2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/vfs/hammer/hammer_io.c,v 1.31 2008/05/15 03:36:40 dillon Exp $
37 * IO Primitives and buffer cache management
39 * All major data-tracking structures in HAMMER contain a struct hammer_io
40 * which is used to manage their backing store. We use filesystem buffers
41 * for backing store and we leave them passively associated with their
44 * If the kernel tries to release a passively associated buf which we cannot
45 * yet let go we set B_LOCKED in the buffer and then actively released it
50 #include <sys/fcntl.h>
51 #include <sys/nlookup.h>
55 static void hammer_io_modify(hammer_io_t io
, int count
);
56 static void hammer_io_deallocate(struct buf
*bp
);
59 * Initialize a new, already-zero'd hammer_io structure, or reinitialize
60 * an existing hammer_io structure which may have switched to another type.
63 hammer_io_init(hammer_io_t io
, hammer_mount_t hmp
, enum hammer_io_type type
)
70 hammer_io_reinit(hammer_io_t io
, enum hammer_io_type type
)
72 hammer_mount_t hmp
= io
->hmp
;
75 KKASSERT(io
->mod_list
!= NULL
);
76 if (io
->mod_list
== &hmp
->volu_list
||
77 io
->mod_list
== &hmp
->meta_list
) {
78 --hmp
->locked_dirty_count
;
79 --hammer_count_dirtybufs
;
81 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
87 case HAMMER_STRUCTURE_VOLUME
:
88 io
->mod_list
= &hmp
->volu_list
;
89 ++hmp
->locked_dirty_count
;
90 ++hammer_count_dirtybufs
;
92 case HAMMER_STRUCTURE_META_BUFFER
:
93 io
->mod_list
= &hmp
->meta_list
;
94 ++hmp
->locked_dirty_count
;
95 ++hammer_count_dirtybufs
;
97 case HAMMER_STRUCTURE_UNDO_BUFFER
:
98 io
->mod_list
= &hmp
->undo_list
;
100 case HAMMER_STRUCTURE_DATA_BUFFER
:
101 io
->mod_list
= &hmp
->data_list
;
104 TAILQ_INSERT_TAIL(io
->mod_list
, io
, mod_entry
);
109 * Helper routine to disassociate a buffer cache buffer from an I/O
110 * structure. Called with the io structure exclusively locked.
112 * The io may have 0 or 1 references depending on who called us. The
113 * caller is responsible for dealing with the refs.
115 * This call can only be made when no action is required on the buffer.
116 * HAMMER must own the buffer (released == 0) since we mess around with it.
119 hammer_io_disassociate(hammer_io_structure_t iou
, int elseit
)
121 struct buf
*bp
= iou
->io
.bp
;
123 KKASSERT(iou
->io
.modified
== 0);
126 bp
->b_flags
&= ~B_LOCKED
;
128 KKASSERT(iou
->io
.released
== 0);
129 iou
->io
.released
= 1;
132 KKASSERT(iou
->io
.released
);
135 switch(iou
->io
.type
) {
136 case HAMMER_STRUCTURE_VOLUME
:
137 iou
->volume
.ondisk
= NULL
;
139 case HAMMER_STRUCTURE_DATA_BUFFER
:
140 case HAMMER_STRUCTURE_META_BUFFER
:
141 case HAMMER_STRUCTURE_UNDO_BUFFER
:
142 iou
->buffer
.ondisk
= NULL
;
148 * Wait for any physical IO to complete
151 hammer_io_wait(hammer_io_t io
)
155 tsleep_interlock(io
);
158 tsleep(io
, 0, "hmrflw", 0);
159 if (io
->running
== 0)
161 tsleep_interlock(io
);
163 if (io
->running
== 0)
171 * Load bp for a HAMMER structure. The io must be exclusively locked by
175 hammer_io_read(struct vnode
*devvp
, struct hammer_io
*io
)
180 if ((bp
= io
->bp
) == NULL
) {
181 error
= bread(devvp
, io
->offset
, HAMMER_BUFSIZE
, &io
->bp
);
184 bp
->b_ops
= &hammer_bioops
;
185 LIST_INSERT_HEAD(&bp
->b_dep
, &io
->worklist
, node
);
188 KKASSERT(io
->modified
== 0);
189 KKASSERT(io
->running
== 0);
190 KKASSERT(io
->waiting
== 0);
191 io
->released
= 0; /* we hold an active lock on bp */
199 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
200 * Must be called with the IO exclusively locked.
202 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
203 * I/O by forcing the buffer to not be in a released state before calling
206 * This function will also mark the IO as modified but it will not
207 * increment the modify_refs count.
210 hammer_io_new(struct vnode
*devvp
, struct hammer_io
*io
)
214 if ((bp
= io
->bp
) == NULL
) {
215 io
->bp
= getblk(devvp
, io
->offset
, HAMMER_BUFSIZE
, 0, 0);
217 bp
->b_ops
= &hammer_bioops
;
218 LIST_INSERT_HEAD(&bp
->b_dep
, &io
->worklist
, node
);
220 KKASSERT(io
->running
== 0);
230 hammer_io_modify(io
, 0);
236 * This routine is called on the last reference to a hammer structure.
237 * The io is usually locked exclusively (but may not be during unmount).
239 * This routine is responsible for the disposition of the buffer cache
240 * buffer backing the IO. Only pure-data and undo buffers can be handed
241 * back to the kernel. Volume and meta-data buffers must be retained
242 * by HAMMER until explicitly flushed by the backend.
245 hammer_io_release(struct hammer_io
*io
, int flush
)
249 if ((bp
= io
->bp
) == NULL
)
253 * Try to flush a dirty IO to disk if asked to by the
254 * caller or if the kernel tried to flush the buffer in the past.
256 * Kernel-initiated flushes are only allowed for pure-data buffers.
257 * meta-data and volume buffers can only be flushed explicitly
263 } else if (bp
->b_flags
& B_LOCKED
) {
265 case HAMMER_STRUCTURE_DATA_BUFFER
:
266 case HAMMER_STRUCTURE_UNDO_BUFFER
:
272 } /* else no explicit request to flush the buffer */
276 * Wait for the IO to complete if asked to.
278 if (io
->waitdep
&& io
->running
) {
283 * Return control of the buffer to the kernel (with the provisio
284 * that our bioops can override kernel decisions with regards to
287 if (flush
&& io
->modified
== 0 && io
->running
== 0) {
289 * Always disassociate the bp if an explicit flush
290 * was requested and the IO completed with no error
291 * (so unmount can really clean up the structure).
298 hammer_io_disassociate((hammer_io_structure_t
)io
, 1);
299 } else if (io
->modified
) {
301 * Only certain IO types can be released to the kernel.
302 * volume and meta-data IO types must be explicitly flushed
306 case HAMMER_STRUCTURE_DATA_BUFFER
:
307 case HAMMER_STRUCTURE_UNDO_BUFFER
:
308 if (io
->released
== 0) {
316 } else if (io
->released
== 0) {
318 * Clean buffers can be generally released to the kernel.
319 * We leave the bp passively associated with the HAMMER
320 * structure and use bioops to disconnect it later on
321 * if the kernel wants to discard the buffer.
329 * This routine is called with a locked IO when a flush is desired and
330 * no other references to the structure exists other then ours. This
331 * routine is ONLY called when HAMMER believes it is safe to flush a
332 * potentially modified buffer out.
335 hammer_io_flush(struct hammer_io
*io
)
340 * Degenerate case - nothing to flush if nothing is dirty.
342 if (io
->modified
== 0) {
347 KKASSERT(io
->modify_refs
== 0);
350 * Acquire ownership of the bp, particularly before we clear our
353 * We are going to bawrite() this bp. Don't leave a window where
354 * io->released is set, we actually own the bp rather then our
360 /* BUF_KERNPROC(io->bp); */
361 /* io->released = 0; */
362 KKASSERT(io
->released
);
363 KKASSERT(io
->bp
== bp
);
368 * Acquire exclusive access to the bp and then clear the modified
369 * state of the buffer prior to issuing I/O to interlock any
370 * modifications made while the I/O is in progress. This shouldn't
371 * happen anyway but losing data would be worse. The modified bit
372 * will be rechecked after the IO completes.
374 * This is only legal when lock.refs == 1 (otherwise we might clear
375 * the modified bit while there are still users of the cluster
376 * modifying the data).
378 * Do this before potentially blocking so any attempt to modify the
379 * ondisk while we are blocked blocks waiting for us.
381 KKASSERT(io
->mod_list
!= NULL
);
382 if (io
->mod_list
== &io
->hmp
->volu_list
||
383 io
->mod_list
== &io
->hmp
->meta_list
) {
384 --io
->hmp
->locked_dirty_count
;
385 --hammer_count_dirtybufs
;
387 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
392 * Transfer ownership to the kernel and initiate I/O.
395 ++io
->hmp
->io_running_count
;
399 /************************************************************************
401 ************************************************************************
403 * These routines deal with dependancies created when IO buffers get
404 * modified. The caller must call hammer_modify_*() on a referenced
405 * HAMMER structure prior to modifying its on-disk data.
407 * Any intent to modify an IO buffer acquires the related bp and imposes
408 * various write ordering dependancies.
412 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
413 * are locked until the flusher can deal with them, pure data buffers
414 * can be written out.
418 hammer_io_modify(hammer_io_t io
, int count
)
420 struct hammer_mount
*hmp
= io
->hmp
;
423 * Shortcut if nothing to do.
425 KKASSERT(io
->lock
.refs
!= 0 && io
->bp
!= NULL
);
426 io
->modify_refs
+= count
;
427 if (io
->modified
&& io
->released
== 0)
430 hammer_lock_ex(&io
->lock
);
431 if (io
->modified
== 0) {
432 KKASSERT(io
->mod_list
== NULL
);
434 case HAMMER_STRUCTURE_VOLUME
:
435 io
->mod_list
= &hmp
->volu_list
;
436 ++hmp
->locked_dirty_count
;
437 ++hammer_count_dirtybufs
;
439 case HAMMER_STRUCTURE_META_BUFFER
:
440 io
->mod_list
= &hmp
->meta_list
;
441 ++hmp
->locked_dirty_count
;
442 ++hammer_count_dirtybufs
;
444 case HAMMER_STRUCTURE_UNDO_BUFFER
:
445 io
->mod_list
= &hmp
->undo_list
;
447 case HAMMER_STRUCTURE_DATA_BUFFER
:
448 io
->mod_list
= &hmp
->data_list
;
451 TAILQ_INSERT_TAIL(io
->mod_list
, io
, mod_entry
);
456 BUF_KERNPROC(io
->bp
);
458 KKASSERT(io
->modified
!= 0);
460 hammer_unlock(&io
->lock
);
465 hammer_io_modify_done(hammer_io_t io
)
467 KKASSERT(io
->modify_refs
> 0);
472 hammer_modify_volume(hammer_transaction_t trans
, hammer_volume_t volume
,
475 hammer_io_modify(&volume
->io
, 1);
478 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)volume
->ondisk
;
479 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
480 hammer_generate_undo(trans
, &volume
->io
,
481 HAMMER_ENCODE_RAW_VOLUME(volume
->vol_no
, rel_offset
),
487 * Caller intends to modify a buffer's ondisk structure. The related
488 * cluster must be marked open prior to being able to flush the modified
489 * buffer so get that I/O going now.
492 hammer_modify_buffer(hammer_transaction_t trans
, hammer_buffer_t buffer
,
495 hammer_io_modify(&buffer
->io
, 1);
497 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)buffer
->ondisk
;
498 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
499 hammer_generate_undo(trans
, &buffer
->io
,
500 buffer
->zone2_offset
+ rel_offset
,
506 hammer_modify_volume_done(hammer_volume_t volume
)
508 hammer_io_modify_done(&volume
->io
);
512 hammer_modify_buffer_done(hammer_buffer_t buffer
)
514 hammer_io_modify_done(&buffer
->io
);
518 * Mark an entity as not being dirty any more -- this usually occurs when
519 * the governing a-list has freed the entire entity.
524 hammer_io_clear_modify(struct hammer_io
*io
)
531 if ((bp
= io
->bp
) != NULL
) {
534 /* BUF_KERNPROC(io->bp); */
538 if (io
->modified
== 0) {
539 hkprintf("hammer_io_clear_modify: cleared %p\n", io
);
549 /************************************************************************
551 ************************************************************************
556 * Pre-IO initiation kernel callback - cluster build only
559 hammer_io_start(struct buf
*bp
)
564 * Post-IO completion kernel callback
566 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
567 * may also be set if we were marking a cluster header open. Only remove
568 * our dependancy if the modified bit is clear.
571 hammer_io_complete(struct buf
*bp
)
573 union hammer_io_structure
*iou
= (void *)LIST_FIRST(&bp
->b_dep
);
575 KKASSERT(iou
->io
.released
== 1);
577 if (iou
->io
.running
) {
578 if (--iou
->io
.hmp
->io_running_count
== 0)
579 wakeup(&iou
->io
.hmp
->io_running_count
);
580 KKASSERT(iou
->io
.hmp
->io_running_count
>= 0);
585 * If no lock references remain and we can acquire the IO lock and
586 * someone at some point wanted us to flush (B_LOCKED test), then
587 * try to dispose of the IO.
589 if (iou
->io
.waiting
) {
595 * Someone wanted us to flush, try to clean out the buffer.
597 if ((bp
->b_flags
& B_LOCKED
) && iou
->io
.lock
.refs
== 0) {
598 KKASSERT(iou
->io
.modified
== 0);
599 bp
->b_flags
&= ~B_LOCKED
;
600 hammer_io_deallocate(bp
);
601 /* structure may be dead now */
606 * Callback from kernel when it wishes to deallocate a passively
607 * associated structure. This mostly occurs with clean buffers
608 * but it may be possible for a holding structure to be marked dirty
609 * while its buffer is passively associated.
611 * If we cannot disassociate we set B_LOCKED to prevent the buffer
612 * from getting reused.
614 * WARNING: Because this can be called directly by getnewbuf we cannot
615 * recurse into the tree. If a bp cannot be immediately disassociated
616 * our only recourse is to set B_LOCKED.
619 hammer_io_deallocate(struct buf
*bp
)
621 hammer_io_structure_t iou
= (void *)LIST_FIRST(&bp
->b_dep
);
623 KKASSERT((bp
->b_flags
& B_LOCKED
) == 0 && iou
->io
.running
== 0);
624 if (iou
->io
.lock
.refs
> 0 || iou
->io
.modified
) {
626 * It is not legal to disassociate a modified buffer. This
627 * case really shouldn't ever occur.
629 bp
->b_flags
|= B_LOCKED
;
632 * Disassociate the BP. If the io has no refs left we
633 * have to add it to the loose list.
635 hammer_io_disassociate(iou
, 0);
636 if (iou
->io
.bp
== NULL
&&
637 iou
->io
.type
!= HAMMER_STRUCTURE_VOLUME
) {
638 KKASSERT(iou
->io
.mod_list
== NULL
);
639 iou
->io
.mod_list
= &iou
->io
.hmp
->lose_list
;
640 TAILQ_INSERT_TAIL(iou
->io
.mod_list
, &iou
->io
, mod_entry
);
646 hammer_io_fsync(struct vnode
*vp
)
652 * NOTE: will not be called unless we tell the kernel about the
653 * bioops. Unused... we use the mount's VFS_SYNC instead.
656 hammer_io_sync(struct mount
*mp
)
662 hammer_io_movedeps(struct buf
*bp1
, struct buf
*bp2
)
667 * I/O pre-check for reading and writing. HAMMER only uses this for
668 * B_CACHE buffers so checkread just shouldn't happen, but if it does
671 * Writing is a different case. We don't want the kernel to try to write
672 * out a buffer that HAMMER may be modifying passively or which has a
673 * dependancy. In addition, kernel-demanded writes can only proceed for
674 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
675 * buffer types can only be explicitly written by the flusher.
677 * checkwrite will only be called for bdwrite()n buffers. If we return
678 * success the kernel is guaranteed to initiate the buffer write.
681 hammer_io_checkread(struct buf
*bp
)
687 hammer_io_checkwrite(struct buf
*bp
)
689 hammer_io_t io
= (void *)LIST_FIRST(&bp
->b_dep
);
692 * This shouldn't happen under normal operation.
694 if (io
->type
== HAMMER_STRUCTURE_VOLUME
||
695 io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
697 panic("hammer_io_checkwrite: illegal buffer");
699 bp
->b_flags
|= B_LOCKED
;
704 * We can only clear the modified bit if the IO is not currently
705 * undergoing modification. Otherwise we may miss changes.
707 if (io
->modify_refs
== 0 && io
->modified
) {
708 KKASSERT(io
->mod_list
!= NULL
);
709 if (io
->mod_list
== &io
->hmp
->volu_list
||
710 io
->mod_list
== &io
->hmp
->meta_list
) {
711 --io
->hmp
->locked_dirty_count
;
712 --hammer_count_dirtybufs
;
714 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
720 * The kernel is going to start the IO, set io->running.
722 KKASSERT(io
->running
== 0);
724 ++io
->hmp
->io_running_count
;
729 * Return non-zero if the caller should flush the structure associated
730 * with this io sub-structure.
733 hammer_io_checkflush(struct hammer_io
*io
)
735 if (io
->bp
== NULL
|| (io
->bp
->b_flags
& B_LOCKED
)) {
742 * Return non-zero if we wish to delay the kernel's attempt to flush
743 * this buffer to disk.
746 hammer_io_countdeps(struct buf
*bp
, int n
)
751 struct bio_ops hammer_bioops
= {
752 .io_start
= hammer_io_start
,
753 .io_complete
= hammer_io_complete
,
754 .io_deallocate
= hammer_io_deallocate
,
755 .io_fsync
= hammer_io_fsync
,
756 .io_sync
= hammer_io_sync
,
757 .io_movedeps
= hammer_io_movedeps
,
758 .io_countdeps
= hammer_io_countdeps
,
759 .io_checkread
= hammer_io_checkread
,
760 .io_checkwrite
= hammer_io_checkwrite
,