xref: /dflybsd-src/sys/vfs/hammer2/hammer2_flush.c (revision d63676ccce44debffd22823892f437449f6acdad)
1 /*
2  * Copyright (c) 2011-2015 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 	int		depth;
70 	int		diddeferral;
71 	int		cache_index;
72 	struct h2_flush_list flushq;
73 	hammer2_chain_t	*debug;
74 };
75 
76 typedef struct hammer2_flush_info hammer2_flush_info_t;
77 
78 static void hammer2_flush_core(hammer2_flush_info_t *info,
79 				hammer2_chain_t *chain, int deleting);
80 static int hammer2_flush_recurse(hammer2_chain_t *child, void *data);
81 
82 /*
83  * Any per-pfs transaction initialization goes here.
84  */
85 void
86 hammer2_trans_manage_init(hammer2_pfs_t *pmp)
87 {
88 }
89 
90 /*
91  * Transaction support for any modifying operation.  Transactions are used
92  * in the pmp layer by the frontend and in the spmp layer by the backend.
93  *
94  * 0			- Normal transaction, interlocked against flush
95  *			  transaction.
96  *
97  * TRANS_ISFLUSH	- Flush transaction, interlocked against normal
98  *			  transaction.
99  *
100  * TRANS_BUFCACHE	- Buffer cache transaction, no interlock.
101  *
102  * Initializing a new transaction allocates a transaction ID.  Typically
103  * passed a pmp (hmp passed as NULL), indicating a cluster transaction.  Can
104  * be passed a NULL pmp and non-NULL hmp to indicate a transaction on a single
105  * media target.  The latter mode is used by the recovery code.
106  *
107  * TWO TRANSACTION IDs can run concurrently, where one is a flush and the
108  * other is a set of any number of concurrent filesystem operations.  We
109  * can either have <running_fs_ops> + <waiting_flush> + <blocked_fs_ops>
110  * or we can have <running_flush> + <concurrent_fs_ops>.
111  *
112  * During a flush, new fs_ops are only blocked until the fs_ops prior to
113  * the flush complete.  The new fs_ops can then run concurrent with the flush.
114  *
115  * Buffer-cache transactions operate as fs_ops but never block.  A
116  * buffer-cache flush will run either before or after the current pending
117  * flush depending on its state.
118  */
119 void
120 hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags)
121 {
122 	uint32_t oflags;
123 	uint32_t nflags;
124 	int dowait;
125 
126 	for (;;) {
127 		oflags = pmp->trans.flags;
128 		cpu_ccfence();
129 		dowait = 0;
130 
131 		if (flags & HAMMER2_TRANS_ISFLUSH) {
132 			/*
133 			 * Requesting flush transaction.  Wait for all
134 			 * currently running transactions to finish.
135 			 */
136 			if (oflags & HAMMER2_TRANS_MASK) {
137 				nflags = oflags | HAMMER2_TRANS_FPENDING |
138 						  HAMMER2_TRANS_WAITING;
139 				dowait = 1;
140 			} else {
141 				nflags = (oflags | flags) + 1;
142 			}
143 			++pmp->modify_tid;
144 		} else if (flags & HAMMER2_TRANS_BUFCACHE) {
145 			/*
146 			 * Requesting strategy transaction.  Generally
147 			 * allowed in all situations unless a flush
148 			 * is running without the preflush flag.
149 			 */
150 			if ((oflags & (HAMMER2_TRANS_ISFLUSH |
151 				       HAMMER2_TRANS_PREFLUSH)) ==
152 			    HAMMER2_TRANS_ISFLUSH) {
153 				nflags = oflags | HAMMER2_TRANS_WAITING;
154 				dowait = 1;
155 			} else {
156 				nflags = (oflags | flags) + 1;
157 			}
158 		} else {
159 			/*
160 			 * Requesting normal transaction.  Wait for any
161 			 * flush to finish before allowing.
162 			 */
163 			if (oflags & HAMMER2_TRANS_ISFLUSH) {
164 				nflags = oflags | HAMMER2_TRANS_WAITING;
165 				dowait = 1;
166 			} else {
167 				nflags = (oflags | flags) + 1;
168 			}
169 		}
170 		if (dowait)
171 			tsleep_interlock(&pmp->trans.sync_wait, 0);
172 		if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
173 			if (dowait == 0)
174 				break;
175 			tsleep(&pmp->trans.sync_wait, PINTERLOCKED,
176 			       "h2trans", hz);
177 		} else {
178 			cpu_pause();
179 		}
180 		/* retry */
181 	}
182 }
183 
184 void
185 hammer2_trans_done(hammer2_pfs_t *pmp)
186 {
187 	uint32_t oflags;
188 	uint32_t nflags;
189 
190 	for (;;) {
191 		oflags = pmp->trans.flags;
192 		cpu_ccfence();
193 		KKASSERT(oflags & HAMMER2_TRANS_MASK);
194 		if ((oflags & HAMMER2_TRANS_MASK) == 1) {
195 			/*
196 			 * This was the last transaction
197 			 */
198 			nflags = (oflags - 1) & ~(HAMMER2_TRANS_ISFLUSH |
199 						  HAMMER2_TRANS_BUFCACHE |
200 						  HAMMER2_TRANS_PREFLUSH |
201 						  HAMMER2_TRANS_FPENDING |
202 						  HAMMER2_TRANS_WAITING);
203 		} else {
204 			/*
205 			 * Still transactions pending
206 			 */
207 			nflags = oflags - 1;
208 		}
209 		if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
210 			if ((nflags & HAMMER2_TRANS_MASK) == 0 &&
211 			    (oflags & HAMMER2_TRANS_WAITING)) {
212 				wakeup(&pmp->trans.sync_wait);
213 			}
214 			break;
215 		} else {
216 			cpu_pause();
217 		}
218 		/* retry */
219 	}
220 }
221 
222 /*
223  * Obtain new, unique inode number (not serialized by caller).
224  */
225 hammer2_tid_t
226 hammer2_trans_newinum(hammer2_pfs_t *pmp)
227 {
228 	hammer2_tid_t tid;
229 
230 	KKASSERT(sizeof(long) == 8);
231 	tid = atomic_fetchadd_long(&pmp->inode_tid, 1);
232 
233 	return tid;
234 }
235 
236 /*
237  * Assert that a strategy call is ok here.  Strategy calls are legal
238  *
239  * (1) In a normal transaction.
240  * (2) In a flush transaction only if PREFLUSH is also set.
241  */
242 void
243 hammer2_trans_assert_strategy(hammer2_pfs_t *pmp)
244 {
245 	KKASSERT((pmp->trans.flags & HAMMER2_TRANS_ISFLUSH) == 0 ||
246 		 (pmp->trans.flags & HAMMER2_TRANS_PREFLUSH));
247 }
248 
249 
250 /*
251  * Chains undergoing destruction are removed from the in-memory topology.
252  * To avoid getting lost these chains are placed on the delayed flush
253  * queue which will properly dispose of them.
254  *
255  * We do this instead of issuing an immediate flush in order to give
256  * recursive deletions (rm -rf, etc) a chance to remove more of the
257  * hierarchy, potentially allowing an enormous amount of write I/O to
258  * be avoided.
259  */
260 void
261 hammer2_delayed_flush(hammer2_chain_t *chain)
262 {
263 	if ((chain->flags & HAMMER2_CHAIN_DELAYED) == 0) {
264 		hammer2_spin_ex(&chain->hmp->list_spin);
265 		if ((chain->flags & (HAMMER2_CHAIN_DELAYED |
266 				     HAMMER2_CHAIN_DEFERRED)) == 0) {
267 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELAYED |
268 						      HAMMER2_CHAIN_DEFERRED);
269 			TAILQ_INSERT_TAIL(&chain->hmp->flushq,
270 					  chain, flush_node);
271 			hammer2_chain_ref(chain);
272 		}
273 		hammer2_spin_unex(&chain->hmp->list_spin);
274 	}
275 }
276 
277 /*
278  * Flush the chain and all modified sub-chains through the specified
279  * synchronization point, propagating parent chain modifications, modify_tid,
280  * and mirror_tid updates back up as needed.
281  *
282  * Caller must have already vetted synchronization points to ensure they
283  * are properly flushed.  Only snapshots and cluster flushes can create
284  * these sorts of synchronization points.
285  *
286  * This routine can be called from several places but the most important
287  * is from VFS_SYNC.
288  *
289  * chain is locked on call and will remain locked on return.  The chain's
290  * UPDATE flag indicates that its parent's block table (which is not yet
291  * part of the flush) should be updated.  The chain may be replaced by
292  * the call if it was modified.
293  */
294 void
295 hammer2_flush(hammer2_chain_t *chain, int istop)
296 {
297 	hammer2_chain_t *scan;
298 	hammer2_flush_info_t info;
299 	hammer2_dev_t *hmp;
300 	int loops;
301 
302 	/*
303 	 * Execute the recursive flush and handle deferrals.
304 	 *
305 	 * Chains can be ridiculously long (thousands deep), so to
306 	 * avoid blowing out the kernel stack the recursive flush has a
307 	 * depth limit.  Elements at the limit are placed on a list
308 	 * for re-execution after the stack has been popped.
309 	 */
310 	bzero(&info, sizeof(info));
311 	TAILQ_INIT(&info.flushq);
312 	info.cache_index = -1;
313 
314 	/*
315 	 * Calculate parent (can be NULL), if not NULL the flush core
316 	 * expects the parent to be referenced so it can easily lock/unlock
317 	 * it without it getting ripped up.
318 	 */
319 	if ((info.parent = chain->parent) != NULL)
320 		hammer2_chain_ref(info.parent);
321 
322 	/*
323 	 * Extra ref needed because flush_core expects it when replacing
324 	 * chain.
325 	 */
326 	hammer2_chain_ref(chain);
327 	hmp = chain->hmp;
328 	loops = 0;
329 
330 	for (;;) {
331 		/*
332 		 * Move hmp->flushq to info.flushq if non-empty so it can
333 		 * be processed.
334 		 */
335 		if (TAILQ_FIRST(&hmp->flushq) != NULL) {
336 			hammer2_spin_ex(&chain->hmp->list_spin);
337 			TAILQ_CONCAT(&info.flushq, &hmp->flushq, flush_node);
338 			hammer2_spin_unex(&chain->hmp->list_spin);
339 		}
340 
341 		/*
342 		 * Unwind deep recursions which had been deferred.  This
343 		 * can leave the FLUSH_* bits set for these chains, which
344 		 * will be handled when we [re]flush chain after the unwind.
345 		 */
346 		while ((scan = TAILQ_FIRST(&info.flushq)) != NULL) {
347 			KKASSERT(scan->flags & HAMMER2_CHAIN_DEFERRED);
348 			TAILQ_REMOVE(&info.flushq, scan, flush_node);
349 			atomic_clear_int(&scan->flags, HAMMER2_CHAIN_DEFERRED |
350 						       HAMMER2_CHAIN_DELAYED);
351 
352 			/*
353 			 * Now that we've popped back up we can do a secondary
354 			 * recursion on the deferred elements.
355 			 *
356 			 * NOTE: hammer2_flush() may replace scan.
357 			 */
358 			if (hammer2_debug & 0x0040)
359 				kprintf("deferred flush %p\n", scan);
360 			hammer2_chain_lock(scan, HAMMER2_RESOLVE_MAYBE);
361 			hammer2_flush(scan, 0);
362 			hammer2_chain_unlock(scan);
363 			hammer2_chain_drop(scan);	/* ref from deferral */
364 		}
365 
366 		/*
367 		 * [re]flush chain.
368 		 */
369 		info.diddeferral = 0;
370 		hammer2_flush_core(&info, chain, istop);
371 
372 		/*
373 		 * Only loop if deep recursions have been deferred.
374 		 */
375 		if (TAILQ_EMPTY(&info.flushq))
376 			break;
377 
378 		if (++loops % 1000 == 0) {
379 			kprintf("hammer2_flush: excessive loops on %p\n",
380 				chain);
381 			if (hammer2_debug & 0x100000)
382 				Debugger("hell4");
383 		}
384 	}
385 	hammer2_chain_drop(chain);
386 	if (info.parent)
387 		hammer2_chain_drop(info.parent);
388 }
389 
390 /*
391  * This is the core of the chain flushing code.  The chain is locked by the
392  * caller and must also have an extra ref on it by the caller, and remains
393  * locked and will have an extra ref on return.  Upon return, the caller can
394  * test the UPDATE bit on the child to determine if the parent needs updating.
395  *
396  * (1) Determine if this node is a candidate for the flush, return if it is
397  *     not.  fchain and vchain are always candidates for the flush.
398  *
399  * (2) If we recurse too deep the chain is entered onto the deferral list and
400  *     the current flush stack is aborted until after the deferral list is
401  *     run.
402  *
403  * (3) Recursively flush live children (rbtree).  This can create deferrals.
404  *     A successful flush clears the MODIFIED and UPDATE bits on the children
405  *     and typically causes the parent to be marked MODIFIED as the children
406  *     update the parent's block table.  A parent might already be marked
407  *     MODIFIED due to a deletion (whos blocktable update in the parent is
408  *     handled by the frontend), or if the parent itself is modified by the
409  *     frontend for other reasons.
410  *
411  * (4) Permanently disconnected sub-trees are cleaned up by the front-end.
412  *     Deleted-but-open inodes can still be individually flushed via the
413  *     filesystem syncer.
414  *
415  * (5) Note that an unmodified child may still need the block table in its
416  *     parent updated (e.g. rename/move).  The child will have UPDATE set
417  *     in this case.
418  *
419  *			WARNING ON BREF MODIFY_TID/MIRROR_TID
420  *
421  * blockref.modify_tid is consistent only within a PFS, and will not be
422  * consistent during synchronization.  mirror_tid is consistent across the
423  * block device regardless of the PFS.
424  */
425 static void
426 hammer2_flush_core(hammer2_flush_info_t *info, hammer2_chain_t *chain,
427 		   int istop)
428 {
429 	hammer2_chain_t *parent;
430 	hammer2_dev_t *hmp;
431 	int diddeferral;
432 
433 	/*
434 	 * (1) Optimize downward recursion to locate nodes needing action.
435 	 *     Nothing to do if none of these flags are set.
436 	 */
437 	if ((chain->flags & HAMMER2_CHAIN_FLUSH_MASK) == 0) {
438 		if (hammer2_debug & 0x200) {
439 			if (info->debug == NULL)
440 				info->debug = chain;
441 		} else {
442 			return;
443 		}
444 	}
445 
446 	hmp = chain->hmp;
447 	diddeferral = info->diddeferral;
448 	parent = info->parent;		/* can be NULL */
449 
450 	/*
451 	 * Downward search recursion
452 	 */
453 	if (chain->flags & (HAMMER2_CHAIN_DEFERRED | HAMMER2_CHAIN_DELAYED)) {
454 		/*
455 		 * Already deferred.
456 		 */
457 		++info->diddeferral;
458 	} else if (info->depth == HAMMER2_FLUSH_DEPTH_LIMIT) {
459 		/*
460 		 * Recursion depth reached.
461 		 */
462 		KKASSERT((chain->flags & HAMMER2_CHAIN_DELAYED) == 0);
463 		hammer2_chain_ref(chain);
464 		TAILQ_INSERT_TAIL(&info->flushq, chain, flush_node);
465 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEFERRED);
466 		++info->diddeferral;
467 	} else if ((chain->flags & HAMMER2_CHAIN_PFSBOUNDARY) && istop == 0) {
468 		/*
469 		 * We do not recurse through PFSROOTs.  PFSROOT flushes are
470 		 * handled by the related pmp's (whether mounted or not,
471 		 * including during recovery).
472 		 *
473 		 * But we must still process the PFSROOT chains for block
474 		 * table updates in their parent (which IS part of our flush).
475 		 *
476 		 * Note that the volume root, vchain, does not set this flag.
477 		 */
478 		;
479 	} else if (chain->flags & HAMMER2_CHAIN_ONFLUSH) {
480 		/*
481 		 * Downward recursion search (actual flush occurs bottom-up).
482 		 * pre-clear ONFLUSH.  It can get set again due to races,
483 		 * which we want so the scan finds us again in the next flush.
484 		 * These races can also include
485 		 *
486 		 * Flush recursions stop at PFSROOT boundaries.  Each PFS
487 		 * must be individually flushed and then the root must
488 		 * be flushed.
489 		 */
490 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
491 		info->parent = chain;
492 		hammer2_spin_ex(&chain->core.spin);
493 		RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
494 			NULL, hammer2_flush_recurse, info);
495 		hammer2_spin_unex(&chain->core.spin);
496 		info->parent = parent;
497 		if (info->diddeferral)
498 			hammer2_chain_setflush(chain);
499 	}
500 
501 	/*
502 	 * Now we are in the bottom-up part of the recursion.
503 	 *
504 	 * Do not update chain if lower layers were deferred.
505 	 */
506 	if (info->diddeferral)
507 		goto done;
508 
509 	/*
510 	 * Propagate the DESTROY flag downwards.  This dummies up the flush
511 	 * code and tries to invalidate related buffer cache buffers to
512 	 * avoid the disk write.
513 	 */
514 	if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
515 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
516 
517 	/*
518 	 * Chain was already modified or has become modified, flush it out.
519 	 */
520 again:
521 	if ((hammer2_debug & 0x200) &&
522 	    info->debug &&
523 	    (chain->flags & (HAMMER2_CHAIN_MODIFIED | HAMMER2_CHAIN_UPDATE))) {
524 		hammer2_chain_t *scan = chain;
525 
526 		kprintf("DISCONNECTED FLUSH %p->%p\n", info->debug, chain);
527 		while (scan) {
528 			kprintf("    chain %p [%08x] bref=%016jx:%02x\n",
529 				scan, scan->flags,
530 				scan->bref.key, scan->bref.type);
531 			if (scan == info->debug)
532 				break;
533 			scan = scan->parent;
534 		}
535 	}
536 
537 	if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
538 		/*
539 		 * Dispose of the modified bit.
540 		 *
541 		 * UPDATE should already be set.
542 		 * bref.mirror_tid should already be set.
543 		 */
544 		KKASSERT((chain->flags & HAMMER2_CHAIN_UPDATE) ||
545 			 chain == &hmp->vchain);
546 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
547 
548 		/*
549 		 * Manage threads waiting for excessive dirty memory to
550 		 * be retired.
551 		 */
552 		if (chain->pmp)
553 			hammer2_pfs_memory_wakeup(chain->pmp);
554 
555 		if ((chain->flags & HAMMER2_CHAIN_UPDATE) ||
556 		    chain == &hmp->vchain ||
557 		    chain == &hmp->fchain) {
558 			/*
559 			 * Drop the ref from the MODIFIED bit we cleared,
560 			 * net -1 ref.
561 			 */
562 			hammer2_chain_drop(chain);
563 		} else {
564 			/*
565 			 * Drop the ref from the MODIFIED bit we cleared and
566 			 * set a ref for the UPDATE bit we are setting.  Net
567 			 * 0 refs.
568 			 */
569 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
570 		}
571 
572 		/*
573 		 * Issue the flush.  This is indirect via the DIO.
574 		 *
575 		 * NOTE: A DELETED node that reaches this point must be
576 		 *	 flushed for synchronization point consistency.
577 		 *
578 		 * NOTE: Even though MODIFIED was already set, the related DIO
579 		 *	 might not be dirty due to a system buffer cache
580 		 *	 flush and must be set dirty if we are going to make
581 		 *	 further modifications to the buffer.  Chains with
582 		 *	 embedded data don't need this.
583 		 */
584 		if (hammer2_debug & 0x1000) {
585 			kprintf("Flush %p.%d %016jx/%d data=%016jx",
586 				chain, chain->bref.type,
587 				(uintmax_t)chain->bref.key,
588 				chain->bref.keybits,
589 				(uintmax_t)chain->bref.data_off);
590 		}
591 		if (hammer2_debug & 0x2000) {
592 			Debugger("Flush hell");
593 		}
594 
595 		/*
596 		 * Update chain CRCs for flush.
597 		 *
598 		 * NOTE: Volume headers are NOT flushed here as they require
599 		 *	 special processing.
600 		 */
601 		switch(chain->bref.type) {
602 		case HAMMER2_BREF_TYPE_FREEMAP:
603 			/*
604 			 * Update the volume header's freemap_tid to the
605 			 * freemap's flushing mirror_tid.
606 			 *
607 			 * (note: embedded data, do not call setdirty)
608 			 */
609 			KKASSERT(hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED);
610 			KKASSERT(chain == &hmp->fchain);
611 			hmp->voldata.freemap_tid = chain->bref.mirror_tid;
612 			kprintf("sync freemap mirror_tid %08jx\n",
613 				(intmax_t)chain->bref.mirror_tid);
614 
615 			/*
616 			 * The freemap can be flushed independently of the
617 			 * main topology, but for the case where it is
618 			 * flushed in the same transaction, and flushed
619 			 * before vchain (a case we want to allow for
620 			 * performance reasons), make sure modifications
621 			 * made during the flush under vchain use a new
622 			 * transaction id.
623 			 *
624 			 * Otherwise the mount recovery code will get confused.
625 			 */
626 			++hmp->voldata.mirror_tid;
627 			break;
628 		case HAMMER2_BREF_TYPE_VOLUME:
629 			/*
630 			 * The free block table is flushed by
631 			 * hammer2_vfs_sync() before it flushes vchain.
632 			 * We must still hold fchain locked while copying
633 			 * voldata to volsync, however.
634 			 *
635 			 * (note: embedded data, do not call setdirty)
636 			 */
637 			hammer2_chain_lock(&hmp->fchain,
638 					   HAMMER2_RESOLVE_ALWAYS);
639 			hammer2_voldata_lock(hmp);
640 			kprintf("sync volume  mirror_tid %08jx\n",
641 				(intmax_t)chain->bref.mirror_tid);
642 
643 			/*
644 			 * Update the volume header's mirror_tid to the
645 			 * main topology's flushing mirror_tid.  It is
646 			 * possible that voldata.mirror_tid is already
647 			 * beyond bref.mirror_tid due to the bump we made
648 			 * above in BREF_TYPE_FREEMAP.
649 			 */
650 			if (hmp->voldata.mirror_tid < chain->bref.mirror_tid) {
651 				hmp->voldata.mirror_tid =
652 					chain->bref.mirror_tid;
653 			}
654 
655 			/*
656 			 * The volume header is flushed manually by the
657 			 * syncer, not here.  All we do here is adjust the
658 			 * crc's.
659 			 */
660 			KKASSERT(chain->data != NULL);
661 			KKASSERT(chain->dio == NULL);
662 
663 			hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT1]=
664 				hammer2_icrc32(
665 					(char *)&hmp->voldata +
666 					 HAMMER2_VOLUME_ICRC1_OFF,
667 					HAMMER2_VOLUME_ICRC1_SIZE);
668 			hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT0]=
669 				hammer2_icrc32(
670 					(char *)&hmp->voldata +
671 					 HAMMER2_VOLUME_ICRC0_OFF,
672 					HAMMER2_VOLUME_ICRC0_SIZE);
673 			hmp->voldata.icrc_volheader =
674 				hammer2_icrc32(
675 					(char *)&hmp->voldata +
676 					 HAMMER2_VOLUME_ICRCVH_OFF,
677 					HAMMER2_VOLUME_ICRCVH_SIZE);
678 
679 			kprintf("syncvolhdr %016jx %016jx\n",
680 				hmp->voldata.mirror_tid,
681 				hmp->vchain.bref.mirror_tid);
682 			hmp->volsync = hmp->voldata;
683 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_VOLUMESYNC);
684 			hammer2_voldata_unlock(hmp);
685 			hammer2_chain_unlock(&hmp->fchain);
686 			break;
687 		case HAMMER2_BREF_TYPE_DATA:
688 			/*
689 			 * Data elements have already been flushed via the
690 			 * logical file buffer cache.  Their hash was set in
691 			 * the bref by the vop_write code.  Do not re-dirty.
692 			 *
693 			 * Make sure any device buffer(s) have been flushed
694 			 * out here (there aren't usually any to flush) XXX.
695 			 */
696 			break;
697 		case HAMMER2_BREF_TYPE_INDIRECT:
698 		case HAMMER2_BREF_TYPE_FREEMAP_NODE:
699 		case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
700 			/*
701 			 * Buffer I/O will be cleaned up when the volume is
702 			 * flushed (but the kernel is free to flush it before
703 			 * then, as well).
704 			 */
705 			KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
706 			hammer2_chain_setcheck(chain, chain->data);
707 			break;
708 		case HAMMER2_BREF_TYPE_INODE:
709 			/*
710 			 * NOTE: We must call io_setdirty() to make any late
711 			 *	 changes to the inode data, the system might
712 			 *	 have already flushed the buffer.
713 			 */
714 			if (chain->data->ipdata.meta.op_flags &
715 			    HAMMER2_OPFLAG_PFSROOT) {
716 				/*
717 				 * non-NULL pmp if mounted as a PFS.  We must
718 				 * sync fields cached in the pmp? XXX
719 				 */
720 				hammer2_inode_data_t *ipdata;
721 
722 				hammer2_io_setdirty(chain->dio);
723 				ipdata = &chain->data->ipdata;
724 				if (chain->pmp) {
725 					ipdata->meta.pfs_inum =
726 						chain->pmp->inode_tid;
727 				}
728 			} else {
729 				/* can't be mounted as a PFS */
730 			}
731 
732 			KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
733 			hammer2_chain_setcheck(chain, chain->data);
734 			break;
735 		default:
736 			KKASSERT(chain->flags & HAMMER2_CHAIN_EMBEDDED);
737 			panic("hammer2_flush_core: unsupported "
738 			      "embedded bref %d",
739 			      chain->bref.type);
740 			/* NOT REACHED */
741 		}
742 
743 		/*
744 		 * If the chain was destroyed try to avoid unnecessary I/O.
745 		 * (this only really works if the DIO system buffer is the
746 		 * same size as chain->bytes).
747 		 */
748 		if ((chain->flags & HAMMER2_CHAIN_DESTROY) && chain->dio) {
749 			hammer2_io_setinval(chain->dio, chain->bytes);
750 		}
751 	}
752 
753 	/*
754 	 * If UPDATE is set the parent block table may need to be updated.
755 	 *
756 	 * NOTE: UPDATE may be set on vchain or fchain in which case
757 	 *	 parent could be NULL.  It's easiest to allow the case
758 	 *	 and test for NULL.  parent can also wind up being NULL
759 	 *	 due to a deletion so we need to handle the case anyway.
760 	 *
761 	 * If no parent exists we can just clear the UPDATE bit.  If the
762 	 * chain gets reattached later on the bit will simply get set
763 	 * again.
764 	 */
765 	if ((chain->flags & HAMMER2_CHAIN_UPDATE) && parent == NULL) {
766 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
767 		hammer2_chain_drop(chain);
768 	}
769 
770 	/*
771 	 * The chain may need its blockrefs updated in the parent.  This
772 	 * requires some fancy footwork.
773 	 */
774 	if (chain->flags & HAMMER2_CHAIN_UPDATE) {
775 		hammer2_blockref_t *base;
776 		int count;
777 
778 		/*
779 		 * Both parent and chain must be locked.  This requires
780 		 * temporarily unlocking the chain.  We have to deal with
781 		 * the case where the chain might be reparented or modified
782 		 * while it was unlocked.
783 		 */
784 		hammer2_chain_unlock(chain);
785 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
786 		hammer2_chain_lock(chain, HAMMER2_RESOLVE_MAYBE);
787 		if (chain->parent != parent) {
788 			kprintf("PARENT MISMATCH ch=%p p=%p/%p\n",
789 				chain, chain->parent, parent);
790 			hammer2_chain_unlock(parent);
791 			goto done;
792 		}
793 
794 		/*
795 		 * Check race condition.  If someone got in and modified
796 		 * it again while it was unlocked, we have to loop up.
797 		 */
798 		if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
799 			hammer2_chain_unlock(parent);
800 			kprintf("hammer2_flush: chain %p flush-mod race\n",
801 				chain);
802 			goto again;
803 		}
804 
805 		/*
806 		 * Clear UPDATE flag, mark parent modified, update its
807 		 * modify_tid if necessary, and adjust the parent blockmap.
808 		 */
809 		if (chain->flags & HAMMER2_CHAIN_UPDATE) {
810 			atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
811 			hammer2_chain_drop(chain);
812 		}
813 
814 		/*
815 		 * (optional code)
816 		 *
817 		 * Avoid actually modifying and updating the parent if it
818 		 * was flagged for destruction.  This can greatly reduce
819 		 * disk I/O in large tree removals because the
820 		 * hammer2_io_setinval() call in the upward recursion
821 		 * (see MODIFIED code above) can only handle a few cases.
822 		 */
823 		if (parent->flags & HAMMER2_CHAIN_DESTROY) {
824 			if (parent->bref.modify_tid < chain->bref.modify_tid) {
825 				parent->bref.modify_tid =
826 					chain->bref.modify_tid;
827 			}
828 			atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
829 							HAMMER2_CHAIN_BMAPUPD);
830 			hammer2_chain_unlock(parent);
831 			goto skipupdate;
832 		}
833 
834 		/*
835 		 * We are updating the parent's blockmap, the parent must
836 		 * be set modified.
837 		 */
838 		hammer2_chain_modify(parent, HAMMER2_MODIFY_KEEPMODIFY);
839 		if (parent->bref.modify_tid < chain->bref.modify_tid)
840 			parent->bref.modify_tid = chain->bref.modify_tid;
841 
842 		/*
843 		 * Calculate blockmap pointer
844 		 */
845 		switch(parent->bref.type) {
846 		case HAMMER2_BREF_TYPE_INODE:
847 			/*
848 			 * Access the inode's block array.  However, there is
849 			 * no block array if the inode is flagged DIRECTDATA.
850 			 */
851 			if (parent->data &&
852 			    (parent->data->ipdata.meta.op_flags &
853 			     HAMMER2_OPFLAG_DIRECTDATA) == 0) {
854 				base = &parent->data->
855 					ipdata.u.blockset.blockref[0];
856 			} else {
857 				base = NULL;
858 			}
859 			count = HAMMER2_SET_COUNT;
860 			break;
861 		case HAMMER2_BREF_TYPE_INDIRECT:
862 		case HAMMER2_BREF_TYPE_FREEMAP_NODE:
863 			if (parent->data)
864 				base = &parent->data->npdata[0];
865 			else
866 				base = NULL;
867 			count = parent->bytes / sizeof(hammer2_blockref_t);
868 			break;
869 		case HAMMER2_BREF_TYPE_VOLUME:
870 			base = &chain->hmp->voldata.sroot_blockset.blockref[0];
871 			count = HAMMER2_SET_COUNT;
872 			break;
873 		case HAMMER2_BREF_TYPE_FREEMAP:
874 			base = &parent->data->npdata[0];
875 			count = HAMMER2_SET_COUNT;
876 			break;
877 		default:
878 			base = NULL;
879 			count = 0;
880 			panic("hammer2_flush_core: "
881 			      "unrecognized blockref type: %d",
882 			      parent->bref.type);
883 		}
884 
885 		/*
886 		 * Blocktable updates
887 		 *
888 		 * We synchronize pending statistics at this time.  Delta
889 		 * adjustments designated for the current and upper level
890 		 * are synchronized.
891 		 */
892 		if (base && (chain->flags & HAMMER2_CHAIN_BMAPUPD)) {
893 			if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
894 				hammer2_spin_ex(&parent->core.spin);
895 				hammer2_base_delete(parent, base, count,
896 						    &info->cache_index, chain);
897 				hammer2_spin_unex(&parent->core.spin);
898 				/* base_delete clears both bits */
899 			} else {
900 				atomic_clear_int(&chain->flags,
901 						 HAMMER2_CHAIN_BMAPUPD);
902 			}
903 		}
904 		if (base && (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
905 			hammer2_spin_ex(&parent->core.spin);
906 			hammer2_base_insert(parent, base, count,
907 					    &info->cache_index, chain);
908 			hammer2_spin_unex(&parent->core.spin);
909 			/* base_insert sets BMAPPED */
910 		}
911 		hammer2_chain_unlock(parent);
912 	}
913 skipupdate:
914 	;
915 
916 	/*
917 	 * Final cleanup after flush
918 	 */
919 done:
920 	KKASSERT(chain->refs > 0);
921 	if (hammer2_debug & 0x200) {
922 		if (info->debug == chain)
923 			info->debug = NULL;
924 	}
925 }
926 
927 /*
928  * Flush recursion helper, called from flush_core, calls flush_core.
929  *
930  * Flushes the children of the caller's chain (info->parent), restricted
931  * by sync_tid.  Set info->domodify if the child's blockref must propagate
932  * back up to the parent.
933  *
934  * Ripouts can move child from rbtree to dbtree or dbq but the caller's
935  * flush scan order prevents any chains from being lost.  A child can be
936  * executes more than once.
937  *
938  * WARNING! If we do not call hammer2_flush_core() we must update
939  *	    bref.mirror_tid ourselves to indicate that the flush has
940  *	    processed the child.
941  *
942  * WARNING! parent->core spinlock is held on entry and return.
943  */
944 static int
945 hammer2_flush_recurse(hammer2_chain_t *child, void *data)
946 {
947 	hammer2_flush_info_t *info = data;
948 	hammer2_chain_t *parent = info->parent;
949 
950 	/*
951 	 * (child can never be fchain or vchain so a special check isn't
952 	 *  needed).
953 	 *
954 	 * We must ref the child before unlocking the spinlock.
955 	 *
956 	 * The caller has added a ref to the parent so we can temporarily
957 	 * unlock it in order to lock the child.
958 	 */
959 	hammer2_chain_ref(child);
960 	hammer2_spin_unex(&parent->core.spin);
961 
962 	hammer2_chain_unlock(parent);
963 	hammer2_chain_lock(child, HAMMER2_RESOLVE_MAYBE);
964 
965 	/*
966 	 * Recurse and collect deferral data.  We're in the media flush,
967 	 * this can cross PFS boundaries.
968 	 */
969 	if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
970 		++info->depth;
971 		hammer2_flush_core(info, child, 0);
972 		--info->depth;
973 	} else if (hammer2_debug & 0x200) {
974 		if (info->debug == NULL)
975 			info->debug = child;
976 		++info->depth;
977 		hammer2_flush_core(info, child, 0);
978 		--info->depth;
979 		if (info->debug == child)
980 			info->debug = NULL;
981 	}
982 
983 	/*
984 	 * Relock to continue the loop
985 	 */
986 	hammer2_chain_unlock(child);
987 	hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
988 	hammer2_chain_drop(child);
989 	KKASSERT(info->parent == parent);
990 	hammer2_spin_ex(&parent->core.spin);
991 
992 	return (0);
993 }
994 
995 /*
996  * flush helper (backend threaded)
997  *
998  * Flushes core chains, issues disk sync, flushes volume roots.
999  *
1000  * Primarily called from vfs_sync().
1001  */
1002 void
1003 hammer2_inode_xop_flush(hammer2_xop_t *arg, int clindex)
1004 {
1005 	hammer2_xop_flush_t *xop = &arg->xop_flush;
1006 	hammer2_chain_t *chain;
1007 	hammer2_chain_t *parent;
1008 	hammer2_dev_t *hmp;
1009 	int error = 0;
1010 	int total_error = 0;
1011 	int j;
1012 
1013 	/*
1014 	 * Flush core chains
1015 	 */
1016 	chain = hammer2_inode_chain(xop->head.ip, clindex,
1017 				    HAMMER2_RESOLVE_ALWAYS);
1018 	if (chain) {
1019 		hmp = chain->hmp;
1020 		if (chain->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1021 			hammer2_flush(chain, 1);
1022 			parent = chain->parent;
1023 			KKASSERT(chain->pmp != parent->pmp);
1024 			hammer2_chain_setflush(parent);
1025 		}
1026 		hammer2_chain_unlock(chain);
1027 		hammer2_chain_drop(chain);
1028 		chain = NULL;
1029 	} else {
1030 		hmp = NULL;
1031 	}
1032 
1033 	/*
1034 	 * Flush volume roots.  Avoid replication, we only want to
1035 	 * flush each hammer2_dev (hmp) once.
1036 	 */
1037 	for (j = clindex - 1; j >= 0; --j) {
1038 		if ((chain = xop->head.ip->cluster.array[j].chain) != NULL) {
1039 			if (chain->hmp == hmp) {
1040 				chain = NULL;	/* safety */
1041 				goto skip;
1042 			}
1043 		}
1044 	}
1045 	chain = NULL;	/* safety */
1046 
1047 	/*
1048 	 * spmp transaction.  The super-root is never directly mounted so
1049 	 * there shouldn't be any vnodes, let alone any dirty vnodes
1050 	 * associated with it.
1051 	 */
1052 	hammer2_trans_init(hmp->spmp, HAMMER2_TRANS_ISFLUSH);
1053 
1054 	/*
1055 	 * Media mounts have two 'roots', vchain for the topology
1056 	 * and fchain for the free block table.  Flush both.
1057 	 *
1058 	 * Note that the topology and free block table are handled
1059 	 * independently, so the free block table can wind up being
1060 	 * ahead of the topology.  We depend on the bulk free scan
1061 	 * code to deal with any loose ends.
1062 	 */
1063 	hammer2_chain_ref(&hmp->vchain);
1064 	hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1065 	hammer2_chain_ref(&hmp->fchain);
1066 	hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1067 	if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1068 		/*
1069 		 * This will also modify vchain as a side effect,
1070 		 * mark vchain as modified now.
1071 		 */
1072 		hammer2_voldata_modify(hmp);
1073 		chain = &hmp->fchain;
1074 		hammer2_flush(chain, 1);
1075 		KKASSERT(chain == &hmp->fchain);
1076 	}
1077 	hammer2_chain_unlock(&hmp->fchain);
1078 	hammer2_chain_unlock(&hmp->vchain);
1079 	hammer2_chain_drop(&hmp->fchain);
1080 	/* vchain dropped down below */
1081 
1082 	hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1083 	if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1084 		chain = &hmp->vchain;
1085 		hammer2_flush(chain, 1);
1086 		KKASSERT(chain == &hmp->vchain);
1087 	}
1088 	hammer2_chain_unlock(&hmp->vchain);
1089 	hammer2_chain_drop(&hmp->vchain);
1090 
1091 	error = 0;
1092 
1093 	/*
1094 	 * We can't safely flush the volume header until we have
1095 	 * flushed any device buffers which have built up.
1096 	 *
1097 	 * XXX this isn't being incremental
1098 	 */
1099 	vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
1100 	error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
1101 	vn_unlock(hmp->devvp);
1102 
1103 	/*
1104 	 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
1105 	 * volume header needs synchronization via hmp->volsync.
1106 	 *
1107 	 * XXX synchronize the flag & data with only this flush XXX
1108 	 */
1109 	if (error == 0 &&
1110 	    (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
1111 		struct buf *bp;
1112 
1113 		/*
1114 		 * Synchronize the disk before flushing the volume
1115 		 * header.
1116 		 */
1117 		bp = getpbuf(NULL);
1118 		bp->b_bio1.bio_offset = 0;
1119 		bp->b_bufsize = 0;
1120 		bp->b_bcount = 0;
1121 		bp->b_cmd = BUF_CMD_FLUSH;
1122 		bp->b_bio1.bio_done = biodone_sync;
1123 		bp->b_bio1.bio_flags |= BIO_SYNC;
1124 		vn_strategy(hmp->devvp, &bp->b_bio1);
1125 		biowait(&bp->b_bio1, "h2vol");
1126 		relpbuf(bp, NULL);
1127 
1128 		/*
1129 		 * Then we can safely flush the version of the
1130 		 * volume header synchronized by the flush code.
1131 		 */
1132 		j = hmp->volhdrno + 1;
1133 		if (j >= HAMMER2_NUM_VOLHDRS)
1134 			j = 0;
1135 		if (j * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
1136 		    hmp->volsync.volu_size) {
1137 			j = 0;
1138 		}
1139 		kprintf("sync volhdr %d %jd\n",
1140 			j, (intmax_t)hmp->volsync.volu_size);
1141 		bp = getblk(hmp->devvp, j * HAMMER2_ZONE_BYTES64,
1142 			    HAMMER2_PBUFSIZE, 0, 0);
1143 		atomic_clear_int(&hmp->vchain.flags,
1144 				 HAMMER2_CHAIN_VOLUMESYNC);
1145 		bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
1146 		bawrite(bp);
1147 		hmp->volhdrno = j;
1148 	}
1149 	if (error)
1150 		total_error = error;
1151 
1152 	hammer2_trans_done(hmp->spmp);  /* spmp trans */
1153 skip:
1154 	error = hammer2_xop_feed(&xop->head, NULL, clindex, total_error);
1155 }
1156