2 * Copyright (c) 2007 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 1.17 2008/01/24 02:14:45 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_deallocate(struct buf *bp);
56 static int hammer_io_checkwrite(struct buf *bp);
59 * Initialize an already-zero'd hammer_io structure
62 hammer_io_init(hammer_io_t io, enum hammer_io_type type)
65 TAILQ_INIT(&io->deplist);
69 * Helper routine to disassociate a buffer cache buffer from an I/O
70 * structure. Called with the io structure exclusively locked.
72 * The io may have 0 or 1 references depending on who called us. The
73 * caller is responsible for dealing with the refs.
75 * This call can only be made when no action is required on the buffer.
76 * HAMMER must own the buffer (released == 0) since we mess around with it.
79 hammer_io_disassociate(hammer_io_structure_t iou, int elseit)
81 struct buf *bp = iou->io.bp;
83 KKASSERT(TAILQ_EMPTY(&iou->io.deplist) && iou->io.modified == 0);
86 bp->b_flags &= ~B_LOCKED;
88 KKASSERT(iou->io.released == 0);
92 KKASSERT(iou->io.released);
95 switch(iou->io.type) {
96 case HAMMER_STRUCTURE_VOLUME:
97 iou->volume.ondisk = NULL;
98 iou->volume.alist.meta = NULL;
100 case HAMMER_STRUCTURE_SUPERCL:
101 iou->supercl.ondisk = NULL;
102 iou->supercl.alist.meta = NULL;
104 case HAMMER_STRUCTURE_CLUSTER:
105 iou->cluster.ondisk = NULL;
106 iou->cluster.alist_master.meta = NULL;
107 iou->cluster.alist_btree.meta = NULL;
108 iou->cluster.alist_record.meta = NULL;
109 iou->cluster.alist_mdata.meta = NULL;
111 case HAMMER_STRUCTURE_BUFFER:
112 iou->buffer.ondisk = NULL;
113 iou->buffer.alist.meta = NULL;
119 * Wait for any physical IO to complete
122 hammer_io_wait(hammer_io_t io)
126 tsleep_interlock(io);
129 tsleep(io, 0, "hmrflw", 0);
130 if (io->running == 0)
132 tsleep_interlock(io);
134 if (io->running == 0)
142 hammer_io_waitdep(hammer_io_t io)
144 while (TAILQ_FIRST(&io->deplist)) {
145 kprintf("waitdep %p\n", io);
146 tsleep(io, 0, "hmrdep", hz);
151 * Load bp for a HAMMER structure. The io is exclusively locked by the
155 hammer_io_read(struct vnode *devvp, struct hammer_io *io)
160 if ((bp = io->bp) == NULL) {
161 error = bread(devvp, io->offset, HAMMER_BUFSIZE, &io->bp);
164 bp->b_ops = &hammer_bioops;
165 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
168 io->modified = 0; /* no new modifications yet */
169 io->released = 0; /* we hold an active lock on bp */
179 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
180 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
181 * I/O so we can call it.
183 * The caller is responsible for calling hammer_modify_*() on the appropriate
187 hammer_io_new(struct vnode *devvp, struct hammer_io *io)
191 if ((bp = io->bp) == NULL) {
192 io->bp = getblk(devvp, io->offset, HAMMER_BUFSIZE, 0, 0);
194 bp->b_ops = &hammer_bioops;
195 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
213 * This routine is called on the last reference to a hammer structure.
214 * The io is usually locked exclusively (but may not be during unmount).
216 * If flush is 1, or B_LOCKED was set indicating that the kernel
217 * wanted to recycle the buffer, and there are no dependancies, this
218 * function will issue an asynchronous write.
220 * If flush is 2 this function waits until all I/O has completed and
221 * disassociates the bp from the IO before returning, unless there
222 * are still other references.
225 hammer_io_release(struct hammer_io *io, int flush)
229 if ((bp = io->bp) == NULL)
234 * If flush is 2 wait for dependancies
236 while (flush == 2 && TAILQ_FIRST(&io->deplist)) {
237 hammer_io_wait(TAILQ_FIRST(&io->deplist));
242 * Try to flush a dirty IO to disk if asked to by the caller
243 * or if the kernel tried to flush the buffer in the past.
245 * The flush will fail if any dependancies are present.
247 if (io->modified && (flush || bp->b_flags & B_LOCKED))
251 * If flush is 2 we wait for the IO to complete.
253 if (flush == 2 && io->running) {
258 * Actively or passively release the buffer. Modified IOs with
259 * dependancies cannot be released.
261 if (flush && io->modified == 0 && io->running == 0) {
262 KKASSERT(TAILQ_EMPTY(&io->deplist));
267 hammer_io_disassociate((hammer_io_structure_t)io, 1);
268 } else if (io->modified) {
269 if (io->released == 0 && TAILQ_EMPTY(&io->deplist)) {
273 } else if (io->released == 0) {
280 * This routine is called with a locked IO when a flush is desired and
281 * no other references to the structure exists other then ours. This
282 * routine is ONLY called when HAMMER believes it is safe to flush a
283 * potentially modified buffer out.
286 hammer_io_flush(struct hammer_io *io)
291 * Can't flush if the IO isn't modified or if it has dependancies.
293 if (io->modified == 0)
295 if (TAILQ_FIRST(&io->deplist))
301 * XXX - umount syncs buffers without referencing them, check for 0
304 KKASSERT(io->lock.refs == 0 || io->lock.refs == 1);
307 * Reset modified to 0 here and re-check it after the IO completes.
308 * This is only legal when lock.refs == 1 (otherwise we might clear
309 * the modified bit while there are still users of the cluster
310 * modifying the data).
312 * NOTE: We have no dependancies so we don't have to worry about
313 * cluster-open's here.
315 * Do this before potentially blocking so any attempt to modify the
316 * ondisk while we are blocked blocks waiting for us.
322 * If we are trying to flush a buffer we have to wait until the
323 * cluster header for the mark-OPEN has completed its I/O.
325 if (io->type == HAMMER_STRUCTURE_BUFFER) {
326 hammer_io_structure_t iou = (void *)io;
327 hammer_cluster_t cluster = iou->buffer.cluster;
329 if (cluster->io.running) {
330 kprintf("WAIT CLUSTER OPEN %d\n", cluster->clu_no);
331 hammer_io_wait(&cluster->io);
332 kprintf("WAIT CLUSTER OPEN OK\n");
335 if (io->type == HAMMER_STRUCTURE_CLUSTER) {
337 * Mark the cluster closed if we can
339 hammer_io_checkwrite(io->bp);
343 /* BUF_KERNPROC(io->bp); */
351 /************************************************************************
353 ************************************************************************
355 * These routines deal with dependancies created when IO buffers get
356 * modified. The caller must call hammer_modify_*() on a referenced
357 * HAMMER structure prior to modifying its on-disk data.
359 * Any intent to modify an IO buffer acquires the related bp and imposes
360 * various write ordering dependancies.
364 * Ensure that the bp is acquired and return non-zero on a 0->1 transition
365 * of the modified bit.
369 hammer_io_modify(hammer_io_t io, struct hammer_io_list *list)
373 KKASSERT(io->lock.refs != 0 && io->bp != NULL);
374 if (io->modified == 0) {
375 hammer_lock_ex(&io->lock);
376 if (io->modified == 0) {
379 BUF_KERNPROC(io->bp);
383 io->entry_list = list;
385 TAILQ_INSERT_TAIL(list, io, entry);
388 hammer_unlock(&io->lock);
389 } else if (io->released) {
391 * Make sure no IO is occuring while we modify the contents
392 * of the buffer. XXX should be able to avoid doing this.
394 hammer_lock_ex(&io->lock);
397 BUF_KERNPROC(io->bp);
401 * The modified bit should still be set because we have
402 * a ref on the structure, so the kernel's checkwrite
403 * should not have cleared it.
405 KKASSERT(io->modified != 0);
406 hammer_unlock(&io->lock);
412 hammer_modify_volume(hammer_volume_t volume)
414 hammer_io_modify(&volume->io, NULL);
418 hammer_modify_supercl(hammer_supercl_t supercl)
420 hammer_io_modify(&supercl->io, &supercl->volume->io.deplist);
424 * Caller intends to modify a cluster's ondisk structure.
427 hammer_modify_cluster(hammer_cluster_t cluster)
429 hammer_io_modify(&cluster->io, &cluster->volume->io.deplist);
433 * Caller intends to modify a buffer's ondisk structure. The related
434 * cluster must be marked open prior to being able to flush the modified
435 * buffer so get that I/O going now.
438 hammer_modify_buffer(hammer_buffer_t buffer)
440 hammer_cluster_t cluster = buffer->cluster;
442 if (hammer_io_modify(&buffer->io, &cluster->io.deplist)) {
443 hammer_modify_cluster(cluster);
444 if ((cluster->ondisk->clu_flags & HAMMER_CLUF_OPEN) == 0) {
445 hammer_lock_ex(&cluster->io.lock);
446 if ((cluster->ondisk->clu_flags & HAMMER_CLUF_OPEN) == 0) {
447 KKASSERT(cluster->io.released == 0);
448 cluster->ondisk->clu_flags |= HAMMER_CLUF_OPEN;
449 cluster->io.released = 1;
450 cluster->io.running = 1;
451 bawrite(cluster->io.bp);
452 kprintf("OPEN CLUSTER %d:%d\n",
453 cluster->volume->vol_no,
456 hammer_unlock(&cluster->io.lock);
462 * Mark an entity as not being dirty any more -- this usually occurs when
463 * the governing a-list has freed the entire entity.
468 hammer_io_clear_modify(struct hammer_io *io)
474 if ((bp = io->bp) != NULL) {
477 /* BUF_KERNPROC(io->bp); */
481 if (io->modified == 0) {
482 kprintf("hammer_io_clear_modify: cleared %p\n", io);
492 /************************************************************************
494 ************************************************************************
499 * Pre-IO initiation kernel callback - cluster build only
502 hammer_io_start(struct buf *bp)
507 * Post-IO completion kernel callback
509 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
510 * may also be set if we were marking a cluster header open. Only remove
511 * our dependancy if the modified bit is clear.
514 hammer_io_complete(struct buf *bp)
516 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
518 KKASSERT(iou->io.released == 1);
521 * If this was a write and the modified bit is still clear we can
522 * remove ourselves from the dependancy list.
524 * If no lock references remain and we can acquire the IO lock and
525 * someone at some point wanted us to flush (B_LOCKED test), then
526 * try to dispose of the IO.
528 if (iou->io.modified == 0 && iou->io.entry_list) {
529 TAILQ_REMOVE(iou->io.entry_list, &iou->io, entry);
530 iou->io.entry_list = NULL;
533 if (iou->io.waiting) {
539 * Someone wanted us to flush, try to clean out the buffer.
541 if ((bp->b_flags & B_LOCKED) && iou->io.lock.refs == 0) {
542 KKASSERT(iou->io.modified == 0);
543 hammer_io_deallocate(bp);
544 /* structure may be dead now */
549 * Callback from kernel when it wishes to deallocate a passively
550 * associated structure. This case can only occur with read-only
553 * If we cannot disassociate we set B_LOCKED to prevent the buffer
554 * from getting reused.
557 hammer_io_deallocate(struct buf *bp)
559 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep);
561 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0);
562 if (iou->io.modified) {
563 bp->b_flags |= B_LOCKED;
566 hammer_ref(&iou->io.lock);
567 if (iou->io.lock.refs > 1 || iou->io.modified) {
568 hammer_unref(&iou->io.lock);
569 bp->b_flags |= B_LOCKED;
571 hammer_io_disassociate(iou, 0);
573 switch(iou->io.type) {
574 case HAMMER_STRUCTURE_VOLUME:
575 hammer_rel_volume(&iou->volume, 1);
577 case HAMMER_STRUCTURE_SUPERCL:
578 hammer_rel_supercl(&iou->supercl, 1);
580 case HAMMER_STRUCTURE_CLUSTER:
581 hammer_rel_cluster(&iou->cluster, 1);
583 case HAMMER_STRUCTURE_BUFFER:
584 hammer_rel_buffer(&iou->buffer, 1);
591 hammer_io_fsync(struct vnode *vp)
597 * NOTE: will not be called unless we tell the kernel about the
598 * bioops. Unused... we use the mount's VFS_SYNC instead.
601 hammer_io_sync(struct mount *mp)
607 hammer_io_movedeps(struct buf *bp1, struct buf *bp2)
612 * I/O pre-check for reading and writing. HAMMER only uses this for
613 * B_CACHE buffers so checkread just shouldn't happen, but if it does
616 * Writing is a different case. We don't want the kernel to try to write
617 * out a buffer that HAMMER may be modifying passively or which has a
620 * This code enforces the following write ordering: buffers, then cluster
621 * headers, then volume headers.
624 hammer_io_checkread(struct buf *bp)
630 hammer_io_checkwrite(struct buf *bp)
632 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
635 * A modified cluster with no dependancies can be closed. It is
636 * possible for the cluster to have been modified by the recovery
637 * code without validation. Only clear the open flag if the
638 * cluster is validated.
640 if (iou->io.type == HAMMER_STRUCTURE_CLUSTER && iou->io.modified) {
641 hammer_cluster_t cluster = &iou->cluster;
643 if (TAILQ_EMPTY(&cluster->io.deplist)) {
644 kprintf("CLOSE CLUSTER %d:%d ",
645 cluster->volume->vol_no,
647 if (cluster->ondisk->clu_flags & HAMMER_CLUF_OPEN) {
648 if (cluster->io.validated) {
649 cluster->ondisk->clu_flags &=
651 kprintf("(closed)\n");
653 kprintf("(leave-open)\n");
656 kprintf("(header-only)\n");
662 * We are called from the kernel on delayed-write buffers, and
663 * called from hammer_io_flush() on flush requests. There should
664 * be no dependancies in either case.
666 * In the case of delayed-writes, the introduction of a dependancy
667 * will block until the bp can be reacquired, and the bp is then
668 * simply not released until the dependancy can be satisfied.
670 * We can only clear the modified bit when entered from the kernel
671 * if io.lock.refs == 0.
673 KKASSERT(TAILQ_EMPTY(&iou->io.deplist));
674 if (iou->io.lock.refs == 0) {
675 iou->io.modified = 0;
681 * Return non-zero if the caller should flush the structure associated
682 * with this io sub-structure.
685 hammer_io_checkflush(struct hammer_io *io)
687 if (io->bp == NULL || (io->bp->b_flags & B_LOCKED)) {
694 * Return non-zero if we wish to delay the kernel's attempt to flush
695 * this buffer to disk.
698 hammer_io_countdeps(struct buf *bp, int n)
703 struct bio_ops hammer_bioops = {
704 .io_start = hammer_io_start,
705 .io_complete = hammer_io_complete,
706 .io_deallocate = hammer_io_deallocate,
707 .io_fsync = hammer_io_fsync,
708 .io_sync = hammer_io_sync,
709 .io_movedeps = hammer_io_movedeps,
710 .io_countdeps = hammer_io_countdeps,
711 .io_checkread = hammer_io_checkread,
712 .io_checkwrite = hammer_io_checkwrite,