HAMMER VFS - REDO implementation base code part 1/many
[dragonfly.git] / sys / vfs / hammer / hammer_flusher.c
1 /*
2  * Copyright (c) 2008 The DragonFly Project.  All rights reserved.
3  * 
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  * 
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  * 
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  * 
34  * $DragonFly: src/sys/vfs/hammer/hammer_flusher.c,v 1.45 2008/07/31 04:42:04 dillon Exp $
35  */
36 /*
37  * HAMMER dependancy flusher thread
38  *
39  * Meta data updates create buffer dependancies which are arranged as a
40  * hierarchy of lists.
41  */
42
43 #include "hammer.h"
44
45 static void hammer_flusher_master_thread(void *arg);
46 static void hammer_flusher_slave_thread(void *arg);
47 static void hammer_flusher_flush(hammer_mount_t hmp);
48 static void hammer_flusher_flush_inode(hammer_inode_t ip,
49                                         hammer_transaction_t trans);
50
51 RB_GENERATE(hammer_fls_rb_tree, hammer_inode, rb_flsnode,
52               hammer_ino_rb_compare);
53
54 /*
55  * Inodes are sorted and assigned to slave threads in groups of 128.
56  * We want a flush group size large enough such that the slave threads
57  * are not likely to interfere with each other when accessing the B-Tree,
58  * but not so large that we lose concurrency.
59  */
60 #define HAMMER_FLUSH_GROUP_SIZE 128
61
62 /*
63  * Support structures for the flusher threads.
64  */
65 struct hammer_flusher_info {
66         TAILQ_ENTRY(hammer_flusher_info) entry;
67         struct hammer_mount *hmp;
68         thread_t        td;
69         int             runstate;
70         int             count;
71         hammer_flush_group_t flg;
72         hammer_inode_t  work_array[HAMMER_FLUSH_GROUP_SIZE];
73 };
74
75 typedef struct hammer_flusher_info *hammer_flusher_info_t;
76
77 /*
78  * Sync all inodes pending on the flusher.
79  *
80  * All flush groups will be flushed.  This does not queue dirty inodes
81  * to the flush groups, it just flushes out what has already been queued!
82  */
83 void
84 hammer_flusher_sync(hammer_mount_t hmp)
85 {
86         int seq;
87
88         seq = hammer_flusher_async(hmp, NULL);
89         hammer_flusher_wait(hmp, seq);
90 }
91
92 /*
93  * Sync all inodes pending on the flusher - return immediately.
94  *
95  * All flush groups will be flushed.
96  */
97 int
98 hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t close_flg)
99 {
100         hammer_flush_group_t flg;
101         int seq = hmp->flusher.next;
102
103         TAILQ_FOREACH(flg, &hmp->flush_group_list, flush_entry) {
104                 if (flg->running == 0)
105                         ++seq;
106                 flg->closed = 1;
107                 if (flg == close_flg)
108                         break;
109         }
110         if (hmp->flusher.td) {
111                 if (hmp->flusher.signal++ == 0)
112                         wakeup(&hmp->flusher.signal);
113         } else {
114                 seq = hmp->flusher.done;
115         }
116         return(seq);
117 }
118
119 int
120 hammer_flusher_async_one(hammer_mount_t hmp)
121 {
122         int seq;
123
124         if (hmp->flusher.td) {
125                 seq = hmp->flusher.next;
126                 if (hmp->flusher.signal++ == 0)
127                         wakeup(&hmp->flusher.signal);
128         } else {
129                 seq = hmp->flusher.done;
130         }
131         return(seq);
132 }
133
134 /*
135  * Wait for the flusher to get to the specified sequence number.
136  * Signal the flusher as often as necessary to keep it going.
137  */
138 void
139 hammer_flusher_wait(hammer_mount_t hmp, int seq)
140 {
141         while ((int)(seq - hmp->flusher.done) > 0) {
142                 if (hmp->flusher.act != seq) {
143                         if (hmp->flusher.signal++ == 0)
144                                 wakeup(&hmp->flusher.signal);
145                 }
146                 tsleep(&hmp->flusher.done, 0, "hmrfls", 0);
147         }
148 }
149
150 void
151 hammer_flusher_wait_next(hammer_mount_t hmp)
152 {
153         int seq;
154
155         seq = hammer_flusher_async_one(hmp);
156         hammer_flusher_wait(hmp, seq);
157 }
158
159 void
160 hammer_flusher_create(hammer_mount_t hmp)
161 {
162         hammer_flusher_info_t info;
163         int i;
164
165         hmp->flusher.signal = 0;
166         hmp->flusher.act = 0;
167         hmp->flusher.done = 0;
168         hmp->flusher.next = 1;
169         hammer_ref(&hmp->flusher.finalize_lock);
170         TAILQ_INIT(&hmp->flusher.run_list);
171         TAILQ_INIT(&hmp->flusher.ready_list);
172
173         lwkt_create(hammer_flusher_master_thread, hmp,
174                     &hmp->flusher.td, NULL, 0, -1, "hammer-M");
175         for (i = 0; i < HAMMER_MAX_FLUSHERS; ++i) {
176                 info = kmalloc(sizeof(*info), hmp->m_misc, M_WAITOK|M_ZERO);
177                 info->hmp = hmp;
178                 TAILQ_INSERT_TAIL(&hmp->flusher.ready_list, info, entry);
179                 lwkt_create(hammer_flusher_slave_thread, info,
180                             &info->td, NULL, 0, -1, "hammer-S%d", i);
181         }
182 }
183
184 void
185 hammer_flusher_destroy(hammer_mount_t hmp)
186 {
187         hammer_flusher_info_t info;
188
189         /*
190          * Kill the master
191          */
192         hmp->flusher.exiting = 1;
193         while (hmp->flusher.td) {
194                 ++hmp->flusher.signal;
195                 wakeup(&hmp->flusher.signal);
196                 tsleep(&hmp->flusher.exiting, 0, "hmrwex", hz);
197         }
198
199         /*
200          * Kill the slaves
201          */
202         while ((info = TAILQ_FIRST(&hmp->flusher.ready_list)) != NULL) {
203                 KKASSERT(info->runstate == 0);
204                 TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry);
205                 info->runstate = -1;
206                 wakeup(&info->runstate);
207                 while (info->td)
208                         tsleep(&info->td, 0, "hmrwwc", 0);
209                 kfree(info, hmp->m_misc);
210         }
211 }
212
213 /*
214  * The master flusher thread manages the flusher sequence id and
215  * synchronization with the slave work threads.
216  */
217 static void
218 hammer_flusher_master_thread(void *arg)
219 {
220         hammer_flush_group_t flg;
221         hammer_mount_t hmp;
222
223         hmp = arg;
224
225         for (;;) {
226                 /*
227                  * Do at least one flush cycle.  We may have to update the
228                  * UNDO FIFO even if no inodes are queued.
229                  */
230                 for (;;) {
231                         while (hmp->flusher.group_lock)
232                                 tsleep(&hmp->flusher.group_lock, 0, "hmrhld", 0);
233                         hmp->flusher.act = hmp->flusher.next;
234                         ++hmp->flusher.next;
235                         hammer_flusher_clean_loose_ios(hmp);
236                         hammer_flusher_flush(hmp);
237                         hmp->flusher.done = hmp->flusher.act;
238                         wakeup(&hmp->flusher.done);
239                         flg = TAILQ_FIRST(&hmp->flush_group_list);
240                         if (flg == NULL || flg->closed == 0)
241                                 break;
242                         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
243                                 break;
244                 }
245
246                 /*
247                  * Wait for activity.
248                  */
249                 if (hmp->flusher.exiting && TAILQ_EMPTY(&hmp->flush_group_list))
250                         break;
251                 while (hmp->flusher.signal == 0)
252                         tsleep(&hmp->flusher.signal, 0, "hmrwwa", 0);
253
254                 /*
255                  * Flush for each count on signal but only allow one extra
256                  * flush request to build up.
257                  */
258                 if (--hmp->flusher.signal != 0)
259                         hmp->flusher.signal = 1;
260         }
261
262         /*
263          * And we are done.
264          */
265         hmp->flusher.td = NULL;
266         wakeup(&hmp->flusher.exiting);
267         lwkt_exit();
268 }
269
270 /*
271  * Flush all inodes in the current flush group.
272  */
273 static void
274 hammer_flusher_flush(hammer_mount_t hmp)
275 {
276         hammer_flusher_info_t info;
277         hammer_flush_group_t flg;
278         hammer_reserve_t resv;
279         hammer_inode_t ip;
280         hammer_inode_t next_ip;
281         int slave_index;
282         int count;
283
284         /*
285          * Just in-case there's a flush race on mount
286          */
287         if (TAILQ_FIRST(&hmp->flusher.ready_list) == NULL)
288                 return;
289
290         /*
291          * We only do one flg but we may have to loop/retry.
292          */
293         count = 0;
294         while ((flg = TAILQ_FIRST(&hmp->flush_group_list)) != NULL) {
295                 ++count;
296                 if (hammer_debug_general & 0x0001) {
297                         kprintf("hammer_flush %d ttl=%d recs=%d\n",
298                                 hmp->flusher.act,
299                                 flg->total_count, flg->refs);
300                 }
301                 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
302                         break;
303                 hammer_start_transaction_fls(&hmp->flusher.trans, hmp);
304
305                 /*
306                  * If the previous flush cycle just about exhausted our
307                  * UNDO space we may have to do a dummy cycle to move the
308                  * first_offset up before actually digging into a new cycle,
309                  * or the new cycle will not have sufficient undo space.
310                  */
311                 if (hammer_flusher_undo_exhausted(&hmp->flusher.trans, 3))
312                         hammer_flusher_finalize(&hmp->flusher.trans, 0);
313
314                 /*
315                  * Ok, we are running this flush group now (this prevents new
316                  * additions to it).
317                  */
318                 flg->running = 1;
319                 if (hmp->next_flush_group == flg)
320                         hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
321
322                 /*
323                  * Iterate the inodes in the flg's flush_tree and assign
324                  * them to slaves.
325                  */
326                 slave_index = 0;
327                 info = TAILQ_FIRST(&hmp->flusher.ready_list);
328                 next_ip = RB_FIRST(hammer_fls_rb_tree, &flg->flush_tree);
329
330                 while ((ip = next_ip) != NULL) {
331                         next_ip = RB_NEXT(hammer_fls_rb_tree,
332                                           &flg->flush_tree, ip);
333
334                         if (++hmp->check_yield > hammer_yield_check) {
335                                 hmp->check_yield = 0;
336                                 lwkt_user_yield();
337                         }
338
339                         /*
340                          * Add ip to the slave's work array.  The slave is
341                          * not currently running.
342                          */
343                         info->work_array[info->count++] = ip;
344                         if (info->count != HAMMER_FLUSH_GROUP_SIZE)
345                                 continue;
346
347                         /*
348                          * Get the slave running
349                          */
350                         TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry);
351                         TAILQ_INSERT_TAIL(&hmp->flusher.run_list, info, entry);
352                         info->flg = flg;
353                         info->runstate = 1;
354                         wakeup(&info->runstate);
355
356                         /*
357                          * Get a new slave.  We may have to wait for one to
358                          * finish running.
359                          */
360                         while ((info = TAILQ_FIRST(&hmp->flusher.ready_list)) == NULL) {
361                                 tsleep(&hmp->flusher.ready_list, 0, "hmrfcc", 0);
362                         }
363                 }
364
365                 /*
366                  * Run the current slave if necessary
367                  */
368                 if (info->count) {
369                         TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry);
370                         TAILQ_INSERT_TAIL(&hmp->flusher.run_list, info, entry);
371                         info->flg = flg;
372                         info->runstate = 1;
373                         wakeup(&info->runstate);
374                 }
375
376                 /*
377                  * Wait for all slaves to finish running
378                  */
379                 while (TAILQ_FIRST(&hmp->flusher.run_list) != NULL)
380                         tsleep(&hmp->flusher.ready_list, 0, "hmrfcc", 0);
381
382                 /*
383                  * Do the final finalization, clean up
384                  */
385                 hammer_flusher_finalize(&hmp->flusher.trans, 1);
386                 hmp->flusher.tid = hmp->flusher.trans.tid;
387
388                 hammer_done_transaction(&hmp->flusher.trans);
389
390                 /*
391                  * Loop up on the same flg.  If the flg is done clean it up
392                  * and break out.  We only flush one flg.
393                  */
394                 if (RB_EMPTY(&flg->flush_tree)) {
395                         KKASSERT(flg->refs == 0);
396                         TAILQ_REMOVE(&hmp->flush_group_list, flg, flush_entry);
397                         kfree(flg, hmp->m_misc);
398                         break;
399                 }
400         }
401
402         /*
403          * We may have pure meta-data to flush, or we may have to finish
404          * cycling the UNDO FIFO, even if there were no flush groups.
405          */
406         if (count == 0 && hammer_flusher_haswork(hmp)) {
407                 hammer_start_transaction_fls(&hmp->flusher.trans, hmp);
408                 hammer_flusher_finalize(&hmp->flusher.trans, 1);
409                 hammer_done_transaction(&hmp->flusher.trans);
410         }
411
412         /*
413          * Clean up any freed big-blocks (typically zone-2). 
414          * resv->flush_group is typically set several flush groups ahead
415          * of the free to ensure that the freed block is not reused until
416          * it can no longer be reused.
417          */
418         while ((resv = TAILQ_FIRST(&hmp->delay_list)) != NULL) {
419                 if (resv->flush_group != hmp->flusher.act)
420                         break;
421                 hammer_reserve_clrdelay(hmp, resv);
422         }
423 }
424
425
426 /*
427  * The slave flusher thread pulls work off the master flush list until no
428  * work is left.
429  */
430 static void
431 hammer_flusher_slave_thread(void *arg)
432 {
433         hammer_flush_group_t flg;
434         hammer_flusher_info_t info;
435         hammer_mount_t hmp;
436         hammer_inode_t ip;
437         int i;
438
439         info = arg;
440         hmp = info->hmp;
441
442         for (;;) {
443                 while (info->runstate == 0)
444                         tsleep(&info->runstate, 0, "hmrssw", 0);
445                 if (info->runstate < 0)
446                         break;
447                 flg = info->flg;
448
449                 for (i = 0; i < info->count; ++i) {
450                         ip = info->work_array[i];
451                         hammer_flusher_flush_inode(ip, &hmp->flusher.trans);
452                         ++hammer_stats_inode_flushes;
453                 }
454                 info->count = 0;
455                 info->runstate = 0;
456                 TAILQ_REMOVE(&hmp->flusher.run_list, info, entry);
457                 TAILQ_INSERT_TAIL(&hmp->flusher.ready_list, info, entry);
458                 wakeup(&hmp->flusher.ready_list);
459         }
460         info->td = NULL;
461         wakeup(&info->td);
462         lwkt_exit();
463 }
464
465 void
466 hammer_flusher_clean_loose_ios(hammer_mount_t hmp)
467 {
468         hammer_buffer_t buffer;
469         hammer_io_t io;
470
471         /*
472          * loose ends - buffers without bp's aren't tracked by the kernel
473          * and can build up, so clean them out.  This can occur when an
474          * IO completes on a buffer with no references left.
475          */
476         if ((io = TAILQ_FIRST(&hmp->lose_list)) != NULL) {
477                 crit_enter();   /* biodone() race */
478                 while ((io = TAILQ_FIRST(&hmp->lose_list)) != NULL) {
479                         KKASSERT(io->mod_list == &hmp->lose_list);
480                         TAILQ_REMOVE(&hmp->lose_list, io, mod_entry);
481                         io->mod_list = NULL;
482                         if (io->lock.refs == 0)
483                                 ++hammer_count_refedbufs;
484                         hammer_ref(&io->lock);
485                         buffer = (void *)io;
486                         hammer_rel_buffer(buffer, 0);
487                 }
488                 crit_exit();
489         }
490 }
491
492 /*
493  * Flush a single inode that is part of a flush group.
494  *
495  * Flusher errors are extremely serious, even ENOSPC shouldn't occur because
496  * the front-end should have reserved sufficient space on the media.  Any
497  * error other then EWOULDBLOCK will force the mount to be read-only.
498  */
499 static
500 void
501 hammer_flusher_flush_inode(hammer_inode_t ip, hammer_transaction_t trans)
502 {
503         hammer_mount_t hmp = ip->hmp;
504         int error;
505
506         hammer_flusher_clean_loose_ios(hmp);
507         error = hammer_sync_inode(trans, ip);
508
509         /*
510          * EWOULDBLOCK can happen under normal operation, all other errors
511          * are considered extremely serious.  We must set WOULDBLOCK
512          * mechanics to deal with the mess left over from the abort of the
513          * previous flush.
514          */
515         if (error) {
516                 ip->flags |= HAMMER_INODE_WOULDBLOCK;
517                 if (error == EWOULDBLOCK)
518                         error = 0;
519         }
520         hammer_flush_inode_done(ip, error);
521         while (hmp->flusher.finalize_want)
522                 tsleep(&hmp->flusher.finalize_want, 0, "hmrsxx", 0);
523         if (hammer_flusher_undo_exhausted(trans, 1)) {
524                 kprintf("HAMMER: Warning: UNDO area too small!\n");
525                 hammer_flusher_finalize(trans, 1);
526         } else if (hammer_flusher_meta_limit(trans->hmp)) {
527                 hammer_flusher_finalize(trans, 0);
528         }
529 }
530
531 /*
532  * Return non-zero if the UNDO area has less then (QUARTER / 4) of its
533  * space left.
534  *
535  * 1/4 - Emergency free undo space level.  Below this point the flusher
536  *       will finalize even if directory dependancies have not been resolved.
537  *
538  * 2/4 - Used by the pruning and reblocking code.  These functions may be
539  *       running in parallel with a flush and cannot be allowed to drop
540  *       available undo space to emergency levels.
541  *
542  * 3/4 - Used at the beginning of a flush to force-sync the volume header
543  *       to give the flush plenty of runway to work in.
544  */
545 int
546 hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter)
547 {
548         if (hammer_undo_space(trans) <
549             hammer_undo_max(trans->hmp) * quarter / 4) {
550                 return(1);
551         } else {
552                 return(0);
553         }
554 }
555
556 /*
557  * Flush all pending UNDOs, wait for write completion, update the volume
558  * header with the new UNDO end position, and flush it.  Then
559  * asynchronously flush the meta-data.
560  *
561  * If this is the last finalization in a flush group we also synchronize
562  * our cached blockmap and set hmp->flusher_undo_start and our cached undo
563  * fifo first_offset so the next flush resets the FIFO pointers.
564  *
565  * If this is not final it is being called because too many dirty meta-data
566  * buffers have built up and must be flushed with UNDO synchronization to
567  * avoid a buffer cache deadlock.
568  */
569 void
570 hammer_flusher_finalize(hammer_transaction_t trans, int final)
571 {
572         hammer_volume_t root_volume;
573         hammer_blockmap_t cundomap, dundomap;
574         hammer_mount_t hmp;
575         hammer_io_t io;
576         int count;
577         int i;
578
579         hmp = trans->hmp;
580         root_volume = trans->rootvol;
581
582         /*
583          * Exclusively lock the flusher.  This guarantees that all dirty
584          * buffers will be idled (have a mod-count of 0).
585          */
586         ++hmp->flusher.finalize_want;
587         hammer_lock_ex(&hmp->flusher.finalize_lock);
588
589         /*
590          * If this isn't the final sync several threads may have hit the
591          * meta-limit at the same time and raced.  Only sync if we really
592          * have to, after acquiring the lock.
593          */
594         if (final == 0 && !hammer_flusher_meta_limit(hmp))
595                 goto done;
596
597         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
598                 goto done;
599
600         /*
601          * Flush data buffers.  This can occur asynchronously and at any
602          * time.  We must interlock against the frontend direct-data write
603          * but do not have to acquire the sync-lock yet.
604          *
605          * These data buffers have already been collected prior to the
606          * related inode(s) getting queued to the flush group.
607          */
608         count = 0;
609         while ((io = TAILQ_FIRST(&hmp->data_list)) != NULL) {
610                 if (io->ioerror)
611                         break;
612                 if (io->lock.refs == 0)
613                         ++hammer_count_refedbufs;
614                 hammer_ref(&io->lock);
615                 hammer_io_write_interlock(io);
616                 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
617                 hammer_io_flush(io, 0);
618                 hammer_io_done_interlock(io);
619                 hammer_rel_buffer((hammer_buffer_t)io, 0);
620                 ++count;
621         }
622
623         /*
624          * The sync-lock is required for the remaining sequence.  This lock
625          * prevents meta-data from being modified.
626          */
627         hammer_sync_lock_ex(trans);
628
629         /*
630          * If we have been asked to finalize the volume header sync the
631          * cached blockmap to the on-disk blockmap.  Generate an UNDO
632          * record for the update.
633          */
634         if (final) {
635                 cundomap = &hmp->blockmap[0];
636                 dundomap = &root_volume->ondisk->vol0_blockmap[0];
637                 if (root_volume->io.modified) {
638                         hammer_modify_volume(trans, root_volume,
639                                              dundomap, sizeof(hmp->blockmap));
640                         for (i = 0; i < HAMMER_MAX_ZONES; ++i)
641                                 hammer_crc_set_blockmap(&cundomap[i]);
642                         bcopy(cundomap, dundomap, sizeof(hmp->blockmap));
643                         hammer_modify_volume_done(root_volume);
644                 }
645         }
646
647         /*
648          * Flush UNDOs.  This also waits for I/Os to complete and flushes
649          * the cache on the target disk.
650          */
651         hammer_flusher_flush_undos(hmp, HAMMER_FLUSH_UNDOS_FORCED);
652
653         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
654                 goto failed;
655
656         /*
657          * HAMMER VERSION < 4:
658          *      Update the on-disk volume header with new UNDO FIFO end
659          *      position (do not generate new UNDO records for this change).
660          *      We have to do this for the UNDO FIFO whether (final) is
661          *      set or not in order for the UNDOs to be recognized on
662          *      recovery.
663          *
664          * HAMMER VERSION >= 4:
665          *      The UNDO FIFO data written above will be recognized on
666          *      recovery without us having to sync the volume header.
667          *
668          * Also update the on-disk next_tid field.  This does not require
669          * an UNDO.  However, because our TID is generated before we get
670          * the sync lock another sync may have beat us to the punch.
671          *
672          * This also has the side effect of updating first_offset based on
673          * a prior finalization when the first finalization of the next flush
674          * cycle occurs, removing any undo info from the prior finalization
675          * from consideration.
676          *
677          * The volume header will be flushed out synchronously.
678          */
679         dundomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
680         cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
681
682         if (dundomap->first_offset != cundomap->first_offset ||
683                    dundomap->next_offset != cundomap->next_offset) {
684                 hammer_modify_volume(NULL, root_volume, NULL, 0);
685                 dundomap->first_offset = cundomap->first_offset;
686                 dundomap->next_offset = cundomap->next_offset;
687                 hammer_crc_set_blockmap(dundomap);
688                 hammer_modify_volume_done(root_volume);
689         }
690
691         /*
692          * vol0_next_tid is used for TID selection and is updated without
693          * an UNDO so we do not reuse a TID that may have been rolled-back.
694          *
695          * vol0_last_tid is the highest fully-synchronized TID.  It is
696          * set-up when the UNDO fifo is fully synced, later on (not here).
697          */
698         if (root_volume->io.modified) {
699                 hammer_modify_volume(NULL, root_volume, NULL, 0);
700                 if (root_volume->ondisk->vol0_next_tid < trans->tid)
701                         root_volume->ondisk->vol0_next_tid = trans->tid;
702                 hammer_crc_set_volume(root_volume->ondisk);
703                 hammer_modify_volume_done(root_volume);
704                 hammer_io_flush(&root_volume->io, 0);
705         }
706
707         /*
708          * Wait for I/Os to complete.
709          *
710          * For HAMMER VERSION 4+ filesystems we do not have to wait for
711          * the I/O to complete as the new UNDO FIFO entries are recognized
712          * even without the volume header update.  This allows the volume
713          * header to flushed along with meta-data, significantly reducing
714          * flush overheads.
715          */
716         hammer_flusher_clean_loose_ios(hmp);
717         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
718                 hammer_io_wait_all(hmp, "hmrfl2");
719
720         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
721                 goto failed;
722
723         /*
724          * Flush meta-data.  The meta-data will be undone if we crash
725          * so we can safely flush it asynchronously.  There is no need
726          * to wait for I/O to complete (or issue a synchronous disk flush).
727          *
728          * In fact, even if we did wait the meta-data will still be undone
729          * by a crash up until the next flush cycle due to the first_offset
730          * in the volume header for the UNDO FIFO not being adjusted until
731          * the following flush cycle.
732          */
733         count = 0;
734         while ((io = TAILQ_FIRST(&hmp->meta_list)) != NULL) {
735                 if (io->ioerror)
736                         break;
737                 KKASSERT(io->modify_refs == 0);
738                 if (io->lock.refs == 0)
739                         ++hammer_count_refedbufs;
740                 hammer_ref(&io->lock);
741                 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
742                 hammer_io_flush(io, 0);
743                 hammer_rel_buffer((hammer_buffer_t)io, 0);
744                 ++count;
745         }
746
747         /*
748          * If this is the final finalization for the flush group set
749          * up for the next sequence by setting a new first_offset in
750          * our cached blockmap and clearing the undo history.
751          *
752          * Even though we have updated our cached first_offset, the on-disk
753          * first_offset still governs available-undo-space calculations.
754          */
755         if (final) {
756                 cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
757                 if (cundomap->first_offset == cundomap->next_offset) {
758                         hmp->hflags &= ~HMNT_UNDO_DIRTY;
759                 } else {
760                         cundomap->first_offset = cundomap->next_offset;
761                         hmp->hflags |= HMNT_UNDO_DIRTY;
762                 }
763                 hammer_clear_undo_history(hmp);
764
765                 /*
766                  * Flush tid sequencing.  flush_tid1 is fully synchronized,
767                  * meaning a crash will not roll it back.  flush_tid2 has
768                  * been written out asynchronously and a crash will roll
769                  * it back.  flush_tid1 is used for all mirroring masters.
770                  */
771                 if (hmp->flush_tid1 != hmp->flush_tid2) {
772                         hmp->flush_tid1 = hmp->flush_tid2;
773                         wakeup(&hmp->flush_tid1);
774                 }
775                 hmp->flush_tid2 = trans->tid;
776         }
777
778         /*
779          * Cleanup.  Report any critical errors.
780          */
781 failed:
782         hammer_sync_unlock(trans);
783
784         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) {
785                 kprintf("HAMMER(%s): Critical write error during flush, "
786                         "refusing to sync UNDO FIFO\n",
787                         root_volume->ondisk->vol_name);
788         }
789
790 done:
791         hammer_unlock(&hmp->flusher.finalize_lock);
792
793         if (--hmp->flusher.finalize_want == 0)
794                 wakeup(&hmp->flusher.finalize_want);
795         hammer_stats_commits += final;
796 }
797
798 /*
799  * Flush UNDOs.
800  */
801 void
802 hammer_flusher_flush_undos(hammer_mount_t hmp, int mode)
803 {
804         hammer_io_t io;
805         int count;
806
807         count = 0;
808         while ((io = TAILQ_FIRST(&hmp->undo_list)) != NULL) {
809                 if (io->ioerror)
810                         break;
811                 KKASSERT(io->modify_refs == 0);
812                 if (io->lock.refs == 0)
813                         ++hammer_count_refedbufs;
814                 hammer_ref(&io->lock);
815                 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
816                 hammer_io_flush(io, hammer_undo_reclaim(io));
817                 hammer_rel_buffer((hammer_buffer_t)io, 0);
818                 ++count;
819         }
820         hammer_flusher_clean_loose_ios(hmp);
821         if (mode == HAMMER_FLUSH_UNDOS_FORCED ||
822             (mode == HAMMER_FLUSH_UNDOS_AUTO && count)) {
823                 hammer_io_wait_all(hmp, "hmrfl1");
824         }
825 }
826
827 /*
828  * Return non-zero if too many dirty meta-data buffers have built up.
829  *
830  * Since we cannot allow such buffers to flush until we have dealt with
831  * the UNDOs, we risk deadlocking the kernel's buffer cache.
832  */
833 int
834 hammer_flusher_meta_limit(hammer_mount_t hmp)
835 {
836         if (hmp->locked_dirty_space + hmp->io_running_space >
837             hammer_limit_dirtybufspace) {
838                 return(1);
839         }
840         return(0);
841 }
842
843 /*
844  * Return non-zero if too many dirty meta-data buffers have built up.
845  *
846  * This version is used by background operations (mirror, prune, reblock)
847  * to leave room for foreground operations.
848  */
849 int
850 hammer_flusher_meta_halflimit(hammer_mount_t hmp)
851 {
852         if (hmp->locked_dirty_space + hmp->io_running_space >
853             hammer_limit_dirtybufspace / 2) {
854                 return(1);
855         }
856         return(0);
857 }
858
859 /*
860  * Return non-zero if the flusher still has something to flush.
861  */
862 int
863 hammer_flusher_haswork(hammer_mount_t hmp)
864 {
865         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
866                 return(0);
867         if (TAILQ_FIRST(&hmp->flush_group_list) ||      /* dirty inodes */
868             TAILQ_FIRST(&hmp->volu_list) ||             /* dirty buffers */
869             TAILQ_FIRST(&hmp->undo_list) ||
870             TAILQ_FIRST(&hmp->data_list) ||
871             TAILQ_FIRST(&hmp->meta_list) ||
872             (hmp->hflags & HMNT_UNDO_DIRTY)             /* UNDO FIFO sync */
873         ) {
874                 return(1);
875         }
876         return(0);
877 }
878