2 * Copyright (c) 2011-2013 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 #include <sys/cdefs.h>
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/types.h>
46 * Recursively flush the specified chain. The chain is locked and
47 * referenced by the caller and will remain so on return. The chain
48 * will remain referenced throughout but can temporarily lose its
49 * lock during the recursion to avoid unnecessarily stalling user
52 struct hammer2_flush_info {
53 hammer2_chain_t *parent;
54 hammer2_trans_t *trans;
59 struct h2_flush_deferral_list flush_list;
60 hammer2_tid_t sync_tid; /* flush synchronization point */
61 hammer2_tid_t mirror_tid; /* collect mirror TID updates */
64 typedef struct hammer2_flush_info hammer2_flush_info_t;
66 static void hammer2_chain_flush_core(hammer2_flush_info_t *info,
67 hammer2_chain_t *chain);
68 static int hammer2_chain_flush_scan1(hammer2_chain_t *child, void *data);
69 static int hammer2_chain_flush_scan2(hammer2_chain_t *child, void *data);
70 static void hammer2_rollup_stats(hammer2_chain_t *parent,
71 hammer2_chain_t *child, int how);
76 hammer2_updatestats(hammer2_flush_info_t *info, hammer2_blockref_t *bref,
81 if (bref->type != 0) {
82 bytes = 1 << (bref->data_off & HAMMER2_OFF_MASK_RADIX);
83 if (bref->type == HAMMER2_BREF_TYPE_INODE)
84 info->inode_count += how;
86 info->data_count -= bytes;
88 info->data_count += bytes;
94 * Transaction support functions for writing to the filesystem.
96 * Initializing a new transaction allocates a transaction ID. We
97 * don't bother marking the volume header MODIFIED. Instead, the volume
98 * will be synchronized at a later time as part of a larger flush sequence.
100 * Non-flush transactions can typically run concurrently. However if
101 * there are non-flush transaction both before AND after a flush trans,
102 * the transactions after stall until the ones before finish.
104 * Non-flush transactions occuring after a flush pointer can run concurrently
105 * with that flush. They only have to wait for transactions prior to the
106 * flush trans to complete before they unstall.
108 * WARNING! Transaction ids are only allocated when the transaction becomes
109 * active, which allows other transactions to insert ahead of us
110 * if we are forced to block (only bioq transactions do that).
112 * WARNING! Modifications to the root volume cannot dup the root volume
113 * header to handle synchronization points, so alloc_tid can
114 * wind up (harmlessly) more advanced on flush.
116 * WARNING! Operations which might call inode_duplicate()/chain_duplicate()
117 * depend heavily on having a unique sync_tid to avoid duplication
118 * collisions (which key off of delete_tid).
121 hammer2_trans_init(hammer2_trans_t *trans, hammer2_pfsmount_t *pmp, int flags)
123 hammer2_cluster_t *cluster;
124 hammer2_mount_t *hmp;
125 hammer2_trans_t *scan;
127 bzero(trans, sizeof(*trans));
129 cluster = pmp->cluster;
132 hammer2_voldata_lock(hmp);
133 trans->flags = flags;
134 trans->td = curthread;
135 /*trans->delete_gen = 0;*/ /* multiple deletions within trans */
137 if (flags & HAMMER2_TRANS_ISFLUSH) {
139 * If multiple flushes are trying to run we have to
140 * wait until it is our turn, then set curflush to
141 * indicate that a flush is now pending (but not
142 * necessarily active yet).
144 * NOTE: Do not set trans->blocked here.
147 while (hmp->curflush != NULL) {
148 lksleep(&hmp->curflush, &hmp->voldatalk,
151 hmp->curflush = trans;
152 TAILQ_INSERT_TAIL(&hmp->transq, trans, entry);
155 * If we are a flush we have to wait for all transactions
156 * prior to our flush synchronization point to complete
157 * before we can start our flush.
159 * Most importantly, this includes bioq flushes.
161 * NOTE: Do not set trans->blocked here.
163 while (TAILQ_FIRST(&hmp->transq) != trans) {
164 lksleep(&trans->sync_tid, &hmp->voldatalk,
169 * don't assign sync_tid until we become the running
170 * flush. topo_flush_tid is used to control when
171 * chain modifications in concurrent transactions are
172 * required to delete-duplicate (so as not to disturb
173 * the state of what is being currently flushed).
175 trans->sync_tid = hmp->voldata.alloc_tid++;
176 hmp->topo_flush_tid = trans->sync_tid;
179 * Once we become the running flush we can wakeup anyone
180 * who blocked on us, up to the next flush. That is,
181 * our flush can run concurrent with frontend operations.
184 while ((scan = TAILQ_NEXT(scan, entry)) != NULL) {
185 if (scan->flags & HAMMER2_TRANS_ISFLUSH)
187 if (scan->blocked == 0)
190 wakeup(&scan->blocked);
192 } else if ((flags & HAMMER2_TRANS_BUFCACHE) && hmp->curflush) {
194 * We cannot block if we are the bioq thread. When a
195 * flush is not pending we can operate normally but
196 * if a flush IS pending the bioq thread's transaction
197 * must be placed either before or after curflush.
199 * If the current flush is waiting the bioq thread's
200 * transaction is placed before. If it is running the
201 * bioq thread's transaction is placed after.
203 scan = TAILQ_FIRST(&hmp->transq);
204 if (scan != hmp->curflush) {
205 TAILQ_INSERT_BEFORE(hmp->curflush, trans, entry);
207 TAILQ_INSERT_TAIL(&hmp->transq, trans, entry);
209 trans->sync_tid = hmp->voldata.alloc_tid++;
212 * If this is a normal transaction and not a flush, or
213 * if this is a bioq transaction and no flush is pending,
214 * we can queue normally.
216 * Normal transactions must block while a pending flush is
217 * waiting for prior transactions to complete. Once the
218 * pending flush becomes active we can run concurrently
221 TAILQ_INSERT_TAIL(&hmp->transq, trans, entry);
222 scan = TAILQ_FIRST(&hmp->transq);
223 if (hmp->curflush && hmp->curflush != scan) {
225 while (trans->blocked) {
226 lksleep(&trans->blocked, &hmp->voldatalk,
230 trans->sync_tid = hmp->voldata.alloc_tid++;
232 hammer2_voldata_unlock(hmp, 0);
236 hammer2_trans_done(hammer2_trans_t *trans)
238 hammer2_cluster_t *cluster;
239 hammer2_mount_t *hmp;
240 hammer2_trans_t *scan;
242 cluster = trans->pmp->cluster;
245 hammer2_voldata_lock(hmp);
246 TAILQ_REMOVE(&hmp->transq, trans, entry);
247 if (trans->flags & HAMMER2_TRANS_ISFLUSH) {
251 * If we were a flush then wakeup anyone waiting on
252 * curflush (i.e. other flushes that want to run).
253 * Leave topo_flush_id set (I think we could probably
254 * clear it to zero here).
256 hmp->curflush = NULL;
257 wakeup(&hmp->curflush);
260 * Theoretically we don't have to clear flush_tid
261 * here since the flush will have synchronized
262 * all operations <= flush_tid already. But for
265 hmp->curflush = NULL;
266 hmp->topo_flush_tid = 0;
270 * If we are not a flush but a flush is now at the head
271 * of the queue and we were previously blocking it,
272 * we can now unblock it.
275 (scan = TAILQ_FIRST(&hmp->transq)) != NULL &&
276 trans->sync_tid < scan->sync_tid &&
277 (scan->flags & HAMMER2_TRANS_ISFLUSH)) {
278 wakeup(&scan->sync_tid);
281 hammer2_voldata_unlock(hmp, 0);
285 * Flush the chain and all modified sub-chains through the specified
286 * synchronization point (sync_tid), propagating parent chain modifications
287 * and mirror_tid updates back up as needed. Since we are recursing downward
288 * we do not have to deal with the complexities of multi-homed chains (chains
289 * with multiple parents).
291 * Caller must have interlocked against any non-flush-related modifying
292 * operations in progress whos modify_tid values are less than or equal
293 * to the passed sync_tid.
295 * Caller must have already vetted synchronization points to ensure they
296 * are properly flushed. Only snapshots and cluster flushes can create
297 * these sorts of synchronization points.
299 * This routine can be called from several places but the most important
300 * is from the hammer2_vop_reclaim() function. We want to try to completely
301 * clean out the inode structure to prevent disconnected inodes from
302 * building up and blowing out the kmalloc pool. However, it is not actually
303 * necessary to flush reclaimed inodes to maintain HAMMER2's crash recovery
306 * chain is locked on call and will remain locked on return. If a flush
307 * occured, the chain's MOVED bit will be set indicating that its parent
308 * (which is not part of the flush) should be updated.
311 hammer2_chain_flush(hammer2_trans_t *trans, hammer2_chain_t *chain)
313 hammer2_chain_t *scan;
314 hammer2_chain_core_t *core;
315 hammer2_flush_info_t info;
318 * Execute the recursive flush and handle deferrals.
320 * Chains can be ridiculously long (thousands deep), so to
321 * avoid blowing out the kernel stack the recursive flush has a
322 * depth limit. Elements at the limit are placed on a list
323 * for re-execution after the stack has been popped.
325 bzero(&info, sizeof(info));
326 TAILQ_INIT(&info.flush_list);
328 info.sync_tid = trans->sync_tid;
330 info.cache_index = -1;
336 * Unwind deep recursions which had been deferred. This
337 * can leave MOVED set for these chains, which will be
338 * handled when we [re]flush chain after the unwind.
340 while ((scan = TAILQ_FIRST(&info.flush_list)) != NULL) {
341 KKASSERT(scan->flags & HAMMER2_CHAIN_DEFERRED);
342 TAILQ_REMOVE(&info.flush_list, scan, flush_node);
343 atomic_clear_int(&scan->flags, HAMMER2_CHAIN_DEFERRED);
346 * Now that we've popped back up we can do a secondary
347 * recursion on the deferred elements.
349 if (hammer2_debug & 0x0040)
350 kprintf("defered flush %p\n", scan);
351 hammer2_chain_lock(scan, HAMMER2_RESOLVE_MAYBE);
352 hammer2_chain_flush(trans, scan);
353 hammer2_chain_unlock(scan);
354 hammer2_chain_drop(scan); /* ref from deferral */
358 * Flush pass1 on root.
360 info.diddeferral = 0;
361 hammer2_chain_flush_core(&info, chain);
363 kprintf("flush_core_done parent=<base> chain=%p.%d %08x\n",
364 chain, chain->bref.type, chain->flags);
368 * Only loop if deep recursions have been deferred.
370 if (TAILQ_EMPTY(&info.flush_list))
376 * This is the core of the chain flushing code. The chain is locked by the
377 * caller and remains locked on return. This function is keyed off of
378 * the SUBMODIFIED bit but must make fine-grained choices based on the
379 * synchronization point we are flushing to.
381 * If the flush accomplished any work chain will be flagged MOVED
382 * indicating a copy-on-write propagation back up is required.
383 * Deep sub-nodes may also have been entered onto the deferral list.
384 * MOVED is never set on the volume root.
386 * NOTE: modify_tid is different from MODIFIED. modify_tid is updated
387 * only when a chain is specifically modified, and not updated
388 * for copy-on-write propagations. MODIFIED is set on any modification
389 * including copy-on-write propagations.
392 hammer2_chain_flush_core(hammer2_flush_info_t *info, hammer2_chain_t *chain)
394 hammer2_mount_t *hmp;
395 hammer2_blockref_t *bref;
398 hammer2_tid_t saved_sync;
399 hammer2_trans_t *trans = info->trans;
400 hammer2_chain_core_t *core;
413 kprintf("flush_core %p->%p.%d %08x (%s)\n",
414 info->parent, chain, chain->bref.type,
416 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE) ?
417 chain->data->ipdata.filename : "?"));
419 kprintf("flush_core NULL->%p.%d %08x (%s)\n",
420 chain, chain->bref.type,
422 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE) ?
423 chain->data->ipdata.filename : "?"));
426 * Ignore chains modified beyond the current flush point. These
427 * will be treated as if they did not exist.
429 if (chain->modify_tid > info->sync_tid)
433 * Deleted chains which have not been destroyed must be retained,
434 * and we probably have to recurse to clean-up any sub-trees.
435 * However, restricted flushes can stop processing here because
436 * the chain cleanup will be handled by a later normal flush.
438 * The MODIFIED bit can likely be cleared in this situation and we
439 * will do so later on in this procedure.
441 if (chain->delete_tid <= info->sync_tid) {
442 if (trans->flags & HAMMER2_TRANS_RESTRICTED)
446 saved_sync = info->sync_tid;
450 * If SUBMODIFIED is set we recurse the flush and adjust the
451 * blockrefs accordingly.
453 * NOTE: Looping on SUBMODIFIED can prevent a flush from ever
454 * finishing in the face of filesystem activity.
456 if (chain->flags & HAMMER2_CHAIN_SUBMODIFIED) {
457 hammer2_chain_t *saved_parent;
458 hammer2_tid_t saved_mirror;
459 hammer2_chain_layer_t *layer;
462 * Clear SUBMODIFIED to catch races. Note that any child
463 * with MODIFIED, DELETED, or MOVED set during scan2, or
464 * which tries to lastdrop but can't free its structures,
465 * or which gets defered, will cause SUBMODIFIED to be set
468 * We don't want to set our chain to MODIFIED gratuitously.
470 * We need an extra ref on chain because we are going to
471 * release its lock temporarily in our child loop.
473 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_SUBMODIFIED);
474 hammer2_chain_ref(chain);
477 * Run two passes. The first pass handles MODIFIED and
478 * SUBMODIFIED chains and recurses while the second pass
479 * handles MOVED chains on the way back up.
481 * If the stack gets too deep we defer scan1, but must
482 * be sure to still run scan2 if on the next loop the
483 * deferred chain has been flushed and now needs MOVED
484 * handling on the way back up.
486 * Scan1 is recursive.
488 * NOTE: The act of handling a modified/submodified chain can
489 * cause the MOVED Flag to be set. It can also be set
490 * via hammer2_chain_delete() and in other situations.
492 * NOTE: RB_SCAN() must be used instead of RB_FOREACH()
493 * because children can be physically removed during
496 saved_parent = info->parent;
497 saved_mirror = info->mirror_tid;
498 info->parent = chain;
499 info->mirror_tid = chain->bref.mirror_tid;
501 if (info->depth == HAMMER2_FLUSH_DEPTH_LIMIT) {
502 if ((chain->flags & HAMMER2_CHAIN_DEFERRED) == 0) {
503 hammer2_chain_ref(chain);
504 TAILQ_INSERT_TAIL(&info->flush_list,
506 atomic_set_int(&chain->flags,
507 HAMMER2_CHAIN_DEFERRED);
511 info->diddeferral = 0;
512 spin_lock(&core->cst.spin);
513 TAILQ_FOREACH_REVERSE(layer, &core->layerq,
514 h2_layer_list, entry) {
516 RB_SCAN(hammer2_chain_tree, &layer->rbtree,
517 NULL, hammer2_chain_flush_scan1, info);
519 diddeferral += info->diddeferral;
521 spin_unlock(&core->cst.spin);
525 * Handle successfully flushed children who are in the MOVED
526 * state on the way back up the recursion. This can have
527 * the side-effect of clearing MOVED.
529 * Scan2 is non-recursive.
532 atomic_set_int(&chain->flags,
533 HAMMER2_CHAIN_SUBMODIFIED);
534 spin_lock(&core->cst.spin);
536 spin_lock(&core->cst.spin);
537 TAILQ_FOREACH_REVERSE(layer, &core->layerq,
538 h2_layer_list, entry) {
541 RB_SCAN(hammer2_chain_tree, &layer->rbtree,
542 NULL, hammer2_chain_flush_scan2, info);
544 RB_SCAN(hammer2_chain_tree, &layer->rbtree,
545 NULL, hammer2_chain_flush_scan2, info);
546 /*diddeferral += info->diddeferral; n/a*/
550 hammer2_chain_layer_check_locked(chain->hmp, core);
551 spin_unlock(&core->cst.spin);
553 chain->bref.mirror_tid = info->mirror_tid;
554 info->mirror_tid = saved_mirror;
555 info->parent = saved_parent;
556 hammer2_chain_drop(chain);
560 * Restore sync_tid in case it was restricted by a delete/duplicate.
562 info->sync_tid = saved_sync;
565 * Rollup diddeferral for caller. Note direct assignment, not +=.
567 info->diddeferral = diddeferral;
570 * Do not flush chain if there were any deferrals. It will be
571 * retried later after the deferrals are independently handled.
574 if (hammer2_debug & 0x0008) {
575 kprintf("%*.*s} %p/%d %04x (deferred)",
576 info->depth, info->depth, "",
577 chain, chain->refs, chain->flags);
583 * If we encounter a deleted chain within our flush we can clear
584 * the MODIFIED bit and avoid flushing it whether it has been
585 * destroyed or not. We must make sure that the chain is flagged
586 * MOVED in this situation so the parent picks up the deletion.
588 * Note that scan2 has already executed above so statistics have
589 * already been rolled up.
591 if (chain->delete_tid <= info->sync_tid) {
592 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
594 if (chain->bytes == chain->bp->b_bufsize)
595 chain->bp->b_flags |= B_INVAL|B_RELBUF;
597 if ((chain->flags & HAMMER2_CHAIN_MOVED) == 0) {
598 hammer2_chain_ref(chain);
599 atomic_set_int(&chain->flags,
600 HAMMER2_CHAIN_MOVED);
602 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
603 hammer2_chain_drop(chain);
608 if ((chain->flags & HAMMER2_CHAIN_DESTROYED) &&
609 (chain->flags & HAMMER2_CHAIN_DELETED) &&
610 (trans->flags & HAMMER2_TRANS_RESTRICTED) == 0) {
612 * Throw-away the MODIFIED flag
614 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
616 if (chain->bytes == chain->bp->b_bufsize)
617 chain->bp->b_flags |= B_INVAL|B_RELBUF;
619 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
620 hammer2_chain_drop(chain);
627 * A degenerate flush might not have flushed anything and thus not
628 * processed modified blocks on the way back up. Detect the case.
630 * Note that MOVED can be set without MODIFIED being set due to
631 * a deletion, in which case it is handled by Scan2 later on.
633 * Both bits can be set along with DELETED due to a deletion if
634 * modified data within the synchronization zone and the chain
635 * was then deleted beyond the zone, in which case we still have
636 * to flush for synchronization point consistency. Otherwise though
637 * DELETED and MODIFIED are treated as separate flags.
639 if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0)
645 * A DESTROYED node that reaches this point must be flushed for
646 * synchronization point consistency.
650 * Update mirror_tid, clear MODIFIED, and set MOVED.
652 * The caller will update the parent's reference to this chain
653 * by testing MOVED as long as the modification was in-bounds.
655 * MOVED is never set on the volume root as there is no parent
658 if (chain->bref.mirror_tid < info->sync_tid)
659 chain->bref.mirror_tid = info->sync_tid;
660 wasmodified = (chain->flags & HAMMER2_CHAIN_MODIFIED) != 0;
661 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
662 if (chain == &hmp->vchain)
663 kprintf("(FLUSHED VOLUME HEADER)\n");
664 if (chain == &hmp->fchain)
665 kprintf("(FLUSHED FREEMAP HEADER)\n");
667 if ((chain->flags & HAMMER2_CHAIN_MOVED) ||
668 chain == &hmp->vchain ||
669 chain == &hmp->fchain) {
671 * Drop the ref from the MODIFIED bit we cleared.
674 hammer2_chain_drop(chain);
677 * If we were MODIFIED we inherit the ref from clearing
678 * that bit, otherwise we need another ref.
680 if (wasmodified == 0)
681 hammer2_chain_ref(chain);
682 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MOVED);
686 * If this is part of a recursive flush we can go ahead and write
687 * out the buffer cache buffer and pass a new bref back up the chain
690 * Volume headers are NOT flushed here as they require special
693 switch(chain->bref.type) {
694 case HAMMER2_BREF_TYPE_FREEMAP:
695 hammer2_modify_volume(hmp);
697 case HAMMER2_BREF_TYPE_VOLUME:
699 * We should flush the free block table before we calculate
700 * CRCs and copy voldata -> volsync.
702 * To prevent SMP races, fchain must remain locked until
703 * voldata is copied to volsync.
705 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
706 if (hmp->fchain.flags & (HAMMER2_CHAIN_MODIFIED |
707 HAMMER2_CHAIN_SUBMODIFIED)) {
708 /* this will modify vchain as a side effect */
709 hammer2_chain_flush(info->trans, &hmp->fchain);
713 * The volume header is flushed manually by the syncer, not
714 * here. All we do is adjust the crc's.
716 KKASSERT(chain->data != NULL);
717 KKASSERT(chain->bp == NULL);
718 kprintf("volume header mirror_tid %jd\n",
719 hmp->voldata.mirror_tid);
721 hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT1]=
723 (char *)&hmp->voldata +
724 HAMMER2_VOLUME_ICRC1_OFF,
725 HAMMER2_VOLUME_ICRC1_SIZE);
726 hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT0]=
728 (char *)&hmp->voldata +
729 HAMMER2_VOLUME_ICRC0_OFF,
730 HAMMER2_VOLUME_ICRC0_SIZE);
731 hmp->voldata.icrc_volheader =
733 (char *)&hmp->voldata +
734 HAMMER2_VOLUME_ICRCVH_OFF,
735 HAMMER2_VOLUME_ICRCVH_SIZE);
736 hmp->volsync = hmp->voldata;
737 atomic_set_int(&chain->flags, HAMMER2_CHAIN_VOLUMESYNC);
738 hammer2_chain_unlock(&hmp->fchain);
740 case HAMMER2_BREF_TYPE_DATA:
742 * Data elements have already been flushed via the logical
743 * file buffer cache. Their hash was set in the bref by
744 * the vop_write code.
746 * Make sure any device buffer(s) have been flushed out here.
747 * (there aren't usually any to flush).
749 psize = hammer2_devblksize(chain->bytes);
750 pmask = (hammer2_off_t)psize - 1;
751 pbase = chain->bref.data_off & ~pmask;
752 boff = chain->bref.data_off & (HAMMER2_OFF_MASK & pmask);
754 bp = getblk(hmp->devvp, pbase, psize, GETBLK_NOWAIT, 0);
756 if ((bp->b_flags & (B_CACHE | B_DIRTY)) ==
757 (B_CACHE | B_DIRTY)) {
760 bp->b_flags |= B_RELBUF;
766 case HAMMER2_BREF_TYPE_INDIRECT:
768 * Indirect blocks may be in an INITIAL state. Use the
769 * chain_lock() call to ensure that the buffer has been
770 * instantiated (even though it is already locked the buffer
771 * might not have been instantiated).
773 * Only write the buffer out if it is dirty, it is possible
774 * the operating system had already written out the buffer.
776 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
777 KKASSERT(chain->bp != NULL);
780 if ((chain->flags & HAMMER2_CHAIN_DIRTYBP) ||
781 (bp->b_flags & B_DIRTY)) {
788 hammer2_chain_unlock(chain);
791 case HAMMER2_BREF_TYPE_INDIRECT:
792 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
794 * Device-backed. Buffer will be flushed by the sync
797 KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
799 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
802 * Embedded elements have to be flushed out.
803 * (Basically just BREF_TYPE_INODE).
805 KKASSERT(chain->flags & HAMMER2_CHAIN_EMBEDDED);
806 KKASSERT(chain->data != NULL);
807 KKASSERT(chain->bp == NULL);
810 KKASSERT((bref->data_off & HAMMER2_OFF_MASK) != 0);
811 KKASSERT(HAMMER2_DEC_CHECK(chain->bref.methods) ==
812 HAMMER2_CHECK_ISCSI32 ||
813 HAMMER2_DEC_CHECK(chain->bref.methods) ==
814 HAMMER2_CHECK_FREEMAP);
817 * The data is embedded, we have to acquire the
818 * buffer cache buffer and copy the data into it.
820 psize = hammer2_devblksize(chain->bytes);
821 pmask = (hammer2_off_t)psize - 1;
822 pbase = bref->data_off & ~pmask;
823 boff = bref->data_off & (HAMMER2_OFF_MASK & pmask);
826 * The getblk() optimization can only be used if the
827 * physical block size matches the request.
829 error = bread(hmp->devvp, pbase, psize, &bp);
830 KKASSERT(error == 0);
832 bdata = (char *)bp->b_data + boff;
835 * Copy the data to the buffer, mark the buffer
836 * dirty, and convert the chain to unmodified.
838 bcopy(chain->data, bdata, chain->bytes);
839 bp->b_flags |= B_CLUSTEROK;
843 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
844 case HAMMER2_CHECK_FREEMAP:
845 chain->bref.check.freemap.icrc32 =
846 hammer2_icrc32(chain->data, chain->bytes);
848 case HAMMER2_CHECK_ISCSI32:
849 chain->bref.check.iscsi32.value =
850 hammer2_icrc32(chain->data, chain->bytes);
853 panic("hammer2_flush_core: bad crc type");
854 break; /* NOT REACHED */
856 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
857 ++hammer2_iod_meta_write;
859 ++hammer2_iod_indr_write;
864 * Flush helper scan1 (recursive)
866 * Flushes the children of the caller's chain (parent) and updates
867 * the blockref, restricted by sync_tid.
869 * Ripouts during the loop should not cause any problems. Because we are
870 * flushing to a synchronization point, modification races will occur after
871 * sync_tid and do not have to be flushed anyway.
873 * It is also ok if the parent is chain_duplicate()'d while unlocked because
874 * the delete/duplication will install a delete_tid that is still larger than
875 * our current sync_tid.
878 hammer2_chain_flush_scan1(hammer2_chain_t *child, void *data)
880 hammer2_flush_info_t *info = data;
881 hammer2_trans_t *trans = info->trans;
882 hammer2_chain_t *parent = info->parent;
886 * We should only need to recurse if SUBMODIFIED is set, but as
887 * a safety also recurse if MODIFIED is also set.
889 * Return early if neither bit is set. We must re-assert the
890 * SUBMODIFIED flag in the parent if any child covered by the
891 * parent (via delete_tid) is skipped.
893 if ((child->flags & (HAMMER2_CHAIN_MODIFIED |
894 HAMMER2_CHAIN_SUBMODIFIED)) == 0) {
897 if (child->modify_tid > trans->sync_tid) {
898 if (parent->delete_tid > trans->sync_tid) {
899 atomic_set_int(&parent->flags,
900 HAMMER2_CHAIN_SUBMODIFIED);
905 hammer2_chain_ref(child);
906 spin_unlock(&parent->core->cst.spin);
909 * The caller has added a ref to the parent so we can temporarily
910 * unlock it in order to lock the child. Re-check the flags before
913 hammer2_chain_unlock(parent);
914 hammer2_chain_lock(child, HAMMER2_RESOLVE_MAYBE);
916 if ((child->flags & (HAMMER2_CHAIN_MODIFIED |
917 HAMMER2_CHAIN_SUBMODIFIED)) == 0) {
918 hammer2_chain_unlock(child);
919 hammer2_chain_drop(child);
920 hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
921 spin_lock(&parent->core->cst.spin);
924 if (child->modify_tid > trans->sync_tid) {
925 hammer2_chain_unlock(child);
926 hammer2_chain_drop(child);
927 hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
928 spin_lock(&parent->core->cst.spin);
929 if (parent->delete_tid > trans->sync_tid) {
930 atomic_set_int(&parent->flags,
931 HAMMER2_CHAIN_SUBMODIFIED);
937 * The DESTROYED flag can only be initially set on an unreferenced
938 * deleted inode and will propagate downward via the mechanic below.
939 * Such inode chains have been deleted for good and should no longer
940 * be subject to delete/duplication.
942 * This optimization allows the inode reclaim (destroy unlinked file
943 * on vnode reclamation after last close) to be flagged by just
944 * setting HAMMER2_CHAIN_DESTROYED at the top level and then will
945 * cause the chains to be terminated and related buffers to be
946 * invalidated and not flushed out.
948 * We have to be careful not to propagate the DESTROYED flag if
949 * the destruction occurred after our flush sync_tid.
951 if ((parent->flags & HAMMER2_CHAIN_DESTROYED) &&
952 (child->flags & HAMMER2_CHAIN_DELETED) &&
953 (child->flags & HAMMER2_CHAIN_DESTROYED) == 0) {
954 atomic_set_int(&child->flags, HAMMER2_CHAIN_DESTROYED |
955 HAMMER2_CHAIN_SUBMODIFIED);
959 * Recurse and collect deferral data.
961 diddeferral = info->diddeferral;
963 hammer2_chain_flush_core(info, child);
965 kprintf("flush_core_done parent=%p flags=%08x child=%p.%d %08x\n",
966 parent, parent->flags, child, child->bref.type, child->flags);
969 info->diddeferral += diddeferral;
971 if (child->flags & HAMMER2_CHAIN_SUBMODIFIED)
972 atomic_set_int(&parent->flags, HAMMER2_CHAIN_SUBMODIFIED);
974 hammer2_chain_unlock(child);
975 hammer2_chain_drop(child);
977 hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
979 spin_lock(&parent->core->cst.spin);
985 * Flush helper scan2 (non-recursive)
987 * This pass on a chain's children propagates any MOVED or DELETED
988 * elements back up the chain towards the root after those elements have
989 * been fully flushed. Unlike scan1, this function is NOT recursive and
990 * the parent remains locked across the entire scan.
992 * SCAN2 is called twice, once with pass set to 1 and once with it set to 2.
993 * We have to do this so base[] elements can be deleted in pass 1 to make
994 * room for adding new elements in pass 2.
996 * This function also rolls up storage statistics.
998 * NOTE! We must re-set SUBMODIFIED on the parent(s) as appropriate, and
999 * due to the above conditions it is possible to do this and still
1000 * have some children flagged MOVED depending on the synchronization.
1002 * NOTE! A deletion is a visbility issue, there can still be references to
1003 * deleted elements (for example, to an unlinked file which is still
1004 * open), and there can also be multiple chains pointing to the same
1005 * bref where some are deleted and some are not (for example due to
1006 * a rename). So a chain marked for deletion is basically considered
1007 * to be live until it is explicitly destroyed or until its ref-count
1008 * reaches zero (also implying that MOVED and MODIFIED are clear).
1011 hammer2_chain_flush_scan2(hammer2_chain_t *child, void *data)
1013 hammer2_flush_info_t *info = data;
1014 hammer2_chain_t *parent = info->parent;
1015 hammer2_chain_core_t *above = child->above;
1016 hammer2_mount_t *hmp = child->hmp;
1017 hammer2_trans_t *trans = info->trans;
1018 hammer2_blockref_t *base;
1023 * Inodes with stale children that have been converted to DIRECTDATA
1024 * mode (file extension or hardlink conversion typically) need to
1025 * skipped right now before we start messing with a non-existant
1029 if (parent->bref.type == HAMMER2_BREF_TYPE_INODE &&
1030 (parent->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA)) {
1036 * Ignore children created after our flush point, treating them as
1037 * if they did not exist). These children will not cause the parent
1040 * When we encounter such children and the parent chain has not been
1041 * deleted, delete/duplicated, or delete/duplicated-for-move, then
1042 * the parent may be used to funnel through several flush points.
1043 * We must re-set the SUBMODIFIED flag in the parent to ensure that
1044 * those flushes have visbility. A simple test of delete_tid suffices
1045 * to determine if the parent spans beyond our current flush.
1047 if (child->modify_tid > trans->sync_tid) {
1052 * Ignore children which have not changed. The parent's block table
1053 * is already correct.
1055 * XXX The MOVED bit is only cleared when all multi-homed parents
1056 * have flushed, creating a situation where a re-flush can occur
1057 * via a parent which has already flushed. The hammer2_base_*()
1058 * functions currently have a hack to deal with this case but
1059 * we need something better.
1061 if ((child->flags & HAMMER2_CHAIN_MOVED) == 0) {
1066 * Make sure child is referenced before we unlock.
1068 hammer2_chain_ref(child);
1069 spin_unlock(&above->cst.spin);
1072 * Parent reflushed after the child has passed them by should skip
1073 * due to the modify_tid test. XXX
1075 hammer2_chain_lock(child, HAMMER2_RESOLVE_NEVER);
1076 KKASSERT(child->above == above);
1077 KKASSERT(parent->core == above);
1080 * The parent's blockref to the child must be deleted or updated.
1082 * This point is not reached on successful DESTROYED optimizations
1083 * but can be reached on recursive deletions and restricted flushes.
1085 * Because flushes are ordered we do not have to make a
1086 * modify/duplicate of indirect blocks. That is, the flush
1087 * code does not have to kmalloc or duplicate anything. We
1088 * can adjust the indirect block table in-place and reuse the
1089 * chain. It IS possible that the chain has already been duplicated
1090 * or may wind up being duplicated on-the-fly by modifying code
1091 * on the frontend. We simply use the original and ignore such
1092 * chains. However, it does mean we can't clear the MOVED bit.
1094 * XXX recursive deletions not optimized.
1096 hammer2_chain_modify(trans, &parent,
1097 HAMMER2_MODIFY_NO_MODIFY_TID |
1098 HAMMER2_MODIFY_ASSERTNOCOPY);
1100 switch(parent->bref.type) {
1101 case HAMMER2_BREF_TYPE_INODE:
1103 * XXX Should assert that OPFLAG_DIRECTDATA is 0 once we
1104 * properly duplicate the inode headers and do proper flush
1105 * range checks (all the children should be beyond the flush
1106 * point). For now just don't sync the non-applicable
1109 * XXX Can also occur due to hardlink consolidation. We
1110 * set OPFLAG_DIRECTDATA to prevent the indirect and data
1111 * blocks from syncing ot the hardlink pointer.
1114 KKASSERT((parent->data->ipdata.op_flags &
1115 HAMMER2_OPFLAG_DIRECTDATA) == 0);
1118 if (parent->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1123 base = &parent->data->ipdata.u.blockset.blockref[0];
1124 count = HAMMER2_SET_COUNT;
1127 case HAMMER2_BREF_TYPE_INDIRECT:
1128 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1130 base = &parent->data->npdata[0];
1133 KKASSERT(child->flags & HAMMER2_CHAIN_DELETED);
1135 count = parent->bytes / sizeof(hammer2_blockref_t);
1137 case HAMMER2_BREF_TYPE_VOLUME:
1138 base = &hmp->voldata.sroot_blockset.blockref[0];
1139 count = HAMMER2_SET_COUNT;
1141 case HAMMER2_BREF_TYPE_FREEMAP:
1142 base = &parent->data->npdata[0];
1143 count = HAMMER2_SET_COUNT;
1148 panic("hammer2_chain_flush_scan2: "
1149 "unrecognized blockref type: %d",
1154 * Don't bother updating a deleted parent's blockrefs (caller will
1155 * optimize-out the disk write). Note that this is not optional,
1156 * a deleted parent's blockref array might not be synchronized at
1157 * all so calling hammer2_base*() functions could result in a panic.
1159 * Otherwise, we need to be COUNTEDBREFS synchronized for the
1160 * hammer2_base_*() functions.
1162 if (parent->delete_tid <= trans->sync_tid)
1164 else if ((above->flags & HAMMER2_CORE_COUNTEDBREFS) == 0)
1165 hammer2_chain_countbrefs(above, base, count);
1168 * Update the parent's blockref table and propagate mirror_tid.
1170 * NOTE! Children with modify_tid's beyond our flush point are
1171 * considered to not exist for the purposes of updating the
1172 * parent's blockref array.
1174 * NOTE! Updates to a parent's blockref table do not adjust the
1175 * parent's bref.modify_tid, only its bref.mirror_tid.
1177 if (info->pass == 1 && child->delete_tid <= trans->sync_tid) {
1179 * Deleting. Only adjust the block array if it contains
1180 * the child's entry (child's REPLACE flag is set). Clear
1181 * the child's REPLACE flag only once all possible parent's
1182 * have been updated.
1185 if (base && (child->flags & HAMMER2_CHAIN_REPLACE)) {
1186 hammer2_rollup_stats(parent, child, -1);
1187 spin_lock(&above->cst.spin);
1188 hammer2_base_delete(base, count, above,
1189 &info->cache_index, &child->bref);
1190 if (TAILQ_NEXT(parent, core_entry) == NULL) {
1191 atomic_clear_int(&child->flags,
1192 HAMMER2_CHAIN_REPLACE);
1194 spin_unlock(&above->cst.spin);
1196 if (info->mirror_tid < child->delete_tid)
1197 info->mirror_tid = child->delete_tid;
1198 } else if (info->pass == 2 && child->delete_tid > trans->sync_tid) {
1200 * Inserting. Only set the child's REPLACE flag indicating
1201 * that the parent's blockref array entry is valid once all
1202 * possible parent's have been updated.
1206 if (child->flags & HAMMER2_CHAIN_REPLACE)
1207 hammer2_rollup_stats(parent, child, 0);
1209 hammer2_rollup_stats(parent, child, 1);
1210 spin_lock(&above->cst.spin);
1211 hammer2_base_insert(base, count, above,
1212 &info->cache_index, &child->bref,
1214 if (TAILQ_NEXT(parent, core_entry) == NULL) {
1215 atomic_set_int(&child->flags,
1216 HAMMER2_CHAIN_REPLACE);
1218 spin_unlock(&above->cst.spin);
1220 if (info->mirror_tid < child->modify_tid)
1221 info->mirror_tid = child->modify_tid;
1226 if (info->mirror_tid < child->bref.mirror_tid) {
1227 info->mirror_tid = child->bref.mirror_tid;
1229 if ((parent->bref.type == HAMMER2_BREF_TYPE_VOLUME ||
1230 parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP) &&
1231 hmp->voldata.mirror_tid < child->bref.mirror_tid) {
1232 hmp->voldata.mirror_tid = child->bref.mirror_tid;
1236 * Only clear MOVED once all possible parents have been flushed.
1238 * When can we safely clear the MOVED flag? Flushes down duplicate
1239 * paths can occur out of order, for example if an inode is moved
1240 * as part of a hardlink consolidation or if an inode is moved into
1241 * an indirect block indexed before the inode.
1243 if (ok && (child->flags & HAMMER2_CHAIN_MOVED)) {
1244 hammer2_chain_t *scan;
1247 spin_lock(&above->cst.spin);
1248 TAILQ_FOREACH(scan, &above->ownerq, core_entry) {
1250 * XXX weird code also checked at the top of scan2,
1251 * I would like to fix this by detaching the core
1252 * on initial hardlink consolidation (1->2 nlinks).
1255 if (scan->bref.type == HAMMER2_BREF_TYPE_INODE &&
1256 (scan->data->ipdata.op_flags &
1257 HAMMER2_OPFLAG_DIRECTDATA)) {
1261 if (scan->flags & HAMMER2_CHAIN_SUBMODIFIED) {
1266 spin_unlock(&above->cst.spin);
1268 atomic_clear_int(&child->flags, HAMMER2_CHAIN_MOVED);
1269 hammer2_chain_drop(child); /* flag */
1274 * Unlock the child. This can wind up dropping the child's
1275 * last ref, removing it from the parent's RB tree, and deallocating
1276 * the structure. The RB_SCAN() our caller is doing handles the
1279 hammer2_chain_unlock(child);
1280 hammer2_chain_drop(child);
1281 spin_lock(&above->cst.spin);
1284 * The parent cleared SUBMODIFIED prior to the scan. If the child
1285 * still requires a flush (possibly due to being outside the current
1286 * synchronization zone), we must re-set SUBMODIFIED on the way back
1295 hammer2_rollup_stats(hammer2_chain_t *parent, hammer2_chain_t *child, int how)
1298 hammer2_chain_t *grandp;
1301 parent->data_count += child->data_count;
1302 parent->inode_count += child->inode_count;
1303 child->data_count = 0;
1304 child->inode_count = 0;
1306 parent->data_count -= child->bytes;
1307 if (child->bref.type == HAMMER2_BREF_TYPE_INODE) {
1308 parent->inode_count -= 1;
1310 /* XXX child->data may be NULL atm */
1311 parent->data_count -= child->data->ipdata.data_count;
1312 parent->inode_count -= child->data->ipdata.inode_count;
1315 } else if (how > 0) {
1316 parent->data_count += child->bytes;
1317 if (child->bref.type == HAMMER2_BREF_TYPE_INODE) {
1318 parent->inode_count += 1;
1320 /* XXX child->data may be NULL atm */
1321 parent->data_count += child->data->ipdata.data_count;
1322 parent->inode_count += child->data->ipdata.inode_count;
1326 if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
1327 parent->data->ipdata.data_count += parent->data_count;
1328 parent->data->ipdata.inode_count += parent->inode_count;
1330 for (grandp = parent->above->first_parent;
1332 grandp = grandp->next_parent) {
1333 grandp->data_count += parent->data_count;
1334 grandp->inode_count += parent->inode_count;
1337 parent->data_count = 0;
1338 parent->inode_count = 0;