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>
51 #include <sys/vnode.h>
56 #define HAMMER2_FLUSH_DEPTH_LIMIT 60 /* stack recursion limit */
60 * Recursively flush the specified chain. The chain is locked and
61 * referenced by the caller and will remain so on return. The chain
62 * will remain referenced throughout but can temporarily lose its
63 * lock during the recursion to avoid unnecessarily stalling user
66 struct hammer2_flush_info {
67 hammer2_chain_t *parent;
69 int error; /* cumulative error */
71 #ifdef HAMMER2_SCAN_DEBUG
81 typedef struct hammer2_flush_info hammer2_flush_info_t;
83 static int hammer2_flush_core(hammer2_flush_info_t *info,
84 hammer2_chain_t *chain, int flags);
85 static int hammer2_flush_recurse(hammer2_chain_t *child, void *data);
88 * Any per-pfs transaction initialization goes here.
91 hammer2_trans_manage_init(hammer2_pfs_t *pmp)
96 * Transaction support for any modifying operation. Transactions are used
97 * in the pmp layer by the frontend and in the spmp layer by the backend.
99 * 0 - Normal transaction. Interlocks against just the
100 * COPYQ portion of an ISFLUSH transaction.
102 * TRANS_ISFLUSH - Flush transaction. Interlocks against other flush
105 * When COPYQ is also specified, waits for the count
108 * TRANS_BUFCACHE - Buffer cache transaction. No interlock.
110 * TRANS_SIDEQ - Run the sideq (only tested in trans_done())
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.
118 hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags)
125 oflags = pmp->trans.flags;
129 if (flags & HAMMER2_TRANS_ISFLUSH) {
131 * Interlock against other flush transactions.
133 if (oflags & HAMMER2_TRANS_ISFLUSH) {
134 nflags = oflags | HAMMER2_TRANS_WAITING;
137 nflags = (oflags | flags) + 1;
139 } else if (flags & HAMMER2_TRANS_BUFCACHE) {
141 * Requesting strategy transaction from buffer-cache,
142 * or a VM getpages/putpages through the buffer cache.
143 * We must allow such transactions in all situations
144 * to avoid deadlocks.
146 nflags = (oflags | flags) + 1;
149 * Normal transaction. We do not interlock against
150 * BUFCACHE or ISFLUSH.
152 * Note that vnode locks may be held going into
155 * NOTE: Remember that non-modifying operations
156 * such as read, stat, readdir, etc, do
157 * not use transactions.
159 nflags = (oflags | flags) + 1;
162 tsleep_interlock(&pmp->trans.sync_wait, 0);
163 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
166 tsleep(&pmp->trans.sync_wait, PINTERLOCKED,
178 * When entering a FLUSH transaction with COPYQ set, wait for the
179 * transaction count to drop to 1 (our flush transaction only)
182 * This waits for all non-flush transactions to complete and blocks
183 * new non-flush transactions from starting until COPYQ is cleared.
184 * (the flush will then proceed after clearing COPYQ). This should
185 * be a very short stall on modifying operations.
187 while ((flags & HAMMER2_TRANS_ISFLUSH) &&
188 (flags & HAMMER2_TRANS_COPYQ)) {
189 oflags = pmp->trans.flags;
191 if ((oflags & HAMMER2_TRANS_MASK) == 1)
193 nflags = oflags | HAMMER2_TRANS_WAITING;
194 tsleep_interlock(&pmp->trans.sync_wait, 0);
195 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
196 tsleep(&pmp->trans.sync_wait, PINTERLOCKED,
204 * Start a sub-transaction, there is no 'subdone' function. This will
205 * issue a new modify_tid (mtid) for the current transaction, which is a
206 * CLC (cluster level change) id and not a per-node id.
208 * This function must be called for each XOP when multiple XOPs are run in
209 * sequence within a transaction.
211 * Callers typically update the inode with the transaction mtid manually
212 * to enforce sequencing.
215 hammer2_trans_sub(hammer2_pfs_t *pmp)
219 mtid = atomic_fetchadd_64(&pmp->modify_tid, 1);
225 hammer2_trans_setflags(hammer2_pfs_t *pmp, uint32_t flags)
227 atomic_set_int(&pmp->trans.flags, flags);
231 * Typically used to clear trans flags asynchronously. If TRANS_WAITING
232 * is in the mask, and was previously set, this function will wake up
236 hammer2_trans_clearflags(hammer2_pfs_t *pmp, uint32_t flags)
242 oflags = pmp->trans.flags;
244 nflags = oflags & ~flags;
245 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
246 if ((oflags ^ nflags) & HAMMER2_TRANS_WAITING)
247 wakeup(&pmp->trans.sync_wait);
256 hammer2_trans_done(hammer2_pfs_t *pmp, uint32_t flags)
263 * Modifying ops on the front-end can cause dirty inodes to
264 * build up in the sideq. We don't flush these on inactive/reclaim
265 * due to potential deadlocks, so we have to deal with them from
266 * inside other nominal modifying front-end transactions.
268 if ((flags & HAMMER2_TRANS_SIDEQ) &&
269 pmp->sideq_count > hammer2_limit_dirty_inodes / 2 &&
270 pmp->sideq_count > (pmp->inum_count >> 3) &&
272 speedup_syncer(pmp->mp);
277 * Clean-up the transaction. Wakeup any waiters when finishing
278 * a flush transaction or transitioning the non-flush transaction
279 * count from 2->1 while a flush transaction is pending.
282 oflags = pmp->trans.flags;
284 KKASSERT(oflags & HAMMER2_TRANS_MASK);
286 nflags = (oflags - 1) & ~flags;
287 if (flags & HAMMER2_TRANS_ISFLUSH) {
288 nflags &= ~HAMMER2_TRANS_WAITING;
290 if ((oflags & (HAMMER2_TRANS_ISFLUSH|HAMMER2_TRANS_MASK)) ==
291 (HAMMER2_TRANS_ISFLUSH|2)) {
292 nflags &= ~HAMMER2_TRANS_WAITING;
294 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
295 if ((oflags ^ nflags) & HAMMER2_TRANS_WAITING)
296 wakeup(&pmp->trans.sync_wait);
305 * Obtain new, unique inode number (not serialized by caller).
308 hammer2_trans_newinum(hammer2_pfs_t *pmp)
312 tid = atomic_fetchadd_64(&pmp->inode_tid, 1);
318 * Assert that a strategy call is ok here. Currently we allow strategy
319 * calls in all situations, including during flushes. Previously:
320 * (old) (1) In a normal transaction.
323 hammer2_trans_assert_strategy(hammer2_pfs_t *pmp)
326 KKASSERT((pmp->trans.flags & HAMMER2_TRANS_ISFLUSH) == 0);
331 * Flush the chain and all modified sub-chains through the specified
332 * synchronization point, propagating blockref updates back up. As
333 * part of this propagation, mirror_tid and inode/data usage statistics
334 * propagates back upward.
336 * Returns a HAMMER2 error code, 0 if no error. Note that I/O errors from
337 * buffers dirtied during the flush operation can occur later.
339 * modify_tid (clc - cluster level change) is not propagated.
341 * update_tid (clc) is used for validation and is not propagated by this
344 * This routine can be called from several places but the most important
345 * is from VFS_SYNC (frontend) via hammer2_xop_inode_flush (backend).
347 * chain is locked on call and will remain locked on return. The chain's
348 * UPDATE flag indicates that its parent's block table (which is not yet
349 * part of the flush) should be updated.
352 * HAMMER2_FLUSH_TOP Indicates that this is the top of the flush.
353 * Is cleared for the recursion.
355 * HAMMER2_FLUSH_ALL Recurse everything
357 * HAMMER2_FLUSH_INODE_STOP
358 * Stop at PFS inode or normal inode boundary
361 hammer2_flush(hammer2_chain_t *chain, int flags)
363 hammer2_flush_info_t info;
367 * Execute the recursive flush and handle deferrals.
369 * Chains can be ridiculously long (thousands deep), so to
370 * avoid blowing out the kernel stack the recursive flush has a
371 * depth limit. Elements at the limit are placed on a list
372 * for re-execution after the stack has been popped.
374 bzero(&info, sizeof(info));
375 info.flags = flags & ~HAMMER2_FLUSH_TOP;
378 * Calculate parent (can be NULL), if not NULL the flush core
379 * expects the parent to be referenced so it can easily lock/unlock
380 * it without it getting ripped up.
382 if ((info.parent = chain->parent) != NULL)
383 hammer2_chain_ref(info.parent);
386 * Extra ref needed because flush_core expects it when replacing
389 hammer2_chain_ref(chain);
394 * [re]flush chain as the deep recursion may have generated
395 * additional modifications.
397 if (info.parent != chain->parent) {
398 if (hammer2_debug & 0x0040) {
399 kprintf("LOST CHILD4 %p->%p "
400 "(actual parent %p)\n",
401 info.parent, chain, chain->parent);
403 hammer2_chain_drop(info.parent);
404 info.parent = chain->parent;
405 hammer2_chain_ref(info.parent);
407 if (hammer2_flush_core(&info, chain, flags) == 0)
410 if (++loops % 1000 == 0) {
411 kprintf("hammer2_flush: excessive loops on %p\n",
413 if (hammer2_debug & 0x100000)
417 #ifdef HAMMER2_SCAN_DEBUG
418 if (info.scan_count >= 10)
419 kprintf("hammer2_flush: scan_count %ld (%ld,%ld,%ld,%ld) "
420 "bt(%ld,%ld,%ld,%ld,%ld,%ld)\n",
433 hammer2_chain_drop(chain);
435 hammer2_chain_drop(info.parent);
440 * This is the core of the chain flushing code. The chain is locked by the
441 * caller and must also have an extra ref on it by the caller, and remains
442 * locked and will have an extra ref on return. info.parent is referenced
445 * Upon return, the caller can test the UPDATE bit on the chain to determine
446 * if the parent needs updating.
448 * If non-zero is returned, the chain's parent changed during the flush and
449 * the caller must retry the operation.
451 * (1) Determine if this node is a candidate for the flush, return if it is
452 * not. fchain and vchain are always candidates for the flush.
454 * (2) If we recurse too deep the chain is entered onto the deferral list and
455 * the current flush stack is aborted until after the deferral list is
458 * (3) Recursively flush live children (rbtree). This can create deferrals.
459 * A successful flush clears the MODIFIED and UPDATE bits on the children
460 * and typically causes the parent to be marked MODIFIED as the children
461 * update the parent's block table. A parent might already be marked
462 * MODIFIED due to a deletion (whos blocktable update in the parent is
463 * handled by the frontend), or if the parent itself is modified by the
464 * frontend for other reasons.
466 * (4) Permanently disconnected sub-trees are cleaned up by the front-end.
467 * Deleted-but-open inodes can still be individually flushed via the
470 * (5) Delete parents on the way back up if they are normal indirect blocks
471 * and have no children.
473 * (6) Note that an unmodified child may still need the block table in its
474 * parent updated (e.g. rename/move). The child will have UPDATE set
477 * WARNING ON BREF MODIFY_TID/MIRROR_TID
479 * blockref.modify_tid is consistent only within a PFS, and will not be
480 * consistent during synchronization. mirror_tid is consistent across the
481 * block device regardless of the PFS.
484 hammer2_flush_core(hammer2_flush_info_t *info, hammer2_chain_t *chain,
487 hammer2_chain_t *parent;
495 * (1) Optimize downward recursion to locate nodes needing action.
496 * Nothing to do if none of these flags are set.
498 if ((chain->flags & HAMMER2_CHAIN_FLUSH_MASK) == 0)
504 * NOTE: parent can be NULL, usually due to destroy races.
506 parent = info->parent;
507 KKASSERT(chain->parent == parent);
510 * Downward search recursion
512 * We must be careful on cold stops, which often occur on inode
513 * boundaries due to the way hammer2_vfs_sync() sequences the flush.
514 * Be sure to issue an appropriate chain_setflush()
516 if ((chain->flags & HAMMER2_CHAIN_PFSBOUNDARY) &&
517 (flags & HAMMER2_FLUSH_ALL) == 0 &&
518 (flags & HAMMER2_FLUSH_TOP) == 0 &&
519 chain->pmp && chain->pmp->mp) {
521 * If FLUSH_ALL is not specified the caller does not want
522 * to recurse through PFS roots that have been mounted.
524 * (If the PFS has not been mounted there may not be
525 * anything monitoring its chains and its up to us
528 * The typical sequence is to flush dirty PFS's starting at
529 * their root downward, then flush the device root (vchain).
530 * It is this second flush that typically leaves out the
533 * However we must still process the PFSROOT chains for block
534 * table updates in their parent (which IS part of our flush).
536 * NOTE: The volume root, vchain, does not set PFSBOUNDARY.
538 * NOTE: We must re-set ONFLUSH in the parent to retain if
539 * this chain (that we are skipping) requires work.
541 if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
542 HAMMER2_CHAIN_DESTROY |
543 HAMMER2_CHAIN_MODIFIED)) {
544 hammer2_chain_setflush(parent);
547 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
548 (flags & HAMMER2_FLUSH_INODE_STOP) &&
549 (flags & HAMMER2_FLUSH_ALL) == 0 &&
550 (flags & HAMMER2_FLUSH_TOP) == 0 &&
551 chain->pmp && chain->pmp->mp) {
553 * When FLUSH_INODE_STOP is specified we are being asked not
554 * to include any inode changes for inodes we encounter,
555 * with the exception of the inode that the flush began with.
556 * So: INODE, INODE_STOP, and TOP==0 basically.
558 * Dirty inodes are flushed based on the hammer2_inode
559 * in-memory structure, issuing a chain_setflush() here
560 * will only cause unnecessary traversals of the topology.
565 * If FLUSH_INODE_STOP is specified and both ALL and TOP
566 * are clear, we must not flush the chain. The chain should
567 * have already been flushed and any further ONFLUSH/UPDATE
568 * setting will be related to the next flush.
570 * This features allows us to flush inodes independently of
571 * each other and meta-data above the inodes separately.
573 if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
574 HAMMER2_CHAIN_DESTROY |
575 HAMMER2_CHAIN_MODIFIED)) {
577 hammer2_chain_setflush(parent);
580 } else if (info->depth == HAMMER2_FLUSH_DEPTH_LIMIT) {
582 * Recursion depth reached.
584 panic("hammer2: flush depth limit");
585 } else if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
586 HAMMER2_CHAIN_DESTROY)) {
588 * Downward recursion search (actual flush occurs bottom-up).
589 * pre-clear ONFLUSH. It can get set again due to races or
590 * flush errors, which we want so the scan finds us again in
593 * We must also recurse if DESTROY is set so we can finally
594 * get rid of the related children, otherwise the node will
595 * just get re-flushed on lastdrop.
597 * WARNING! The recursion will unlock/relock info->parent
598 * (which is 'chain'), potentially allowing it
601 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
602 save_error = info->error;
604 info->parent = chain;
607 * We may have to do this twice to catch any indirect
608 * block maintenance that occurs.
610 hammer2_spin_ex(&chain->core.spin);
611 RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
612 NULL, hammer2_flush_recurse, info);
613 if (chain->flags & HAMMER2_CHAIN_ONFLUSH) {
614 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
615 RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
616 NULL, hammer2_flush_recurse, info);
618 hammer2_spin_unex(&chain->core.spin);
619 info->parent = parent;
622 * Re-set the flush bits if the flush was incomplete or
623 * an error occurred. If an error occurs it is typically
624 * an allocation error. Errors do not cause deferrals.
627 hammer2_chain_setflush(chain);
628 info->error |= save_error;
631 * If we lost the parent->chain association we have to
632 * stop processing this chain because it is no longer
633 * in this recursion. If it moved, it will be handled
634 * by the ONFLUSH flag elsewhere.
636 if (chain->parent != parent) {
637 kprintf("LOST CHILD2 %p->%p (actual parent %p)\n",
638 parent, chain, chain->parent);
644 * Now we are in the bottom-up part of the recursion.
646 * We continue to try to update the chain on lower-level errors, but
647 * the flush code may decide not to flush the volume root.
649 * XXX should we continue to try to update the chain if an error
654 * Both parent and chain must be locked in order to flush chain,
655 * in order to properly update the parent under certain conditions.
657 * In addition, we can't safely unlock/relock the chain once we
658 * start flushing the chain itself, which we would have to do later
659 * on in order to lock the parent if we didn't do that now.
661 hammer2_chain_ref_hold(chain);
662 hammer2_chain_unlock(chain);
664 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
665 hammer2_chain_lock(chain, HAMMER2_RESOLVE_MAYBE);
666 hammer2_chain_drop_unhold(chain);
669 * Can't process if we can't access their content.
671 if ((parent && parent->error) || chain->error) {
672 kprintf("hammer2: chain error during flush\n");
673 info->error |= chain->error;
675 info->error |= parent->error;
676 hammer2_chain_unlock(parent);
681 if (chain->parent != parent) {
682 if (hammer2_debug & 0x0040) {
683 kprintf("LOST CHILD3 %p->%p (actual parent %p)\n",
684 parent, chain, chain->parent);
686 KKASSERT(parent != NULL);
687 hammer2_chain_unlock(parent);
693 * Propagate the DESTROY flag downwards. This dummies up the flush
694 * code and tries to invalidate related buffer cache buffers to
695 * avoid the disk write.
697 if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
698 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
701 * Dispose of the modified bit.
703 * If parent is present, the UPDATE bit should already be set.
704 * UPDATE should already be set.
705 * bref.mirror_tid should already be set.
707 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
708 KKASSERT((chain->flags & HAMMER2_CHAIN_UPDATE) ||
709 chain->parent == NULL);
710 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
711 atomic_add_long(&hammer2_count_modified_chains, -1);
714 * Manage threads waiting for excessive dirty memory to
718 hammer2_pfs_memory_wakeup(chain->pmp, -1);
721 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0 &&
722 chain != &hmp->vchain &&
723 chain != &hmp->fchain) {
725 * Set UPDATE bit indicating that the parent block
726 * table requires updating.
728 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
733 * Issue the flush. This is indirect via the DIO.
735 * NOTE: A DELETED node that reaches this point must be
736 * flushed for synchronization point consistency.
738 * NOTE: Even though MODIFIED was already set, the related DIO
739 * might not be dirty due to a system buffer cache
740 * flush and must be set dirty if we are going to make
741 * further modifications to the buffer. Chains with
742 * embedded data don't need this.
744 if (hammer2_debug & 0x1000) {
745 kprintf("Flush %p.%d %016jx/%d data=%016jx\n",
746 chain, chain->bref.type,
747 (uintmax_t)chain->bref.key,
749 (uintmax_t)chain->bref.data_off);
753 * Update chain CRCs for flush.
755 * NOTE: Volume headers are NOT flushed here as they require
756 * special processing.
758 switch(chain->bref.type) {
759 case HAMMER2_BREF_TYPE_FREEMAP:
761 * Update the volume header's freemap_tid to the
762 * freemap's flushing mirror_tid.
764 * (note: embedded data, do not call setdirty)
766 KKASSERT(hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED);
767 KKASSERT(chain == &hmp->fchain);
768 hmp->voldata.freemap_tid = chain->bref.mirror_tid;
769 if (hammer2_debug & 0x8000) {
770 /* debug only, avoid syslogd loop */
771 kprintf("sync freemap mirror_tid %08jx\n",
772 (intmax_t)chain->bref.mirror_tid);
776 * The freemap can be flushed independently of the
777 * main topology, but for the case where it is
778 * flushed in the same transaction, and flushed
779 * before vchain (a case we want to allow for
780 * performance reasons), make sure modifications
781 * made during the flush under vchain use a new
784 * Otherwise the mount recovery code will get confused.
786 ++hmp->voldata.mirror_tid;
788 case HAMMER2_BREF_TYPE_VOLUME:
790 * The free block table is flushed by
791 * hammer2_vfs_sync() before it flushes vchain.
792 * We must still hold fchain locked while copying
793 * voldata to volsync, however.
795 * These do not error per-say since their data does
796 * not need to be re-read from media on lock.
798 * (note: embedded data, do not call setdirty)
800 hammer2_chain_lock(&hmp->fchain,
801 HAMMER2_RESOLVE_ALWAYS);
802 hammer2_voldata_lock(hmp);
803 if (hammer2_debug & 0x8000) {
804 /* debug only, avoid syslogd loop */
805 kprintf("sync volume mirror_tid %08jx\n",
806 (intmax_t)chain->bref.mirror_tid);
810 * Update the volume header's mirror_tid to the
811 * main topology's flushing mirror_tid. It is
812 * possible that voldata.mirror_tid is already
813 * beyond bref.mirror_tid due to the bump we made
814 * above in BREF_TYPE_FREEMAP.
816 if (hmp->voldata.mirror_tid < chain->bref.mirror_tid) {
817 hmp->voldata.mirror_tid =
818 chain->bref.mirror_tid;
822 * The volume header is flushed manually by the
823 * syncer, not here. All we do here is adjust the
826 KKASSERT(chain->data != NULL);
827 KKASSERT(chain->dio == NULL);
829 hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT1]=
831 (char *)&hmp->voldata +
832 HAMMER2_VOLUME_ICRC1_OFF,
833 HAMMER2_VOLUME_ICRC1_SIZE);
834 hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT0]=
836 (char *)&hmp->voldata +
837 HAMMER2_VOLUME_ICRC0_OFF,
838 HAMMER2_VOLUME_ICRC0_SIZE);
839 hmp->voldata.icrc_volheader =
841 (char *)&hmp->voldata +
842 HAMMER2_VOLUME_ICRCVH_OFF,
843 HAMMER2_VOLUME_ICRCVH_SIZE);
845 if (hammer2_debug & 0x8000) {
846 /* debug only, avoid syslogd loop */
847 kprintf("syncvolhdr %016jx %016jx\n",
848 hmp->voldata.mirror_tid,
849 hmp->vchain.bref.mirror_tid);
851 hmp->volsync = hmp->voldata;
852 atomic_set_int(&chain->flags, HAMMER2_CHAIN_VOLUMESYNC);
853 hammer2_voldata_unlock(hmp);
854 hammer2_chain_unlock(&hmp->fchain);
856 case HAMMER2_BREF_TYPE_DATA:
858 * Data elements have already been flushed via the
859 * logical file buffer cache. Their hash was set in
860 * the bref by the vop_write code. Do not re-dirty.
862 * Make sure any device buffer(s) have been flushed
863 * out here (there aren't usually any to flush) XXX.
866 case HAMMER2_BREF_TYPE_INDIRECT:
867 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
868 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
870 * Buffer I/O will be cleaned up when the volume is
871 * flushed (but the kernel is free to flush it before
874 hammer2_chain_setcheck(chain, chain->data);
876 case HAMMER2_BREF_TYPE_DIRENT:
878 * A directory entry can use the check area to store
879 * the filename for filenames <= 64 bytes, don't blow
883 hammer2_chain_setcheck(chain, chain->data);
885 case HAMMER2_BREF_TYPE_INODE:
887 * NOTE: We must call io_setdirty() to make any late
888 * changes to the inode data, the system might
889 * have already flushed the buffer.
891 if (chain->data->ipdata.meta.op_flags &
892 HAMMER2_OPFLAG_PFSROOT) {
894 * non-NULL pmp if mounted as a PFS. We must
895 * sync fields cached in the pmp? XXX
897 hammer2_inode_data_t *ipdata;
899 hammer2_io_setdirty(chain->dio);
900 ipdata = &chain->data->ipdata;
902 ipdata->meta.pfs_inum =
903 chain->pmp->inode_tid;
906 /* can't be mounted as a PFS */
909 hammer2_chain_setcheck(chain, chain->data);
912 panic("hammer2_flush_core: unsupported "
919 * If the chain was destroyed try to avoid unnecessary I/O
920 * that might not have yet occurred. Remove the data range
921 * from dedup candidacy and attempt to invalidation that
922 * potentially dirty portion of the I/O buffer.
924 if (chain->flags & HAMMER2_CHAIN_DESTROY) {
925 hammer2_io_dedup_delete(hmp,
927 chain->bref.data_off,
932 hammer2_io_inval(chain->dio,
933 chain->bref.data_off,
935 } else if ((dio = hammer2_io_getquick(hmp,
936 chain->bref.data_off,
939 hammer2_io_inval(dio,
940 chain->bref.data_off,
942 hammer2_io_putblk(&dio);
949 * If UPDATE is set the parent block table may need to be updated.
950 * This can fail if the hammer2_chain_modify() fails.
952 * NOTE: UPDATE may be set on vchain or fchain in which case
953 * parent could be NULL, or on an inode that has not yet
954 * been inserted into the radix tree. It's easiest to allow
955 * the case and test for NULL. parent can also wind up being
956 * NULL due to a deletion so we need to handle the case anyway.
958 * NOTE: UPDATE can be set when chains are renamed into or out of
959 * an indirect block, without the chain itself being flagged
962 * If no parent exists we can just clear the UPDATE bit. If the
963 * chain gets reattached later on the bit will simply get set
966 if ((chain->flags & HAMMER2_CHAIN_UPDATE) && parent == NULL)
967 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
970 * When flushing an inode outside of a FLUSH_FSSYNC we must NOT
971 * update the parent block table to point at the flushed inode.
972 * The block table should only ever be updated by the filesystem
973 * sync code. If we do, inode<->inode dependencies (such as
974 * directory entries vs inode nlink count) can wind up not being
975 * flushed together and result in a broken topology if a crash/reboot
976 * occurs at the wrong time.
978 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
979 (flags & HAMMER2_FLUSH_INODE_STOP) &&
980 (flags & HAMMER2_FLUSH_FSSYNC) == 0 &&
981 (flags & HAMMER2_FLUSH_ALL) == 0 &&
982 chain->pmp && chain->pmp->mp) {
983 #ifdef HAMMER2_DEBUG_SYNC
984 kprintf("inum %ld do not update parent, non-fssync\n",
985 (long)chain->bref.key);
989 #ifdef HAMMER2_DEBUG_SYNC
990 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
991 kprintf("inum %ld update parent\n", (long)chain->bref.key);
995 * The chain may need its blockrefs updated in the parent, normal
998 if (chain->flags & HAMMER2_CHAIN_UPDATE) {
999 hammer2_blockref_t *base;
1003 * Clear UPDATE flag, mark parent modified, update its
1004 * modify_tid if necessary, and adjust the parent blockmap.
1006 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1011 * Avoid actually modifying and updating the parent if it
1012 * was flagged for destruction. This can greatly reduce
1013 * disk I/O in large tree removals because the
1014 * hammer2_io_setinval() call in the upward recursion
1015 * (see MODIFIED code above) can only handle a few cases.
1017 if (parent->flags & HAMMER2_CHAIN_DESTROY) {
1018 if (parent->bref.modify_tid < chain->bref.modify_tid) {
1019 parent->bref.modify_tid =
1020 chain->bref.modify_tid;
1022 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BLKMAPPED |
1023 HAMMER2_CHAIN_BLKMAPUPD);
1028 * The flusher is responsible for deleting empty indirect
1029 * blocks at this point. If we don't do this, no major harm
1030 * will be done but the empty indirect blocks will stay in
1031 * the topology and make it a messy and inefficient.
1033 * The flusher is also responsible for collapsing the
1034 * content of an indirect block into its parent whenever
1035 * possible (with some hysteresis). Not doing this will also
1036 * not harm the topology, but would make it messy and
1039 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT) {
1040 if (hammer2_chain_indirect_maintenance(parent, chain))
1045 * We are updating the parent's blockmap, the parent must
1046 * be set modified. If this fails we re-set the UPDATE flag
1049 * NOTE! A modification error can be ENOSPC. We still want
1050 * to flush modified chains recursively, not break out,
1051 * so we just skip the update in this situation and
1052 * continue. That is, we still need to try to clean
1053 * out dirty chains and buffers.
1055 * This may not help bulkfree though. XXX
1057 save_error = hammer2_chain_modify(parent, 0, 0, 0);
1059 info->error |= save_error;
1060 kprintf("hammer2_flush: %016jx.%02x error=%08x\n",
1061 parent->bref.data_off, parent->bref.type,
1063 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1066 if (parent->bref.modify_tid < chain->bref.modify_tid)
1067 parent->bref.modify_tid = chain->bref.modify_tid;
1070 * Calculate blockmap pointer
1072 switch(parent->bref.type) {
1073 case HAMMER2_BREF_TYPE_INODE:
1075 * Access the inode's block array. However, there is
1076 * no block array if the inode is flagged DIRECTDATA.
1079 (parent->data->ipdata.meta.op_flags &
1080 HAMMER2_OPFLAG_DIRECTDATA) == 0) {
1081 base = &parent->data->
1082 ipdata.u.blockset.blockref[0];
1086 count = HAMMER2_SET_COUNT;
1088 case HAMMER2_BREF_TYPE_INDIRECT:
1089 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1091 base = &parent->data->npdata[0];
1094 count = parent->bytes / sizeof(hammer2_blockref_t);
1096 case HAMMER2_BREF_TYPE_VOLUME:
1097 base = &chain->hmp->voldata.sroot_blockset.blockref[0];
1098 count = HAMMER2_SET_COUNT;
1100 case HAMMER2_BREF_TYPE_FREEMAP:
1101 base = &parent->data->npdata[0];
1102 count = HAMMER2_SET_COUNT;
1107 panic("hammer2_flush_core: "
1108 "unrecognized blockref type: %d",
1114 * Blocktable updates
1116 * We synchronize pending statistics at this time. Delta
1117 * adjustments designated for the current and upper level
1120 if (base && (chain->flags & HAMMER2_CHAIN_BLKMAPUPD)) {
1121 if (chain->flags & HAMMER2_CHAIN_BLKMAPPED) {
1122 hammer2_spin_ex(&parent->core.spin);
1123 hammer2_base_delete(parent, base, count, chain,
1125 hammer2_spin_unex(&parent->core.spin);
1126 /* base_delete clears both bits */
1128 atomic_clear_int(&chain->flags,
1129 HAMMER2_CHAIN_BLKMAPUPD);
1132 if (base && (chain->flags & HAMMER2_CHAIN_BLKMAPPED) == 0) {
1133 hammer2_spin_ex(&parent->core.spin);
1134 hammer2_base_insert(parent, base, count,
1135 chain, &chain->bref);
1136 hammer2_spin_unex(&parent->core.spin);
1137 /* base_insert sets BLKMAPPED */
1142 hammer2_chain_unlock(parent);
1145 * Final cleanup after flush
1148 KKASSERT(chain->refs > 0);
1154 * Flush recursion helper, called from flush_core, calls flush_core.
1156 * Flushes the children of the caller's chain (info->parent), restricted
1159 * This function may set info->error as a side effect.
1161 * WARNING! If we do not call hammer2_flush_core() we must update
1162 * bref.mirror_tid ourselves to indicate that the flush has
1163 * processed the child.
1165 * WARNING! parent->core spinlock is held on entry and return.
1168 hammer2_flush_recurse(hammer2_chain_t *child, void *data)
1170 hammer2_flush_info_t *info = data;
1171 hammer2_chain_t *parent = info->parent;
1173 #ifdef HAMMER2_SCAN_DEBUG
1175 if (child->flags & HAMMER2_CHAIN_MODIFIED)
1176 ++info->scan_mod_count;
1177 if (child->flags & HAMMER2_CHAIN_UPDATE)
1178 ++info->scan_upd_count;
1179 if (child->flags & HAMMER2_CHAIN_ONFLUSH)
1180 ++info->scan_onf_count;
1184 * (child can never be fchain or vchain so a special check isn't
1187 * We must ref the child before unlocking the spinlock.
1189 * The caller has added a ref to the parent so we can temporarily
1190 * unlock it in order to lock the child. However, if it no longer
1191 * winds up being the child of the parent we must skip this child.
1193 * NOTE! chain locking errors are fatal. They are never out-of-space
1196 hammer2_chain_ref(child);
1197 hammer2_spin_unex(&parent->core.spin);
1199 hammer2_chain_ref_hold(parent);
1200 hammer2_chain_unlock(parent);
1201 hammer2_chain_lock(child, HAMMER2_RESOLVE_MAYBE);
1202 if (child->parent != parent) {
1203 kprintf("LOST CHILD1 %p->%p (actual parent %p)\n",
1204 parent, child, child->parent);
1208 kprintf("CHILD ERROR DURING FLUSH LOCK %p->%p\n",
1210 info->error |= child->error;
1215 * Must propagate the DESTROY flag downwards, otherwise the
1216 * parent could end up never being removed because it will
1217 * be requeued to the flusher if it survives this run due to
1220 if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
1221 atomic_set_int(&child->flags, HAMMER2_CHAIN_DESTROY);
1222 #ifdef HAMMER2_SCAN_DEBUG
1223 if (child->flags & HAMMER2_CHAIN_DESTROY)
1224 ++info->scan_del_count;
1227 * Special handling of the root inode. Because the root inode
1228 * contains an index of all the inodes in the PFS in addition to
1229 * its normal directory entries, any flush that is not part of a
1230 * filesystem sync must only flush the directory entries, and not
1233 * The child might be an indirect block, but H2 guarantees that
1234 * the key-range will fully partition the inode index from the
1235 * directory entries so the case just works naturally.
1237 if ((parent->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
1238 (child->flags & HAMMER2_CHAIN_DESTROY) == 0 &&
1239 parent->bref.type == HAMMER2_BREF_TYPE_INODE &&
1240 (info->flags & HAMMER2_FLUSH_FSSYNC) == 0) {
1241 if ((child->bref.key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1242 if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1243 hammer2_chain_setflush(parent);
1250 * Recurse and collect deferral data. We're in the media flush,
1251 * this can cross PFS boundaries.
1253 if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1254 #ifdef HAMMER2_SCAN_DEBUG
1255 if (child->bref.type < 7)
1256 ++info->scan_btype[child->bref.type];
1259 hammer2_flush_core(info, child, info->flags);
1265 * Relock to continue the loop.
1267 hammer2_chain_unlock(child);
1268 hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
1269 hammer2_chain_drop_unhold(parent);
1270 if (parent->error) {
1271 kprintf("PARENT ERROR DURING FLUSH LOCK %p->%p\n",
1273 info->error |= parent->error;
1275 hammer2_chain_drop(child);
1276 KKASSERT(info->parent == parent);
1277 hammer2_spin_ex(&parent->core.spin);
1283 * flush helper (backend threaded)
1285 * Flushes chain topology for the specified inode.
1287 * HAMMER2_XOP_INODE_STOP The flush recursion stops at inode boundaries.
1288 * Inodes belonging to the same flush are flushed
1291 * chain->parent can be NULL, usually due to destroy races or detached inodes.
1293 * Primarily called from vfs_sync().
1296 hammer2_xop_inode_flush(hammer2_xop_t *arg, void *scratch __unused, int clindex)
1298 hammer2_xop_flush_t *xop = &arg->xop_flush;
1299 hammer2_chain_t *chain;
1300 hammer2_inode_t *ip;
1304 struct vnode *devvp;
1305 int flush_error = 0;
1306 int fsync_error = 0;
1307 int total_error = 0;
1312 xflags = HAMMER2_FLUSH_TOP;
1313 if (xop->head.flags & HAMMER2_XOP_INODE_STOP)
1314 xflags |= HAMMER2_FLUSH_INODE_STOP;
1315 if (xop->head.flags & HAMMER2_XOP_FSSYNC)
1316 xflags |= HAMMER2_FLUSH_FSSYNC;
1323 chain = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
1326 if (chain->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1328 * Due to flush partitioning the chain topology
1329 * above the inode's chain may no longer be flagged.
1330 * When asked to flush an inode, remark the topology
1331 * leading to that inode.
1334 hammer2_chain_setflush(chain->parent);
1335 hammer2_flush(chain, xflags);
1338 if (ip == pmp->iroot && pmp != hmp->spmp) {
1339 hammer2_spin_ex(&pmp->inum_spin);
1340 pmp->pfs_iroot_blocksets[clindex] =
1341 chain->data->ipdata.u.blockset;
1342 hammer2_spin_unex(&pmp->inum_spin);
1347 * Propogate upwards but only cross an inode boundary
1348 * for inodes associated with the current filesystem
1351 if ((xop->head.flags & HAMMER2_XOP_PARENTONFLUSH) ||
1352 chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
1353 parent = chain->parent;
1355 hammer2_chain_setflush(parent);
1359 if (chain->flags & HAMMER2_CHAIN_PFSBOUNDARY)
1361 hammer2_chain_unlock(chain);
1362 hammer2_chain_drop(chain);
1369 * Only flush the volume header if asked to, plus the inode must also
1372 if ((xop->head.flags & HAMMER2_XOP_VOLHDR) == 0)
1378 * Flush volume roots. Avoid replication, we only want to
1379 * flush each hammer2_dev (hmp) once.
1381 for (j = clindex - 1; j >= 0; --j) {
1382 if ((chain = ip->cluster.array[j].chain) != NULL) {
1383 if (chain->hmp == hmp) {
1384 chain = NULL; /* safety */
1389 chain = NULL; /* safety */
1392 * spmp transaction. The super-root is never directly mounted so
1393 * there shouldn't be any vnodes, let alone any dirty vnodes
1394 * associated with it, so we shouldn't have to mess around with any
1395 * vnode flushes here.
1397 hammer2_trans_init(hmp->spmp, HAMMER2_TRANS_ISFLUSH);
1400 * We must flush the superroot down to the PFS iroot. Remember
1401 * that hammer2_chain_setflush() stops at inode boundaries, so
1402 * the pmp->iroot has been flushed and flagged down to the superroot,
1403 * but the volume root (vchain) probably has not yet been flagged.
1405 if (hmp->spmp->iroot) {
1406 chain = hmp->spmp->iroot->cluster.array[0].chain;
1408 hammer2_chain_ref(chain);
1409 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
1411 hammer2_flush(chain,
1413 HAMMER2_FLUSH_INODE_STOP |
1414 HAMMER2_FLUSH_FSSYNC);
1415 hammer2_chain_unlock(chain);
1416 hammer2_chain_drop(chain);
1421 * Media mounts have two 'roots', vchain for the topology
1422 * and fchain for the free block table. Flush both.
1424 * Note that the topology and free block table are handled
1425 * independently, so the free block table can wind up being
1426 * ahead of the topology. We depend on the bulk free scan
1427 * code to deal with any loose ends.
1429 * vchain and fchain do not error on-lock since their data does
1430 * not have to be re-read from media.
1432 hammer2_chain_ref(&hmp->vchain);
1433 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1434 hammer2_chain_ref(&hmp->fchain);
1435 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1436 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1438 * This will also modify vchain as a side effect,
1439 * mark vchain as modified now.
1441 hammer2_voldata_modify(hmp);
1442 chain = &hmp->fchain;
1443 flush_error |= hammer2_flush(chain, HAMMER2_FLUSH_TOP);
1444 KKASSERT(chain == &hmp->fchain);
1446 hammer2_chain_unlock(&hmp->fchain);
1447 hammer2_chain_unlock(&hmp->vchain);
1448 hammer2_chain_drop(&hmp->fchain);
1449 /* vchain dropped down below */
1451 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1452 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1453 chain = &hmp->vchain;
1454 flush_error |= hammer2_flush(chain, HAMMER2_FLUSH_TOP);
1455 KKASSERT(chain == &hmp->vchain);
1457 hammer2_chain_unlock(&hmp->vchain);
1458 hammer2_chain_drop(&hmp->vchain);
1461 * We can't safely flush the volume header until we have
1462 * flushed any device buffers which have built up.
1464 * XXX this isn't being incremental
1466 TAILQ_FOREACH(e, &hmp->devvpl, entry) {
1469 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1470 fsync_error = VOP_FSYNC(devvp, MNT_WAIT, 0);
1472 if (fsync_error || flush_error) {
1473 kprintf("hammer2: sync error fsync=%d h2flush=0x%04x dev=%s\n",
1474 fsync_error, flush_error, e->path);
1479 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
1480 * volume header needs synchronization via hmp->volsync.
1482 * XXX synchronize the flag & data with only this flush XXX
1484 if (fsync_error == 0 && flush_error == 0 &&
1485 (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
1490 * Synchronize the disk before flushing the volume
1494 bp->b_bio1.bio_offset = 0;
1497 bp->b_cmd = BUF_CMD_FLUSH;
1498 bp->b_bio1.bio_done = biodone_sync;
1499 bp->b_bio1.bio_flags |= BIO_SYNC;
1500 vn_strategy(hmp->devvp, &bp->b_bio1);
1501 fsync_error = biowait(&bp->b_bio1, "h2vol");
1505 * Then we can safely flush the version of the
1506 * volume header synchronized by the flush code.
1508 j = hmp->volhdrno + 1;
1511 if (j >= HAMMER2_NUM_VOLHDRS)
1513 if (j * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
1514 hmp->volsync.volu_size) {
1517 if (hammer2_debug & 0x8000) {
1518 /* debug only, avoid syslogd loop */
1519 kprintf("sync volhdr %d %jd\n",
1520 j, (intmax_t)hmp->volsync.volu_size);
1522 bp = getblk(hmp->devvp, j * HAMMER2_ZONE_BYTES64,
1523 HAMMER2_PBUFSIZE, GETBLK_KVABIO, 0);
1524 atomic_clear_int(&hmp->vchain.flags,
1525 HAMMER2_CHAIN_VOLUMESYNC);
1527 bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
1528 vol_error = bwrite(bp);
1531 fsync_error = vol_error;
1534 total_error = flush_error;
1536 total_error = hammer2_errno_to_error(fsync_error);
1539 hammer2_trans_done(hmp->spmp, HAMMER2_TRANS_ISFLUSH);
1541 hammer2_xop_feed(&xop->head, NULL, clindex, total_error);