HAMMER: MFC to 2.0
[dragonfly.git] / sys / vfs / hammer / hammer_io.c
blob6ca02b94405f4ff0e54dd46636b69aa67b79e8da
1 /*
2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
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
16 * distribution.
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
32 * SUCH DAMAGE.
34 * $DragonFly: src/sys/vfs/hammer/hammer_io.c,v 1.49.2.3 2008/08/02 21:24:28 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
42 * HAMMER structures.
44 * If the kernel tries to destroy a passively associated buf which we cannot
45 * yet let go we set B_LOCKED in the buffer and then actively released it
46 * later when we can.
49 #include "hammer.h"
50 #include <sys/fcntl.h>
51 #include <sys/nlookup.h>
52 #include <sys/buf.h>
53 #include <sys/buf2.h>
55 static void hammer_io_modify(hammer_io_t io, int count);
56 static void hammer_io_deallocate(struct buf *bp);
57 #if 0
58 static void hammer_io_direct_read_complete(struct bio *nbio);
59 #endif
60 static void hammer_io_direct_write_complete(struct bio *nbio);
61 static int hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data);
62 static void hammer_io_set_modlist(struct hammer_io *io);
65 * Initialize a new, already-zero'd hammer_io structure, or reinitialize
66 * an existing hammer_io structure which may have switched to another type.
68 void
69 hammer_io_init(hammer_io_t io, hammer_mount_t hmp, enum hammer_io_type type)
71 io->hmp = hmp;
72 io->type = type;
76 * Helper routine to disassociate a buffer cache buffer from an I/O
77 * structure. The buffer is unlocked and marked appropriate for reclamation.
79 * The io may have 0 or 1 references depending on who called us. The
80 * caller is responsible for dealing with the refs.
82 * This call can only be made when no action is required on the buffer.
84 * The caller must own the buffer and the IO must indicate that the
85 * structure no longer owns it (io.released != 0).
87 static void
88 hammer_io_disassociate(hammer_io_structure_t iou)
90 struct buf *bp = iou->io.bp;
92 KKASSERT(iou->io.released);
93 KKASSERT(iou->io.modified == 0);
94 KKASSERT(LIST_FIRST(&bp->b_dep) == (void *)iou);
95 buf_dep_init(bp);
96 iou->io.bp = NULL;
99 * If the buffer was locked someone wanted to get rid of it.
101 if (bp->b_flags & B_LOCKED) {
102 --hammer_count_io_locked;
103 bp->b_flags &= ~B_LOCKED;
105 if (iou->io.reclaim) {
106 bp->b_flags |= B_NOCACHE|B_RELBUF;
107 iou->io.reclaim = 0;
110 switch(iou->io.type) {
111 case HAMMER_STRUCTURE_VOLUME:
112 iou->volume.ondisk = NULL;
113 break;
114 case HAMMER_STRUCTURE_DATA_BUFFER:
115 case HAMMER_STRUCTURE_META_BUFFER:
116 case HAMMER_STRUCTURE_UNDO_BUFFER:
117 iou->buffer.ondisk = NULL;
118 break;
123 * Wait for any physical IO to complete
125 void
126 hammer_io_wait(hammer_io_t io)
128 if (io->running) {
129 crit_enter();
130 tsleep_interlock(io);
131 io->waiting = 1;
132 for (;;) {
133 tsleep(io, 0, "hmrflw", 0);
134 if (io->running == 0)
135 break;
136 tsleep_interlock(io);
137 io->waiting = 1;
138 if (io->running == 0)
139 break;
141 crit_exit();
146 * Wait for all hammer_io-initated write I/O's to complete. This is not
147 * supposed to count direct I/O's but some can leak through (for
148 * non-full-sized direct I/Os).
150 void
151 hammer_io_wait_all(hammer_mount_t hmp, const char *ident)
153 crit_enter();
154 while (hmp->io_running_space)
155 tsleep(&hmp->io_running_space, 0, ident, 0);
156 crit_exit();
159 #define HAMMER_MAXRA 4
162 * Load bp for a HAMMER structure. The io must be exclusively locked by
163 * the caller.
165 * This routine is mostly used on meta-data and small-data blocks. Generally
166 * speaking HAMMER assumes some locality of reference and will cluster
167 * a 64K read.
169 * Note that clustering occurs at the device layer, not the logical layer.
170 * If the buffers do not apply to the current operation they may apply to
171 * some other.
174 hammer_io_read(struct vnode *devvp, struct hammer_io *io, hammer_off_t limit)
176 struct buf *bp;
177 int error;
179 if ((bp = io->bp) == NULL) {
180 hammer_count_io_running_read += io->bytes;
181 if (hammer_cluster_enable) {
182 error = cluster_read(devvp, limit,
183 io->offset, io->bytes,
184 HAMMER_CLUSTER_SIZE,
185 HAMMER_CLUSTER_BUFS, &io->bp);
186 } else {
187 error = bread(devvp, io->offset, io->bytes, &io->bp);
189 hammer_stats_disk_read += io->bytes;
190 hammer_count_io_running_read -= io->bytes;
193 * The code generally assumes b_ops/b_dep has been set-up,
194 * even if we error out here.
196 bp = io->bp;
197 bp->b_ops = &hammer_bioops;
198 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
199 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
200 BUF_KERNPROC(bp);
201 KKASSERT(io->modified == 0);
202 KKASSERT(io->running == 0);
203 KKASSERT(io->waiting == 0);
204 io->released = 0; /* we hold an active lock on bp */
205 } else {
206 error = 0;
208 return(error);
212 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
213 * Must be called with the IO exclusively locked.
215 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
216 * I/O by forcing the buffer to not be in a released state before calling
217 * it.
219 * This function will also mark the IO as modified but it will not
220 * increment the modify_refs count.
223 hammer_io_new(struct vnode *devvp, struct hammer_io *io)
225 struct buf *bp;
227 if ((bp = io->bp) == NULL) {
228 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0);
229 bp = io->bp;
230 bp->b_ops = &hammer_bioops;
231 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
232 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
233 io->released = 0;
234 KKASSERT(io->running == 0);
235 io->waiting = 0;
236 BUF_KERNPROC(bp);
237 } else {
238 if (io->released) {
239 regetblk(bp);
240 BUF_KERNPROC(bp);
241 io->released = 0;
244 hammer_io_modify(io, 0);
245 vfs_bio_clrbuf(bp);
246 return(0);
250 * Remove potential device level aliases against buffers managed by high level
251 * vnodes. Aliases can also be created due to mixed buffer sizes.
253 * This is nasty because the buffers are also VMIO-backed. Even if a buffer
254 * does not exist its backing VM pages might, and we have to invalidate
255 * those as well or a getblk() will reinstate them.
257 void
258 hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset)
260 hammer_io_structure_t iou;
261 hammer_off_t phys_offset;
262 struct buf *bp;
264 phys_offset = volume->ondisk->vol_buf_beg +
265 (zone2_offset & HAMMER_OFF_SHORT_MASK);
266 crit_enter();
267 if ((bp = findblk(volume->devvp, phys_offset)) != NULL)
268 bp = getblk(volume->devvp, phys_offset, bp->b_bufsize, 0, 0);
269 else
270 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0);
271 if ((iou = (void *)LIST_FIRST(&bp->b_dep)) != NULL) {
272 hammer_io_clear_modify(&iou->io, 1);
273 bundirty(bp);
274 iou->io.reclaim = 1;
275 hammer_io_deallocate(bp);
276 } else {
277 KKASSERT((bp->b_flags & B_LOCKED) == 0);
278 bundirty(bp);
279 bp->b_flags |= B_NOCACHE|B_RELBUF;
281 brelse(bp);
282 crit_exit();
286 * This routine is called on the last reference to a hammer structure.
287 * The io is usually interlocked with io.loading and io.refs must be 1.
289 * This routine may return a non-NULL bp to the caller for dispoal. Disposal
290 * simply means the caller finishes decrementing the ref-count on the
291 * IO structure then brelse()'s the bp. The bp may or may not still be
292 * passively associated with the IO.
294 * The only requirement here is that modified meta-data and volume-header
295 * buffer may NOT be disassociated from the IO structure, and consequently
296 * we also leave such buffers actively associated with the IO if they already
297 * are (since the kernel can't do anything with them anyway). Only the
298 * flusher is allowed to write such buffers out. Modified pure-data and
299 * undo buffers are returned to the kernel but left passively associated
300 * so we can track when the kernel writes the bp out.
302 struct buf *
303 hammer_io_release(struct hammer_io *io, int flush)
305 union hammer_io_structure *iou = (void *)io;
306 struct buf *bp;
308 if ((bp = io->bp) == NULL)
309 return(NULL);
312 * Try to flush a dirty IO to disk if asked to by the
313 * caller or if the kernel tried to flush the buffer in the past.
315 * Kernel-initiated flushes are only allowed for pure-data buffers.
316 * meta-data and volume buffers can only be flushed explicitly
317 * by HAMMER.
319 if (io->modified) {
320 if (flush) {
321 hammer_io_flush(io);
322 } else if (bp->b_flags & B_LOCKED) {
323 switch(io->type) {
324 case HAMMER_STRUCTURE_DATA_BUFFER:
325 case HAMMER_STRUCTURE_UNDO_BUFFER:
326 hammer_io_flush(io);
327 break;
328 default:
329 break;
331 } /* else no explicit request to flush the buffer */
335 * Wait for the IO to complete if asked to.
337 if (io->waitdep && io->running) {
338 hammer_io_wait(io);
342 * Return control of the buffer to the kernel (with the provisio
343 * that our bioops can override kernel decisions with regards to
344 * the buffer).
346 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) {
348 * Always disassociate the bp if an explicit flush
349 * was requested and the IO completed with no error
350 * (so unmount can really clean up the structure).
352 if (io->released) {
353 regetblk(bp);
354 BUF_KERNPROC(bp);
355 } else {
356 io->released = 1;
358 hammer_io_disassociate((hammer_io_structure_t)io);
359 /* return the bp */
360 } else if (io->modified) {
362 * Only certain IO types can be released to the kernel if
363 * the buffer has been modified.
365 * volume and meta-data IO types may only be explicitly
366 * flushed by HAMMER.
368 switch(io->type) {
369 case HAMMER_STRUCTURE_DATA_BUFFER:
370 case HAMMER_STRUCTURE_UNDO_BUFFER:
371 if (io->released == 0) {
372 io->released = 1;
373 bdwrite(bp);
375 break;
376 default:
377 break;
379 bp = NULL; /* bp left associated */
380 } else if (io->released == 0) {
382 * Clean buffers can be generally released to the kernel.
383 * We leave the bp passively associated with the HAMMER
384 * structure and use bioops to disconnect it later on
385 * if the kernel wants to discard the buffer.
387 * We can steal the structure's ownership of the bp.
389 io->released = 1;
390 if (bp->b_flags & B_LOCKED) {
391 hammer_io_disassociate(iou);
392 /* return the bp */
393 } else {
394 if (io->reclaim) {
395 hammer_io_disassociate(iou);
396 /* return the bp */
397 } else {
398 /* return the bp (bp passively associated) */
401 } else {
403 * A released buffer is passively associate with our
404 * hammer_io structure. The kernel cannot destroy it
405 * without making a bioops call. If the kernel (B_LOCKED)
406 * or we (reclaim) requested that the buffer be destroyed
407 * we destroy it, otherwise we do a quick get/release to
408 * reset its position in the kernel's LRU list.
410 * Leaving the buffer passively associated allows us to
411 * use the kernel's LRU buffer flushing mechanisms rather
412 * then rolling our own.
414 * XXX there are two ways of doing this. We can re-acquire
415 * and passively release to reset the LRU, or not.
417 if (io->running == 0) {
418 regetblk(bp);
419 if ((bp->b_flags & B_LOCKED) || io->reclaim) {
420 hammer_io_disassociate(iou);
421 /* return the bp */
422 } else {
423 /* return the bp (bp passively associated) */
425 } else {
427 * bp is left passively associated but we do not
428 * try to reacquire it. Interactions with the io
429 * structure will occur on completion of the bp's
430 * I/O.
432 bp = NULL;
435 return(bp);
439 * This routine is called with a locked IO when a flush is desired and
440 * no other references to the structure exists other then ours. This
441 * routine is ONLY called when HAMMER believes it is safe to flush a
442 * potentially modified buffer out.
444 void
445 hammer_io_flush(struct hammer_io *io)
447 struct buf *bp;
450 * Degenerate case - nothing to flush if nothing is dirty.
452 if (io->modified == 0) {
453 return;
456 KKASSERT(io->bp);
457 KKASSERT(io->modify_refs <= 0);
460 * Acquire ownership of the bp, particularly before we clear our
461 * modified flag.
463 * We are going to bawrite() this bp. Don't leave a window where
464 * io->released is set, we actually own the bp rather then our
465 * buffer.
467 bp = io->bp;
468 if (io->released) {
469 regetblk(bp);
470 /* BUF_KERNPROC(io->bp); */
471 /* io->released = 0; */
472 KKASSERT(io->released);
473 KKASSERT(io->bp == bp);
475 io->released = 1;
478 * Acquire exclusive access to the bp and then clear the modified
479 * state of the buffer prior to issuing I/O to interlock any
480 * modifications made while the I/O is in progress. This shouldn't
481 * happen anyway but losing data would be worse. The modified bit
482 * will be rechecked after the IO completes.
484 * NOTE: This call also finalizes the buffer's content (inval == 0).
486 * This is only legal when lock.refs == 1 (otherwise we might clear
487 * the modified bit while there are still users of the cluster
488 * modifying the data).
490 * Do this before potentially blocking so any attempt to modify the
491 * ondisk while we are blocked blocks waiting for us.
493 hammer_io_clear_modify(io, 0);
496 * Transfer ownership to the kernel and initiate I/O.
498 io->running = 1;
499 io->hmp->io_running_space += io->bytes;
500 hammer_count_io_running_write += io->bytes;
501 bawrite(bp);
504 /************************************************************************
505 * BUFFER DIRTYING *
506 ************************************************************************
508 * These routines deal with dependancies created when IO buffers get
509 * modified. The caller must call hammer_modify_*() on a referenced
510 * HAMMER structure prior to modifying its on-disk data.
512 * Any intent to modify an IO buffer acquires the related bp and imposes
513 * various write ordering dependancies.
517 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
518 * are locked until the flusher can deal with them, pure data buffers
519 * can be written out.
521 static
522 void
523 hammer_io_modify(hammer_io_t io, int count)
526 * io->modify_refs must be >= 0
528 while (io->modify_refs < 0) {
529 io->waitmod = 1;
530 tsleep(io, 0, "hmrmod", 0);
534 * Shortcut if nothing to do.
536 KKASSERT(io->lock.refs != 0 && io->bp != NULL);
537 io->modify_refs += count;
538 if (io->modified && io->released == 0)
539 return;
541 hammer_lock_ex(&io->lock);
542 if (io->modified == 0) {
543 hammer_io_set_modlist(io);
544 io->modified = 1;
546 if (io->released) {
547 regetblk(io->bp);
548 BUF_KERNPROC(io->bp);
549 io->released = 0;
550 KKASSERT(io->modified != 0);
552 hammer_unlock(&io->lock);
555 static __inline
556 void
557 hammer_io_modify_done(hammer_io_t io)
559 KKASSERT(io->modify_refs > 0);
560 --io->modify_refs;
561 if (io->modify_refs == 0 && io->waitmod) {
562 io->waitmod = 0;
563 wakeup(io);
567 void
568 hammer_io_write_interlock(hammer_io_t io)
570 while (io->modify_refs != 0) {
571 io->waitmod = 1;
572 tsleep(io, 0, "hmrmod", 0);
574 io->modify_refs = -1;
577 void
578 hammer_io_done_interlock(hammer_io_t io)
580 KKASSERT(io->modify_refs == -1);
581 io->modify_refs = 0;
582 if (io->waitmod) {
583 io->waitmod = 0;
584 wakeup(io);
589 * Caller intends to modify a volume's ondisk structure.
591 * This is only allowed if we are the flusher or we have a ref on the
592 * sync_lock.
594 void
595 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
596 void *base, int len)
598 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
600 hammer_io_modify(&volume->io, 1);
601 if (len) {
602 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk;
603 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
604 hammer_generate_undo(trans, &volume->io,
605 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset),
606 base, len);
611 * Caller intends to modify a buffer's ondisk structure.
613 * This is only allowed if we are the flusher or we have a ref on the
614 * sync_lock.
616 void
617 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
618 void *base, int len)
620 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
622 hammer_io_modify(&buffer->io, 1);
623 if (len) {
624 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk;
625 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
626 hammer_generate_undo(trans, &buffer->io,
627 buffer->zone2_offset + rel_offset,
628 base, len);
632 void
633 hammer_modify_volume_done(hammer_volume_t volume)
635 hammer_io_modify_done(&volume->io);
638 void
639 hammer_modify_buffer_done(hammer_buffer_t buffer)
641 hammer_io_modify_done(&buffer->io);
645 * Mark an entity as not being dirty any more and finalize any
646 * delayed adjustments to the buffer.
648 * Delayed adjustments are an important performance enhancement, allowing
649 * us to avoid recalculating B-Tree node CRCs over and over again when
650 * making bulk-modifications to the B-Tree.
652 * If inval is non-zero delayed adjustments are ignored.
654 void
655 hammer_io_clear_modify(struct hammer_io *io, int inval)
657 if (io->modified == 0)
658 return;
661 * Take us off the mod-list and clear the modified bit.
663 KKASSERT(io->mod_list != NULL);
664 if (io->mod_list == &io->hmp->volu_list ||
665 io->mod_list == &io->hmp->meta_list) {
666 io->hmp->locked_dirty_space -= io->bytes;
667 hammer_count_dirtybufspace -= io->bytes;
669 TAILQ_REMOVE(io->mod_list, io, mod_entry);
670 io->mod_list = NULL;
671 io->modified = 0;
674 * If this bit is not set there are no delayed adjustments.
676 if (io->gencrc == 0)
677 return;
678 io->gencrc = 0;
681 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
682 * on the node (& underlying buffer). Release the node after clearing
683 * the flag.
685 if (io->type == HAMMER_STRUCTURE_META_BUFFER) {
686 hammer_buffer_t buffer = (void *)io;
687 hammer_node_t node;
689 restart:
690 TAILQ_FOREACH(node, &buffer->clist, entry) {
691 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0)
692 continue;
693 node->flags &= ~HAMMER_NODE_NEEDSCRC;
694 KKASSERT(node->ondisk);
695 if (inval == 0)
696 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE);
697 hammer_rel_node(node);
698 goto restart;
705 * Clear the IO's modify list. Even though the IO is no longer modified
706 * it may still be on the lose_list. This routine is called just before
707 * the governing hammer_buffer is destroyed.
709 void
710 hammer_io_clear_modlist(struct hammer_io *io)
712 KKASSERT(io->modified == 0);
713 if (io->mod_list) {
714 crit_enter(); /* biodone race against list */
715 KKASSERT(io->mod_list == &io->hmp->lose_list);
716 TAILQ_REMOVE(io->mod_list, io, mod_entry);
717 io->mod_list = NULL;
718 crit_exit();
722 static void
723 hammer_io_set_modlist(struct hammer_io *io)
725 struct hammer_mount *hmp = io->hmp;
727 KKASSERT(io->mod_list == NULL);
729 switch(io->type) {
730 case HAMMER_STRUCTURE_VOLUME:
731 io->mod_list = &hmp->volu_list;
732 hmp->locked_dirty_space += io->bytes;
733 hammer_count_dirtybufspace += io->bytes;
734 break;
735 case HAMMER_STRUCTURE_META_BUFFER:
736 io->mod_list = &hmp->meta_list;
737 hmp->locked_dirty_space += io->bytes;
738 hammer_count_dirtybufspace += io->bytes;
739 break;
740 case HAMMER_STRUCTURE_UNDO_BUFFER:
741 io->mod_list = &hmp->undo_list;
742 break;
743 case HAMMER_STRUCTURE_DATA_BUFFER:
744 io->mod_list = &hmp->data_list;
745 break;
747 TAILQ_INSERT_TAIL(io->mod_list, io, mod_entry);
750 /************************************************************************
751 * HAMMER_BIOOPS *
752 ************************************************************************
757 * Pre-IO initiation kernel callback - cluster build only
759 static void
760 hammer_io_start(struct buf *bp)
765 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
767 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
768 * may also be set if we were marking a cluster header open. Only remove
769 * our dependancy if the modified bit is clear.
771 static void
772 hammer_io_complete(struct buf *bp)
774 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
776 KKASSERT(iou->io.released == 1);
779 * Deal with people waiting for I/O to drain
781 if (iou->io.running) {
783 * Deal with critical write errors. Once a critical error
784 * has been flagged in hmp the UNDO FIFO will not be updated.
785 * That way crash recover will give us a consistent
786 * filesystem.
788 * Because of this we can throw away failed UNDO buffers. If
789 * we throw away META or DATA buffers we risk corrupting
790 * the now read-only version of the filesystem visible to
791 * the user. Clear B_ERROR so the buffer is not re-dirtied
792 * by the kernel and ref the io so it doesn't get thrown
793 * away.
795 if (bp->b_flags & B_ERROR) {
796 hammer_critical_error(iou->io.hmp, NULL, bp->b_error,
797 "while flushing meta-data");
798 switch(iou->io.type) {
799 case HAMMER_STRUCTURE_UNDO_BUFFER:
800 break;
801 default:
802 if (iou->io.ioerror == 0) {
803 iou->io.ioerror = 1;
804 if (iou->io.lock.refs == 0)
805 ++hammer_count_refedbufs;
806 hammer_ref(&iou->io.lock);
808 break;
810 bp->b_flags &= ~B_ERROR;
811 bundirty(bp);
812 #if 0
813 hammer_io_set_modlist(&iou->io);
814 iou->io.modified = 1;
815 #endif
817 hammer_stats_disk_write += iou->io.bytes;
818 hammer_count_io_running_write -= iou->io.bytes;
819 iou->io.hmp->io_running_space -= iou->io.bytes;
820 if (iou->io.hmp->io_running_space == 0)
821 wakeup(&iou->io.hmp->io_running_space);
822 KKASSERT(iou->io.hmp->io_running_space >= 0);
823 iou->io.running = 0;
824 } else {
825 hammer_stats_disk_read += iou->io.bytes;
828 if (iou->io.waiting) {
829 iou->io.waiting = 0;
830 wakeup(iou);
834 * If B_LOCKED is set someone wanted to deallocate the bp at some
835 * point, do it now if refs has become zero.
837 if ((bp->b_flags & B_LOCKED) && iou->io.lock.refs == 0) {
838 KKASSERT(iou->io.modified == 0);
839 --hammer_count_io_locked;
840 bp->b_flags &= ~B_LOCKED;
841 hammer_io_deallocate(bp);
842 /* structure may be dead now */
847 * Callback from kernel when it wishes to deallocate a passively
848 * associated structure. This mostly occurs with clean buffers
849 * but it may be possible for a holding structure to be marked dirty
850 * while its buffer is passively associated. The caller owns the bp.
852 * If we cannot disassociate we set B_LOCKED to prevent the buffer
853 * from getting reused.
855 * WARNING: Because this can be called directly by getnewbuf we cannot
856 * recurse into the tree. If a bp cannot be immediately disassociated
857 * our only recourse is to set B_LOCKED.
859 * WARNING: This may be called from an interrupt via hammer_io_complete()
861 static void
862 hammer_io_deallocate(struct buf *bp)
864 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep);
866 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0);
867 if (iou->io.lock.refs > 0 || iou->io.modified) {
869 * It is not legal to disassociate a modified buffer. This
870 * case really shouldn't ever occur.
872 bp->b_flags |= B_LOCKED;
873 ++hammer_count_io_locked;
874 } else {
876 * Disassociate the BP. If the io has no refs left we
877 * have to add it to the loose list.
879 hammer_io_disassociate(iou);
880 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) {
881 KKASSERT(iou->io.bp == NULL);
882 KKASSERT(iou->io.mod_list == NULL);
883 crit_enter(); /* biodone race against list */
884 iou->io.mod_list = &iou->io.hmp->lose_list;
885 TAILQ_INSERT_TAIL(iou->io.mod_list, &iou->io, mod_entry);
886 crit_exit();
891 static int
892 hammer_io_fsync(struct vnode *vp)
894 return(0);
898 * NOTE: will not be called unless we tell the kernel about the
899 * bioops. Unused... we use the mount's VFS_SYNC instead.
901 static int
902 hammer_io_sync(struct mount *mp)
904 return(0);
907 static void
908 hammer_io_movedeps(struct buf *bp1, struct buf *bp2)
913 * I/O pre-check for reading and writing. HAMMER only uses this for
914 * B_CACHE buffers so checkread just shouldn't happen, but if it does
915 * allow it.
917 * Writing is a different case. We don't want the kernel to try to write
918 * out a buffer that HAMMER may be modifying passively or which has a
919 * dependancy. In addition, kernel-demanded writes can only proceed for
920 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
921 * buffer types can only be explicitly written by the flusher.
923 * checkwrite will only be called for bdwrite()n buffers. If we return
924 * success the kernel is guaranteed to initiate the buffer write.
926 static int
927 hammer_io_checkread(struct buf *bp)
929 return(0);
932 static int
933 hammer_io_checkwrite(struct buf *bp)
935 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep);
938 * This shouldn't happen under normal operation.
940 if (io->type == HAMMER_STRUCTURE_VOLUME ||
941 io->type == HAMMER_STRUCTURE_META_BUFFER) {
942 if (!panicstr)
943 panic("hammer_io_checkwrite: illegal buffer");
944 if ((bp->b_flags & B_LOCKED) == 0) {
945 bp->b_flags |= B_LOCKED;
946 ++hammer_count_io_locked;
948 return(1);
952 * We can only clear the modified bit if the IO is not currently
953 * undergoing modification. Otherwise we may miss changes.
955 if (io->modify_refs == 0 && io->modified)
956 hammer_io_clear_modify(io, 0);
959 * The kernel is going to start the IO, set io->running.
961 KKASSERT(io->running == 0);
962 io->running = 1;
963 io->hmp->io_running_space += io->bytes;
964 hammer_count_io_running_write += io->bytes;
965 return(0);
969 * Return non-zero if we wish to delay the kernel's attempt to flush
970 * this buffer to disk.
972 static int
973 hammer_io_countdeps(struct buf *bp, int n)
975 return(0);
978 struct bio_ops hammer_bioops = {
979 .io_start = hammer_io_start,
980 .io_complete = hammer_io_complete,
981 .io_deallocate = hammer_io_deallocate,
982 .io_fsync = hammer_io_fsync,
983 .io_sync = hammer_io_sync,
984 .io_movedeps = hammer_io_movedeps,
985 .io_countdeps = hammer_io_countdeps,
986 .io_checkread = hammer_io_checkread,
987 .io_checkwrite = hammer_io_checkwrite,
990 /************************************************************************
991 * DIRECT IO OPS *
992 ************************************************************************
994 * These functions operate directly on the buffer cache buffer associated
995 * with a front-end vnode rather then a back-end device vnode.
999 * Read a buffer associated with a front-end vnode directly from the
1000 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
1001 * we validate the CRC.
1003 * We must check for the presence of a HAMMER buffer to handle the case
1004 * where the reblocker has rewritten the data (which it does via the HAMMER
1005 * buffer system, not via the high-level vnode buffer cache), but not yet
1006 * committed the buffer to the media.
1009 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1010 hammer_btree_leaf_elm_t leaf)
1012 hammer_off_t buf_offset;
1013 hammer_off_t zone2_offset;
1014 hammer_volume_t volume;
1015 struct buf *bp;
1016 struct bio *nbio;
1017 int vol_no;
1018 int error;
1020 buf_offset = bio->bio_offset;
1021 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) ==
1022 HAMMER_ZONE_LARGE_DATA);
1025 * The buffer cache may have an aliased buffer (the reblocker can
1026 * write them). If it does we have to sync any dirty data before
1027 * we can build our direct-read. This is a non-critical code path.
1029 bp = bio->bio_buf;
1030 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize);
1033 * Resolve to a zone-2 offset. The conversion just requires
1034 * munging the top 4 bits but we want to abstract it anyway
1035 * so the blockmap code can verify the zone assignment.
1037 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1038 if (error)
1039 goto done;
1040 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
1041 HAMMER_ZONE_RAW_BUFFER);
1044 * Resolve volume and raw-offset for 3rd level bio. The
1045 * offset will be specific to the volume.
1047 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1048 volume = hammer_get_volume(hmp, vol_no, &error);
1049 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1050 error = EIO;
1052 if (error == 0) {
1054 * 3rd level bio
1056 nbio = push_bio(bio);
1057 nbio->bio_offset = volume->ondisk->vol_buf_beg +
1058 (zone2_offset & HAMMER_OFF_SHORT_MASK);
1059 #if 0
1061 * XXX disabled - our CRC check doesn't work if the OS
1062 * does bogus_page replacement on the direct-read.
1064 if (leaf && hammer_verify_data) {
1065 nbio->bio_done = hammer_io_direct_read_complete;
1066 nbio->bio_caller_info1.uvalue32 = leaf->data_crc;
1068 #endif
1069 hammer_stats_disk_read += bp->b_bufsize;
1070 vn_strategy(volume->devvp, nbio);
1072 hammer_rel_volume(volume, 0);
1073 done:
1074 if (error) {
1075 kprintf("hammer_direct_read: failed @ %016llx\n",
1076 zone2_offset);
1077 bp->b_error = error;
1078 bp->b_flags |= B_ERROR;
1079 biodone(bio);
1081 return(error);
1084 #if 0
1086 * On completion of the BIO this callback must check the data CRC
1087 * and chain to the previous bio.
1089 static
1090 void
1091 hammer_io_direct_read_complete(struct bio *nbio)
1093 struct bio *obio;
1094 struct buf *bp;
1095 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32;
1097 bp = nbio->bio_buf;
1098 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) {
1099 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1100 nbio->bio_offset, bp->b_bufsize);
1101 if (hammer_debug_debug)
1102 Debugger("");
1103 bp->b_flags |= B_ERROR;
1104 bp->b_error = EIO;
1106 obio = pop_bio(nbio);
1107 biodone(obio);
1109 #endif
1112 * Write a buffer associated with a front-end vnode directly to the
1113 * disk media. The bio may be issued asynchronously.
1115 * The BIO is associated with the specified record and RECF_DIRECT_IO
1116 * is set. The recorded is added to its object.
1119 hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
1120 struct bio *bio)
1122 hammer_btree_leaf_elm_t leaf = &record->leaf;
1123 hammer_off_t buf_offset;
1124 hammer_off_t zone2_offset;
1125 hammer_volume_t volume;
1126 hammer_buffer_t buffer;
1127 struct buf *bp;
1128 struct bio *nbio;
1129 char *ptr;
1130 int vol_no;
1131 int error;
1133 buf_offset = leaf->data_offset;
1135 KKASSERT(buf_offset > HAMMER_ZONE_BTREE);
1136 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE);
1138 if ((buf_offset & HAMMER_BUFMASK) == 0 &&
1139 leaf->data_len >= HAMMER_BUFSIZE) {
1141 * We are using the vnode's bio to write directly to the
1142 * media, any hammer_buffer at the same zone-X offset will
1143 * now have stale data.
1145 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1146 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1147 volume = hammer_get_volume(hmp, vol_no, &error);
1149 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1150 error = EIO;
1151 if (error == 0) {
1152 bp = bio->bio_buf;
1153 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0);
1155 hammer_del_buffers(hmp, buf_offset,
1156 zone2_offset, bp->b_bufsize);
1160 * Second level bio - cached zone2 offset.
1162 * (We can put our bio_done function in either the
1163 * 2nd or 3rd level).
1165 nbio = push_bio(bio);
1166 nbio->bio_offset = zone2_offset;
1167 nbio->bio_done = hammer_io_direct_write_complete;
1168 nbio->bio_caller_info1.ptr = record;
1169 record->zone2_offset = zone2_offset;
1170 record->flags |= HAMMER_RECF_DIRECT_IO |
1171 HAMMER_RECF_DIRECT_INVAL;
1174 * Third level bio - raw offset specific to the
1175 * correct volume.
1177 zone2_offset &= HAMMER_OFF_SHORT_MASK;
1178 nbio = push_bio(nbio);
1179 nbio->bio_offset = volume->ondisk->vol_buf_beg +
1180 zone2_offset;
1181 hammer_stats_disk_write += bp->b_bufsize;
1182 vn_strategy(volume->devvp, nbio);
1184 hammer_rel_volume(volume, 0);
1185 } else {
1187 * Must fit in a standard HAMMER buffer. In this case all
1188 * consumers use the HAMMER buffer system and RECF_DIRECT_IO
1189 * does not need to be set-up.
1191 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0);
1192 buffer = NULL;
1193 ptr = hammer_bread(hmp, buf_offset, &error, &buffer);
1194 if (error == 0) {
1195 bp = bio->bio_buf;
1196 bp->b_flags |= B_AGE;
1197 hammer_io_modify(&buffer->io, 1);
1198 bcopy(bp->b_data, ptr, leaf->data_len);
1199 hammer_io_modify_done(&buffer->io);
1200 hammer_rel_buffer(buffer, 0);
1201 bp->b_resid = 0;
1202 biodone(bio);
1205 if (error == 0) {
1207 * The record is all setup now, add it. Potential conflics
1208 * have already been dealt with.
1210 error = hammer_mem_add(record);
1211 KKASSERT(error == 0);
1212 } else {
1214 * Major suckage occured.
1216 kprintf("hammer_direct_write: failed @ %016llx\n",
1217 leaf->data_offset);
1218 bp = bio->bio_buf;
1219 bp->b_resid = 0;
1220 bp->b_error = EIO;
1221 bp->b_flags |= B_ERROR;
1222 biodone(bio);
1223 record->flags |= HAMMER_RECF_DELETED_FE;
1224 hammer_rel_mem_record(record);
1226 return(error);
1230 * On completion of the BIO this callback must disconnect
1231 * it from the hammer_record and chain to the previous bio.
1233 * An I/O error forces the mount to read-only. Data buffers
1234 * are not B_LOCKED like meta-data buffers are, so we have to
1235 * throw the buffer away to prevent the kernel from retrying.
1237 static
1238 void
1239 hammer_io_direct_write_complete(struct bio *nbio)
1241 struct bio *obio;
1242 struct buf *bp;
1243 hammer_record_t record = nbio->bio_caller_info1.ptr;
1245 bp = nbio->bio_buf;
1246 obio = pop_bio(nbio);
1247 if (bp->b_flags & B_ERROR) {
1248 hammer_critical_error(record->ip->hmp, record->ip,
1249 bp->b_error,
1250 "while writing bulk data");
1251 bp->b_flags |= B_INVAL;
1253 biodone(obio);
1255 KKASSERT(record != NULL);
1256 KKASSERT(record->flags & HAMMER_RECF_DIRECT_IO);
1257 record->flags &= ~HAMMER_RECF_DIRECT_IO;
1258 if (record->flags & HAMMER_RECF_DIRECT_WAIT) {
1259 record->flags &= ~HAMMER_RECF_DIRECT_WAIT;
1260 wakeup(&record->flags);
1266 * This is called before a record is either committed to the B-Tree
1267 * or destroyed, to resolve any associated direct-IO.
1269 * (1) We must wait for any direct-IO related to the record to complete.
1271 * (2) We must remove any buffer cache aliases for data accessed via
1272 * leaf->data_offset or zone2_offset so non-direct-IO consumers
1273 * (the mirroring and reblocking code) do not see stale data.
1275 void
1276 hammer_io_direct_wait(hammer_record_t record)
1279 * Wait for I/O to complete
1281 if (record->flags & HAMMER_RECF_DIRECT_IO) {
1282 crit_enter();
1283 while (record->flags & HAMMER_RECF_DIRECT_IO) {
1284 record->flags |= HAMMER_RECF_DIRECT_WAIT;
1285 tsleep(&record->flags, 0, "hmdiow", 0);
1287 crit_exit();
1291 * Invalidate any related buffer cache aliases.
1293 if (record->flags & HAMMER_RECF_DIRECT_INVAL) {
1294 KKASSERT(record->leaf.data_offset);
1295 hammer_del_buffers(record->ip->hmp,
1296 record->leaf.data_offset,
1297 record->zone2_offset,
1298 record->leaf.data_len);
1299 record->flags &= ~HAMMER_RECF_DIRECT_INVAL;
1304 * This is called to remove the second-level cached zone-2 offset from
1305 * frontend buffer cache buffers, now stale due to a data relocation.
1306 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1307 * by hammer_vop_strategy_read().
1309 * This is rather nasty because here we have something like the reblocker
1310 * scanning the raw B-Tree with no held references on anything, really,
1311 * other then a shared lock on the B-Tree node, and we have to access the
1312 * frontend's buffer cache to check for and clean out the association.
1313 * Specifically, if the reblocker is moving data on the disk, these cached
1314 * offsets will become invalid.
1316 * Only data record types associated with the large-data zone are subject
1317 * to direct-io and need to be checked.
1320 void
1321 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf)
1323 struct hammer_inode_info iinfo;
1324 int zone;
1326 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA)
1327 return;
1328 zone = HAMMER_ZONE_DECODE(leaf->data_offset);
1329 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX)
1330 return;
1331 iinfo.obj_id = leaf->base.obj_id;
1332 iinfo.obj_asof = 0; /* unused */
1333 iinfo.obj_localization = leaf->base.localization &
1334 HAMMER_LOCALIZE_PSEUDOFS_MASK;
1335 iinfo.u.leaf = leaf;
1336 hammer_scan_inode_snapshots(hmp, &iinfo,
1337 hammer_io_direct_uncache_callback,
1338 leaf);
1341 static int
1342 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data)
1344 hammer_inode_info_t iinfo = data;
1345 hammer_off_t data_offset;
1346 hammer_off_t file_offset;
1347 struct vnode *vp;
1348 struct buf *bp;
1349 int blksize;
1351 if (ip->vp == NULL)
1352 return(0);
1353 data_offset = iinfo->u.leaf->data_offset;
1354 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len;
1355 blksize = iinfo->u.leaf->data_len;
1356 KKASSERT((blksize & HAMMER_BUFMASK) == 0);
1358 hammer_ref(&ip->lock);
1359 if (hammer_get_vnode(ip, &vp) == 0) {
1360 if ((bp = findblk(ip->vp, file_offset)) != NULL &&
1361 bp->b_bio2.bio_offset != NOOFFSET) {
1362 bp = getblk(ip->vp, file_offset, blksize, 0, 0);
1363 bp->b_bio2.bio_offset = NOOFFSET;
1364 brelse(bp);
1366 vput(vp);
1368 hammer_rel_inode(ip, 0);
1369 return(0);