2 * Copyright (c) 2011-2018 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
6 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * TRANSACTION AND FLUSH HANDLING
38 * Deceptively simple but actually fairly difficult to implement properly is
39 * how I would describe it.
41 * Flushing generally occurs bottom-up but requires a top-down scan to
42 * locate chains with MODIFIED and/or UPDATE bits set. The ONFLUSH flag
43 * tells how to recurse downward to find these chains.
46 #include <sys/cdefs.h>
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/types.h>
57 #define HAMMER2_FLUSH_DEPTH_LIMIT 60 /* stack recursion limit */
61 * Recursively flush the specified chain. The chain is locked and
62 * referenced by the caller and will remain so on return. The chain
63 * will remain referenced throughout but can temporarily lose its
64 * lock during the recursion to avoid unnecessarily stalling user
67 struct hammer2_flush_info {
68 hammer2_chain_t *parent;
71 int error; /* cumulative error */
73 #ifdef HAMMER2_SCAN_DEBUG
82 struct h2_flush_list flushq;
83 hammer2_chain_t *debug;
86 typedef struct hammer2_flush_info hammer2_flush_info_t;
88 static void hammer2_flush_core(hammer2_flush_info_t *info,
89 hammer2_chain_t *chain, int flags);
90 static int hammer2_flush_recurse(hammer2_chain_t *child, void *data);
93 * Any per-pfs transaction initialization goes here.
96 hammer2_trans_manage_init(hammer2_pfs_t *pmp)
101 * Transaction support for any modifying operation. Transactions are used
102 * in the pmp layer by the frontend and in the spmp layer by the backend.
104 * 0 - Normal transaction, interlocked against flush
107 * TRANS_ISFLUSH - Flush transaction, interlocked against normal
110 * TRANS_BUFCACHE - Buffer cache transaction, no interlock.
112 * Initializing a new transaction allocates a transaction ID. Typically
113 * passed a pmp (hmp passed as NULL), indicating a cluster transaction. Can
114 * be passed a NULL pmp and non-NULL hmp to indicate a transaction on a single
115 * media target. The latter mode is used by the recovery code.
117 * TWO TRANSACTION IDs can run concurrently, where one is a flush and the
118 * other is a set of any number of concurrent filesystem operations. We
119 * can either have <running_fs_ops> + <waiting_flush> + <blocked_fs_ops>
120 * or we can have <running_flush> + <concurrent_fs_ops>.
122 * During a flush, new fs_ops are only blocked until the fs_ops prior to
123 * the flush complete. The new fs_ops can then run concurrent with the flush.
125 * Buffer-cache transactions operate as fs_ops but never block. A
126 * buffer-cache flush will run either before or after the current pending
127 * flush depending on its state.
130 hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags)
137 oflags = pmp->trans.flags;
141 if (flags & HAMMER2_TRANS_ISFLUSH) {
143 * Requesting flush transaction. This interlocks
144 * only with other flush transactions. Note that
145 * non-flush modifying transactions can run
146 * concurrently, but will interlock on any inode
147 * that are on the SYNCQ.
149 if (oflags & HAMMER2_TRANS_ISFLUSH) {
150 nflags = oflags | HAMMER2_TRANS_WAITING;
153 nflags = (oflags | flags) + 1;
156 if (oflags & HAMMER2_TRANS_MASK) {
157 nflags = oflags | HAMMER2_TRANS_FPENDING |
158 HAMMER2_TRANS_WAITING;
161 nflags = (oflags | flags) + 1;
164 } else if (flags & HAMMER2_TRANS_BUFCACHE) {
166 * Requesting strategy transaction from buffer-cache,
167 * or a VM getpages/putpages through the buffer cache.
168 * We must allow such transactions in all situations
169 * to avoid deadlocks.
171 nflags = (oflags | flags) + 1;
174 * Requesting a normal modifying transaction.
175 * Does not interlock with flushes. Multiple
176 * modifying transactions can run concurrently.
177 * These do not mess with the on-media topology
180 * If a flush is pending for more than one second
181 * but can't run because many modifying transactions
182 * are active, we wait for the flush to be granted.
184 * NOTE: Remember that non-modifying operations
185 * such as read, stat, readdir, etc, do
186 * not use transactions.
189 if ((oflags & HAMMER2_TRANS_FPENDING) &&
190 (u_int)(ticks - pmp->trans.fticks) >= (u_int)hz) {
191 nflags = oflags | HAMMER2_TRANS_WAITING;
193 } else if (oflags & HAMMER2_TRANS_ISFLUSH) {
194 nflags = oflags | HAMMER2_TRANS_WAITING;
199 nflags = (oflags | flags) + 1;
203 tsleep_interlock(&pmp->trans.sync_wait, 0);
204 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
205 if ((oflags & HAMMER2_TRANS_FPENDING) == 0 &&
206 (nflags & HAMMER2_TRANS_FPENDING)) {
207 pmp->trans.fticks = ticks;
211 tsleep(&pmp->trans.sync_wait, PINTERLOCKED,
221 * Start a sub-transaction, there is no 'subdone' function. This will
222 * issue a new modify_tid (mtid) for the current transaction, which is a
223 * CLC (cluster level change) id and not a per-node id.
225 * This function must be called for each XOP when multiple XOPs are run in
226 * sequence within a transaction.
228 * Callers typically update the inode with the transaction mtid manually
229 * to enforce sequencing.
232 hammer2_trans_sub(hammer2_pfs_t *pmp)
236 mtid = atomic_fetchadd_64(&pmp->modify_tid, 1);
242 hammer2_trans_done(hammer2_pfs_t *pmp, int quicksideq)
248 * Modifying ops on the front-end can cause dirty inodes to
249 * build up in the sideq. We don't flush these on inactive/reclaim
250 * due to potential deadlocks, so we have to deal with them from
251 * inside other nominal modifying front-end transactions.
253 if (quicksideq && pmp->sideq_count > (pmp->inum_count >> 3) && pmp->mp)
254 speedup_syncer(pmp->mp);
256 hammer2_inode_run_sideq(pmp, 0);
260 * Clean-up the transaction
263 oflags = pmp->trans.flags;
265 KKASSERT(oflags & HAMMER2_TRANS_MASK);
266 if ((oflags & HAMMER2_TRANS_MASK) == 1) {
268 * This was the last transaction
270 nflags = (oflags - 1) & ~(HAMMER2_TRANS_ISFLUSH |
271 HAMMER2_TRANS_BUFCACHE |
272 HAMMER2_TRANS_FPENDING |
273 HAMMER2_TRANS_WAITING);
276 * Still transactions pending
280 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
281 if ((nflags & HAMMER2_TRANS_MASK) == 0 &&
282 (oflags & HAMMER2_TRANS_WAITING)) {
283 wakeup(&pmp->trans.sync_wait);
294 * Obtain new, unique inode number (not serialized by caller).
297 hammer2_trans_newinum(hammer2_pfs_t *pmp)
301 tid = atomic_fetchadd_64(&pmp->inode_tid, 1);
307 * Assert that a strategy call is ok here. Currently we allow strategy
308 * calls in all situations, including during flushes. Previously:
309 * (old) (1) In a normal transaction.
310 * (old) (2) In a flush transaction only if PREFLUSH is also set.
313 hammer2_trans_assert_strategy(hammer2_pfs_t *pmp)
316 KKASSERT((pmp->trans.flags & HAMMER2_TRANS_ISFLUSH) == 0 ||
317 (pmp->trans.flags & HAMMER2_TRANS_PREFLUSH));
323 * Chains undergoing destruction are removed from the in-memory topology.
324 * To avoid getting lost these chains are placed on the delayed flush
325 * queue which will properly dispose of them.
327 * We do this instead of issuing an immediate flush in order to give
328 * recursive deletions (rm -rf, etc) a chance to remove more of the
329 * hierarchy, potentially allowing an enormous amount of write I/O to
332 * NOTE: The flush code tests HAMMER2_CHAIN_DESTROY to differentiate
333 * between these chains and the deep-recursion requeue.
336 hammer2_delayed_flush(hammer2_chain_t *chain)
338 if ((chain->flags & HAMMER2_CHAIN_DELAYED) == 0) {
339 hammer2_spin_ex(&chain->hmp->list_spin);
340 if ((chain->flags & (HAMMER2_CHAIN_DELAYED |
341 HAMMER2_CHAIN_DEFERRED)) == 0) {
342 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELAYED |
343 HAMMER2_CHAIN_DEFERRED);
344 TAILQ_INSERT_TAIL(&chain->hmp->flushq,
346 hammer2_chain_ref(chain);
348 hammer2_spin_unex(&chain->hmp->list_spin);
349 hammer2_voldata_modify(chain->hmp);
354 * Flush the chain and all modified sub-chains through the specified
355 * synchronization point, propagating blockref updates back up. As
356 * part of this propagation, mirror_tid and inode/data usage statistics
357 * propagates back upward.
359 * Returns a HAMMER2 error code, 0 if no error. Note that I/O errors from
360 * buffers dirtied during the flush operation can occur later.
362 * modify_tid (clc - cluster level change) is not propagated.
364 * update_tid (clc) is used for validation and is not propagated by this
367 * This routine can be called from several places but the most important
368 * is from VFS_SYNC (frontend) via hammer2_xop_inode_flush (backend).
370 * chain is locked on call and will remain locked on return. The chain's
371 * UPDATE flag indicates that its parent's block table (which is not yet
372 * part of the flush) should be updated.
375 * HAMMER2_FLUSH_TOP Indicates that this is the top of the flush.
376 * Is cleared for the recursion.
378 * HAMMER2_FLUSH_ALL Recurse everything
380 * HAMMER2_FLUSH_INODE_STOP
381 * Stop at PFS inode or normal inode boundary
384 hammer2_flush(hammer2_chain_t *chain, int flags)
386 hammer2_chain_t *scan;
387 hammer2_flush_info_t info;
392 * Execute the recursive flush and handle deferrals.
394 * Chains can be ridiculously long (thousands deep), so to
395 * avoid blowing out the kernel stack the recursive flush has a
396 * depth limit. Elements at the limit are placed on a list
397 * for re-execution after the stack has been popped.
399 bzero(&info, sizeof(info));
400 TAILQ_INIT(&info.flushq);
401 info.flags = flags & ~HAMMER2_FLUSH_TOP;
404 * Calculate parent (can be NULL), if not NULL the flush core
405 * expects the parent to be referenced so it can easily lock/unlock
406 * it without it getting ripped up.
408 if ((info.parent = chain->parent) != NULL)
409 hammer2_chain_ref(info.parent);
412 * Extra ref needed because flush_core expects it when replacing
415 hammer2_chain_ref(chain);
421 * Move hmp->flushq to info.flushq if non-empty so it can
424 if (TAILQ_FIRST(&hmp->flushq) != NULL) {
425 hammer2_spin_ex(&chain->hmp->list_spin);
426 TAILQ_CONCAT(&info.flushq, &hmp->flushq, flush_node);
427 hammer2_spin_unex(&chain->hmp->list_spin);
431 * Unwind deep recursions which had been deferred. This
432 * can leave the FLUSH_* bits set for these chains, which
433 * will be handled when we [re]flush chain after the unwind.
435 while ((scan = TAILQ_FIRST(&info.flushq)) != NULL) {
436 KKASSERT(scan->flags & HAMMER2_CHAIN_DEFERRED);
437 TAILQ_REMOVE(&info.flushq, scan, flush_node);
438 #ifdef HAMMER2_SCAN_DEBUG
441 atomic_clear_int(&scan->flags, HAMMER2_CHAIN_DEFERRED |
442 HAMMER2_CHAIN_DELAYED);
445 * Now that we've popped back up we can do a secondary
446 * recursion on the deferred elements.
448 * NOTE: hmp->flushq chains (marked DESTROY) must be
449 * handled unconditionally so they can be cleaned
452 * NOTE: hammer2_flush() may replace scan.
454 if (hammer2_debug & 0x0040)
455 kprintf("deferred flush %p\n", scan);
456 hammer2_chain_lock(scan, HAMMER2_RESOLVE_MAYBE);
457 if (scan->error == 0) {
458 if (scan->flags & HAMMER2_CHAIN_DESTROY) {
465 flags & ~HAMMER2_FLUSH_TOP);
468 info.error |= scan->error;
470 hammer2_chain_unlock(scan);
471 hammer2_chain_drop(scan);/* ref from defer */
475 * [re]flush chain as the deep recursion may have generated
476 * additional modifications.
478 info.diddeferral = 0;
479 if (info.parent != chain->parent) {
480 if (hammer2_debug & 0x0040) {
481 kprintf("LOST CHILD4 %p->%p "
482 "(actual parent %p)\n",
483 info.parent, chain, chain->parent);
485 hammer2_chain_drop(info.parent);
486 info.parent = chain->parent;
487 hammer2_chain_ref(info.parent);
489 hammer2_flush_core(&info, chain, flags);
492 * Only loop if deep recursions have been deferred.
494 if (TAILQ_EMPTY(&info.flushq))
497 if (++loops % 1000 == 0) {
498 kprintf("hammer2_flush: excessive loops on %p\n",
500 if (hammer2_debug & 0x100000)
504 #ifdef HAMMER2_SCAN_DEBUG
505 if (info.scan_count >= 10)
506 kprintf("hammer2_flush: scan_count %ld (%ld,%ld,%ld,%ld) "
507 "bt(%ld,%ld,%ld,%ld,%ld,%ld) flushq %ld\n",
521 hammer2_chain_drop(chain);
523 hammer2_chain_drop(info.parent);
528 * This is the core of the chain flushing code. The chain is locked by the
529 * caller and must also have an extra ref on it by the caller, and remains
530 * locked and will have an extra ref on return. info.parent is referenced
533 * Upon return, the caller can test the UPDATE bit on the chain to determine
534 * if the parent needs updating.
536 * (1) Determine if this node is a candidate for the flush, return if it is
537 * not. fchain and vchain are always candidates for the flush.
539 * (2) If we recurse too deep the chain is entered onto the deferral list and
540 * the current flush stack is aborted until after the deferral list is
543 * (3) Recursively flush live children (rbtree). This can create deferrals.
544 * A successful flush clears the MODIFIED and UPDATE bits on the children
545 * and typically causes the parent to be marked MODIFIED as the children
546 * update the parent's block table. A parent might already be marked
547 * MODIFIED due to a deletion (whos blocktable update in the parent is
548 * handled by the frontend), or if the parent itself is modified by the
549 * frontend for other reasons.
551 * (4) Permanently disconnected sub-trees are cleaned up by the front-end.
552 * Deleted-but-open inodes can still be individually flushed via the
555 * (5) Delete parents on the way back up if they are normal indirect blocks
556 * and have no children.
558 * (6) Note that an unmodified child may still need the block table in its
559 * parent updated (e.g. rename/move). The child will have UPDATE set
562 * WARNING ON BREF MODIFY_TID/MIRROR_TID
564 * blockref.modify_tid is consistent only within a PFS, and will not be
565 * consistent during synchronization. mirror_tid is consistent across the
566 * block device regardless of the PFS.
569 hammer2_flush_core(hammer2_flush_info_t *info, hammer2_chain_t *chain,
572 hammer2_chain_t *parent;
577 * (1) Optimize downward recursion to locate nodes needing action.
578 * Nothing to do if none of these flags are set.
580 if ((chain->flags & HAMMER2_CHAIN_FLUSH_MASK) == 0) {
581 if (hammer2_debug & 0x200) {
582 if (info->debug == NULL)
592 * NOTE: parent can be NULL, usually due to destroy races.
594 parent = info->parent;
595 KKASSERT(chain->parent == parent);
598 * Downward search recursion
600 * We must be careful on cold stops. If CHAIN_UPDATE is set and
601 * we stop cold (verses a deferral which will re-run the chain later),
602 * the update can wind up never being applied. This situation most
603 * typically occurs on inode boundaries due to the way
604 * hammer2_vfs_sync() breaks-up the flush. As a safety, we
605 * flush-through such situations.
607 if (chain->flags & (HAMMER2_CHAIN_DEFERRED | HAMMER2_CHAIN_DELAYED)) {
612 } else if ((chain->flags & HAMMER2_CHAIN_PFSBOUNDARY) &&
613 (chain->flags & HAMMER2_CHAIN_UPDATE) == 0 &&
614 (flags & HAMMER2_FLUSH_ALL) == 0 &&
615 (flags & HAMMER2_FLUSH_TOP) == 0 &&
616 chain->pmp && chain->pmp->mp) {
618 * If FLUSH_ALL is not specified the caller does not want
619 * to recurse through PFS roots that have been mounted.
621 * (If the PFS has not been mounted there may not be
622 * anything monitoring its chains and its up to us
625 * The typical sequence is to flush dirty PFS's starting at
626 * their root downward, then flush the device root (vchain).
627 * It is this second flush that typically leaves out the
630 * However we must still process the PFSROOT chains for block
631 * table updates in their parent (which IS part of our flush).
633 * NOTE: The volume root, vchain, does not set PFSBOUNDARY.
635 * NOTE: This test must be done before the depth-limit test,
636 * else it might become the top on a flushq iteration.
638 * NOTE: We must re-set ONFLUSH in the parent to retain if
639 * this chain (that we are skipping) requires work.
641 if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
642 HAMMER2_CHAIN_DESTROY |
643 HAMMER2_CHAIN_MODIFIED)) {
644 hammer2_chain_setflush(parent);
646 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
647 (chain->flags & HAMMER2_CHAIN_UPDATE) == 0 &&
648 (flags & HAMMER2_FLUSH_INODE_STOP) &&
649 (flags & HAMMER2_FLUSH_ALL) == 0 &&
650 (flags & HAMMER2_FLUSH_TOP) == 0 &&
651 chain->pmp && chain->pmp->mp) {
653 * If FLUSH_INODE_STOP is specified and both ALL and TOP
654 * are clear, we must not flush the chain. The chain should
655 * have already been flushed and any further ONFLUSH/UPDATE
656 * setting will be related to the next flush.
658 * This features allows us to flush inodes independently of
659 * each other and meta-data above the inodes separately.
661 if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
662 HAMMER2_CHAIN_DESTROY |
663 HAMMER2_CHAIN_MODIFIED)) {
665 hammer2_chain_setflush(parent);
667 } else if (info->depth == HAMMER2_FLUSH_DEPTH_LIMIT) {
669 * Recursion depth reached.
671 KKASSERT((chain->flags & HAMMER2_CHAIN_DELAYED) == 0);
672 hammer2_chain_ref(chain);
673 TAILQ_INSERT_TAIL(&info->flushq, chain, flush_node);
674 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEFERRED);
676 } else if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
677 HAMMER2_CHAIN_DESTROY)) {
679 * Downward recursion search (actual flush occurs bottom-up).
680 * pre-clear ONFLUSH. It can get set again due to races or
681 * flush errors, which we want so the scan finds us again in
684 * We must also recurse if DESTROY is set so we can finally
685 * get rid of the related children, otherwise the node will
686 * just get re-flushed on lastdrop.
688 * WARNING! The recursion will unlock/relock info->parent
689 * (which is 'chain'), potentially allowing it
692 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
693 save_error = info->error;
695 info->parent = chain;
698 * We may have to do this twice to catch any indirect
699 * block maintenance that occurs. Other conditions which
700 * can keep setting ONFLUSH (such as deferrals) ought to
701 * be handled by the flushq code. XXX needs more help
703 hammer2_spin_ex(&chain->core.spin);
704 RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
705 NULL, hammer2_flush_recurse, info);
706 if (chain->flags & HAMMER2_CHAIN_ONFLUSH) {
707 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
708 RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
709 NULL, hammer2_flush_recurse, info);
711 hammer2_spin_unex(&chain->core.spin);
712 info->parent = parent;
715 * Re-set the flush bits if the flush was incomplete or
716 * an error occurred. If an error occurs it is typically
717 * an allocation error. Errors do not cause deferrals.
720 hammer2_chain_setflush(chain);
721 info->error |= save_error;
722 if (info->diddeferral)
723 hammer2_chain_setflush(chain);
726 * If we lost the parent->chain association we have to
727 * stop processing this chain because it is no longer
728 * in this recursion. If it moved, it will be handled
729 * by the ONFLUSH flag elsewhere.
731 if (chain->parent != parent) {
732 kprintf("LOST CHILD2 %p->%p (actual parent %p)\n",
733 parent, chain, chain->parent);
739 * Now we are in the bottom-up part of the recursion.
741 * Do not update chain if lower layers were deferred. We continue
742 * to try to update the chain on lower-level errors, but the flush
743 * code may decide not to flush the volume root.
745 * XXX should we continue to try to update the chain if an error
748 if (info->diddeferral)
752 * Both parent and chain must be locked in order to flush chain,
753 * in order to properly update the parent under certain conditions.
755 * In addition, we can't safely unlock/relock the chain once we
756 * start flushing the chain itself, which we would have to do later
757 * on in order to lock the parent if we didn't do that now.
759 hammer2_chain_ref_hold(chain);
760 hammer2_chain_unlock(chain);
762 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
763 hammer2_chain_lock(chain, HAMMER2_RESOLVE_MAYBE);
764 hammer2_chain_drop_unhold(chain);
767 * Can't process if we can't access their content.
769 if ((parent && parent->error) || chain->error) {
770 kprintf("hammer2: chain error during flush\n");
771 info->error |= chain->error;
773 info->error |= parent->error;
774 hammer2_chain_unlock(parent);
779 if (chain->parent != parent) {
780 if (hammer2_debug & 0x0040) {
781 kprintf("LOST CHILD3 %p->%p (actual parent %p)\n",
782 parent, chain, chain->parent);
784 KKASSERT(parent != NULL);
785 hammer2_chain_unlock(parent);
786 if ((chain->flags & HAMMER2_CHAIN_DELAYED) == 0) {
787 hammer2_chain_ref(chain);
788 TAILQ_INSERT_TAIL(&info->flushq, chain, flush_node);
789 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEFERRED);
796 * Propagate the DESTROY flag downwards. This dummies up the flush
797 * code and tries to invalidate related buffer cache buffers to
798 * avoid the disk write.
800 if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
801 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
804 * Dispose of the modified bit.
806 * If parent is present, the UPDATE bit should already be set.
807 * UPDATE should already be set.
808 * bref.mirror_tid should already be set.
810 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
811 KKASSERT((chain->flags & HAMMER2_CHAIN_UPDATE) ||
812 chain->parent == NULL);
813 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
814 atomic_add_long(&hammer2_count_modified_chains, -1);
817 * Manage threads waiting for excessive dirty memory to
821 hammer2_pfs_memory_wakeup(chain->pmp);
824 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0 &&
825 chain != &hmp->vchain &&
826 chain != &hmp->fchain) {
828 * Set UPDATE bit indicating that the parent block
829 * table requires updating.
831 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
836 * Issue the flush. This is indirect via the DIO.
838 * NOTE: A DELETED node that reaches this point must be
839 * flushed for synchronization point consistency.
841 * NOTE: Even though MODIFIED was already set, the related DIO
842 * might not be dirty due to a system buffer cache
843 * flush and must be set dirty if we are going to make
844 * further modifications to the buffer. Chains with
845 * embedded data don't need this.
847 if (hammer2_debug & 0x1000) {
848 kprintf("Flush %p.%d %016jx/%d data=%016jx\n",
849 chain, chain->bref.type,
850 (uintmax_t)chain->bref.key,
852 (uintmax_t)chain->bref.data_off);
854 if (hammer2_debug & 0x2000) {
855 Debugger("Flush hell");
859 * Update chain CRCs for flush.
861 * NOTE: Volume headers are NOT flushed here as they require
862 * special processing.
864 switch(chain->bref.type) {
865 case HAMMER2_BREF_TYPE_FREEMAP:
867 * Update the volume header's freemap_tid to the
868 * freemap's flushing mirror_tid.
870 * (note: embedded data, do not call setdirty)
872 KKASSERT(hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED);
873 KKASSERT(chain == &hmp->fchain);
874 hmp->voldata.freemap_tid = chain->bref.mirror_tid;
875 if (hammer2_debug & 0x8000) {
876 /* debug only, avoid syslogd loop */
877 kprintf("sync freemap mirror_tid %08jx\n",
878 (intmax_t)chain->bref.mirror_tid);
882 * The freemap can be flushed independently of the
883 * main topology, but for the case where it is
884 * flushed in the same transaction, and flushed
885 * before vchain (a case we want to allow for
886 * performance reasons), make sure modifications
887 * made during the flush under vchain use a new
890 * Otherwise the mount recovery code will get confused.
892 ++hmp->voldata.mirror_tid;
894 case HAMMER2_BREF_TYPE_VOLUME:
896 * The free block table is flushed by
897 * hammer2_vfs_sync() before it flushes vchain.
898 * We must still hold fchain locked while copying
899 * voldata to volsync, however.
901 * These do not error per-say since their data does
902 * not need to be re-read from media on lock.
904 * (note: embedded data, do not call setdirty)
906 hammer2_chain_lock(&hmp->fchain,
907 HAMMER2_RESOLVE_ALWAYS);
908 hammer2_voldata_lock(hmp);
909 if (hammer2_debug & 0x8000) {
910 /* debug only, avoid syslogd loop */
911 kprintf("sync volume mirror_tid %08jx\n",
912 (intmax_t)chain->bref.mirror_tid);
916 * Update the volume header's mirror_tid to the
917 * main topology's flushing mirror_tid. It is
918 * possible that voldata.mirror_tid is already
919 * beyond bref.mirror_tid due to the bump we made
920 * above in BREF_TYPE_FREEMAP.
922 if (hmp->voldata.mirror_tid < chain->bref.mirror_tid) {
923 hmp->voldata.mirror_tid =
924 chain->bref.mirror_tid;
928 * The volume header is flushed manually by the
929 * syncer, not here. All we do here is adjust the
932 KKASSERT(chain->data != NULL);
933 KKASSERT(chain->dio == NULL);
935 hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT1]=
937 (char *)&hmp->voldata +
938 HAMMER2_VOLUME_ICRC1_OFF,
939 HAMMER2_VOLUME_ICRC1_SIZE);
940 hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT0]=
942 (char *)&hmp->voldata +
943 HAMMER2_VOLUME_ICRC0_OFF,
944 HAMMER2_VOLUME_ICRC0_SIZE);
945 hmp->voldata.icrc_volheader =
947 (char *)&hmp->voldata +
948 HAMMER2_VOLUME_ICRCVH_OFF,
949 HAMMER2_VOLUME_ICRCVH_SIZE);
951 if (hammer2_debug & 0x8000) {
952 /* debug only, avoid syslogd loop */
953 kprintf("syncvolhdr %016jx %016jx\n",
954 hmp->voldata.mirror_tid,
955 hmp->vchain.bref.mirror_tid);
957 hmp->volsync = hmp->voldata;
958 atomic_set_int(&chain->flags, HAMMER2_CHAIN_VOLUMESYNC);
959 hammer2_voldata_unlock(hmp);
960 hammer2_chain_unlock(&hmp->fchain);
962 case HAMMER2_BREF_TYPE_DATA:
964 * Data elements have already been flushed via the
965 * logical file buffer cache. Their hash was set in
966 * the bref by the vop_write code. Do not re-dirty.
968 * Make sure any device buffer(s) have been flushed
969 * out here (there aren't usually any to flush) XXX.
972 case HAMMER2_BREF_TYPE_INDIRECT:
973 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
974 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
976 * Buffer I/O will be cleaned up when the volume is
977 * flushed (but the kernel is free to flush it before
980 KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
981 hammer2_chain_setcheck(chain, chain->data);
983 case HAMMER2_BREF_TYPE_DIRENT:
985 * A directory entry can use the check area to store
986 * the filename for filenames <= 64 bytes, don't blow
989 KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
991 hammer2_chain_setcheck(chain, chain->data);
993 case HAMMER2_BREF_TYPE_INODE:
995 * NOTE: We must call io_setdirty() to make any late
996 * changes to the inode data, the system might
997 * have already flushed the buffer.
999 if (chain->data->ipdata.meta.op_flags &
1000 HAMMER2_OPFLAG_PFSROOT) {
1002 * non-NULL pmp if mounted as a PFS. We must
1003 * sync fields cached in the pmp? XXX
1005 hammer2_inode_data_t *ipdata;
1007 hammer2_io_setdirty(chain->dio);
1008 ipdata = &chain->data->ipdata;
1010 ipdata->meta.pfs_inum =
1011 chain->pmp->inode_tid;
1014 /* can't be mounted as a PFS */
1017 KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
1018 hammer2_chain_setcheck(chain, chain->data);
1020 hammer2_inode_data_t *ipdata;
1021 ipdata = &chain->data->ipdata;
1024 KKASSERT(chain->flags & HAMMER2_CHAIN_EMBEDDED);
1025 panic("hammer2_flush_core: unsupported "
1032 * If the chain was destroyed try to avoid unnecessary I/O
1033 * that might not have yet occurred. Remove the data range
1034 * from dedup candidacy and attempt to invalidation that
1035 * potentially dirty portion of the I/O buffer.
1037 if (chain->flags & HAMMER2_CHAIN_DESTROY) {
1038 hammer2_io_dedup_delete(hmp,
1040 chain->bref.data_off,
1045 hammer2_io_inval(chain->dio,
1046 chain->bref.data_off,
1048 } else if ((dio = hammer2_io_getquick(hmp,
1049 chain->bref.data_off,
1052 hammer2_io_inval(dio,
1053 chain->bref.data_off,
1055 hammer2_io_putblk(&dio);
1062 * If UPDATE is set the parent block table may need to be updated.
1063 * This can fail if the hammer2_chain_modify() fails.
1065 * NOTE: UPDATE may be set on vchain or fchain in which case
1066 * parent could be NULL. It's easiest to allow the case
1067 * and test for NULL. parent can also wind up being NULL
1068 * due to a deletion so we need to handle the case anyway.
1070 * If no parent exists we can just clear the UPDATE bit. If the
1071 * chain gets reattached later on the bit will simply get set
1074 if ((chain->flags & HAMMER2_CHAIN_UPDATE) && parent == NULL)
1075 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1078 * The chain may need its blockrefs updated in the parent.
1080 if (chain->flags & HAMMER2_CHAIN_UPDATE) {
1081 hammer2_blockref_t *base;
1085 * Clear UPDATE flag, mark parent modified, update its
1086 * modify_tid if necessary, and adjust the parent blockmap.
1088 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1093 * Avoid actually modifying and updating the parent if it
1094 * was flagged for destruction. This can greatly reduce
1095 * disk I/O in large tree removals because the
1096 * hammer2_io_setinval() call in the upward recursion
1097 * (see MODIFIED code above) can only handle a few cases.
1099 if (parent->flags & HAMMER2_CHAIN_DESTROY) {
1100 if (parent->bref.modify_tid < chain->bref.modify_tid) {
1101 parent->bref.modify_tid =
1102 chain->bref.modify_tid;
1104 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
1105 HAMMER2_CHAIN_BMAPUPD);
1110 * The flusher is responsible for deleting empty indirect
1111 * blocks at this point. If we don't do this, no major harm
1112 * will be done but the empty indirect blocks will stay in
1113 * the topology and make it a messy and inefficient.
1115 * The flusher is also responsible for collapsing the
1116 * content of an indirect block into its parent whenever
1117 * possible (with some hysteresis). Not doing this will also
1118 * not harm the topology, but would make it messy and
1121 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT) {
1122 if (hammer2_chain_indirect_maintenance(parent, chain))
1127 * We are updating the parent's blockmap, the parent must
1128 * be set modified. If this fails we re-set the UPDATE flag
1131 * NOTE! A modification error can be ENOSPC. We still want
1132 * to flush modified chains recursively, not break out,
1133 * so we just skip the update in this situation and
1134 * continue. That is, we still need to try to clean
1135 * out dirty chains and buffers.
1137 * This may not help bulkfree though. XXX
1139 save_error = hammer2_chain_modify(parent, 0, 0, 0);
1141 info->error |= save_error;
1142 kprintf("hammer2_flush: %016jx.%02x error=%08x\n",
1143 parent->bref.data_off, parent->bref.type,
1145 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1148 if (parent->bref.modify_tid < chain->bref.modify_tid)
1149 parent->bref.modify_tid = chain->bref.modify_tid;
1152 * Calculate blockmap pointer
1154 switch(parent->bref.type) {
1155 case HAMMER2_BREF_TYPE_INODE:
1157 * Access the inode's block array. However, there is
1158 * no block array if the inode is flagged DIRECTDATA.
1161 (parent->data->ipdata.meta.op_flags &
1162 HAMMER2_OPFLAG_DIRECTDATA) == 0) {
1163 base = &parent->data->
1164 ipdata.u.blockset.blockref[0];
1168 count = HAMMER2_SET_COUNT;
1170 case HAMMER2_BREF_TYPE_INDIRECT:
1171 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1173 base = &parent->data->npdata[0];
1176 count = parent->bytes / sizeof(hammer2_blockref_t);
1178 case HAMMER2_BREF_TYPE_VOLUME:
1179 base = &chain->hmp->voldata.sroot_blockset.blockref[0];
1180 count = HAMMER2_SET_COUNT;
1182 case HAMMER2_BREF_TYPE_FREEMAP:
1183 base = &parent->data->npdata[0];
1184 count = HAMMER2_SET_COUNT;
1189 panic("hammer2_flush_core: "
1190 "unrecognized blockref type: %d",
1195 * Blocktable updates
1197 * We synchronize pending statistics at this time. Delta
1198 * adjustments designated for the current and upper level
1201 if (base && (chain->flags & HAMMER2_CHAIN_BMAPUPD)) {
1202 if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
1203 hammer2_spin_ex(&parent->core.spin);
1204 hammer2_base_delete(parent, base, count, chain);
1205 hammer2_spin_unex(&parent->core.spin);
1206 /* base_delete clears both bits */
1208 atomic_clear_int(&chain->flags,
1209 HAMMER2_CHAIN_BMAPUPD);
1212 if (base && (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
1213 hammer2_spin_ex(&parent->core.spin);
1214 hammer2_base_insert(parent, base, count,
1215 chain, &chain->bref);
1216 hammer2_spin_unex(&parent->core.spin);
1217 /* base_insert sets BMAPPED */
1222 hammer2_chain_unlock(parent);
1225 * Final cleanup after flush
1228 KKASSERT(chain->refs > 0);
1229 if (hammer2_debug & 0x200) {
1230 if (info->debug == chain)
1236 * Flush recursion helper, called from flush_core, calls flush_core.
1238 * Flushes the children of the caller's chain (info->parent), restricted
1239 * by sync_tid. Set info->domodify if the child's blockref must propagate
1240 * back up to the parent.
1242 * This function may set info->error as a side effect.
1244 * Ripouts can move child from rbtree to dbtree or dbq but the caller's
1245 * flush scan order prevents any chains from being lost. A child can be
1246 * executes more than once.
1248 * WARNING! If we do not call hammer2_flush_core() we must update
1249 * bref.mirror_tid ourselves to indicate that the flush has
1250 * processed the child.
1252 * WARNING! parent->core spinlock is held on entry and return.
1255 hammer2_flush_recurse(hammer2_chain_t *child, void *data)
1257 hammer2_flush_info_t *info = data;
1258 hammer2_chain_t *parent = info->parent;
1260 #ifdef HAMMER2_SCAN_DEBUG
1262 if (child->flags & HAMMER2_CHAIN_MODIFIED)
1263 ++info->scan_mod_count;
1264 if (child->flags & HAMMER2_CHAIN_UPDATE)
1265 ++info->scan_upd_count;
1266 if (child->flags & HAMMER2_CHAIN_ONFLUSH)
1267 ++info->scan_onf_count;
1271 * (child can never be fchain or vchain so a special check isn't
1274 * We must ref the child before unlocking the spinlock.
1276 * The caller has added a ref to the parent so we can temporarily
1277 * unlock it in order to lock the child. However, if it no longer
1278 * winds up being the child of the parent we must skip this child.
1280 * NOTE! chain locking errors are fatal. They are never out-of-space
1283 hammer2_chain_ref(child);
1284 hammer2_spin_unex(&parent->core.spin);
1286 hammer2_chain_ref_hold(parent);
1287 hammer2_chain_unlock(parent);
1288 hammer2_chain_lock(child, HAMMER2_RESOLVE_MAYBE);
1289 if (child->parent != parent) {
1290 kprintf("LOST CHILD1 %p->%p (actual parent %p)\n",
1291 parent, child, child->parent);
1295 kprintf("CHILD ERROR DURING FLUSH LOCK %p->%p\n",
1297 info->error |= child->error;
1302 * Must propagate the DESTROY flag downwards, otherwise the
1303 * parent could end up never being removed because it will
1304 * be requeued to the flusher if it survives this run due to
1307 if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
1308 atomic_set_int(&child->flags, HAMMER2_CHAIN_DESTROY);
1309 #ifdef HAMMER2_SCAN_DEBUG
1310 if (child->flags & HAMMER2_CHAIN_DESTROY)
1311 ++info->scan_del_count;
1315 * Recurse and collect deferral data. We're in the media flush,
1316 * this can cross PFS boundaries.
1318 if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1319 #ifdef HAMMER2_SCAN_DEBUG
1320 if (child->bref.type < 7)
1321 ++info->scan_btype[child->bref.type];
1324 hammer2_flush_core(info, child, info->flags);
1326 } else if (hammer2_debug & 0x200) {
1327 if (info->debug == NULL)
1328 info->debug = child;
1330 hammer2_flush_core(info, child, info->flags);
1332 if (info->debug == child)
1338 * Relock to continue the loop.
1340 hammer2_chain_unlock(child);
1341 hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
1342 hammer2_chain_drop_unhold(parent);
1343 if (parent->error) {
1344 kprintf("PARENT ERROR DURING FLUSH LOCK %p->%p\n",
1346 info->error |= parent->error;
1348 hammer2_chain_drop(child);
1349 KKASSERT(info->parent == parent);
1350 hammer2_spin_ex(&parent->core.spin);
1356 * flush helper (backend threaded)
1358 * Flushes chain topology for the specified inode.
1360 * If HAMMER2_XOP_FLUSH is set we flush all chains from the current inode
1361 * through but stop at sub-inodes (we flush the inode chains for sub-inodes,
1362 * but do not go further as deeper modifications do not belong to the current
1365 * If HAMMER2_XOP_FLUSH is not set we flush the current inode's chains only
1366 * and do not recurse through sub-inodes, including not including those
1369 * Remember that HAMMER2 is currently using a flat inode model, so directory
1370 * hierarchies do not translate to inode hierarchies. PFS ROOTs, however,
1373 * chain->parent can be NULL, usually due to destroy races.
1375 * Primarily called from vfs_sync().
1378 hammer2_xop_inode_flush(hammer2_xop_t *arg, void *scratch __unused, int clindex)
1380 hammer2_xop_flush_t *xop = &arg->xop_flush;
1381 hammer2_chain_t *chain;
1382 hammer2_chain_t *parent;
1384 int flush_error = 0;
1385 int fsync_error = 0;
1386 int total_error = 0;
1391 xflags = HAMMER2_FLUSH_TOP;
1392 if (xop->head.flags & HAMMER2_XOP_INODE_STOP)
1393 xflags |= HAMMER2_FLUSH_INODE_STOP;
1398 chain = hammer2_inode_chain(xop->head.ip1, clindex,
1399 HAMMER2_RESOLVE_ALWAYS);
1402 if ((chain->flags & HAMMER2_CHAIN_FLUSH_MASK) ||
1403 TAILQ_FIRST(&hmp->flushq) != NULL) {
1404 hammer2_flush(chain, xflags);
1405 parent = chain->parent;
1407 hammer2_chain_setflush(parent);
1409 if (chain->flags & HAMMER2_CHAIN_PFSBOUNDARY)
1411 hammer2_chain_unlock(chain);
1412 hammer2_chain_drop(chain);
1419 * Only flush the volume header if asked to, plus the inode must also
1422 if ((xop->head.flags & HAMMER2_XOP_VOLHDR) == 0)
1428 * Flush volume roots. Avoid replication, we only want to
1429 * flush each hammer2_dev (hmp) once.
1431 for (j = clindex - 1; j >= 0; --j) {
1432 if ((chain = xop->head.ip1->cluster.array[j].chain) != NULL) {
1433 if (chain->hmp == hmp) {
1434 chain = NULL; /* safety */
1439 chain = NULL; /* safety */
1442 * spmp transaction. The super-root is never directly mounted so
1443 * there shouldn't be any vnodes, let alone any dirty vnodes
1444 * associated with it, so we shouldn't have to mess around with any
1445 * vnode flushes here.
1447 hammer2_trans_init(hmp->spmp, HAMMER2_TRANS_ISFLUSH);
1450 * Media mounts have two 'roots', vchain for the topology
1451 * and fchain for the free block table. Flush both.
1453 * Note that the topology and free block table are handled
1454 * independently, so the free block table can wind up being
1455 * ahead of the topology. We depend on the bulk free scan
1456 * code to deal with any loose ends.
1458 * vchain and fchain do not error on-lock since their data does
1459 * not have to be re-read from media.
1461 hammer2_chain_ref(&hmp->vchain);
1462 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1463 hammer2_chain_ref(&hmp->fchain);
1464 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1465 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1467 * This will also modify vchain as a side effect,
1468 * mark vchain as modified now.
1470 hammer2_voldata_modify(hmp);
1471 chain = &hmp->fchain;
1472 flush_error |= hammer2_flush(chain, HAMMER2_FLUSH_TOP);
1473 KKASSERT(chain == &hmp->fchain);
1475 hammer2_chain_unlock(&hmp->fchain);
1476 hammer2_chain_unlock(&hmp->vchain);
1477 hammer2_chain_drop(&hmp->fchain);
1478 /* vchain dropped down below */
1480 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1481 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1482 chain = &hmp->vchain;
1483 flush_error |= hammer2_flush(chain, HAMMER2_FLUSH_TOP);
1484 KKASSERT(chain == &hmp->vchain);
1486 hammer2_chain_unlock(&hmp->vchain);
1487 hammer2_chain_drop(&hmp->vchain);
1490 * We can't safely flush the volume header until we have
1491 * flushed any device buffers which have built up.
1493 * XXX this isn't being incremental
1495 vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
1496 fsync_error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
1497 vn_unlock(hmp->devvp);
1498 if (fsync_error || flush_error) {
1499 kprintf("hammer2: sync error fsync=%d h2flush=0x%04x dev=%s\n",
1500 fsync_error, flush_error, hmp->devrepname);
1504 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
1505 * volume header needs synchronization via hmp->volsync.
1507 * XXX synchronize the flag & data with only this flush XXX
1509 if (fsync_error == 0 && flush_error == 0 &&
1510 (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
1515 * Synchronize the disk before flushing the volume
1519 bp->b_bio1.bio_offset = 0;
1522 bp->b_cmd = BUF_CMD_FLUSH;
1523 bp->b_bio1.bio_done = biodone_sync;
1524 bp->b_bio1.bio_flags |= BIO_SYNC;
1525 vn_strategy(hmp->devvp, &bp->b_bio1);
1526 fsync_error = biowait(&bp->b_bio1, "h2vol");
1530 * Then we can safely flush the version of the
1531 * volume header synchronized by the flush code.
1533 j = hmp->volhdrno + 1;
1536 if (j >= HAMMER2_NUM_VOLHDRS)
1538 if (j * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
1539 hmp->volsync.volu_size) {
1542 if (hammer2_debug & 0x8000) {
1543 /* debug only, avoid syslogd loop */
1544 kprintf("sync volhdr %d %jd\n",
1545 j, (intmax_t)hmp->volsync.volu_size);
1547 bp = getblk(hmp->devvp, j * HAMMER2_ZONE_BYTES64,
1548 HAMMER2_PBUFSIZE, GETBLK_KVABIO, 0);
1549 atomic_clear_int(&hmp->vchain.flags,
1550 HAMMER2_CHAIN_VOLUMESYNC);
1552 bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
1553 vol_error = bwrite(bp);
1556 fsync_error = vol_error;
1559 total_error = flush_error;
1561 total_error = hammer2_errno_to_error(fsync_error);
1563 hammer2_trans_done(hmp->spmp, 0); /* spmp trans */
1565 hammer2_xop_feed(&xop->head, NULL, clindex, total_error);