh
->b_next_free
;
512 bh
->b_next_free
= bh
->b_prev_free
= NULL
;
515 /* The following two functions must operate atomically
516 * because they control the visibility of a buffer head
517 * to the rest of the kernel.
519 static __inline__
void __remove_from_queues(struct buffer_head
*bh
)
521 write_lock(&hash_table_lock
);
524 __remove_from_lru_list(bh
, bh
->b_list
);
525 write_unlock(&hash_table_lock
);
528 static void insert_into_queues(struct buffer_head
*bh
)
530 struct buffer_head
**head
= &hash(bh
->b_dev
, bh
->b_blocknr
);
532 spin_lock(&lru_list_lock
);
533 write_lock(&hash_table_lock
);
534 __hash_link(bh
, head
);
535 __insert_into_lru_list(bh
, bh
->b_list
);
536 write_unlock(&hash_table_lock
);
537 spin_unlock(&lru_list_lock
);
540 /* This function must only run if there are no other
541 * references _anywhere_ to this buffer head.
543 static void put_last_free(struct buffer_head
* bh
)
545 struct bh_free_head
*head
= &free_list
[BUFSIZE_INDEX(bh
->b_size
)];
546 struct buffer_head
**bhp
= &head
->list
;
548 spin_lock(&head
->lock
);
552 bh
->b_prev_free
= bh
;
554 bh
->b_next_free
= *bhp
;
555 bh
->b_prev_free
= (*bhp
)->b_prev_free
;
556 (*bhp
)->b_prev_free
->b_next_free
= bh
;
557 (*bhp
)->b_prev_free
= bh
;
558 spin_unlock(&head
->lock
);
562 * Why like this, I hear you say... The reason is race-conditions.
563 * As we don't lock buffers (unless we are reading them, that is),
564 * something might happen to it while we sleep (ie a read-error
565 * will force it bad). This shouldn't really happen currently, but
568 struct buffer_head
* get_hash_table(kdev_t dev
, int block
, int size
)
570 struct buffer_head
**head
= &hash(dev
, block
);
571 struct buffer_head
*bh
;
573 read_lock(&hash_table_lock
);
574 for(bh
= *head
; bh
; bh
= bh
->b_next
)
575 if (bh
->b_blocknr
== block
&&
576 bh
->b_size
== size
&&
580 atomic_inc(&bh
->b_count
);
581 read_unlock(&hash_table_lock
);
586 unsigned int get_hardblocksize(kdev_t dev
)
589 * Get the hard sector size for the given device. If we don't know
590 * what it is, return 0.
592 if (hardsect_size
[MAJOR(dev
)] != NULL
) {
593 int blksize
= hardsect_size
[MAJOR(dev
)][MINOR(dev
)];
599 * We don't know what the hardware sector size for this device is.
600 * Return 0 indicating that we don't know.
605 void set_blocksize(kdev_t dev
, int size
)
607 extern int *blksize_size
[];
609 struct buffer_head
* bh
, *bhnext
;
611 if (!blksize_size
[MAJOR(dev
)])
614 /* Size must be a power of two, and between 512 and PAGE_SIZE */
615 if (size
> PAGE_SIZE
|| size
< 512 || (size
& (size
-1)))
616 panic("Invalid blocksize passed to set_blocksize");
618 if (blksize_size
[MAJOR(dev
)][MINOR(dev
)] == 0 && size
== BLOCK_SIZE
) {
619 blksize_size
[MAJOR(dev
)][MINOR(dev
)] = size
;
622 if (blksize_size
[MAJOR(dev
)][MINOR(dev
)] == size
)
624 sync_buffers(dev
, 2);
625 blksize_size
[MAJOR(dev
)][MINOR(dev
)] = size
;
627 /* We need to be quite careful how we do this - we are moving entries
628 * around on the free list, and we can get in a loop if we are not careful.
630 for(nlist
= 0; nlist
< NR_LIST
; nlist
++) {
632 spin_lock(&lru_list_lock
);
633 bh
= lru_list
[nlist
];
634 for (i
= nr_buffers_type
[nlist
]*2 ; --i
> 0 ; bh
= bhnext
) {
638 bhnext
= bh
->b_next_free
;
639 if (bh
->b_dev
!= dev
)
641 if (bh
->b_size
== size
)
643 if (buffer_locked(bh
)) {
644 atomic_inc(&bh
->b_count
);
645 spin_unlock(&lru_list_lock
);
647 atomic_dec(&bh
->b_count
);
650 if (bh
->b_dev
== dev
&& bh
->b_size
!= size
) {
651 clear_bit(BH_Dirty
, &bh
->b_state
);
652 clear_bit(BH_Uptodate
, &bh
->b_state
);
653 clear_bit(BH_Req
, &bh
->b_state
);
656 if (atomic_read(&bh
->b_count
) == 0) {
657 __remove_from_queues(bh
);
661 spin_unlock(&lru_list_lock
);
666 * We used to try various strange things. Let's not.
668 static void refill_freelist(int size
)
670 if (!grow_buffers(size
)) {
672 current
->policy
|= SCHED_YIELD
;
677 void init_buffer(struct buffer_head
*bh
, bh_end_io_t
*handler
, void *dev_id
)
679 bh
->b_list
= BUF_CLEAN
;
681 bh
->b_end_io
= handler
;
682 bh
->b_dev_id
= dev_id
;
685 static void end_buffer_io_sync(struct buffer_head
*bh
, int uptodate
)
687 mark_buffer_uptodate(bh
, uptodate
);
691 static void end_buffer_io_bad(struct buffer_head
*bh
, int uptodate
)
693 mark_buffer_uptodate(bh
, uptodate
);
698 static void end_buffer_io_async(struct buffer_head
* bh
, int uptodate
)
700 static spinlock_t page_uptodate_lock
= SPIN_LOCK_UNLOCKED
;
702 struct buffer_head
*tmp
;
706 mark_buffer_uptodate(bh
, uptodate
);
708 /* This is a temporary buffer used for page I/O. */
709 page
= mem_map
+ MAP_NR(bh
->b_data
);
715 * Be _very_ careful from here on. Bad things can happen if
716 * two buffer heads end IO at almost the same time and both
717 * decide that the page is now completely done.
719 * Async buffer_heads are here only as labels for IO, and get
720 * thrown away once the IO for this page is complete. IO is
721 * deemed complete once all buffers have been visited
722 * (b_count==0) and are now unlocked. We must make sure that
723 * only the _last_ buffer that decrements its count is the one
724 * that free's the page..
726 spin_lock_irqsave(&page_uptodate_lock
, flags
);
728 atomic_dec(&bh
->b_count
);
729 tmp
= bh
->b_this_page
;
731 if (atomic_read(&tmp
->b_count
) &&
732 (tmp
->b_end_io
== end_buffer_io_async
))
734 tmp
= tmp
->b_this_page
;
737 /* OK, the async IO on this page is complete. */
738 spin_unlock_irqrestore(&page_uptodate_lock
, flags
);
741 * if none of the buffers had errors then we can set the
744 if (!PageError(page
))
745 SetPageUptodate(page
);
748 * Run the hooks that have to be done when a page I/O has completed.
750 * Note - we need to test the flags before we unlock the page, but
751 * we must not actually free the page until after the unlock!
753 if (test_and_clear_bit(PG_decr_after
, &page
->flags
))
754 atomic_dec(&nr_async_pages
);
756 if (test_and_clear_bit(PG_free_swap_after
, &page
->flags
))
757 swap_free(page
->offset
);
759 free
= test_and_clear_bit(PG_free_after
, &page
->flags
);
761 if (page
->owner
!= (void *)-1)
763 page
->owner
= current
;
772 spin_unlock_irqrestore(&page_uptodate_lock
, flags
);
778 * Ok, this is getblk, and it isn't very clear, again to hinder
779 * race-conditions. Most of the code is seldom used, (ie repeating),
780 * so it should be much more efficient than it looks.
782 * The algorithm is changed: hopefully better, and an elusive bug removed.
784 * 14.02.92: changed it to sync dirty buffers a bit: better performance
785 * when the filesystem starts to get full of dirty blocks (I hope).
787 struct buffer_head
* getblk(kdev_t dev
, int block
, int size
)
789 struct buffer_head
* bh
;
793 bh
= get_hash_table(dev
, block
, size
);
795 if (!buffer_dirty(bh
)) {
801 isize
= BUFSIZE_INDEX(size
);
802 spin_lock(&free_list
[isize
].lock
);
803 bh
= free_list
[isize
].list
;
805 __remove_from_free_list(bh
, isize
);
806 atomic_set(&bh
->b_count
, 1);
808 spin_unlock(&free_list
[isize
].lock
);
812 /* OK, FINALLY we know that this buffer is the only one of its kind,
813 * we hold a reference (b_count>0), it is unlocked, and it is clean.
815 init_buffer(bh
, end_buffer_io_sync
, NULL
);
817 bh
->b_blocknr
= block
;
818 bh
->b_state
= 1 << BH_Mapped
;
820 /* Insert the buffer into the regular lists */
821 insert_into_queues(bh
);
825 * If we block while refilling the free list, somebody may
826 * create the buffer first ... search the hashes again.
829 refill_freelist(size
);
836 * if a new dirty buffer is created we need to balance bdflush.
838 * in the future we might want to make bdflush aware of different
839 * pressures on different devices - thus the (currently unused)
842 int too_many_dirty_buffers
;
844 void balance_dirty(kdev_t dev
)
846 int dirty
= nr_buffers_type
[BUF_DIRTY
];
847 int ndirty
= bdf_prm
.b_un
.ndirty
;
849 if (dirty
> ndirty
) {
850 if (dirty
> 2*ndirty
) {
851 too_many_dirty_buffers
= 1;
857 too_many_dirty_buffers
= 0;
861 static inline void __mark_dirty(struct buffer_head
*bh
, int flag
)
863 bh
->b_flushtime
= jiffies
+ (flag
? bdf_prm
.b_un
.age_super
: bdf_prm
.b_un
.age_buffer
);
864 clear_bit(BH_New
, &bh
->b_state
);
868 void __mark_buffer_dirty(struct buffer_head
*bh
, int flag
)
870 __mark_dirty(bh
, flag
);
874 * A buffer may need to be moved from one buffer list to another
875 * (e.g. in case it is not shared any more). Handle this.
877 static __inline__
void __refile_buffer(struct buffer_head
*bh
)
879 int dispose
= BUF_CLEAN
;
880 if (buffer_locked(bh
))
881 dispose
= BUF_LOCKED
;
882 if (buffer_dirty(bh
))
884 if (dispose
!= bh
->b_list
) {
885 __remove_from_lru_list(bh
, bh
->b_list
);
886 bh
->b_list
= dispose
;
887 __insert_into_lru_list(bh
, dispose
);
891 void refile_buffer(struct buffer_head
*bh
)
893 spin_lock(&lru_list_lock
);
895 spin_unlock(&lru_list_lock
);
899 * Release a buffer head
901 void __brelse(struct buffer_head
* buf
)
905 if (atomic_read(&buf
->b_count
)) {
906 atomic_dec(&buf
->b_count
);
909 printk("VFS: brelse: Trying to free free buffer\n");
913 * bforget() is like brelse(), except it puts the buffer on the
914 * free list if it can.. We can NOT free the buffer if:
915 * - there are other users of it
916 * - it is locked and thus can have active IO
918 void __bforget(struct buffer_head
* buf
)
920 spin_lock(&lru_list_lock
);
921 write_lock(&hash_table_lock
);
922 if (atomic_read(&buf
->b_count
) != 1 || buffer_locked(buf
)) {
924 atomic_dec(&buf
->b_count
);
926 atomic_set(&buf
->b_count
, 0);
930 __remove_from_lru_list(buf
, buf
->b_list
);
933 write_unlock(&hash_table_lock
);
934 spin_unlock(&lru_list_lock
);
938 * bread() reads a specified block and returns the buffer that contains
939 * it. It returns NULL if the block was unreadable.
941 struct buffer_head
* bread(kdev_t dev
, int block
, int size
)
943 struct buffer_head
* bh
;
945 bh
= getblk(dev
, block
, size
);
946 if (buffer_uptodate(bh
))
948 ll_rw_block(READ
, 1, &bh
);
950 if (buffer_uptodate(bh
))
957 * Ok, breada can be used as bread, but additionally to mark other
958 * blocks for reading as well. End the argument list with a negative
964 struct buffer_head
* breada(kdev_t dev
, int block
, int bufsize
,
965 unsigned int pos
, unsigned int filesize
)
967 struct buffer_head
* bhlist
[NBUF
];
969 struct buffer_head
* bh
;
979 bh
= getblk(dev
, block
, bufsize
);
980 index
= BUFSIZE_INDEX(bh
->b_size
);
982 if (buffer_uptodate(bh
))
984 else ll_rw_block(READ
, 1, &bh
);
986 blocks
= (filesize
- pos
) >> (9+index
);
988 if (blocks
< (read_ahead
[MAJOR(dev
)] >> index
))
989 blocks
= read_ahead
[MAJOR(dev
)] >> index
;
993 /* if (blocks) printk("breada (new) %d blocks\n",blocks); */
997 for(i
=1; i
<blocks
; i
++) {
998 bh
= getblk(dev
,block
+i
,bufsize
);
999 if (buffer_uptodate(bh
)) {
1003 else bhlist
[j
++] = bh
;
1006 /* Request the read for these buffers, and then release them. */
1008 ll_rw_block(READA
, (j
-1), bhlist
+1);
1012 /* Wait for this buffer, and then continue on. */
1015 if (buffer_uptodate(bh
))
1022 * Note: the caller should wake up the buffer_wait list if needed.
1024 static __inline__
void __put_unused_buffer_head(struct buffer_head
* bh
)
1026 if (nr_unused_buffer_heads
>= MAX_UNUSED_BUFFERS
) {
1027 kmem_cache_free(bh_cachep
, bh
);
1030 init_waitqueue_head(&bh
->b_wait
);
1031 nr_unused_buffer_heads
++;
1032 bh
->b_next_free
= unused_list
;
1033 bh
->b_this_page
= NULL
;
1038 static void put_unused_buffer_head(struct buffer_head
*bh
)
1040 spin_lock(&unused_list_lock
);
1041 __put_unused_buffer_head(bh
);
1042 spin_unlock(&unused_list_lock
);
1046 * Reserve NR_RESERVED buffer heads for async IO requests to avoid
1047 * no-buffer-head deadlock. Return NULL on failure; waiting for
1048 * buffer heads is now handled in create_buffers().
1050 static struct buffer_head
* get_unused_buffer_head(int async
)
1052 struct buffer_head
* bh
;
1054 spin_lock(&unused_list_lock
);
1055 if (nr_unused_buffer_heads
> NR_RESERVED
) {
1057 unused_list
= bh
->b_next_free
;
1058 nr_unused_buffer_heads
--;
1059 spin_unlock(&unused_list_lock
);
1062 spin_unlock(&unused_list_lock
);
1064 /* This is critical. We can't swap out pages to get
1065 * more buffer heads, because the swap-out may need
1066 * more buffer-heads itself. Thus SLAB_BUFFER.
1068 if((bh
= kmem_cache_alloc(bh_cachep
, SLAB_BUFFER
)) != NULL
) {
1069 memset(bh
, 0, sizeof(*bh
));
1070 init_waitqueue_head(&bh
->b_wait
);
1075 * If we need an async buffer, use the reserved buffer heads.
1078 spin_lock(&unused_list_lock
);
1081 unused_list
= bh
->b_next_free
;
1082 nr_unused_buffer_heads
--;
1083 spin_unlock(&unused_list_lock
);
1086 spin_unlock(&unused_list_lock
);
1090 * (Pending further analysis ...)
1091 * Ordinary (non-async) requests can use a different memory priority
1092 * to free up pages. Any swapping thus generated will use async
1096 (bh
= kmem_cache_alloc(bh_cachep
, SLAB_KERNEL
)) != NULL
) {
1097 memset(bh
, 0, sizeof(*bh
));
1098 init_waitqueue_head(&bh
->b_wait
);
1107 * Create the appropriate buffers when given a page for data area and
1108 * the size of each buffer.. Use the bh->b_this_page linked list to
1109 * follow the buffers created. Return NULL if unable to create more
1111 * The async flag is used to differentiate async IO (paging, swapping)
1112 * from ordinary buffer allocations, and only async requests are allowed
1113 * to sleep waiting for buffer heads.
1115 static struct buffer_head
* create_buffers(unsigned long page
, unsigned long size
, int async
)
1117 DECLARE_WAITQUEUE(wait
, current
);
1118 struct buffer_head
*bh
, *head
;
1124 while ((offset
-= size
) >= 0) {
1125 bh
= get_unused_buffer_head(async
);
1129 bh
->b_dev
= B_FREE
; /* Flag as unused */
1130 bh
->b_this_page
= head
;
1134 bh
->b_next_free
= NULL
;
1136 atomic_set(&bh
->b_count
, 0);
1139 bh
->b_data
= (char *) (page
+offset
);
1140 bh
->b_list
= BUF_CLEAN
;
1141 bh
->b_flushtime
= 0;
1142 bh
->b_end_io
= end_buffer_io_bad
;
1146 * In case anything failed, we just free everything we got.
1152 head
= head
->b_this_page
;
1153 put_unused_buffer_head(bh
);
1156 /* Wake up any waiters ... */
1157 wake_up(&buffer_wait
);
1161 * Return failure for non-async IO requests. Async IO requests
1162 * are not allowed to fail, so we have to wait until buffer heads
1163 * become available. But we don't want tasks sleeping with
1164 * partially complete buffers, so all were released above.
1169 /* We're _really_ low on memory. Now we just
1170 * wait for old buffer heads to become free due to
1171 * finishing IO. Since this is an async request and
1172 * the reserve list is empty, we're sure there are
1173 * async buffer heads in use.
1175 run_task_queue(&tq_disk
);
1178 * Set our state for sleeping, then check again for buffer heads.
1179 * This ensures we won't miss a wake_up from an interrupt.
1181 add_wait_queue(&buffer_wait
, &wait
);
1182 current
->state
= TASK_UNINTERRUPTIBLE
;
1183 if (nr_unused_buffer_heads
< MAX_BUF_PER_PAGE
) {
1184 current
->policy
|= SCHED_YIELD
;
1187 remove_wait_queue(&buffer_wait
, &wait
);
1188 current
->state
= TASK_RUNNING
;
1192 static int create_page_buffers(int rw
, struct page
*page
, kdev_t dev
, int b
[], int size
, int bmap
)
1194 struct buffer_head
*head
, *bh
, *tail
;
1197 if (!PageLocked(page
))
1199 if (page
->owner
!= current
)
1202 * Allocate async buffer heads pointing to this page, just for I/O.
1203 * They show up in the buffer hash table and are registered in
1206 head
= create_buffers(page_address(page
), size
, 1);
1212 for (bh
= head
; bh
; bh
= bh
->b_this_page
) {
1216 init_buffer(bh
, end_buffer_io_async
, NULL
);
1218 bh
->b_blocknr
= block
;
1221 * When we use bmap, we define block zero to represent
1222 * a hole. ll_rw_page, however, may legitimately
1223 * access block zero, and we need to distinguish the
1226 if (bmap
&& !block
) {
1227 memset(bh
->b_data
, 0, size
);
1228 set_bit(BH_Uptodate
, &bh
->b_state
);
1231 set_bit(BH_Mapped
, &bh
->b_state
);
1233 tail
->b_this_page
= head
;
1235 page
->buffers
= head
;
1240 * We don't have to release all buffers here, but
1241 * we have to be sure that no dirty buffer is left
1242 * and no IO is going on (no buffer is locked), because
1243 * we have truncated the file and are going to free the
1246 int block_flushpage(struct inode
*inode
, struct page
*page
, unsigned long offset
)
1248 struct buffer_head
*head
, *bh
, *next
;
1249 unsigned int curr_off
= 0;
1251 if (!PageLocked(page
))
1256 head
= page
->buffers
;
1259 unsigned int next_off
= curr_off
+ bh
->b_size
;
1260 next
= bh
->b_this_page
;
1263 * is this block fully flushed?
1265 if (offset
<= curr_off
) {
1266 if (buffer_mapped(bh
)) {
1267 atomic_inc(&bh
->b_count
);
1269 if (bh
->b_dev
== B_FREE
)
1271 mark_buffer_clean(bh
);
1272 clear_bit(BH_Uptodate
, &bh
->b_state
);
1273 clear_bit(BH_Mapped
, &bh
->b_state
);
1274 clear_bit(BH_Req
, &bh
->b_state
);
1276 atomic_dec(&bh
->b_count
);
1279 curr_off
= next_off
;
1281 } while (bh
!= head
);
1284 * subtle. We release buffer-heads only if this is
1285 * the 'final' flushpage. We have invalidated the bmap
1286 * cached value unconditionally, so real IO is not
1289 * If the free doesn't work out, the buffers can be
1290 * left around - they just turn into anonymous buffers
1294 if (!try_to_free_buffers(page
))
1295 atomic_add(PAGE_CACHE_SIZE
, &buffermem
);
1301 static void create_empty_buffers(struct page
*page
, struct inode
*inode
, unsigned long blocksize
)
1303 struct buffer_head
*bh
, *head
, *tail
;
1305 head
= create_buffers(page_address(page
), blocksize
, 1);
1311 bh
->b_dev
= inode
->i_dev
;
1313 bh
->b_end_io
= end_buffer_io_bad
;
1315 bh
= bh
->b_this_page
;
1317 tail
->b_this_page
= head
;
1318 page
->buffers
= head
;
1323 * block_write_full_page() is SMP-safe - currently it's still
1324 * being called with the kernel lock held, but the code is ready.
1326 int block_write_full_page(struct file
*file
, struct page
*page
)
1328 struct dentry
*dentry
= file
->f_dentry
;
1329 struct inode
*inode
= dentry
->d_inode
;
1331 unsigned long block
, offset
;
1332 struct buffer_head
*bh
, *head
;
1334 if (!PageLocked(page
))
1338 create_empty_buffers(page
, inode
, inode
->i_sb
->s_blocksize
);
1339 head
= page
->buffers
;
1341 offset
= page
->offset
;
1342 block
= offset
>> inode
->i_sb
->s_blocksize_bits
;
1344 // FIXME: currently we assume page alignment.
1345 if (offset
& (PAGE_SIZE
-1))
1355 * If the buffer isn't up-to-date, we can't be sure
1356 * that the buffer has been initialized with the proper
1357 * block number information etc..
1359 * Leave it to the low-level FS to make all those
1360 * decisions (block #0 may actually be a valid block)
1362 bh
->b_end_io
= end_buffer_io_sync
;
1363 if (!buffer_mapped(bh
)) {
1364 err
= inode
->i_op
->get_block(inode
, block
, bh
, 1);
1368 set_bit(BH_Uptodate
, &bh
->b_state
);
1369 mark_buffer_dirty(bh
,0);
1371 bh
= bh
->b_this_page
;
1373 } while (bh
!= head
);
1375 SetPageUptodate(page
);
1378 ClearPageUptodate(page
);
1382 int block_write_partial_page(struct file
*file
, struct page
*page
, unsigned long offset
, unsigned long bytes
, const char * buf
)
1384 struct dentry
*dentry
= file
->f_dentry
;
1385 struct inode
*inode
= dentry
->d_inode
;
1386 unsigned long block
;
1388 unsigned long blocksize
, start_block
, end_block
;
1389 unsigned long start_offset
, start_bytes
, end_bytes
;
1390 unsigned long bbits
, blocks
, i
, len
;
1391 struct buffer_head
*bh
, *head
;
1394 target_buf
= (char *)page_address(page
) + offset
;
1396 if (!PageLocked(page
))
1399 blocksize
= inode
->i_sb
->s_blocksize
;
1401 create_empty_buffers(page
, inode
, blocksize
);
1402 head
= page
->buffers
;
1404 bbits
= inode
->i_sb
->s_blocksize_bits
;
1405 block
= page
->offset
>> bbits
;
1406 blocks
= PAGE_SIZE
>> bbits
;
1407 start_block
= offset
>> bbits
;
1408 end_block
= (offset
+ bytes
- 1) >> bbits
;
1409 start_offset
= offset
& (blocksize
- 1);
1410 start_bytes
= blocksize
- start_offset
;
1411 if (start_bytes
> bytes
)
1412 start_bytes
= bytes
;
1413 end_bytes
= (offset
+bytes
) & (blocksize
- 1);
1414 if (end_bytes
> bytes
)
1417 if (offset
< 0 || offset
>= PAGE_SIZE
)
1419 if (bytes
+offset
< 0 || bytes
+offset
> PAGE_SIZE
)
1421 if (start_block
< 0 || start_block
>= blocks
)
1423 if (end_block
< 0 || end_block
>= blocks
)
1425 // FIXME: currently we assume page alignment.
1426 if (page
->offset
& (PAGE_SIZE
-1))
1436 if ((i
< start_block
) || (i
> end_block
)) {
1437 if (!buffer_uptodate(bh
))
1443 * If the buffer is not up-to-date, we need to ask the low-level
1444 * FS to do something for us (we used to have assumptions about
1445 * the meaning of b_blocknr etc, that's bad).
1447 * If "update" is set, that means that the low-level FS should
1448 * try to make sure that the block is up-to-date because we're
1449 * not going to fill it completely.
1451 bh
->b_end_io
= end_buffer_io_sync
;
1452 if (!buffer_mapped(bh
)) {
1453 err
= inode
->i_op
->get_block(inode
, block
, bh
, 1);
1458 if (!buffer_uptodate(bh
) && (start_offset
|| (end_bytes
&& (i
== end_block
)))) {
1459 if (buffer_new(bh
)) {
1460 memset(bh
->b_data
, 0, bh
->b_size
);
1462 ll_rw_block(READ
, 1, &bh
);
1465 if (!buffer_uptodate(bh
))
1474 } else if (end_bytes
&& (i
== end_block
)) {
1478 err
= copy_from_user(target_buf
, buf
, len
);
1483 * we dirty buffers only after copying the data into
1484 * the page - this way we can dirty the buffer even if
1485 * the bh is still doing IO.
1487 * NOTE! This also does a direct dirty balace check,
1488 * rather than relying on bdflush just waking up every
1489 * once in a while. This is to catch (and slow down)
1490 * the processes that write tons of buffer..
1492 * Note how we do NOT want to do this in the full block
1493 * case: full pages are flushed not by the people who
1494 * dirtied them, but by people who need memory. And we
1495 * should not penalize them for somebody else writing
1496 * lots of dirty pages.
1498 set_bit(BH_Uptodate
, &bh
->b_state
);
1499 if (!test_and_set_bit(BH_Dirty
, &bh
->b_state
)) {
1500 __mark_dirty(bh
, 0);
1501 if (too_many_dirty_buffers
)
1502 balance_dirty(bh
->b_dev
);
1513 bh
= bh
->b_this_page
;
1514 } while (bh
!= head
);
1517 * is this a partial write that happened to make all buffers
1518 * uptodate then we can optimize away a bogus readpage() for
1519 * the next read(). Here we 'discover' wether the page went
1520 * uptodate as a result of this (potentially partial) write.
1523 SetPageUptodate(page
);
1526 ClearPageUptodate(page
);
1531 * Start I/O on a page.
1532 * This function expects the page to be locked and may return
1533 * before I/O is complete. You then have to check page->locked,
1534 * page->uptodate, and maybe wait on page->wait.
1536 * brw_page() is SMP-safe, although it's being called with the
1537 * kernel lock held - but the code is ready.
1539 * FIXME: we need a swapper_inode->get_block function to remove
1540 * some of the bmap kludges and interface ugliness here.
1542 int brw_page(int rw
, struct page
*page
, kdev_t dev
, int b
[], int size
, int bmap
)
1544 struct buffer_head
*head
, *bh
, *arr
[MAX_BUF_PER_PAGE
];
1545 int nr
, fresh
/* temporary debugging flag */, block
;
1547 if (!PageLocked(page
))
1548 panic("brw_page: page not locked for I/O");
1549 // clear_bit(PG_error, &page->flags);
1551 * We pretty much rely on the page lock for this, because
1552 * create_page_buffers() might sleep.
1555 if (!page
->buffers
) {
1556 create_page_buffers(rw
, page
, dev
, b
, size
, bmap
);
1561 page
->owner
= (void *)-1;
1563 head
= page
->buffers
;
1569 if (fresh
&& (atomic_read(&bh
->b_count
) != 0))
1574 if (bmap
&& !block
) {
1580 if (!buffer_uptodate(bh
)) {
1582 atomic_inc(&bh
->b_count
);
1585 } else { /* WRITE */
1586 if (!bh
->b_blocknr
) {
1589 bh
->b_blocknr
= block
;
1594 set_bit(BH_Uptodate
, &bh
->b_state
);
1595 set_bit(BH_Dirty
, &bh
->b_state
);
1597 atomic_inc(&bh
->b_count
);
1599 bh
= bh
->b_this_page
;
1600 } while (bh
!= head
);
1603 if ((rw
== READ
) && nr
) {
1604 if (Page_Uptodate(page
))
1606 ll_rw_block(rw
, nr
, arr
);
1608 if (!nr
&& rw
== READ
) {
1609 SetPageUptodate(page
);
1610 page
->owner
= current
;
1613 if (nr
&& (rw
== WRITE
))
1614 ll_rw_block(rw
, nr
, arr
);
1620 * Generic "read page" function for block devices that have the normal
1621 * bmap functionality. This is most of the block device filesystems.
1622 * Reads the page asynchronously --- the unlock_buffer() and
1623 * mark_buffer_uptodate() functions propagate buffer state into the
1624 * page struct once IO has completed.
1626 int block_read_full_page(struct file
* file
, struct page
* page
)
1628 struct dentry
*dentry
= file
->f_dentry
;
1629 struct inode
*inode
= dentry
->d_inode
;
1630 unsigned long iblock
;
1631 struct buffer_head
*bh
, *head
, *arr
[MAX_BUF_PER_PAGE
];
1632 unsigned int blocksize
, blocks
;
1635 if (!PageLocked(page
))
1637 blocksize
= inode
->i_sb
->s_blocksize
;
1639 create_empty_buffers(page
, inode
, blocksize
);
1640 head
= page
->buffers
;
1642 blocks
= PAGE_SIZE
>> inode
->i_sb
->s_blocksize_bits
;
1643 iblock
= page
->offset
>> inode
->i_sb
->s_blocksize_bits
;
1644 page
->owner
= (void *)-1;
1645 head
= page
->buffers
;
1650 if (buffer_uptodate(bh
))
1653 if (!buffer_mapped(bh
)) {
1654 inode
->i_op
->get_block(inode
, iblock
, bh
, 0);
1655 if (!buffer_mapped(bh
)) {
1656 memset(bh
->b_data
, 0, blocksize
);
1657 set_bit(BH_Uptodate
, &bh
->b_state
);
1662 init_buffer(bh
, end_buffer_io_async
, NULL
);
1663 atomic_inc(&bh
->b_count
);
1666 } while (iblock
++, (bh
= bh
->b_this_page
) != head
);
1670 if (Page_Uptodate(page
))
1672 ll_rw_block(READ
, nr
, arr
);
1675 * all buffers are uptodate - we can set the page
1678 SetPageUptodate(page
);
1679 page
->owner
= current
;
1686 * Try to increase the number of buffers available: the size argument
1687 * is used to determine what kind of buffers we want.
1689 static int grow_buffers(int size
)
1692 struct buffer_head
*bh
, *tmp
;
1693 struct buffer_head
* insert_point
;
1696 if ((size
& 511) || (size
> PAGE_SIZE
)) {
1697 printk("VFS: grow_buffers: size = %d\n",size
);
1701 if (!(page
= __get_free_page(GFP_BUFFER
)))
1703 bh
= create_buffers(page
, size
, 0);
1709 isize
= BUFSIZE_INDEX(size
);
1711 spin_lock(&free_list
[isize
].lock
);
1712 insert_point
= free_list
[isize
].list
;
1716 tmp
->b_next_free
= insert_point
->b_next_free
;
1717 tmp
->b_prev_free
= insert_point
;
1718 insert_point
->b_next_free
->b_prev_free
= tmp
;
1719 insert_point
->b_next_free
= tmp
;
1721 tmp
->b_prev_free
= tmp
;
1722 tmp
->b_next_free
= tmp
;
1725 if (tmp
->b_this_page
)
1726 tmp
= tmp
->b_this_page
;
1730 tmp
->b_this_page
= bh
;
1731 free_list
[isize
].list
= bh
;
1732 spin_unlock(&free_list
[isize
].lock
);
1734 mem_map
[MAP_NR(page
)].buffers
= bh
;
1735 atomic_add(PAGE_SIZE
, &buffermem
);
1740 * Can the buffer be thrown out?
1742 #define BUFFER_BUSY_BITS ((1<<BH_Dirty) | (1<<BH_Lock) | (1<<BH_Protected))
1743 #define buffer_busy(bh) (atomic_read(&(bh)->b_count) | ((bh)->b_state & BUFFER_BUSY_BITS))
1746 * try_to_free_buffers() checks if all the buffers on this particular page
1747 * are unused, and free's the page if so.
1749 * Wake up bdflush() if this fails - if we're running low on memory due
1750 * to dirty buffers, we need to flush them out as quickly as possible.
1752 * NOTE: There are quite a number of ways that threads of control can
1753 * obtain a reference to a buffer head within a page. So we must
1754 * lock out all of these paths to cleanly toss the page.
1756 int try_to_free_buffers(struct page
* page
)
1758 struct buffer_head
* tmp
, * bh
= page
->buffers
;
1759 int index
= BUFSIZE_INDEX(bh
->b_size
);
1762 spin_lock(&lru_list_lock
);
1763 write_lock(&hash_table_lock
);
1764 spin_lock(&free_list
[index
].lock
);
1767 struct buffer_head
* p
= tmp
;
1769 tmp
= tmp
->b_this_page
;
1771 goto busy_buffer_page
;
1772 } while (tmp
!= bh
);
1774 spin_lock(&unused_list_lock
);
1777 struct buffer_head
* p
= tmp
;
1778 tmp
= tmp
->b_this_page
;
1780 /* The buffer can be either on the regular
1781 * queues or on the free list..
1783 if (p
->b_dev
== B_FREE
) {
1784 __remove_from_free_list(p
, index
);
1788 __remove_from_lru_list(p
, p
->b_list
);
1790 __put_unused_buffer_head(p
);
1791 } while (tmp
!= bh
);
1792 spin_unlock(&unused_list_lock
);
1794 /* Wake up anyone waiting for buffer heads */
1795 wake_up(&buffer_wait
);
1797 /* And free the page */
1798 page
->buffers
= NULL
;
1802 spin_unlock(&free_list
[index
].lock
);
1803 write_unlock(&hash_table_lock
);
1804 spin_unlock(&lru_list_lock
);
1808 /* Uhhuh, start writeback so that we don't end up with all dirty pages */
1809 too_many_dirty_buffers
= 1;
1815 /* ===================== Init ======================= */
1818 * allocate the hash table and init the free list
1819 * Use gfp() for the hash table to decrease TLB misses, use
1820 * SLAB cache for buffer heads.
1822 void __init
buffer_init(unsigned long memory_size
)
1825 unsigned int nr_hash
;
1827 /* The buffer cache hash table is less important these days,
1831 memory_size
*= sizeof(struct buffer_head
*);
1832 for (order
= 0; (PAGE_SIZE
<< order
) < memory_size
; order
++)
1835 /* try to allocate something until we get it or we're asking
1836 for something that is really too small */
1841 nr_hash
= (PAGE_SIZE
<< order
) / sizeof(struct buffer_head
*);
1842 bh_hash_mask
= (nr_hash
- 1);
1846 while((tmp
>>= 1UL) != 0UL)
1849 hash_table
= (struct buffer_head
**)
1850 __get_free_pages(GFP_ATOMIC
, order
);
1851 } while (hash_table
== NULL
&& --order
> 0);
1852 printk("Buffer-cache hash table entries: %d (order: %d, %ld bytes)\n",
1853 nr_hash
, order
, (1UL<<order
) * PAGE_SIZE
);
1856 panic("Failed to allocate buffer hash table\n");
1858 /* Setup hash chains. */
1859 for(i
= 0; i
< nr_hash
; i
++)
1860 hash_table
[i
] = NULL
;
1862 /* Setup free lists. */
1863 for(i
= 0; i
< NR_SIZES
; i
++) {
1864 free_list
[i
].list
= NULL
;
1865 free_list
[i
].lock
= SPIN_LOCK_UNLOCKED
;
1868 /* Setup lru lists. */
1869 for(i
= 0; i
< NR_LIST
; i
++)
1872 bh_cachep
= kmem_cache_create("buffer_head",
1873 sizeof(struct buffer_head
),
1875 SLAB_HWCACHE_ALIGN
, NULL
, NULL
);
1877 panic("Cannot create buffer head SLAB cache\n");
1881 /* ====================== bdflush support =================== */
1883 /* This is a simple kernel daemon, whose job it is to provide a dynamic
1884 * response to dirty buffers. Once this process is activated, we write back
1885 * a limited number of buffers to the disks and then go back to sleep again.
1887 static DECLARE_WAIT_QUEUE_HEAD(bdflush_wait
);
1888 static DECLARE_WAIT_QUEUE_HEAD(bdflush_done
);
1889 struct task_struct
*bdflush_tsk
= 0;
1891 void wakeup_bdflush(int wait
)
1893 if (current
== bdflush_tsk
)
1896 run_task_queue(&tq_disk
);
1897 wake_up(&bdflush_wait
);
1899 sleep_on(&bdflush_done
);
1904 * Here we attempt to write back old buffers. We also try to flush inodes
1905 * and supers as well, since this function is essentially "update", and
1906 * otherwise there would be no way of ensuring that these quantities ever
1907 * get written back. Ideally, we would have a timestamp on the inodes
1908 * and superblocks so that we could write back only the old ones as well
1911 static int sync_old_buffers(void)
1920 for(nlist
= BUF_LOCKED
; nlist
<= BUF_DIRTY
; nlist
++) {
1921 struct buffer_head
*bh
;
1923 spin_lock(&lru_list_lock
);
1924 bh
= lru_list
[nlist
];
1926 struct buffer_head
*next
;
1928 for (i
= nr_buffers_type
[nlist
]; i
-- > 0; bh
= next
) {
1929 next
= bh
->b_next_free
;
1931 /* If the buffer is not on the proper list,
1934 if ((nlist
== BUF_DIRTY
&&
1935 (!buffer_dirty(bh
) && !buffer_locked(bh
))) ||
1936 (nlist
== BUF_LOCKED
&& !buffer_locked(bh
))) {
1937 __refile_buffer(bh
);
1941 if (buffer_locked(bh
) || !buffer_dirty(bh
))
1944 /* OK, now we are committed to write it out. */
1945 bh
->b_flushtime
= 0;
1946 atomic_inc(&bh
->b_count
);
1947 spin_unlock(&lru_list_lock
);
1948 ll_rw_block(WRITE
, 1, &bh
);
1949 atomic_dec(&bh
->b_count
);
1953 spin_unlock(&lru_list_lock
);
1955 run_task_queue(&tq_disk
);
1959 struct mm_struct
* start_lazy_tlb(void)
1961 struct mm_struct
*mm
= current
->mm
;
1962 atomic_inc(&mm
->mm_count
);
1964 /* active_mm is still 'mm' */
1968 void end_lazy_tlb(struct mm_struct
*mm
)
1970 struct mm_struct
*active_mm
= current
->active_mm
;
1973 if (mm
!= active_mm
) {
1974 current
->active_mm
= mm
;
1980 /* This is the interface to bdflush. As we get more sophisticated, we can
1981 * pass tuning parameters to this "process", to adjust how it behaves.
1982 * We would want to verify each parameter, however, to make sure that it
1985 asmlinkage
int sys_bdflush(int func
, long data
)
1987 if (!capable(CAP_SYS_ADMIN
))
1992 struct mm_struct
*user_mm
;
1995 * bdflush will spend all of it's time in kernel-space,
1996 * without touching user-space, so we can switch it into
1997 * 'lazy TLB mode' to reduce the cost of context-switches
1998 * to and from bdflush.
2000 user_mm
= start_lazy_tlb();
2001 error
= sync_old_buffers();
2002 end_lazy_tlb(user_mm
);
2006 /* Basically func 1 means read param 1, 2 means write param 1, etc */
2008 int i
= (func
-2) >> 1;
2009 if (i
>= 0 && i
< N_PARAM
) {
2010 if ((func
& 1) == 0)
2011 return put_user(bdf_prm
.data
[i
], (int*)data
);
2013 if (data
>= bdflush_min
[i
] && data
<= bdflush_max
[i
]) {
2014 bdf_prm
.data
[i
] = data
;
2021 /* Having func 0 used to launch the actual bdflush and then never
2022 * return (unless explicitly killed). We return zero here to
2023 * remain semi-compatible with present update(8) programs.
2029 * This is the actual bdflush daemon itself. It used to be started from
2030 * the syscall above, but now we launch it ourselves internally with
2031 * kernel_thread(...) directly after the first thread in init/main.c
2033 int bdflush(void * unused
)
2036 * We have a bare-bones task_struct, and really should fill
2037 * in a few more things so "top" and /proc/2/{exe,root,cwd}
2038 * display semi-sane things. Not real crucial though...
2041 current
->session
= 1;
2043 sprintf(current
->comm
, "kflushd");
2044 bdflush_tsk
= current
;
2049 CHECK_EMERGENCY_SYNC
2051 for(nlist
= BUF_LOCKED
; nlist
<= BUF_DIRTY
; nlist
++) {
2052 int nr
, major
, written
= 0;
2053 struct buffer_head
*next
;
2056 spin_lock(&lru_list_lock
);
2057 next
= lru_list
[nlist
];
2058 nr
= nr_buffers_type
[nlist
];
2060 struct buffer_head
*bh
= next
;
2062 next
= next
->b_next_free
;
2064 /* If the buffer is not on the correct list,
2067 if ((nlist
== BUF_DIRTY
&&
2068 (!buffer_dirty(bh
) && !buffer_locked(bh
))) ||
2069 (nlist
== BUF_LOCKED
&& !buffer_locked(bh
))) {
2070 __refile_buffer(bh
);
2074 /* If we aren't in panic mode, don't write out too much
2075 * at a time. Also, don't write out buffers we don't
2076 * really have to write out yet..
2078 if (!too_many_dirty_buffers
) {
2079 if (written
> bdf_prm
.b_un
.ndirty
)
2081 if (time_before(jiffies
, bh
->b_flushtime
))
2085 if (buffer_locked(bh
) || !buffer_dirty(bh
))
2088 major
= MAJOR(bh
->b_dev
);
2090 bh
->b_flushtime
= 0;
2093 * For the loop major we can try to do asynchronous writes,
2094 * but we have to guarantee that we're making some progress..
2096 atomic_inc(&bh
->b_count
);
2097 spin_unlock(&lru_list_lock
);
2098 if (major
== LOOP_MAJOR
&& written
> 1) {
2099 ll_rw_block(WRITEA
, 1, &bh
);
2100 if (buffer_dirty(bh
))
2103 ll_rw_block(WRITE
, 1, &bh
);
2104 atomic_dec(&bh
->b_count
);
2107 spin_unlock(&lru_list_lock
);
2109 run_task_queue(&tq_disk
);
2110 wake_up(&bdflush_done
);
2113 * If there are still a lot of dirty buffers around,
2114 * skip the sleep and flush some more. Otherwise, we
2115 * sleep for a while and mark us as not being in panic
2118 if (!too_many_dirty_buffers
|| nr_buffers_type
[BUF_DIRTY
] < bdf_prm
.b_un
.ndirty
) {
2119 too_many_dirty_buffers
= 0;
2120 spin_lock_irq(¤t
->sigmask_lock
);
2121 flush_signals(current
);
2122 spin_unlock_irq(¤t
->sigmask_lock
);
2123 interruptible_sleep_on_timeout(&bdflush_wait
, 5*HZ
);