2 * Copyright (c) 2011-2015 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 10 /* 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;
73 struct h2_flush_list flushq;
74 hammer2_chain_t *debug;
77 typedef struct hammer2_flush_info hammer2_flush_info_t;
79 static void hammer2_flush_core(hammer2_flush_info_t *info,
80 hammer2_chain_t *chain, int deleting);
81 static int hammer2_flush_recurse(hammer2_chain_t *child, void *data);
84 * Any per-pfs transaction initialization goes here.
87 hammer2_trans_manage_init(hammer2_pfs_t *pmp)
92 * Transaction support for any modifying operation. Transactions are used
93 * in the pmp layer by the frontend and in the spmp layer by the backend.
95 * 0 - Normal transaction, interlocked against flush
98 * TRANS_ISFLUSH - Flush transaction, interlocked against normal
101 * TRANS_BUFCACHE - Buffer cache transaction, no interlock.
103 * Initializing a new transaction allocates a transaction ID. Typically
104 * passed a pmp (hmp passed as NULL), indicating a cluster transaction. Can
105 * be passed a NULL pmp and non-NULL hmp to indicate a transaction on a single
106 * media target. The latter mode is used by the recovery code.
108 * TWO TRANSACTION IDs can run concurrently, where one is a flush and the
109 * other is a set of any number of concurrent filesystem operations. We
110 * can either have <running_fs_ops> + <waiting_flush> + <blocked_fs_ops>
111 * or we can have <running_flush> + <concurrent_fs_ops>.
113 * During a flush, new fs_ops are only blocked until the fs_ops prior to
114 * the flush complete. The new fs_ops can then run concurrent with the flush.
116 * Buffer-cache transactions operate as fs_ops but never block. A
117 * buffer-cache flush will run either before or after the current pending
118 * flush depending on its state.
121 hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags)
128 oflags = pmp->trans.flags;
132 if (flags & HAMMER2_TRANS_ISFLUSH) {
134 * Requesting flush transaction. Wait for all
135 * currently running transactions to finish.
137 if (oflags & HAMMER2_TRANS_MASK) {
138 nflags = oflags | HAMMER2_TRANS_FPENDING |
139 HAMMER2_TRANS_WAITING;
142 nflags = (oflags | flags) + 1;
144 } else if (flags & HAMMER2_TRANS_BUFCACHE) {
146 * Requesting strategy transaction. Generally
147 * allowed in all situations unless a flush
148 * is running without the preflush flag.
150 if ((oflags & (HAMMER2_TRANS_ISFLUSH |
151 HAMMER2_TRANS_PREFLUSH)) ==
152 HAMMER2_TRANS_ISFLUSH) {
153 nflags = oflags | HAMMER2_TRANS_WAITING;
156 nflags = (oflags | flags) + 1;
160 * Requesting normal transaction. Wait for any
161 * flush to finish before allowing.
163 if (oflags & HAMMER2_TRANS_ISFLUSH) {
164 nflags = oflags | HAMMER2_TRANS_WAITING;
167 nflags = (oflags | flags) + 1;
171 tsleep_interlock(&pmp->trans.sync_wait, 0);
172 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
175 tsleep(&pmp->trans.sync_wait, PINTERLOCKED,
185 * Start a sub-transaction, there is no 'subdone' function. This will
186 * issue a new modify_tid (mtid) for the current transaction and must
187 * be called for each XOP when multiple XOPs are run in sequence.
190 hammer2_trans_sub(hammer2_pfs_t *pmp)
194 mtid = atomic_fetchadd_64(&pmp->modify_tid, 1);
200 * Clears the PREFLUSH stage, called during a flush transaction after all
201 * logical buffer I/O has completed.
204 hammer2_trans_clear_preflush(hammer2_pfs_t *pmp)
206 atomic_clear_int(&pmp->trans.flags, HAMMER2_TRANS_PREFLUSH);
210 hammer2_trans_done(hammer2_pfs_t *pmp)
216 oflags = pmp->trans.flags;
218 KKASSERT(oflags & HAMMER2_TRANS_MASK);
219 if ((oflags & HAMMER2_TRANS_MASK) == 1) {
221 * This was the last transaction
223 nflags = (oflags - 1) & ~(HAMMER2_TRANS_ISFLUSH |
224 HAMMER2_TRANS_BUFCACHE |
225 HAMMER2_TRANS_PREFLUSH |
226 HAMMER2_TRANS_FPENDING |
227 HAMMER2_TRANS_WAITING);
230 * Still transactions pending
234 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
235 if ((nflags & HAMMER2_TRANS_MASK) == 0 &&
236 (oflags & HAMMER2_TRANS_WAITING)) {
237 wakeup(&pmp->trans.sync_wait);
248 * Obtain new, unique inode number (not serialized by caller).
251 hammer2_trans_newinum(hammer2_pfs_t *pmp)
255 tid = atomic_fetchadd_64(&pmp->inode_tid, 1);
261 * Assert that a strategy call is ok here. Strategy calls are legal
263 * (1) In a normal transaction.
264 * (2) In a flush transaction only if PREFLUSH is also set.
267 hammer2_trans_assert_strategy(hammer2_pfs_t *pmp)
269 KKASSERT((pmp->trans.flags & HAMMER2_TRANS_ISFLUSH) == 0 ||
270 (pmp->trans.flags & HAMMER2_TRANS_PREFLUSH));
275 * Chains undergoing destruction are removed from the in-memory topology.
276 * To avoid getting lost these chains are placed on the delayed flush
277 * queue which will properly dispose of them.
279 * We do this instead of issuing an immediate flush in order to give
280 * recursive deletions (rm -rf, etc) a chance to remove more of the
281 * hierarchy, potentially allowing an enormous amount of write I/O to
285 hammer2_delayed_flush(hammer2_chain_t *chain)
287 if ((chain->flags & HAMMER2_CHAIN_DELAYED) == 0) {
288 hammer2_spin_ex(&chain->hmp->list_spin);
289 if ((chain->flags & (HAMMER2_CHAIN_DELAYED |
290 HAMMER2_CHAIN_DEFERRED)) == 0) {
291 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELAYED |
292 HAMMER2_CHAIN_DEFERRED);
293 TAILQ_INSERT_TAIL(&chain->hmp->flushq,
295 hammer2_chain_ref(chain);
297 hammer2_spin_unex(&chain->hmp->list_spin);
302 * Flush the chain and all modified sub-chains through the specified
303 * synchronization point, propagating parent chain modifications, modify_tid,
304 * and mirror_tid updates back up as needed.
306 * Caller must have already vetted synchronization points to ensure they
307 * are properly flushed. Only snapshots and cluster flushes can create
308 * these sorts of synchronization points.
310 * This routine can be called from several places but the most important
313 * chain is locked on call and will remain locked on return. The chain's
314 * UPDATE flag indicates that its parent's block table (which is not yet
315 * part of the flush) should be updated. The chain may be replaced by
316 * the call if it was modified.
319 hammer2_flush(hammer2_chain_t *chain, hammer2_tid_t mtid, int istop)
321 hammer2_chain_t *scan;
322 hammer2_flush_info_t info;
327 * Execute the recursive flush and handle deferrals.
329 * Chains can be ridiculously long (thousands deep), so to
330 * avoid blowing out the kernel stack the recursive flush has a
331 * depth limit. Elements at the limit are placed on a list
332 * for re-execution after the stack has been popped.
334 bzero(&info, sizeof(info));
335 TAILQ_INIT(&info.flushq);
336 info.cache_index = -1;
340 * Calculate parent (can be NULL), if not NULL the flush core
341 * expects the parent to be referenced so it can easily lock/unlock
342 * it without it getting ripped up.
344 if ((info.parent = chain->parent) != NULL)
345 hammer2_chain_ref(info.parent);
348 * Extra ref needed because flush_core expects it when replacing
351 hammer2_chain_ref(chain);
357 * Move hmp->flushq to info.flushq if non-empty so it can
360 if (TAILQ_FIRST(&hmp->flushq) != NULL) {
361 hammer2_spin_ex(&chain->hmp->list_spin);
362 TAILQ_CONCAT(&info.flushq, &hmp->flushq, flush_node);
363 hammer2_spin_unex(&chain->hmp->list_spin);
367 * Unwind deep recursions which had been deferred. This
368 * can leave the FLUSH_* bits set for these chains, which
369 * will be handled when we [re]flush chain after the unwind.
371 while ((scan = TAILQ_FIRST(&info.flushq)) != NULL) {
372 KKASSERT(scan->flags & HAMMER2_CHAIN_DEFERRED);
373 TAILQ_REMOVE(&info.flushq, scan, flush_node);
374 atomic_clear_int(&scan->flags, HAMMER2_CHAIN_DEFERRED |
375 HAMMER2_CHAIN_DELAYED);
378 * Now that we've popped back up we can do a secondary
379 * recursion on the deferred elements.
381 * NOTE: hammer2_flush() may replace scan.
383 if (hammer2_debug & 0x0040)
384 kprintf("deferred flush %p\n", scan);
385 hammer2_chain_lock(scan, HAMMER2_RESOLVE_MAYBE);
386 hammer2_flush(scan, mtid, 0);
387 hammer2_chain_unlock(scan);
388 hammer2_chain_drop(scan); /* ref from deferral */
394 info.diddeferral = 0;
395 hammer2_flush_core(&info, chain, istop);
398 * Only loop if deep recursions have been deferred.
400 if (TAILQ_EMPTY(&info.flushq))
403 if (++loops % 1000 == 0) {
404 kprintf("hammer2_flush: excessive loops on %p\n",
406 if (hammer2_debug & 0x100000)
410 hammer2_chain_drop(chain);
412 hammer2_chain_drop(info.parent);
416 * This is the core of the chain flushing code. The chain is locked by the
417 * caller and must also have an extra ref on it by the caller, and remains
418 * locked and will have an extra ref on return. Upon return, the caller can
419 * test the UPDATE bit on the child to determine if the parent needs updating.
421 * (1) Determine if this node is a candidate for the flush, return if it is
422 * not. fchain and vchain are always candidates for the flush.
424 * (2) If we recurse too deep the chain is entered onto the deferral list and
425 * the current flush stack is aborted until after the deferral list is
428 * (3) Recursively flush live children (rbtree). This can create deferrals.
429 * A successful flush clears the MODIFIED and UPDATE bits on the children
430 * and typically causes the parent to be marked MODIFIED as the children
431 * update the parent's block table. A parent might already be marked
432 * MODIFIED due to a deletion (whos blocktable update in the parent is
433 * handled by the frontend), or if the parent itself is modified by the
434 * frontend for other reasons.
436 * (4) Permanently disconnected sub-trees are cleaned up by the front-end.
437 * Deleted-but-open inodes can still be individually flushed via the
440 * (5) Note that an unmodified child may still need the block table in its
441 * parent updated (e.g. rename/move). The child will have UPDATE set
444 * WARNING ON BREF MODIFY_TID/MIRROR_TID
446 * blockref.modify_tid is consistent only within a PFS, and will not be
447 * consistent during synchronization. mirror_tid is consistent across the
448 * block device regardless of the PFS.
451 hammer2_flush_core(hammer2_flush_info_t *info, hammer2_chain_t *chain,
454 hammer2_chain_t *parent;
459 * (1) Optimize downward recursion to locate nodes needing action.
460 * Nothing to do if none of these flags are set.
462 if ((chain->flags & HAMMER2_CHAIN_FLUSH_MASK) == 0) {
463 if (hammer2_debug & 0x200) {
464 if (info->debug == NULL)
472 diddeferral = info->diddeferral;
473 parent = info->parent; /* can be NULL */
476 * Downward search recursion
478 if (chain->flags & (HAMMER2_CHAIN_DEFERRED | HAMMER2_CHAIN_DELAYED)) {
483 } else if (info->depth == HAMMER2_FLUSH_DEPTH_LIMIT) {
485 * Recursion depth reached.
487 KKASSERT((chain->flags & HAMMER2_CHAIN_DELAYED) == 0);
488 hammer2_chain_ref(chain);
489 TAILQ_INSERT_TAIL(&info->flushq, chain, flush_node);
490 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEFERRED);
492 } else if ((chain->flags & HAMMER2_CHAIN_PFSBOUNDARY) && istop == 0) {
494 * We do not recurse through PFSROOTs. PFSROOT flushes are
495 * handled by the related pmp's (whether mounted or not,
496 * including during recovery).
498 * But we must still process the PFSROOT chains for block
499 * table updates in their parent (which IS part of our flush).
501 * Note that the volume root, vchain, does not set this flag.
504 } else if (chain->flags & HAMMER2_CHAIN_ONFLUSH) {
506 * Downward recursion search (actual flush occurs bottom-up).
507 * pre-clear ONFLUSH. It can get set again due to races,
508 * which we want so the scan finds us again in the next flush.
509 * These races can also include
511 * Flush recursions stop at PFSROOT boundaries. Each PFS
512 * must be individually flushed and then the root must
515 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
516 info->parent = chain;
517 hammer2_spin_ex(&chain->core.spin);
518 RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
519 NULL, hammer2_flush_recurse, info);
520 hammer2_spin_unex(&chain->core.spin);
521 info->parent = parent;
522 if (info->diddeferral)
523 hammer2_chain_setflush(chain);
527 * Now we are in the bottom-up part of the recursion.
529 * Do not update chain if lower layers were deferred.
531 if (info->diddeferral)
535 * Propagate the DESTROY flag downwards. This dummies up the flush
536 * code and tries to invalidate related buffer cache buffers to
537 * avoid the disk write.
539 if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
540 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
543 * Chain was already modified or has become modified, flush it out.
546 if ((hammer2_debug & 0x200) &&
548 (chain->flags & (HAMMER2_CHAIN_MODIFIED | HAMMER2_CHAIN_UPDATE))) {
549 hammer2_chain_t *scan = chain;
551 kprintf("DISCONNECTED FLUSH %p->%p\n", info->debug, chain);
553 kprintf(" chain %p [%08x] bref=%016jx:%02x\n",
555 scan->bref.key, scan->bref.type);
556 if (scan == info->debug)
562 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
564 * Dispose of the modified bit.
566 * UPDATE should already be set.
567 * bref.mirror_tid should already be set.
569 KKASSERT((chain->flags & HAMMER2_CHAIN_UPDATE) ||
570 chain == &hmp->vchain);
571 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
574 * Manage threads waiting for excessive dirty memory to
578 hammer2_pfs_memory_wakeup(chain->pmp);
580 if ((chain->flags & HAMMER2_CHAIN_UPDATE) ||
581 chain == &hmp->vchain ||
582 chain == &hmp->fchain) {
584 * Drop the ref from the MODIFIED bit we cleared,
587 hammer2_chain_drop(chain);
590 * Drop the ref from the MODIFIED bit we cleared and
591 * set a ref for the UPDATE bit we are setting. Net
594 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
598 * Issue the flush. This is indirect via the DIO.
600 * NOTE: A DELETED node that reaches this point must be
601 * flushed for synchronization point consistency.
603 * NOTE: Even though MODIFIED was already set, the related DIO
604 * might not be dirty due to a system buffer cache
605 * flush and must be set dirty if we are going to make
606 * further modifications to the buffer. Chains with
607 * embedded data don't need this.
609 if (hammer2_debug & 0x1000) {
610 kprintf("Flush %p.%d %016jx/%d data=%016jx",
611 chain, chain->bref.type,
612 (uintmax_t)chain->bref.key,
614 (uintmax_t)chain->bref.data_off);
616 if (hammer2_debug & 0x2000) {
617 Debugger("Flush hell");
621 * Update chain CRCs for flush.
623 * NOTE: Volume headers are NOT flushed here as they require
624 * special processing.
626 switch(chain->bref.type) {
627 case HAMMER2_BREF_TYPE_FREEMAP:
629 * Update the volume header's freemap_tid to the
630 * freemap's flushing mirror_tid.
632 * (note: embedded data, do not call setdirty)
634 KKASSERT(hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED);
635 KKASSERT(chain == &hmp->fchain);
636 hmp->voldata.freemap_tid = chain->bref.mirror_tid;
637 if (hammer2_debug & 0x8000) {
638 /* debug only, avoid syslogd loop */
639 kprintf("sync freemap mirror_tid %08jx\n",
640 (intmax_t)chain->bref.mirror_tid);
644 * The freemap can be flushed independently of the
645 * main topology, but for the case where it is
646 * flushed in the same transaction, and flushed
647 * before vchain (a case we want to allow for
648 * performance reasons), make sure modifications
649 * made during the flush under vchain use a new
652 * Otherwise the mount recovery code will get confused.
654 ++hmp->voldata.mirror_tid;
656 case HAMMER2_BREF_TYPE_VOLUME:
658 * The free block table is flushed by
659 * hammer2_vfs_sync() before it flushes vchain.
660 * We must still hold fchain locked while copying
661 * voldata to volsync, however.
663 * (note: embedded data, do not call setdirty)
665 hammer2_chain_lock(&hmp->fchain,
666 HAMMER2_RESOLVE_ALWAYS);
667 hammer2_voldata_lock(hmp);
668 if (hammer2_debug & 0x8000) {
669 /* debug only, avoid syslogd loop */
670 kprintf("sync volume mirror_tid %08jx\n",
671 (intmax_t)chain->bref.mirror_tid);
675 * Update the volume header's mirror_tid to the
676 * main topology's flushing mirror_tid. It is
677 * possible that voldata.mirror_tid is already
678 * beyond bref.mirror_tid due to the bump we made
679 * above in BREF_TYPE_FREEMAP.
681 if (hmp->voldata.mirror_tid < chain->bref.mirror_tid) {
682 hmp->voldata.mirror_tid =
683 chain->bref.mirror_tid;
687 * The volume header is flushed manually by the
688 * syncer, not here. All we do here is adjust the
691 KKASSERT(chain->data != NULL);
692 KKASSERT(chain->dio == NULL);
694 hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT1]=
696 (char *)&hmp->voldata +
697 HAMMER2_VOLUME_ICRC1_OFF,
698 HAMMER2_VOLUME_ICRC1_SIZE);
699 hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT0]=
701 (char *)&hmp->voldata +
702 HAMMER2_VOLUME_ICRC0_OFF,
703 HAMMER2_VOLUME_ICRC0_SIZE);
704 hmp->voldata.icrc_volheader =
706 (char *)&hmp->voldata +
707 HAMMER2_VOLUME_ICRCVH_OFF,
708 HAMMER2_VOLUME_ICRCVH_SIZE);
710 if (hammer2_debug & 0x8000) {
711 /* debug only, avoid syslogd loop */
712 kprintf("syncvolhdr %016jx %016jx\n",
713 hmp->voldata.mirror_tid,
714 hmp->vchain.bref.mirror_tid);
716 hmp->volsync = hmp->voldata;
717 atomic_set_int(&chain->flags, HAMMER2_CHAIN_VOLUMESYNC);
718 hammer2_voldata_unlock(hmp);
719 hammer2_chain_unlock(&hmp->fchain);
721 case HAMMER2_BREF_TYPE_DATA:
723 * Data elements have already been flushed via the
724 * logical file buffer cache. Their hash was set in
725 * the bref by the vop_write code. Do not re-dirty.
727 * Make sure any device buffer(s) have been flushed
728 * out here (there aren't usually any to flush) XXX.
731 case HAMMER2_BREF_TYPE_INDIRECT:
732 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
733 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
735 * Buffer I/O will be cleaned up when the volume is
736 * flushed (but the kernel is free to flush it before
739 KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
740 hammer2_chain_setcheck(chain, chain->data);
742 case HAMMER2_BREF_TYPE_INODE:
744 * NOTE: We must call io_setdirty() to make any late
745 * changes to the inode data, the system might
746 * have already flushed the buffer.
748 if (chain->data->ipdata.meta.op_flags &
749 HAMMER2_OPFLAG_PFSROOT) {
751 * non-NULL pmp if mounted as a PFS. We must
752 * sync fields cached in the pmp? XXX
754 hammer2_inode_data_t *ipdata;
756 hammer2_io_setdirty(chain->dio);
757 ipdata = &chain->data->ipdata;
759 ipdata->meta.pfs_inum =
760 chain->pmp->inode_tid;
763 /* can't be mounted as a PFS */
766 KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
767 hammer2_chain_setcheck(chain, chain->data);
770 KKASSERT(chain->flags & HAMMER2_CHAIN_EMBEDDED);
771 panic("hammer2_flush_core: unsupported "
778 * If the chain was destroyed try to avoid unnecessary I/O.
779 * (this only really works if the DIO system buffer is the
780 * same size as chain->bytes).
782 if ((chain->flags & HAMMER2_CHAIN_DESTROY) && chain->dio) {
783 hammer2_io_setinval(chain->dio, chain->bytes);
788 * If UPDATE is set the parent block table may need to be updated.
790 * NOTE: UPDATE may be set on vchain or fchain in which case
791 * parent could be NULL. It's easiest to allow the case
792 * and test for NULL. parent can also wind up being NULL
793 * due to a deletion so we need to handle the case anyway.
795 * If no parent exists we can just clear the UPDATE bit. If the
796 * chain gets reattached later on the bit will simply get set
799 if ((chain->flags & HAMMER2_CHAIN_UPDATE) && parent == NULL) {
800 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
801 hammer2_chain_drop(chain);
805 * The chain may need its blockrefs updated in the parent. This
806 * requires some fancy footwork.
808 if (chain->flags & HAMMER2_CHAIN_UPDATE) {
809 hammer2_blockref_t *base;
813 * Both parent and chain must be locked. This requires
814 * temporarily unlocking the chain. We have to deal with
815 * the case where the chain might be reparented or modified
816 * while it was unlocked.
818 hammer2_chain_unlock(chain);
819 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
820 hammer2_chain_lock(chain, HAMMER2_RESOLVE_MAYBE);
821 if (chain->parent != parent) {
822 kprintf("PARENT MISMATCH ch=%p p=%p/%p\n",
823 chain, chain->parent, parent);
824 hammer2_chain_unlock(parent);
829 * Check race condition. If someone got in and modified
830 * it again while it was unlocked, we have to loop up.
832 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
833 hammer2_chain_unlock(parent);
834 kprintf("hammer2_flush: chain %p flush-mod race\n",
840 * Clear UPDATE flag, mark parent modified, update its
841 * modify_tid if necessary, and adjust the parent blockmap.
843 if (chain->flags & HAMMER2_CHAIN_UPDATE) {
844 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
845 hammer2_chain_drop(chain);
851 * Avoid actually modifying and updating the parent if it
852 * was flagged for destruction. This can greatly reduce
853 * disk I/O in large tree removals because the
854 * hammer2_io_setinval() call in the upward recursion
855 * (see MODIFIED code above) can only handle a few cases.
857 if (parent->flags & HAMMER2_CHAIN_DESTROY) {
858 if (parent->bref.modify_tid < chain->bref.modify_tid) {
859 parent->bref.modify_tid =
860 chain->bref.modify_tid;
862 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
863 HAMMER2_CHAIN_BMAPUPD);
864 hammer2_chain_unlock(parent);
869 * We are updating the parent's blockmap, the parent must
872 hammer2_chain_modify(parent, info->mtid, 0);
873 if (parent->bref.modify_tid < chain->bref.modify_tid)
874 parent->bref.modify_tid = chain->bref.modify_tid;
877 * Calculate blockmap pointer
879 switch(parent->bref.type) {
880 case HAMMER2_BREF_TYPE_INODE:
882 * Access the inode's block array. However, there is
883 * no block array if the inode is flagged DIRECTDATA.
886 (parent->data->ipdata.meta.op_flags &
887 HAMMER2_OPFLAG_DIRECTDATA) == 0) {
888 base = &parent->data->
889 ipdata.u.blockset.blockref[0];
893 count = HAMMER2_SET_COUNT;
895 case HAMMER2_BREF_TYPE_INDIRECT:
896 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
898 base = &parent->data->npdata[0];
901 count = parent->bytes / sizeof(hammer2_blockref_t);
903 case HAMMER2_BREF_TYPE_VOLUME:
904 base = &chain->hmp->voldata.sroot_blockset.blockref[0];
905 count = HAMMER2_SET_COUNT;
907 case HAMMER2_BREF_TYPE_FREEMAP:
908 base = &parent->data->npdata[0];
909 count = HAMMER2_SET_COUNT;
914 panic("hammer2_flush_core: "
915 "unrecognized blockref type: %d",
922 * We synchronize pending statistics at this time. Delta
923 * adjustments designated for the current and upper level
926 if (base && (chain->flags & HAMMER2_CHAIN_BMAPUPD)) {
927 if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
928 hammer2_spin_ex(&parent->core.spin);
929 hammer2_base_delete(parent, base, count,
930 &info->cache_index, chain);
931 hammer2_spin_unex(&parent->core.spin);
932 /* base_delete clears both bits */
934 atomic_clear_int(&chain->flags,
935 HAMMER2_CHAIN_BMAPUPD);
938 if (base && (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
939 hammer2_spin_ex(&parent->core.spin);
940 hammer2_base_insert(parent, base, count,
941 &info->cache_index, chain);
942 hammer2_spin_unex(&parent->core.spin);
943 /* base_insert sets BMAPPED */
945 hammer2_chain_unlock(parent);
951 * Final cleanup after flush
954 KKASSERT(chain->refs > 0);
955 if (hammer2_debug & 0x200) {
956 if (info->debug == chain)
962 * Flush recursion helper, called from flush_core, calls flush_core.
964 * Flushes the children of the caller's chain (info->parent), restricted
965 * by sync_tid. Set info->domodify if the child's blockref must propagate
966 * back up to the parent.
968 * Ripouts can move child from rbtree to dbtree or dbq but the caller's
969 * flush scan order prevents any chains from being lost. A child can be
970 * executes more than once.
972 * WARNING! If we do not call hammer2_flush_core() we must update
973 * bref.mirror_tid ourselves to indicate that the flush has
974 * processed the child.
976 * WARNING! parent->core spinlock is held on entry and return.
979 hammer2_flush_recurse(hammer2_chain_t *child, void *data)
981 hammer2_flush_info_t *info = data;
982 hammer2_chain_t *parent = info->parent;
985 * (child can never be fchain or vchain so a special check isn't
988 * We must ref the child before unlocking the spinlock.
990 * The caller has added a ref to the parent so we can temporarily
991 * unlock it in order to lock the child.
993 hammer2_chain_ref(child);
994 hammer2_spin_unex(&parent->core.spin);
996 hammer2_chain_unlock(parent);
997 hammer2_chain_lock(child, HAMMER2_RESOLVE_MAYBE);
1000 * Recurse and collect deferral data. We're in the media flush,
1001 * this can cross PFS boundaries.
1003 if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1005 hammer2_flush_core(info, child, 0);
1007 } else if (hammer2_debug & 0x200) {
1008 if (info->debug == NULL)
1009 info->debug = child;
1011 hammer2_flush_core(info, child, 0);
1013 if (info->debug == child)
1018 * Relock to continue the loop
1020 hammer2_chain_unlock(child);
1021 hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
1022 hammer2_chain_drop(child);
1023 KKASSERT(info->parent == parent);
1024 hammer2_spin_ex(&parent->core.spin);
1030 * flush helper (backend threaded)
1032 * Flushes core chains, issues disk sync, flushes volume roots.
1034 * Primarily called from vfs_sync().
1037 hammer2_inode_xop_flush(hammer2_xop_t *arg, int clindex)
1039 hammer2_xop_flush_t *xop = &arg->xop_flush;
1040 hammer2_chain_t *chain;
1041 hammer2_chain_t *parent;
1044 int total_error = 0;
1050 chain = hammer2_inode_chain(xop->head.ip1, clindex,
1051 HAMMER2_RESOLVE_ALWAYS);
1054 if (chain->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1055 hammer2_flush(chain, xop->head.mtid, 1);
1056 parent = chain->parent;
1057 KKASSERT(chain->pmp != parent->pmp);
1058 hammer2_chain_setflush(parent);
1060 hammer2_chain_unlock(chain);
1061 hammer2_chain_drop(chain);
1068 * Flush volume roots. Avoid replication, we only want to
1069 * flush each hammer2_dev (hmp) once.
1071 for (j = clindex - 1; j >= 0; --j) {
1072 if ((chain = xop->head.ip1->cluster.array[j].chain) != NULL) {
1073 if (chain->hmp == hmp) {
1074 chain = NULL; /* safety */
1079 chain = NULL; /* safety */
1082 * spmp transaction. The super-root is never directly mounted so
1083 * there shouldn't be any vnodes, let alone any dirty vnodes
1084 * associated with it.
1086 hammer2_trans_init(hmp->spmp, HAMMER2_TRANS_ISFLUSH);
1089 * Media mounts have two 'roots', vchain for the topology
1090 * and fchain for the free block table. Flush both.
1092 * Note that the topology and free block table are handled
1093 * independently, so the free block table can wind up being
1094 * ahead of the topology. We depend on the bulk free scan
1095 * code to deal with any loose ends.
1097 hammer2_chain_ref(&hmp->vchain);
1098 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1099 hammer2_chain_ref(&hmp->fchain);
1100 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1101 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1103 * This will also modify vchain as a side effect,
1104 * mark vchain as modified now.
1106 hammer2_voldata_modify(hmp);
1107 chain = &hmp->fchain;
1108 hammer2_flush(chain, xop->head.mtid, 1);
1109 KKASSERT(chain == &hmp->fchain);
1111 hammer2_chain_unlock(&hmp->fchain);
1112 hammer2_chain_unlock(&hmp->vchain);
1113 hammer2_chain_drop(&hmp->fchain);
1114 /* vchain dropped down below */
1116 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1117 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1118 chain = &hmp->vchain;
1119 hammer2_flush(chain, xop->head.mtid, 1);
1120 KKASSERT(chain == &hmp->vchain);
1122 hammer2_chain_unlock(&hmp->vchain);
1123 hammer2_chain_drop(&hmp->vchain);
1128 * We can't safely flush the volume header until we have
1129 * flushed any device buffers which have built up.
1131 * XXX this isn't being incremental
1133 vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
1134 error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
1135 vn_unlock(hmp->devvp);
1138 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
1139 * volume header needs synchronization via hmp->volsync.
1141 * XXX synchronize the flag & data with only this flush XXX
1144 (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
1148 * Synchronize the disk before flushing the volume
1152 bp->b_bio1.bio_offset = 0;
1155 bp->b_cmd = BUF_CMD_FLUSH;
1156 bp->b_bio1.bio_done = biodone_sync;
1157 bp->b_bio1.bio_flags |= BIO_SYNC;
1158 vn_strategy(hmp->devvp, &bp->b_bio1);
1159 biowait(&bp->b_bio1, "h2vol");
1163 * Then we can safely flush the version of the
1164 * volume header synchronized by the flush code.
1166 j = hmp->volhdrno + 1;
1167 if (j >= HAMMER2_NUM_VOLHDRS)
1169 if (j * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
1170 hmp->volsync.volu_size) {
1173 if (hammer2_debug & 0x8000) {
1174 /* debug only, avoid syslogd loop */
1175 kprintf("sync volhdr %d %jd\n",
1176 j, (intmax_t)hmp->volsync.volu_size);
1178 bp = getblk(hmp->devvp, j * HAMMER2_ZONE_BYTES64,
1179 HAMMER2_PBUFSIZE, 0, 0);
1180 atomic_clear_int(&hmp->vchain.flags,
1181 HAMMER2_CHAIN_VOLUMESYNC);
1182 bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
1187 total_error = error;
1189 hammer2_trans_done(hmp->spmp); /* spmp trans */
1191 error = hammer2_xop_feed(&xop->head, NULL, clindex, total_error);