xref: /dflybsd-src/sys/vfs/hammer2/hammer2_flush.c (revision 25ca8c794df06033ca37cb3fdf9b2992a7b5a9e6)
1 /*
2  * Copyright (c) 2011-2014 The DragonFly Project.  All rights reserved.
3  *
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>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  *
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
17  *    distribution.
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.
21  *
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
33  * SUCH DAMAGE.
34  */
35 /*
36  *			TRANSACTION AND FLUSH HANDLING
37  *
38  * Deceptively simple but actually fairly difficult to implement properly is
39  * how I would describe it.
40  *
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.
44  */
45 
46 #include <sys/cdefs.h>
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/types.h>
50 #include <sys/lock.h>
51 #include <sys/uuid.h>
52 
53 #include "hammer2.h"
54 
55 #define FLUSH_DEBUG 0
56 
57 #define HAMMER2_FLUSH_DEPTH_LIMIT       10      /* stack recursion limit */
58 
59 
60 /*
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
65  * processes.
66  */
67 struct hammer2_flush_info {
68 	hammer2_chain_t *parent;
69 	hammer2_trans_t	*trans;
70 	int		depth;
71 	int		diddeferral;
72 	int		cache_index;
73 	struct h2_flush_list flushq;
74 	hammer2_xid_t	sync_xid;	/* memory synchronization point */
75 	hammer2_chain_t	*debug;
76 };
77 
78 typedef struct hammer2_flush_info hammer2_flush_info_t;
79 
80 static void hammer2_flush_core(hammer2_flush_info_t *info,
81 				hammer2_chain_t *chain, int deleting);
82 static int hammer2_flush_recurse(hammer2_chain_t *child, void *data);
83 
84 /*
85  * For now use a global transaction manager.  What we ultimately want to do
86  * is give each non-overlapping hmp/pmp group its own transaction manager.
87  *
88  * Transactions govern XID tracking on the physical media (the hmp), but they
89  * also govern TID tracking which is per-PFS and thus might cross multiple
90  * hmp's.  So we can't just stuff tmanage into hammer2_dev or
91  * hammer2_pfs.
92  */
93 void
94 hammer2_trans_manage_init(hammer2_trans_manage_t *tman)
95 {
96 	lockinit(&tman->translk, "h2trans", 0, 0);
97 	TAILQ_INIT(&tman->transq);
98 	tman->flush_xid = 1;
99 	tman->alloc_xid = tman->flush_xid + 1;
100 }
101 
102 hammer2_xid_t
103 hammer2_trans_newxid(hammer2_pfs_t *pmp)
104 {
105 	hammer2_xid_t xid;
106 
107 	for (;;) {
108 		xid = atomic_fetchadd_int(&pmp->tmanage.alloc_xid, 1);
109 		if (xid)
110 			break;
111 	}
112 	return xid;
113 }
114 
115 /*
116  * Transaction support functions for writing to the filesystem.
117  *
118  * Initializing a new transaction allocates a transaction ID.  Typically
119  * passed a pmp (hmp passed as NULL), indicating a cluster transaction.  Can
120  * be passed a NULL pmp and non-NULL hmp to indicate a transaction on a single
121  * media target.  The latter mode is used by the recovery code.
122  *
123  * TWO TRANSACTION IDs can run concurrently, where one is a flush and the
124  * other is a set of any number of concurrent filesystem operations.  We
125  * can either have <running_fs_ops> + <waiting_flush> + <blocked_fs_ops>
126  * or we can have <running_flush> + <concurrent_fs_ops>.
127  *
128  * During a flush, new fs_ops are only blocked until the fs_ops prior to
129  * the flush complete.  The new fs_ops can then run concurrent with the flush.
130  *
131  * Buffer-cache transactions operate as fs_ops but never block.  A
132  * buffer-cache flush will run either before or after the current pending
133  * flush depending on its state.
134  */
135 void
136 hammer2_trans_init(hammer2_trans_t *trans, hammer2_pfs_t *pmp, int flags)
137 {
138 	hammer2_trans_manage_t *tman;
139 	hammer2_trans_t *head;
140 
141 	tman = &pmp->tmanage;
142 
143 	bzero(trans, sizeof(*trans));
144 	trans->pmp = pmp;
145 	trans->flags = flags;
146 	trans->td = curthread;
147 
148 	lockmgr(&tman->translk, LK_EXCLUSIVE);
149 
150 	if (flags & HAMMER2_TRANS_ISFLUSH) {
151 		/*
152 		 * If multiple flushes are trying to run we have to
153 		 * wait until it is our turn.  All flushes are serialized.
154 		 *
155 		 * We queue ourselves and then wait to become the head
156 		 * of the queue, allowing all prior flushes to complete.
157 		 *
158 		 * Multiple normal transactions can share the current
159 		 * transaction id but a flush transaction needs its own
160 		 * unique TID for proper block table update accounting.
161 		 */
162 		++tman->flushcnt;
163 		++pmp->modify_tid;
164 		tman->flush_xid = hammer2_trans_newxid(pmp);
165 		trans->sync_xid = tman->flush_xid;
166 		trans->modify_tid = pmp->modify_tid;
167 		TAILQ_INSERT_TAIL(&tman->transq, trans, entry);
168 		if (TAILQ_FIRST(&tman->transq) != trans) {
169 			trans->blocked = 1;
170 			while (trans->blocked) {
171 				lksleep(&trans->sync_xid, &tman->translk,
172 					0, "h2multf", hz);
173 			}
174 		}
175 	} else if (tman->flushcnt == 0) {
176 		/*
177 		 * No flushes are pending, we can go.  Use prior flush_xid + 1.
178 		 *
179 		 * WARNING!  Also see hammer2_chain_setflush()
180 		 */
181 		TAILQ_INSERT_TAIL(&tman->transq, trans, entry);
182 		trans->sync_xid = tman->flush_xid + 1;
183 
184 		/* XXX improve/optimize inode allocation */
185 	} else if (trans->flags & HAMMER2_TRANS_BUFCACHE) {
186 		/*
187 		 * A buffer cache transaction is requested while a flush
188 		 * is in progress.  The flush's PREFLUSH flag must be set
189 		 * in this situation.
190 		 *
191 		 * The buffer cache flush takes on the main flush's
192 		 * transaction id.
193 		 */
194 		TAILQ_FOREACH(head, &tman->transq, entry) {
195 			if (head->flags & HAMMER2_TRANS_ISFLUSH)
196 				break;
197 		}
198 		KKASSERT(head);
199 		KKASSERT(head->flags & HAMMER2_TRANS_PREFLUSH);
200 		trans->flags |= HAMMER2_TRANS_PREFLUSH;
201 		TAILQ_INSERT_AFTER(&tman->transq, head, trans, entry);
202 		trans->sync_xid = head->sync_xid;
203 		trans->modify_tid = head->modify_tid;
204 		trans->flags |= HAMMER2_TRANS_CONCURRENT;
205 		/* not allowed to block */
206 	} else {
207 		/*
208 		 * A normal transaction is requested while a flush is in
209 		 * progress.  We insert after the current flush and may
210 		 * block.
211 		 *
212 		 * WARNING!  Also see hammer2_chain_setflush()
213 		 */
214 		TAILQ_FOREACH(head, &tman->transq, entry) {
215 			if (head->flags & HAMMER2_TRANS_ISFLUSH)
216 				break;
217 		}
218 		KKASSERT(head);
219 		TAILQ_INSERT_AFTER(&tman->transq, head, trans, entry);
220 		trans->sync_xid = head->sync_xid + 1;
221 		trans->flags |= HAMMER2_TRANS_CONCURRENT;
222 
223 		/*
224 		 * XXX for now we must block new transactions, synchronous
225 		 * flush mode is on by default.
226 		 *
227 		 * If synchronous flush mode is enabled concurrent
228 		 * frontend transactions during the flush are not
229 		 * allowed (except we don't have a choice for buffer
230 		 * cache ops).
231 		 */
232 		if (hammer2_synchronous_flush > 0 ||
233 		    TAILQ_FIRST(&tman->transq) != head) {
234 			trans->blocked = 1;
235 			while (trans->blocked) {
236 				lksleep(&trans->sync_xid, &tman->translk,
237 					0, "h2multf", hz);
238 			}
239 		}
240 	}
241 	if (flags & HAMMER2_TRANS_NEWINODE) {
242 		if (pmp->spmp_hmp) {
243 			/*
244 			 * Super-root transaction, all new inodes have an
245 			 * inode number of 1.  Normal pfs inode cache
246 			 * semantics are not used.
247 			 */
248 			trans->inode_tid = 1;
249 		} else {
250 			/*
251 			 * Normal transaction
252 			 */
253 			if (pmp->inode_tid < HAMMER2_INODE_START)
254 				pmp->inode_tid = HAMMER2_INODE_START;
255 			trans->inode_tid = pmp->inode_tid++;
256 		}
257 	}
258 
259 	lockmgr(&tman->translk, LK_RELEASE);
260 }
261 
262 void
263 hammer2_trans_assert_strategy(hammer2_pfs_t *pmp)
264 {
265 	hammer2_trans_manage_t *tman;
266 	hammer2_trans_t *head;
267 
268 	tman = &pmp->tmanage;
269 	lockmgr(&tman->translk, LK_EXCLUSIVE);
270 	if (tman->flushcnt) {
271 		TAILQ_FOREACH(head, &tman->transq, entry) {
272 			if (head->flags & HAMMER2_TRANS_ISFLUSH)
273 				break;
274 		}
275 		KKASSERT(head);
276 		KKASSERT(head->flags & HAMMER2_TRANS_PREFLUSH);
277 	}
278 	lockmgr(&tman->translk, LK_RELEASE);
279 }
280 
281 void
282 hammer2_trans_done(hammer2_trans_t *trans)
283 {
284 	hammer2_trans_manage_t *tman;
285 	hammer2_trans_t *head;
286 	hammer2_trans_t *scan;
287 
288 	tman = &trans->pmp->tmanage;
289 
290 	/*
291 	 * Remove.
292 	 */
293 	lockmgr(&tman->translk, LK_EXCLUSIVE);
294 	TAILQ_REMOVE(&tman->transq, trans, entry);
295 	head = TAILQ_FIRST(&tman->transq);
296 
297 	/*
298 	 * Adjust flushcnt if this was a flush, clear TRANS_CONCURRENT
299 	 * up through the next flush.  (If the head is a flush then we
300 	 * stop there, unlike the unblock code following this section).
301 	 */
302 	if (trans->flags & HAMMER2_TRANS_ISFLUSH) {
303 		--tman->flushcnt;
304 		scan = head;
305 		while (scan && (scan->flags & HAMMER2_TRANS_ISFLUSH) == 0) {
306 			atomic_clear_int(&scan->flags,
307 					 HAMMER2_TRANS_CONCURRENT);
308 			scan = TAILQ_NEXT(scan, entry);
309 		}
310 	}
311 
312 	/*
313 	 * Unblock the head of the queue and any additional transactions
314 	 * up to the next flush.  The head can be a flush and it will be
315 	 * unblocked along with the non-flush transactions following it
316 	 * (which are allowed to run concurrently with it).
317 	 *
318 	 * In synchronous flush mode we stop if the head transaction is
319 	 * a flush.
320 	 */
321 	if (head && head->blocked) {
322 		head->blocked = 0;
323 		wakeup(&head->sync_xid);
324 
325 		if (hammer2_synchronous_flush > 0)
326 			scan = head;
327 		else
328 			scan = TAILQ_NEXT(head, entry);
329 		while (scan && (scan->flags & HAMMER2_TRANS_ISFLUSH) == 0) {
330 			if (scan->blocked) {
331 				scan->blocked = 0;
332 				wakeup(&scan->sync_xid);
333 			}
334 			scan = TAILQ_NEXT(scan, entry);
335 		}
336 	}
337 	lockmgr(&tman->translk, LK_RELEASE);
338 }
339 
340 /*
341  * Chains undergoing destruction are removed from the in-memory topology.
342  * To avoid getting lost these chains are placed on the delayed flush
343  * queue which will properly dispose of them.
344  *
345  * We do this instead of issuing an immediate flush in order to give
346  * recursive deletions (rm -rf, etc) a chance to remove more of the
347  * hierarchy, potentially allowing an enormous amount of write I/O to
348  * be avoided.
349  */
350 void
351 hammer2_delayed_flush(hammer2_trans_t *trans, hammer2_chain_t *chain)
352 {
353 	if ((chain->flags & HAMMER2_CHAIN_DELAYED) == 0) {
354 		hammer2_spin_ex(&chain->hmp->list_spin);
355 		if ((chain->flags & (HAMMER2_CHAIN_DELAYED |
356 				     HAMMER2_CHAIN_DEFERRED)) == 0) {
357 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELAYED |
358 						      HAMMER2_CHAIN_DEFERRED);
359 			TAILQ_INSERT_TAIL(&chain->hmp->flushq,
360 					  chain, flush_node);
361 			hammer2_chain_ref(chain);
362 		}
363 		hammer2_spin_unex(&chain->hmp->list_spin);
364 	}
365 }
366 
367 /*
368  * Flush the chain and all modified sub-chains through the specified
369  * synchronization point, propagating parent chain modifications, modify_tid,
370  * and mirror_tid updates back up as needed.
371  *
372  * Caller must have interlocked against any non-flush-related modifying
373  * operations in progress whos XXX values are less than or equal
374  * to the passed sync_xid.
375  *
376  * Caller must have already vetted synchronization points to ensure they
377  * are properly flushed.  Only snapshots and cluster flushes can create
378  * these sorts of synchronization points.
379  *
380  * This routine can be called from several places but the most important
381  * is from VFS_SYNC.
382  *
383  * chain is locked on call and will remain locked on return.  The chain's
384  * UPDATE flag indicates that its parent's block table (which is not yet
385  * part of the flush) should be updated.  The chain may be replaced by
386  * the call if it was modified.
387  */
388 void
389 hammer2_flush(hammer2_trans_t *trans, hammer2_chain_t *chain, int istop)
390 {
391 	hammer2_chain_t *scan;
392 	hammer2_flush_info_t info;
393 	hammer2_dev_t *hmp;
394 	int loops;
395 
396 	/*
397 	 * Execute the recursive flush and handle deferrals.
398 	 *
399 	 * Chains can be ridiculously long (thousands deep), so to
400 	 * avoid blowing out the kernel stack the recursive flush has a
401 	 * depth limit.  Elements at the limit are placed on a list
402 	 * for re-execution after the stack has been popped.
403 	 */
404 	bzero(&info, sizeof(info));
405 	TAILQ_INIT(&info.flushq);
406 	info.trans = trans;
407 	info.sync_xid = trans->sync_xid;
408 	info.cache_index = -1;
409 
410 	/*
411 	 * Calculate parent (can be NULL), if not NULL the flush core
412 	 * expects the parent to be referenced so it can easily lock/unlock
413 	 * it without it getting ripped up.
414 	 */
415 	if ((info.parent = chain->parent) != NULL)
416 		hammer2_chain_ref(info.parent);
417 
418 	/*
419 	 * Extra ref needed because flush_core expects it when replacing
420 	 * chain.
421 	 */
422 	hammer2_chain_ref(chain);
423 	hmp = chain->hmp;
424 	loops = 0;
425 
426 	for (;;) {
427 		/*
428 		 * Move hmp->flushq to info.flushq if non-empty so it can
429 		 * be processed.
430 		 */
431 		if (TAILQ_FIRST(&hmp->flushq) != NULL) {
432 			hammer2_spin_ex(&chain->hmp->list_spin);
433 			TAILQ_CONCAT(&info.flushq, &hmp->flushq, flush_node);
434 			hammer2_spin_unex(&chain->hmp->list_spin);
435 		}
436 
437 		/*
438 		 * Unwind deep recursions which had been deferred.  This
439 		 * can leave the FLUSH_* bits set for these chains, which
440 		 * will be handled when we [re]flush chain after the unwind.
441 		 */
442 		while ((scan = TAILQ_FIRST(&info.flushq)) != NULL) {
443 			KKASSERT(scan->flags & HAMMER2_CHAIN_DEFERRED);
444 			TAILQ_REMOVE(&info.flushq, scan, flush_node);
445 			atomic_clear_int(&scan->flags, HAMMER2_CHAIN_DEFERRED |
446 						       HAMMER2_CHAIN_DELAYED);
447 
448 			/*
449 			 * Now that we've popped back up we can do a secondary
450 			 * recursion on the deferred elements.
451 			 *
452 			 * NOTE: hammer2_flush() may replace scan.
453 			 */
454 			if (hammer2_debug & 0x0040)
455 				kprintf("deferred flush %p\n", scan);
456 			hammer2_chain_lock(scan, HAMMER2_RESOLVE_MAYBE);
457 			hammer2_flush(trans, scan, 0);
458 			hammer2_chain_unlock(scan);
459 			hammer2_chain_drop(scan);	/* ref from deferral */
460 		}
461 
462 		/*
463 		 * [re]flush chain.
464 		 */
465 		info.diddeferral = 0;
466 		hammer2_flush_core(&info, chain, istop);
467 
468 		/*
469 		 * Only loop if deep recursions have been deferred.
470 		 */
471 		if (TAILQ_EMPTY(&info.flushq))
472 			break;
473 
474 		if (++loops % 1000 == 0) {
475 			kprintf("hammer2_flush: excessive loops on %p\n",
476 				chain);
477 			if (hammer2_debug & 0x100000)
478 				Debugger("hell4");
479 		}
480 	}
481 	hammer2_chain_drop(chain);
482 	if (info.parent)
483 		hammer2_chain_drop(info.parent);
484 }
485 
486 /*
487  * This is the core of the chain flushing code.  The chain is locked by the
488  * caller and must also have an extra ref on it by the caller, and remains
489  * locked and will have an extra ref on return.  Upon return, the caller can
490  * test the UPDATE bit on the child to determine if the parent needs updating.
491  *
492  * (1) Determine if this node is a candidate for the flush, return if it is
493  *     not.  fchain and vchain are always candidates for the flush.
494  *
495  * (2) If we recurse too deep the chain is entered onto the deferral list and
496  *     the current flush stack is aborted until after the deferral list is
497  *     run.
498  *
499  * (3) Recursively flush live children (rbtree).  This can create deferrals.
500  *     A successful flush clears the MODIFIED and UPDATE bits on the children
501  *     and typically causes the parent to be marked MODIFIED as the children
502  *     update the parent's block table.  A parent might already be marked
503  *     MODIFIED due to a deletion (whos blocktable update in the parent is
504  *     handled by the frontend), or if the parent itself is modified by the
505  *     frontend for other reasons.
506  *
507  * (4) Permanently disconnected sub-trees are cleaned up by the front-end.
508  *     Deleted-but-open inodes can still be individually flushed via the
509  *     filesystem syncer.
510  *
511  * (5) Note that an unmodified child may still need the block table in its
512  *     parent updated (e.g. rename/move).  The child will have UPDATE set
513  *     in this case.
514  *
515  *			WARNING ON BREF MODIFY_TID/MIRROR_TID
516  *
517  * blockref.modify_tid is consistent only within a PFS, and will not be
518  * consistent during synchronization.  mirror_tid is consistent across the
519  * block device regardless of the PFS.
520  */
521 static void
522 hammer2_flush_core(hammer2_flush_info_t *info, hammer2_chain_t *chain,
523 		   int istop)
524 {
525 	hammer2_chain_t *parent;
526 	hammer2_dev_t *hmp;
527 	int diddeferral;
528 
529 	/*
530 	 * (1) Optimize downward recursion to locate nodes needing action.
531 	 *     Nothing to do if none of these flags are set.
532 	 */
533 	if ((chain->flags & HAMMER2_CHAIN_FLUSH_MASK) == 0) {
534 		if (hammer2_debug & 0x200) {
535 			if (info->debug == NULL)
536 				info->debug = chain;
537 		} else {
538 			return;
539 		}
540 	}
541 
542 	hmp = chain->hmp;
543 	diddeferral = info->diddeferral;
544 	parent = info->parent;		/* can be NULL */
545 
546 	/*
547 	 * Downward search recursion
548 	 */
549 	if (chain->flags & (HAMMER2_CHAIN_DEFERRED | HAMMER2_CHAIN_DELAYED)) {
550 		/*
551 		 * Already deferred.
552 		 */
553 		++info->diddeferral;
554 	} else if (info->depth == HAMMER2_FLUSH_DEPTH_LIMIT) {
555 		/*
556 		 * Recursion depth reached.
557 		 */
558 		KKASSERT((chain->flags & HAMMER2_CHAIN_DELAYED) == 0);
559 		hammer2_chain_ref(chain);
560 		TAILQ_INSERT_TAIL(&info->flushq, chain, flush_node);
561 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEFERRED);
562 		++info->diddeferral;
563 	} else if ((chain->flags & HAMMER2_CHAIN_PFSBOUNDARY) && istop == 0) {
564 		/*
565 		 * We do not recurse through PFSROOTs.  PFSROOT flushes are
566 		 * handled by the related pmp's (whether mounted or not,
567 		 * including during recovery).
568 		 *
569 		 * But we must still process the PFSROOT chains for block
570 		 * table updates in their parent (which IS part of our flush).
571 		 *
572 		 * Note that the volume root, vchain, does not set this flag.
573 		 */
574 		;
575 	} else if (chain->flags & HAMMER2_CHAIN_ONFLUSH) {
576 		/*
577 		 * Downward recursion search (actual flush occurs bottom-up).
578 		 * pre-clear ONFLUSH.  It can get set again due to races,
579 		 * which we want so the scan finds us again in the next flush.
580 		 * These races can also include
581 		 *
582 		 * Flush recursions stop at PFSROOT boundaries.  Each PFS
583 		 * must be individually flushed and then the root must
584 		 * be flushed.
585 		 */
586 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
587 		info->parent = chain;
588 		hammer2_spin_ex(&chain->core.spin);
589 		RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
590 			NULL, hammer2_flush_recurse, info);
591 		hammer2_spin_unex(&chain->core.spin);
592 		info->parent = parent;
593 		if (info->diddeferral)
594 			hammer2_chain_setflush(info->trans, chain);
595 	}
596 
597 	/*
598 	 * Now we are in the bottom-up part of the recursion.
599 	 *
600 	 * Do not update chain if lower layers were deferred.
601 	 */
602 	if (info->diddeferral)
603 		goto done;
604 
605 	/*
606 	 * Propagate the DESTROY flag downwards.  This dummies up the flush
607 	 * code and tries to invalidate related buffer cache buffers to
608 	 * avoid the disk write.
609 	 */
610 	if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
611 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
612 
613 	/*
614 	 * Chain was already modified or has become modified, flush it out.
615 	 */
616 again:
617 	if ((hammer2_debug & 0x200) &&
618 	    info->debug &&
619 	    (chain->flags & (HAMMER2_CHAIN_MODIFIED | HAMMER2_CHAIN_UPDATE))) {
620 		hammer2_chain_t *scan = chain;
621 
622 		kprintf("DISCONNECTED FLUSH %p->%p\n", info->debug, chain);
623 		while (scan) {
624 			kprintf("    chain %p [%08x] bref=%016jx:%02x\n",
625 				scan, scan->flags,
626 				scan->bref.key, scan->bref.type);
627 			if (scan == info->debug)
628 				break;
629 			scan = scan->parent;
630 		}
631 	}
632 
633 	if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
634 		/*
635 		 * Dispose of the modified bit.
636 		 *
637 		 * UPDATE should already be set.
638 		 * bref.mirror_tid should already be set.
639 		 */
640 		KKASSERT((chain->flags & HAMMER2_CHAIN_UPDATE) ||
641 			 chain == &hmp->vchain);
642 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
643 
644 		/*
645 		 * Manage threads waiting for excessive dirty memory to
646 		 * be retired.
647 		 */
648 		if (chain->pmp)
649 			hammer2_pfs_memory_wakeup(chain->pmp);
650 
651 		if ((chain->flags & HAMMER2_CHAIN_UPDATE) ||
652 		    chain == &hmp->vchain ||
653 		    chain == &hmp->fchain) {
654 			/*
655 			 * Drop the ref from the MODIFIED bit we cleared,
656 			 * net -1 ref.
657 			 */
658 			hammer2_chain_drop(chain);
659 		} else {
660 			/*
661 			 * Drop the ref from the MODIFIED bit we cleared and
662 			 * set a ref for the UPDATE bit we are setting.  Net
663 			 * 0 refs.
664 			 */
665 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
666 		}
667 
668 		/*
669 		 * Issue the flush.  This is indirect via the DIO.
670 		 *
671 		 * NOTE: A DELETED node that reaches this point must be
672 		 *	 flushed for synchronization point consistency.
673 		 *
674 		 * NOTE: Even though MODIFIED was already set, the related DIO
675 		 *	 might not be dirty due to a system buffer cache
676 		 *	 flush and must be set dirty if we are going to make
677 		 *	 further modifications to the buffer.  Chains with
678 		 *	 embedded data don't need this.
679 		 */
680 		if (hammer2_debug & 0x1000) {
681 			kprintf("Flush %p.%d %016jx/%d sync_xid=%08x "
682 				"data=%016jx\n",
683 				chain, chain->bref.type,
684 				chain->bref.key, chain->bref.keybits,
685 				info->sync_xid,
686 				chain->bref.data_off);
687 		}
688 		if (hammer2_debug & 0x2000) {
689 			Debugger("Flush hell");
690 		}
691 
692 		/*
693 		 * Update chain CRCs for flush.
694 		 *
695 		 * NOTE: Volume headers are NOT flushed here as they require
696 		 *	 special processing.
697 		 */
698 		switch(chain->bref.type) {
699 		case HAMMER2_BREF_TYPE_FREEMAP:
700 			/*
701 			 * Update the volume header's freemap_tid to the
702 			 * freemap's flushing mirror_tid.
703 			 *
704 			 * (note: embedded data, do not call setdirty)
705 			 */
706 			KKASSERT(hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED);
707 			KKASSERT(chain == &hmp->fchain);
708 			hmp->voldata.freemap_tid = chain->bref.mirror_tid;
709 			kprintf("sync freemap mirror_tid %08jx\n",
710 				(intmax_t)chain->bref.mirror_tid);
711 
712 			/*
713 			 * The freemap can be flushed independently of the
714 			 * main topology, but for the case where it is
715 			 * flushed in the same transaction, and flushed
716 			 * before vchain (a case we want to allow for
717 			 * performance reasons), make sure modifications
718 			 * made during the flush under vchain use a new
719 			 * transaction id.
720 			 *
721 			 * Otherwise the mount recovery code will get confused.
722 			 */
723 			++hmp->voldata.mirror_tid;
724 			break;
725 		case HAMMER2_BREF_TYPE_VOLUME:
726 			/*
727 			 * The free block table is flushed by
728 			 * hammer2_vfs_sync() before it flushes vchain.
729 			 * We must still hold fchain locked while copying
730 			 * voldata to volsync, however.
731 			 *
732 			 * (note: embedded data, do not call setdirty)
733 			 */
734 			hammer2_chain_lock(&hmp->fchain,
735 					   HAMMER2_RESOLVE_ALWAYS);
736 			hammer2_voldata_lock(hmp);
737 			kprintf("sync volume  mirror_tid %08jx\n",
738 				(intmax_t)chain->bref.mirror_tid);
739 
740 			/*
741 			 * Update the volume header's mirror_tid to the
742 			 * main topology's flushing mirror_tid.  It is
743 			 * possible that voldata.mirror_tid is already
744 			 * beyond bref.mirror_tid due to the bump we made
745 			 * above in BREF_TYPE_FREEMAP.
746 			 */
747 			if (hmp->voldata.mirror_tid < chain->bref.mirror_tid) {
748 				hmp->voldata.mirror_tid =
749 					chain->bref.mirror_tid;
750 			}
751 
752 			/*
753 			 * The volume header is flushed manually by the
754 			 * syncer, not here.  All we do here is adjust the
755 			 * crc's.
756 			 */
757 			KKASSERT(chain->data != NULL);
758 			KKASSERT(chain->dio == NULL);
759 
760 			hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT1]=
761 				hammer2_icrc32(
762 					(char *)&hmp->voldata +
763 					 HAMMER2_VOLUME_ICRC1_OFF,
764 					HAMMER2_VOLUME_ICRC1_SIZE);
765 			hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT0]=
766 				hammer2_icrc32(
767 					(char *)&hmp->voldata +
768 					 HAMMER2_VOLUME_ICRC0_OFF,
769 					HAMMER2_VOLUME_ICRC0_SIZE);
770 			hmp->voldata.icrc_volheader =
771 				hammer2_icrc32(
772 					(char *)&hmp->voldata +
773 					 HAMMER2_VOLUME_ICRCVH_OFF,
774 					HAMMER2_VOLUME_ICRCVH_SIZE);
775 
776 			kprintf("syncvolhdr %016jx %016jx\n",
777 				hmp->voldata.mirror_tid,
778 				hmp->vchain.bref.mirror_tid);
779 			hmp->volsync = hmp->voldata;
780 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_VOLUMESYNC);
781 			hammer2_voldata_unlock(hmp);
782 			hammer2_chain_unlock(&hmp->fchain);
783 			break;
784 		case HAMMER2_BREF_TYPE_DATA:
785 			/*
786 			 * Data elements have already been flushed via the
787 			 * logical file buffer cache.  Their hash was set in
788 			 * the bref by the vop_write code.  Do not re-dirty.
789 			 *
790 			 * Make sure any device buffer(s) have been flushed
791 			 * out here (there aren't usually any to flush) XXX.
792 			 */
793 			break;
794 		case HAMMER2_BREF_TYPE_INDIRECT:
795 		case HAMMER2_BREF_TYPE_FREEMAP_NODE:
796 		case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
797 			/*
798 			 * Buffer I/O will be cleaned up when the volume is
799 			 * flushed (but the kernel is free to flush it before
800 			 * then, as well).
801 			 */
802 			KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
803 			hammer2_chain_setcheck(chain, chain->data);
804 			break;
805 		case HAMMER2_BREF_TYPE_INODE:
806 			/*
807 			 * NOTE: We must call io_setdirty() to make any late
808 			 *	 changes to the inode data, the system might
809 			 *	 have already flushed the buffer.
810 			 */
811 			if (chain->data->ipdata.meta.op_flags &
812 			    HAMMER2_OPFLAG_PFSROOT) {
813 				/*
814 				 * non-NULL pmp if mounted as a PFS.  We must
815 				 * sync fields cached in the pmp? XXX
816 				 */
817 				hammer2_inode_data_t *ipdata;
818 
819 				hammer2_io_setdirty(chain->dio);
820 				ipdata = &chain->data->ipdata;
821 				if (chain->pmp) {
822 					ipdata->meta.pfs_inum =
823 						chain->pmp->inode_tid;
824 				}
825 			} else {
826 				/* can't be mounted as a PFS */
827 			}
828 
829 			KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
830 			hammer2_chain_setcheck(chain, chain->data);
831 			break;
832 		default:
833 			KKASSERT(chain->flags & HAMMER2_CHAIN_EMBEDDED);
834 			panic("hammer2_flush_core: unsupported "
835 			      "embedded bref %d",
836 			      chain->bref.type);
837 			/* NOT REACHED */
838 		}
839 
840 		/*
841 		 * If the chain was destroyed try to avoid unnecessary I/O.
842 		 * (this only really works if the DIO system buffer is the
843 		 * same size as chain->bytes).
844 		 */
845 		if ((chain->flags & HAMMER2_CHAIN_DESTROY) && chain->dio) {
846 			hammer2_io_setinval(chain->dio, chain->bytes);
847 		}
848 	}
849 
850 	/*
851 	 * If UPDATE is set the parent block table may need to be updated.
852 	 *
853 	 * NOTE: UPDATE may be set on vchain or fchain in which case
854 	 *	 parent could be NULL.  It's easiest to allow the case
855 	 *	 and test for NULL.  parent can also wind up being NULL
856 	 *	 due to a deletion so we need to handle the case anyway.
857 	 *
858 	 * If no parent exists we can just clear the UPDATE bit.  If the
859 	 * chain gets reattached later on the bit will simply get set
860 	 * again.
861 	 */
862 	if ((chain->flags & HAMMER2_CHAIN_UPDATE) && parent == NULL) {
863 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
864 		hammer2_chain_drop(chain);
865 	}
866 
867 	/*
868 	 * The chain may need its blockrefs updated in the parent.  This
869 	 * requires some fancy footwork.
870 	 */
871 	if (chain->flags & HAMMER2_CHAIN_UPDATE) {
872 		hammer2_blockref_t *base;
873 		int count;
874 
875 		/*
876 		 * Both parent and chain must be locked.  This requires
877 		 * temporarily unlocking the chain.  We have to deal with
878 		 * the case where the chain might be reparented or modified
879 		 * while it was unlocked.
880 		 */
881 		hammer2_chain_unlock(chain);
882 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
883 		hammer2_chain_lock(chain, HAMMER2_RESOLVE_MAYBE);
884 		if (chain->parent != parent) {
885 			kprintf("PARENT MISMATCH ch=%p p=%p/%p\n",
886 				chain, chain->parent, parent);
887 			hammer2_chain_unlock(parent);
888 			goto done;
889 		}
890 
891 		/*
892 		 * Check race condition.  If someone got in and modified
893 		 * it again while it was unlocked, we have to loop up.
894 		 */
895 		if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
896 			hammer2_chain_unlock(parent);
897 			kprintf("hammer2_flush: chain %p flush-mod race\n",
898 				chain);
899 			goto again;
900 		}
901 
902 		/*
903 		 * Clear UPDATE flag, mark parent modified, update its
904 		 * modify_tid if necessary, and adjust the parent blockmap.
905 		 */
906 		if (chain->flags & HAMMER2_CHAIN_UPDATE) {
907 			atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
908 			hammer2_chain_drop(chain);
909 		}
910 
911 		/*
912 		 * (optional code)
913 		 *
914 		 * Avoid actually modifying and updating the parent if it
915 		 * was flagged for destruction.  This can greatly reduce
916 		 * disk I/O in large tree removals because the
917 		 * hammer2_io_setinval() call in the upward recursion
918 		 * (see MODIFIED code above) can only handle a few cases.
919 		 */
920 		if (parent->flags & HAMMER2_CHAIN_DESTROY) {
921 			if (parent->bref.modify_tid < chain->bref.modify_tid) {
922 				parent->bref.modify_tid =
923 					chain->bref.modify_tid;
924 			}
925 			atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
926 							HAMMER2_CHAIN_BMAPUPD);
927 			hammer2_chain_unlock(parent);
928 			goto skipupdate;
929 		}
930 
931 		/*
932 		 * We are updating the parent's blockmap, the parent must
933 		 * be set modified.
934 		 */
935 		hammer2_chain_modify(info->trans, parent, 0);
936 		if (parent->bref.modify_tid < chain->bref.modify_tid)
937 			parent->bref.modify_tid = chain->bref.modify_tid;
938 
939 		/*
940 		 * Calculate blockmap pointer
941 		 */
942 		switch(parent->bref.type) {
943 		case HAMMER2_BREF_TYPE_INODE:
944 			/*
945 			 * Access the inode's block array.  However, there is
946 			 * no block array if the inode is flagged DIRECTDATA.
947 			 */
948 			if (parent->data &&
949 			    (parent->data->ipdata.meta.op_flags &
950 			     HAMMER2_OPFLAG_DIRECTDATA) == 0) {
951 				base = &parent->data->
952 					ipdata.u.blockset.blockref[0];
953 			} else {
954 				base = NULL;
955 			}
956 			count = HAMMER2_SET_COUNT;
957 			break;
958 		case HAMMER2_BREF_TYPE_INDIRECT:
959 		case HAMMER2_BREF_TYPE_FREEMAP_NODE:
960 			if (parent->data)
961 				base = &parent->data->npdata[0];
962 			else
963 				base = NULL;
964 			count = parent->bytes / sizeof(hammer2_blockref_t);
965 			break;
966 		case HAMMER2_BREF_TYPE_VOLUME:
967 			base = &chain->hmp->voldata.sroot_blockset.blockref[0];
968 			count = HAMMER2_SET_COUNT;
969 			break;
970 		case HAMMER2_BREF_TYPE_FREEMAP:
971 			base = &parent->data->npdata[0];
972 			count = HAMMER2_SET_COUNT;
973 			break;
974 		default:
975 			base = NULL;
976 			count = 0;
977 			panic("hammer2_flush_core: "
978 			      "unrecognized blockref type: %d",
979 			      parent->bref.type);
980 		}
981 
982 		/*
983 		 * Blocktable updates
984 		 *
985 		 * We synchronize pending statistics at this time.  Delta
986 		 * adjustments designated for the current and upper level
987 		 * are synchronized.
988 		 */
989 		if (base && (chain->flags & HAMMER2_CHAIN_BMAPUPD)) {
990 			if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
991 				hammer2_spin_ex(&parent->core.spin);
992 				hammer2_base_delete(info->trans, parent,
993 						    base, count,
994 						    &info->cache_index, chain);
995 				hammer2_spin_unex(&parent->core.spin);
996 				/* base_delete clears both bits */
997 			} else {
998 				atomic_clear_int(&chain->flags,
999 						 HAMMER2_CHAIN_BMAPUPD);
1000 			}
1001 		}
1002 		if (base && (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
1003 			hammer2_spin_ex(&parent->core.spin);
1004 			hammer2_base_insert(info->trans, parent,
1005 					    base, count,
1006 					    &info->cache_index, chain);
1007 			hammer2_spin_unex(&parent->core.spin);
1008 			/* base_insert sets BMAPPED */
1009 		}
1010 		hammer2_chain_unlock(parent);
1011 	}
1012 skipupdate:
1013 	;
1014 
1015 	/*
1016 	 * Final cleanup after flush
1017 	 */
1018 done:
1019 	KKASSERT(chain->refs > 0);
1020 	if (hammer2_debug & 0x200) {
1021 		if (info->debug == chain)
1022 			info->debug = NULL;
1023 	}
1024 }
1025 
1026 /*
1027  * Flush recursion helper, called from flush_core, calls flush_core.
1028  *
1029  * Flushes the children of the caller's chain (info->parent), restricted
1030  * by sync_tid.  Set info->domodify if the child's blockref must propagate
1031  * back up to the parent.
1032  *
1033  * Ripouts can move child from rbtree to dbtree or dbq but the caller's
1034  * flush scan order prevents any chains from being lost.  A child can be
1035  * executes more than once.
1036  *
1037  * WARNING! If we do not call hammer2_flush_core() we must update
1038  *	    bref.mirror_tid ourselves to indicate that the flush has
1039  *	    processed the child.
1040  *
1041  * WARNING! parent->core spinlock is held on entry and return.
1042  */
1043 static int
1044 hammer2_flush_recurse(hammer2_chain_t *child, void *data)
1045 {
1046 	hammer2_flush_info_t *info = data;
1047 	/*hammer2_trans_t *trans = info->trans;*/
1048 	hammer2_chain_t *parent = info->parent;
1049 
1050 	/*
1051 	 * (child can never be fchain or vchain so a special check isn't
1052 	 *  needed).
1053 	 *
1054 	 * We must ref the child before unlocking the spinlock.
1055 	 *
1056 	 * The caller has added a ref to the parent so we can temporarily
1057 	 * unlock it in order to lock the child.
1058 	 */
1059 	hammer2_chain_ref(child);
1060 	hammer2_spin_unex(&parent->core.spin);
1061 
1062 	hammer2_chain_unlock(parent);
1063 	hammer2_chain_lock(child, HAMMER2_RESOLVE_MAYBE);
1064 
1065 	/*
1066 	 * Recurse and collect deferral data.  We're in the media flush,
1067 	 * this can cross PFS boundaries.
1068 	 */
1069 	if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1070 		++info->depth;
1071 		hammer2_flush_core(info, child, 0);
1072 		--info->depth;
1073 	} else if (hammer2_debug & 0x200) {
1074 		if (info->debug == NULL)
1075 			info->debug = child;
1076 		++info->depth;
1077 		hammer2_flush_core(info, child, 0);
1078 		--info->depth;
1079 		if (info->debug == child)
1080 			info->debug = NULL;
1081 	}
1082 
1083 	/*
1084 	 * Relock to continue the loop
1085 	 */
1086 	hammer2_chain_unlock(child);
1087 	hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
1088 	hammer2_chain_drop(child);
1089 	KKASSERT(info->parent == parent);
1090 	hammer2_spin_ex(&parent->core.spin);
1091 
1092 	return (0);
1093 }
1094