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 1.22 2008/03/19 20:18:17 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;
99 case HAMMER_STRUCTURE_BUFFER:
100 iou->buffer.ondisk = NULL;
106 * Wait for any physical IO to complete
109 hammer_io_wait(hammer_io_t io)
113 tsleep_interlock(io);
116 tsleep(io, 0, "hmrflw", 0);
117 if (io->running == 0)
119 tsleep_interlock(io);
121 if (io->running == 0)
129 hammer_io_waitdep(hammer_io_t io)
131 while (TAILQ_FIRST(&io->deplist)) {
132 kprintf("waitdep %p\n", io);
133 tsleep(io, 0, "hmrdep", hz);
138 * Load bp for a HAMMER structure. The io is exclusively locked by the
142 hammer_io_read(struct vnode *devvp, struct hammer_io *io)
147 if ((bp = io->bp) == NULL) {
148 error = bread(devvp, io->offset, HAMMER_BUFSIZE, &io->bp);
151 bp->b_ops = &hammer_bioops;
152 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
155 io->modified = 0; /* no new modifications yet */
156 io->released = 0; /* we hold an active lock on bp */
166 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
167 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
168 * I/O so we can call it.
170 * The caller is responsible for calling hammer_modify_*() on the appropriate
174 hammer_io_new(struct vnode *devvp, struct hammer_io *io)
178 if ((bp = io->bp) == NULL) {
179 io->bp = getblk(devvp, io->offset, HAMMER_BUFSIZE, 0, 0);
181 bp->b_ops = &hammer_bioops;
182 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
200 * This routine is called on the last reference to a hammer structure.
201 * The io is usually locked exclusively (but may not be during unmount).
203 * If flush is 1, or B_LOCKED was set indicating that the kernel
204 * wanted to recycle the buffer, and there are no dependancies, this
205 * function will issue an asynchronous write.
207 * If flush is 2 this function waits until all I/O has completed and
208 * disassociates the bp from the IO before returning, unless there
209 * are still other references.
212 hammer_io_release(struct hammer_io *io, int flush)
216 if ((bp = io->bp) == NULL)
221 * If flush is 2 wait for dependancies
223 while (flush == 2 && TAILQ_FIRST(&io->deplist)) {
224 hammer_io_wait(TAILQ_FIRST(&io->deplist));
229 * Try to flush a dirty IO to disk if asked to by the caller
230 * or if the kernel tried to flush the buffer in the past.
232 * The flush will fail if any dependancies are present.
234 if (io->modified && (flush || bp->b_flags & B_LOCKED))
238 * If flush is 2 we wait for the IO to complete.
240 if (flush == 2 && io->running) {
245 * Actively or passively release the buffer. Modified IOs with
246 * dependancies cannot be released.
248 if (flush && io->modified == 0 && io->running == 0) {
249 KKASSERT(TAILQ_EMPTY(&io->deplist));
255 hammer_io_disassociate((hammer_io_structure_t)io, 1);
256 } else if (io->modified) {
257 if (io->released == 0 && TAILQ_EMPTY(&io->deplist)) {
261 } else if (io->released == 0) {
268 * This routine is called with a locked IO when a flush is desired and
269 * no other references to the structure exists other then ours. This
270 * routine is ONLY called when HAMMER believes it is safe to flush a
271 * potentially modified buffer out.
274 hammer_io_flush(struct hammer_io *io)
279 * Can't flush if the IO isn't modified or if it has dependancies.
281 if (io->modified == 0)
283 if (TAILQ_FIRST(&io->deplist))
289 * XXX - umount syncs buffers without referencing them, check for 0
292 KKASSERT(io->lock.refs == 0 || io->lock.refs == 1);
295 * Reset modified to 0 here and re-check it after the IO completes.
296 * This is only legal when lock.refs == 1 (otherwise we might clear
297 * the modified bit while there are still users of the cluster
298 * modifying the data).
300 * NOTE: We have no dependancies so we don't have to worry about
301 * cluster-open's here.
303 * Do this before potentially blocking so any attempt to modify the
304 * ondisk while we are blocked blocks waiting for us.
306 io->modified = 0; /* force interlock */
311 /* BUF_KERNPROC(io->bp); */
319 /************************************************************************
321 ************************************************************************
323 * These routines deal with dependancies created when IO buffers get
324 * modified. The caller must call hammer_modify_*() on a referenced
325 * HAMMER structure prior to modifying its on-disk data.
327 * Any intent to modify an IO buffer acquires the related bp and imposes
328 * various write ordering dependancies.
332 * Mark a HAMMER structure as undergoing modification. Return 1 when applying
333 * a non-NULL ordering dependancy for the first time, 0 otherwise.
335 * list can be NULL, indicating that a structural modification is being made
336 * without creating an ordering dependancy.
340 hammer_io_modify(hammer_io_t io, struct hammer_io_list *list)
345 * Shortcut if nothing to do.
347 KKASSERT(io->lock.refs != 0 && io->bp != NULL);
348 if (io->modified && io->released == 0 &&
349 (io->entry_list || list == NULL)) {
353 hammer_lock_ex(&io->lock);
357 BUF_KERNPROC(io->bp);
359 KKASSERT(io->modified != 0);
361 if (io->entry_list == NULL) {
362 io->entry_list = list;
364 TAILQ_INSERT_TAIL(list, io, entry);
370 /* only one dependancy is allowed */
371 KKASSERT(list == NULL || io->entry_list == list);
374 hammer_unlock(&io->lock);
379 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
382 hammer_io_modify(&volume->io, NULL);
385 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk;
386 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
387 hammer_generate_undo(trans,
388 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset),
394 * Caller intends to modify a buffer's ondisk structure. The related
395 * cluster must be marked open prior to being able to flush the modified
396 * buffer so get that I/O going now.
399 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
402 hammer_io_modify(&buffer->io, NULL);
404 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk;
405 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
406 hammer_generate_undo(trans,
407 buffer->zone2_offset + rel_offset,
413 * Mark an entity as not being dirty any more -- this usually occurs when
414 * the governing a-list has freed the entire entity.
419 hammer_io_clear_modify(struct hammer_io *io)
425 if ((bp = io->bp) != NULL) {
428 /* BUF_KERNPROC(io->bp); */
432 if (io->modified == 0) {
433 kprintf("hammer_io_clear_modify: cleared %p\n", io);
443 /************************************************************************
445 ************************************************************************
450 * Pre-IO initiation kernel callback - cluster build only
453 hammer_io_start(struct buf *bp)
458 * Post-IO completion kernel callback
460 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
461 * may also be set if we were marking a cluster header open. Only remove
462 * our dependancy if the modified bit is clear.
465 hammer_io_complete(struct buf *bp)
467 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
469 KKASSERT(iou->io.released == 1);
472 * If this was a write and the modified bit is still clear we can
473 * remove ourselves from the dependancy list.
475 * If no lock references remain and we can acquire the IO lock and
476 * someone at some point wanted us to flush (B_LOCKED test), then
477 * try to dispose of the IO.
479 if (iou->io.modified == 0 && iou->io.entry_list) {
480 TAILQ_REMOVE(iou->io.entry_list, &iou->io, entry);
481 iou->io.entry_list = NULL;
484 if (iou->io.waiting) {
490 * Someone wanted us to flush, try to clean out the buffer.
492 if ((bp->b_flags & B_LOCKED) && iou->io.lock.refs == 0) {
493 KKASSERT(iou->io.modified == 0);
494 bp->b_flags &= ~B_LOCKED;
495 hammer_io_deallocate(bp);
496 /* structure may be dead now */
501 * Callback from kernel when it wishes to deallocate a passively
502 * associated structure. This case can only occur with read-only
505 * If we cannot disassociate we set B_LOCKED to prevent the buffer
506 * from getting reused.
508 * WARNING: Because this can be called directly by getnewbuf we cannot
509 * recurse into the tree. If a bp cannot be immediately disassociated
510 * our only recourse is to set B_LOCKED.
512 * WARNING: If the HAMMER structure is passively cached we have to
516 hammer_io_deallocate(struct buf *bp)
518 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep);
520 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0);
521 if (iou->io.lock.refs > 0 || iou->io.modified) {
522 bp->b_flags |= B_LOCKED;
524 /* XXX interlock against ref or another disassociate */
525 /* XXX this can leave HAMMER structures lying around */
526 hammer_io_disassociate(iou, 0);
528 switch(iou->io.type) {
529 case HAMMER_STRUCTURE_VOLUME:
530 hammer_rel_volume(&iou->volume, 1);
532 case HAMMER_STRUCTURE_BUFFER:
533 hammer_rel_buffer(&iou->buffer, 1);
541 hammer_io_fsync(struct vnode *vp)
547 * NOTE: will not be called unless we tell the kernel about the
548 * bioops. Unused... we use the mount's VFS_SYNC instead.
551 hammer_io_sync(struct mount *mp)
557 hammer_io_movedeps(struct buf *bp1, struct buf *bp2)
562 * I/O pre-check for reading and writing. HAMMER only uses this for
563 * B_CACHE buffers so checkread just shouldn't happen, but if it does
566 * Writing is a different case. We don't want the kernel to try to write
567 * out a buffer that HAMMER may be modifying passively or which has a
570 * This code enforces the following write ordering: buffers, then cluster
571 * headers, then volume headers.
574 hammer_io_checkread(struct buf *bp)
580 hammer_io_checkwrite(struct buf *bp)
582 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
584 KKASSERT(TAILQ_EMPTY(&iou->io.deplist));
587 * We are called from the kernel on delayed-write buffers, and
588 * called from hammer_io_flush() on flush requests. There should
589 * be no dependancies in either case.
591 * In the case of delayed-writes, the introduction of a dependancy
592 * will block until the bp can be reacquired, and the bp is then
593 * simply not released until the dependancy can be satisfied.
595 * We can only clear the modified bit when entered from the kernel
596 * if io.lock.refs == 0.
598 if (iou->io.lock.refs == 0) {
599 iou->io.modified = 0;
605 * Return non-zero if the caller should flush the structure associated
606 * with this io sub-structure.
609 hammer_io_checkflush(struct hammer_io *io)
611 if (io->bp == NULL || (io->bp->b_flags & B_LOCKED)) {
618 * Return non-zero if we wish to delay the kernel's attempt to flush
619 * this buffer to disk.
622 hammer_io_countdeps(struct buf *bp, int n)
627 struct bio_ops hammer_bioops = {
628 .io_start = hammer_io_start,
629 .io_complete = hammer_io_complete,
630 .io_deallocate = hammer_io_deallocate,
631 .io_fsync = hammer_io_fsync,
632 .io_sync = hammer_io_sync,
633 .io_movedeps = hammer_io_movedeps,
634 .io_countdeps = hammer_io_countdeps,
635 .io_checkread = hammer_io_checkread,
636 .io_checkwrite = hammer_io_checkwrite,