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 1.32 2008/05/18 01:48:50 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)
170 #define HAMMER_MAXRA 4
173 * Load bp for a HAMMER structure. The io must be exclusively locked by
176 * Generally speaking HAMMER assumes either an optimized layout or that
177 * typical access patterns will be close to the original layout when the
178 * information was written. For this reason we try to cluster all reads.
181 hammer_io_read(struct vnode *devvp, struct hammer_io *io, hammer_off_t limit)
186 if ((bp = io->bp) == NULL) {
188 error = cluster_read(devvp, limit, io->offset,
189 HAMMER_BUFSIZE, MAXBSIZE, 16, &io->bp);
191 error = bread(devvp, io->offset, HAMMER_BUFSIZE, &io->bp);
196 bp->b_ops = &hammer_bioops;
197 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
200 KKASSERT(io->modified == 0);
201 KKASSERT(io->running == 0);
202 KKASSERT(io->waiting == 0);
203 io->released = 0; /* we hold an active lock on bp */
211 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
212 * Must be called with the IO exclusively locked.
214 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
215 * I/O by forcing the buffer to not be in a released state before calling
218 * This function will also mark the IO as modified but it will not
219 * increment the modify_refs count.
222 hammer_io_new(struct vnode *devvp, struct hammer_io *io)
226 if ((bp = io->bp) == NULL) {
227 io->bp = getblk(devvp, io->offset, HAMMER_BUFSIZE, 0, 0);
229 bp->b_ops = &hammer_bioops;
230 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
232 KKASSERT(io->running == 0);
242 hammer_io_modify(io, 0);
248 * This routine is called on the last reference to a hammer structure.
249 * The io is usually locked exclusively (but may not be during unmount).
251 * This routine is responsible for the disposition of the buffer cache
252 * buffer backing the IO. Only pure-data and undo buffers can be handed
253 * back to the kernel. Volume and meta-data buffers must be retained
254 * by HAMMER until explicitly flushed by the backend.
257 hammer_io_release(struct hammer_io *io, int flush)
261 if ((bp = io->bp) == NULL)
265 * Try to flush a dirty IO to disk if asked to by the
266 * caller or if the kernel tried to flush the buffer in the past.
268 * Kernel-initiated flushes are only allowed for pure-data buffers.
269 * meta-data and volume buffers can only be flushed explicitly
275 } else if (bp->b_flags & B_LOCKED) {
277 case HAMMER_STRUCTURE_DATA_BUFFER:
278 case HAMMER_STRUCTURE_UNDO_BUFFER:
284 } /* else no explicit request to flush the buffer */
288 * Wait for the IO to complete if asked to.
290 if (io->waitdep && io->running) {
295 * Return control of the buffer to the kernel (with the provisio
296 * that our bioops can override kernel decisions with regards to
299 if (flush && io->modified == 0 && io->running == 0) {
301 * Always disassociate the bp if an explicit flush
302 * was requested and the IO completed with no error
303 * (so unmount can really clean up the structure).
310 hammer_io_disassociate((hammer_io_structure_t)io, 1);
311 } else if (io->modified) {
313 * Only certain IO types can be released to the kernel.
314 * volume and meta-data IO types must be explicitly flushed
318 case HAMMER_STRUCTURE_DATA_BUFFER:
319 case HAMMER_STRUCTURE_UNDO_BUFFER:
320 if (io->released == 0) {
328 } else if (io->released == 0) {
330 * Clean buffers can be generally released to the kernel.
331 * We leave the bp passively associated with the HAMMER
332 * structure and use bioops to disconnect it later on
333 * if the kernel wants to discard the buffer.
341 * This routine is called with a locked IO when a flush is desired and
342 * no other references to the structure exists other then ours. This
343 * routine is ONLY called when HAMMER believes it is safe to flush a
344 * potentially modified buffer out.
347 hammer_io_flush(struct hammer_io *io)
352 * Degenerate case - nothing to flush if nothing is dirty.
354 if (io->modified == 0) {
359 KKASSERT(io->modify_refs == 0);
362 * Acquire ownership of the bp, particularly before we clear our
365 * We are going to bawrite() this bp. Don't leave a window where
366 * io->released is set, we actually own the bp rather then our
372 /* BUF_KERNPROC(io->bp); */
373 /* io->released = 0; */
374 KKASSERT(io->released);
375 KKASSERT(io->bp == bp);
380 * Acquire exclusive access to the bp and then clear the modified
381 * state of the buffer prior to issuing I/O to interlock any
382 * modifications made while the I/O is in progress. This shouldn't
383 * happen anyway but losing data would be worse. The modified bit
384 * will be rechecked after the IO completes.
386 * This is only legal when lock.refs == 1 (otherwise we might clear
387 * the modified bit while there are still users of the cluster
388 * modifying the data).
390 * Do this before potentially blocking so any attempt to modify the
391 * ondisk while we are blocked blocks waiting for us.
393 KKASSERT(io->mod_list != NULL);
394 if (io->mod_list == &io->hmp->volu_list ||
395 io->mod_list == &io->hmp->meta_list) {
396 --io->hmp->locked_dirty_count;
397 --hammer_count_dirtybufs;
399 TAILQ_REMOVE(io->mod_list, io, mod_entry);
404 * Transfer ownership to the kernel and initiate I/O.
407 ++io->hmp->io_running_count;
411 /************************************************************************
413 ************************************************************************
415 * These routines deal with dependancies created when IO buffers get
416 * modified. The caller must call hammer_modify_*() on a referenced
417 * HAMMER structure prior to modifying its on-disk data.
419 * Any intent to modify an IO buffer acquires the related bp and imposes
420 * various write ordering dependancies.
424 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
425 * are locked until the flusher can deal with them, pure data buffers
426 * can be written out.
430 hammer_io_modify(hammer_io_t io, int count)
432 struct hammer_mount *hmp = io->hmp;
435 * Shortcut if nothing to do.
437 KKASSERT(io->lock.refs != 0 && io->bp != NULL);
438 io->modify_refs += count;
439 if (io->modified && io->released == 0)
442 hammer_lock_ex(&io->lock);
443 if (io->modified == 0) {
444 KKASSERT(io->mod_list == NULL);
446 case HAMMER_STRUCTURE_VOLUME:
447 io->mod_list = &hmp->volu_list;
448 ++hmp->locked_dirty_count;
449 ++hammer_count_dirtybufs;
451 case HAMMER_STRUCTURE_META_BUFFER:
452 io->mod_list = &hmp->meta_list;
453 ++hmp->locked_dirty_count;
454 ++hammer_count_dirtybufs;
456 case HAMMER_STRUCTURE_UNDO_BUFFER:
457 io->mod_list = &hmp->undo_list;
459 case HAMMER_STRUCTURE_DATA_BUFFER:
460 io->mod_list = &hmp->data_list;
463 TAILQ_INSERT_TAIL(io->mod_list, io, mod_entry);
468 BUF_KERNPROC(io->bp);
470 KKASSERT(io->modified != 0);
472 hammer_unlock(&io->lock);
477 hammer_io_modify_done(hammer_io_t io)
479 KKASSERT(io->modify_refs > 0);
484 * Caller intends to modify a volume's ondisk structure.
486 * This is only allowed if we are the flusher or we have a ref on the
490 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
493 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
495 hammer_io_modify(&volume->io, 1);
497 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk;
498 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
499 hammer_generate_undo(trans, &volume->io,
500 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset),
506 * Caller intends to modify a buffer's ondisk structure.
508 * This is only allowed if we are the flusher or we have a ref on the
512 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
515 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
517 hammer_io_modify(&buffer->io, 1);
519 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk;
520 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
521 hammer_generate_undo(trans, &buffer->io,
522 buffer->zone2_offset + rel_offset,
528 hammer_modify_volume_done(hammer_volume_t volume)
530 hammer_io_modify_done(&volume->io);
534 hammer_modify_buffer_done(hammer_buffer_t buffer)
536 hammer_io_modify_done(&buffer->io);
540 * Mark an entity as not being dirty any more -- this usually occurs when
541 * the governing a-list has freed the entire entity.
546 hammer_io_clear_modify(struct hammer_io *io)
553 if ((bp = io->bp) != NULL) {
556 /* BUF_KERNPROC(io->bp); */
560 if (io->modified == 0) {
561 hkprintf("hammer_io_clear_modify: cleared %p\n", io);
571 /************************************************************************
573 ************************************************************************
578 * Pre-IO initiation kernel callback - cluster build only
581 hammer_io_start(struct buf *bp)
586 * Post-IO completion kernel callback
588 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
589 * may also be set if we were marking a cluster header open. Only remove
590 * our dependancy if the modified bit is clear.
593 hammer_io_complete(struct buf *bp)
595 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
597 KKASSERT(iou->io.released == 1);
599 if (iou->io.running) {
600 if (--iou->io.hmp->io_running_count == 0)
601 wakeup(&iou->io.hmp->io_running_count);
602 KKASSERT(iou->io.hmp->io_running_count >= 0);
607 * If no lock references remain and we can acquire the IO lock and
608 * someone at some point wanted us to flush (B_LOCKED test), then
609 * try to dispose of the IO.
611 if (iou->io.waiting) {
617 * Someone wanted us to flush, try to clean out the buffer.
619 if ((bp->b_flags & B_LOCKED) && iou->io.lock.refs == 0) {
620 KKASSERT(iou->io.modified == 0);
621 bp->b_flags &= ~B_LOCKED;
622 hammer_io_deallocate(bp);
623 /* structure may be dead now */
628 * Callback from kernel when it wishes to deallocate a passively
629 * associated structure. This mostly occurs with clean buffers
630 * but it may be possible for a holding structure to be marked dirty
631 * while its buffer is passively associated.
633 * If we cannot disassociate we set B_LOCKED to prevent the buffer
634 * from getting reused.
636 * WARNING: Because this can be called directly by getnewbuf we cannot
637 * recurse into the tree. If a bp cannot be immediately disassociated
638 * our only recourse is to set B_LOCKED.
641 hammer_io_deallocate(struct buf *bp)
643 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep);
645 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0);
646 if (iou->io.lock.refs > 0 || iou->io.modified) {
648 * It is not legal to disassociate a modified buffer. This
649 * case really shouldn't ever occur.
651 bp->b_flags |= B_LOCKED;
654 * Disassociate the BP. If the io has no refs left we
655 * have to add it to the loose list.
657 hammer_io_disassociate(iou, 0);
658 if (iou->io.bp == NULL &&
659 iou->io.type != HAMMER_STRUCTURE_VOLUME) {
660 KKASSERT(iou->io.mod_list == NULL);
661 iou->io.mod_list = &iou->io.hmp->lose_list;
662 TAILQ_INSERT_TAIL(iou->io.mod_list, &iou->io, mod_entry);
668 hammer_io_fsync(struct vnode *vp)
674 * NOTE: will not be called unless we tell the kernel about the
675 * bioops. Unused... we use the mount's VFS_SYNC instead.
678 hammer_io_sync(struct mount *mp)
684 hammer_io_movedeps(struct buf *bp1, struct buf *bp2)
689 * I/O pre-check for reading and writing. HAMMER only uses this for
690 * B_CACHE buffers so checkread just shouldn't happen, but if it does
693 * Writing is a different case. We don't want the kernel to try to write
694 * out a buffer that HAMMER may be modifying passively or which has a
695 * dependancy. In addition, kernel-demanded writes can only proceed for
696 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
697 * buffer types can only be explicitly written by the flusher.
699 * checkwrite will only be called for bdwrite()n buffers. If we return
700 * success the kernel is guaranteed to initiate the buffer write.
703 hammer_io_checkread(struct buf *bp)
709 hammer_io_checkwrite(struct buf *bp)
711 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep);
714 * This shouldn't happen under normal operation.
716 if (io->type == HAMMER_STRUCTURE_VOLUME ||
717 io->type == HAMMER_STRUCTURE_META_BUFFER) {
719 panic("hammer_io_checkwrite: illegal buffer");
721 bp->b_flags |= B_LOCKED;
726 * We can only clear the modified bit if the IO is not currently
727 * undergoing modification. Otherwise we may miss changes.
729 if (io->modify_refs == 0 && io->modified) {
730 KKASSERT(io->mod_list != NULL);
731 if (io->mod_list == &io->hmp->volu_list ||
732 io->mod_list == &io->hmp->meta_list) {
733 --io->hmp->locked_dirty_count;
734 --hammer_count_dirtybufs;
736 TAILQ_REMOVE(io->mod_list, io, mod_entry);
742 * The kernel is going to start the IO, set io->running.
744 KKASSERT(io->running == 0);
746 ++io->hmp->io_running_count;
751 * Return non-zero if the caller should flush the structure associated
752 * with this io sub-structure.
755 hammer_io_checkflush(struct hammer_io *io)
757 if (io->bp == NULL || (io->bp->b_flags & B_LOCKED)) {
764 * Return non-zero if we wish to delay the kernel's attempt to flush
765 * this buffer to disk.
768 hammer_io_countdeps(struct buf *bp, int n)
773 struct bio_ops hammer_bioops = {
774 .io_start = hammer_io_start,
775 .io_complete = hammer_io_complete,
776 .io_deallocate = hammer_io_deallocate,
777 .io_fsync = hammer_io_fsync,
778 .io_sync = hammer_io_sync,
779 .io_movedeps = hammer_io_movedeps,
780 .io_countdeps = hammer_io_countdeps,
781 .io_checkread = hammer_io_checkread,
782 .io_checkwrite = hammer_io_checkwrite,