HAMMER VFS - The backend flusher now sorts inodes
[dragonfly.git] / sys / vfs / hammer / hammer_flusher.c
CommitLineData
059819e3
MD
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 *
4889cbd4 34 * $DragonFly: src/sys/vfs/hammer/hammer_flusher.c,v 1.45 2008/07/31 04:42:04 dillon Exp $
059819e3
MD
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
da2da375
MD
45static void hammer_flusher_master_thread(void *arg);
46static void hammer_flusher_slave_thread(void *arg);
059819e3 47static void hammer_flusher_flush(hammer_mount_t hmp);
9f5097dc
MD
48static void hammer_flusher_flush_inode(hammer_inode_t ip,
49 hammer_transaction_t trans);
c9b9e29d 50
ff003b11
MD
51RB_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
af209b0f
MD
62/*
63 * Support structures for the flusher threads.
64 */
65struct hammer_flusher_info {
7a61b85d 66 TAILQ_ENTRY(hammer_flusher_info) entry;
af209b0f
MD
67 struct hammer_mount *hmp;
68 thread_t td;
7a61b85d
MD
69 int runstate;
70 int count;
71 hammer_flush_group_t flg;
cb51be26 72 hammer_inode_t work_array[HAMMER_FLUSH_GROUP_SIZE];
af209b0f
MD
73};
74
75typedef struct hammer_flusher_info *hammer_flusher_info_t;
059819e3 76
7bc5b8c2 77/*
7a61b85d
MD
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!
7bc5b8c2 82 */
059819e3
MD
83void
84hammer_flusher_sync(hammer_mount_t hmp)
85{
86 int seq;
87
7a61b85d 88 seq = hammer_flusher_async(hmp, NULL);
f437a2ab 89 hammer_flusher_wait(hmp, seq);
059819e3
MD
90}
91
7bc5b8c2
MD
92/*
93 * Sync all inodes pending on the flusher - return immediately.
7a61b85d
MD
94 *
95 * All flush groups will be flushed.
7bc5b8c2 96 */
93291532 97int
7a61b85d 98hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t close_flg)
059819e3 99{
7a61b85d
MD
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 }
da2da375
MD
110 if (hmp->flusher.td) {
111 if (hmp->flusher.signal++ == 0)
112 wakeup(&hmp->flusher.signal);
93291532
MD
113 } else {
114 seq = hmp->flusher.done;
f90dde4c 115 }
93291532
MD
116 return(seq);
117}
118
15e75dab
MD
119int
120hammer_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
f437a2ab
MD
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 */
93291532
MD
138void
139hammer_flusher_wait(hammer_mount_t hmp, int seq)
140{
cdb6e4e6 141 while ((int)(seq - hmp->flusher.done) > 0) {
f437a2ab
MD
142 if (hmp->flusher.act != seq) {
143 if (hmp->flusher.signal++ == 0)
144 wakeup(&hmp->flusher.signal);
145 }
93291532 146 tsleep(&hmp->flusher.done, 0, "hmrfls", 0);
cdb6e4e6 147 }
059819e3
MD
148}
149
150void
82010f9f
MD
151hammer_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
159void
059819e3
MD
160hammer_flusher_create(hammer_mount_t hmp)
161{
da2da375
MD
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;
da2da375 169 hammer_ref(&hmp->flusher.finalize_lock);
7a61b85d
MD
170 TAILQ_INIT(&hmp->flusher.run_list);
171 TAILQ_INIT(&hmp->flusher.ready_list);
da2da375
MD
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) {
bac808fe 176 info = kmalloc(sizeof(*info), hmp->m_misc, M_WAITOK|M_ZERO);
da2da375 177 info->hmp = hmp;
7a61b85d 178 TAILQ_INSERT_TAIL(&hmp->flusher.ready_list, info, entry);
da2da375
MD
179 lwkt_create(hammer_flusher_slave_thread, info,
180 &info->td, NULL, 0, -1, "hammer-S%d", i);
181 }
059819e3
MD
182}
183
184void
185hammer_flusher_destroy(hammer_mount_t hmp)
186{
da2da375 187 hammer_flusher_info_t info;
da2da375
MD
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 */
7a61b85d
MD
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);
bac808fe 209 kfree(info, hmp->m_misc);
f90dde4c 210 }
059819e3
MD
211}
212
af209b0f
MD
213/*
214 * The master flusher thread manages the flusher sequence id and
215 * synchronization with the slave work threads.
216 */
059819e3 217static void
da2da375 218hammer_flusher_master_thread(void *arg)
059819e3 219{
7a61b85d
MD
220 hammer_flush_group_t flg;
221 hammer_mount_t hmp;
0729c8c8 222
7a61b85d 223 hmp = arg;
c32a6806 224
7a61b85d 225 for (;;) {
c32a6806 226 /*
7a61b85d
MD
227 * Do at least one flush cycle. We may have to update the
228 * UNDO FIFO even if no inodes are queued.
c32a6806 229 */
7a61b85d
MD
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;
cdb6e4e6
MD
242 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
243 break;
7a61b85d 244 }
1f07f686 245
c9b9e29d 246 /*
7a61b85d 247 * Wait for activity.
c9b9e29d 248 */
7a61b85d
MD
249 if (hmp->flusher.exiting && TAILQ_EMPTY(&hmp->flush_group_list))
250 break;
da2da375
MD
251 while (hmp->flusher.signal == 0)
252 tsleep(&hmp->flusher.signal, 0, "hmrwwa", 0);
4889cbd4
MD
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;
059819e3 260 }
da2da375
MD
261
262 /*
263 * And we are done.
264 */
265 hmp->flusher.td = NULL;
266 wakeup(&hmp->flusher.exiting);
267 lwkt_exit();
268}
269
af209b0f 270/*
7a61b85d
MD
271 * Flush all inodes in the current flush group.
272 */
273static void
274hammer_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;
15e75dab 282 int count;
7a61b85d
MD
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 */
15e75dab 293 count = 0;
7a61b85d 294 while ((flg = TAILQ_FIRST(&hmp->flush_group_list)) != NULL) {
15e75dab 295 ++count;
7a61b85d
MD
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 }
cdb6e4e6
MD
301 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
302 break;
7a61b85d
MD
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 /*
7b6ccb11
MD
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 /*
ff003b11 323 * Iterate the inodes in the flg's flush_tree and assign
7a61b85d
MD
324 * them to slaves.
325 */
7a61b85d
MD
326 slave_index = 0;
327 info = TAILQ_FIRST(&hmp->flusher.ready_list);
ff003b11 328 next_ip = RB_FIRST(hammer_fls_rb_tree, &flg->flush_tree);
7a61b85d
MD
329
330 while ((ip = next_ip) != NULL) {
ff003b11
MD
331 next_ip = RB_NEXT(hammer_fls_rb_tree,
332 &flg->flush_tree, ip);
7a61b85d 333
3e583440
MD
334 if (++hmp->check_yield > hammer_yield_check) {
335 hmp->check_yield = 0;
336 lwkt_user_yield();
337 }
338
7a61b85d
MD
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 */
ff003b11 394 if (RB_EMPTY(&flg->flush_tree)) {
7a61b85d
MD
395 KKASSERT(flg->refs == 0);
396 TAILQ_REMOVE(&hmp->flush_group_list, flg, flush_entry);
bac808fe 397 kfree(flg, hmp->m_misc);
7a61b85d
MD
398 break;
399 }
400 }
401
402 /*
1b0ab2c3
MD
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.
15e75dab 405 */
1b0ab2c3 406 if (count == 0 && hammer_flusher_haswork(hmp)) {
15e75dab
MD
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 /*
7a61b85d
MD
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/*
ff003b11 427 * The slave flusher thread pulls work off the master flush list until no
af209b0f
MD
428 * work is left.
429 */
da2da375
MD
430static void
431hammer_flusher_slave_thread(void *arg)
432{
7a61b85d 433 hammer_flush_group_t flg;
da2da375
MD
434 hammer_flusher_info_t info;
435 hammer_mount_t hmp;
436 hammer_inode_t ip;
cb51be26 437 int i;
da2da375
MD
438
439 info = arg;
440 hmp = info->hmp;
441
442 for (;;) {
7a61b85d
MD
443 while (info->runstate == 0)
444 tsleep(&info->runstate, 0, "hmrssw", 0);
445 if (info->runstate < 0)
da2da375 446 break;
7a61b85d 447 flg = info->flg;
cb51be26 448
7a61b85d
MD
449 for (i = 0; i < info->count; ++i) {
450 ip = info->work_array[i];
451 hammer_flusher_flush_inode(ip, &hmp->flusher.trans);
ce0138a6 452 ++hammer_stats_inode_flushes;
cb51be26 453 }
7a61b85d
MD
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);
da2da375
MD
459 }
460 info->td = NULL;
461 wakeup(&info->td);
059819e3
MD
462 lwkt_exit();
463}
464
525aad3a 465void
10a5d1ba
MD
466hammer_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 */
525aad3a
MD
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();
10a5d1ba
MD
489 }
490}
491
059819e3 492/*
9f5097dc 493 * Flush a single inode that is part of a flush group.
06ad81ff 494 *
cdb6e4e6
MD
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.
9f5097dc
MD
498 */
499static
500void
501hammer_flusher_flush_inode(hammer_inode_t ip, hammer_transaction_t trans)
502{
503 hammer_mount_t hmp = ip->hmp;
06ad81ff 504 int error;
9f5097dc 505
525aad3a 506 hammer_flusher_clean_loose_ios(hmp);
02325004 507 error = hammer_sync_inode(trans, ip);
cdb6e4e6
MD
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);
da2da375
MD
521 while (hmp->flusher.finalize_want)
522 tsleep(&hmp->flusher.finalize_want, 0, "hmrsxx", 0);
06ad81ff 523 if (hammer_flusher_undo_exhausted(trans, 1)) {
5a930e66 524 kprintf("HAMMER: Warning: UNDO area too small!\n");
9f5097dc 525 hammer_flusher_finalize(trans, 1);
06ad81ff 526 } else if (hammer_flusher_meta_limit(trans->hmp)) {
9f5097dc
MD
527 hammer_flusher_finalize(trans, 0);
528 }
529}
530
531/*
06ad81ff
MD
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.
ec4e8497 544 */
ec4e8497 545int
06ad81ff 546hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter)
ec4e8497 547{
06ad81ff
MD
548 if (hammer_undo_space(trans) <
549 hammer_undo_max(trans->hmp) * quarter / 4) {
1f07f686 550 return(1);
ec4e8497 551 } else {
1f07f686 552 return(0);
ec4e8497 553 }
ec4e8497
MD
554}
555
556/*
9f5097dc
MD
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.
10a5d1ba 560 *
9f5097dc
MD
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.
6c1f89f4
MD
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.
10a5d1ba 568 */
10a5d1ba 569void
9f5097dc 570hammer_flusher_finalize(hammer_transaction_t trans, int final)
059819e3 571{
9f5097dc
MD
572 hammer_volume_t root_volume;
573 hammer_blockmap_t cundomap, dundomap;
574 hammer_mount_t hmp;
10a5d1ba 575 hammer_io_t io;
c9b9e29d 576 int count;
19619882 577 int i;
059819e3 578
9f5097dc
MD
579 hmp = trans->hmp;
580 root_volume = trans->rootvol;
581
47637bff 582 /*
6c1f89f4
MD
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
cdb6e4e6
MD
597 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
598 goto done;
599
6c1f89f4 600 /*
47637bff 601 * Flush data buffers. This can occur asynchronously and at any
9f5097dc
MD
602 * time. We must interlock against the frontend direct-data write
603 * but do not have to acquire the sync-lock yet.
47637bff
MD
604 */
605 count = 0;
606 while ((io = TAILQ_FIRST(&hmp->data_list)) != NULL) {
cdb6e4e6
MD
607 if (io->ioerror)
608 break;
a99b9ea2
MD
609 if (io->lock.refs == 0)
610 ++hammer_count_refedbufs;
47637bff 611 hammer_ref(&io->lock);
9f5097dc 612 hammer_io_write_interlock(io);
47637bff 613 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
710733a6 614 hammer_io_flush(io, 0);
9f5097dc 615 hammer_io_done_interlock(io);
47637bff
MD
616 hammer_rel_buffer((hammer_buffer_t)io, 0);
617 ++count;
618 }
47637bff 619
9f5097dc
MD
620 /*
621 * The sync-lock is required for the remaining sequence. This lock
622 * prevents meta-data from being modified.
623 */
2f85fa4d 624 hammer_sync_lock_ex(trans);
9480ff55 625
059819e3 626 /*
9f5097dc
MD
627 * If we have been asked to finalize the volume header sync the
628 * cached blockmap to the on-disk blockmap. Generate an UNDO
629 * record for the update.
e8599db1 630 */
9f5097dc
MD
631 if (final) {
632 cundomap = &hmp->blockmap[0];
633 dundomap = &root_volume->ondisk->vol0_blockmap[0];
634 if (root_volume->io.modified) {
635 hammer_modify_volume(trans, root_volume,
636 dundomap, sizeof(hmp->blockmap));
637 for (i = 0; i < HAMMER_MAX_ZONES; ++i)
638 hammer_crc_set_blockmap(&cundomap[i]);
639 bcopy(cundomap, dundomap, sizeof(hmp->blockmap));
640 hammer_modify_volume_done(root_volume);
641 }
e8599db1
MD
642 }
643
644 /*
9f5097dc 645 * Flush UNDOs
059819e3 646 */
c9b9e29d 647 count = 0;
10a5d1ba 648 while ((io = TAILQ_FIRST(&hmp->undo_list)) != NULL) {
cdb6e4e6
MD
649 if (io->ioerror)
650 break;
10a5d1ba 651 KKASSERT(io->modify_refs == 0);
a99b9ea2
MD
652 if (io->lock.refs == 0)
653 ++hammer_count_refedbufs;
10a5d1ba
MD
654 hammer_ref(&io->lock);
655 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
710733a6 656 hammer_io_flush(io, hammer_undo_reclaim(io));
09ac686b 657 hammer_rel_buffer((hammer_buffer_t)io, 0);
c9b9e29d 658 ++count;
059819e3
MD
659 }
660
661 /*
02428fb6 662 * Wait for I/Os to complete and flush the cache on the target disk.
059819e3 663 */
a99b9ea2 664 hammer_flusher_clean_loose_ios(hmp);
af209b0f 665 hammer_io_wait_all(hmp, "hmrfl1");
059819e3 666
cdb6e4e6
MD
667 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
668 goto failed;
669
059819e3 670 /*
02428fb6
MD
671 * HAMMER VERSION < 4:
672 * Update the on-disk volume header with new UNDO FIFO end
673 * position (do not generate new UNDO records for this change).
674 * We have to do this for the UNDO FIFO whether (final) is
675 * set or not in order for the UNDOs to be recognized on
676 * recovery.
677 *
678 * HAMMER VERSION >= 4:
679 * The UNDO FIFO data written above will be recognized on
680 * recovery without us having to sync the volume header.
9f5097dc
MD
681 *
682 * Also update the on-disk next_tid field. This does not require
683 * an UNDO. However, because our TID is generated before we get
684 * the sync lock another sync may have beat us to the punch.
685 *
06ad81ff
MD
686 * This also has the side effect of updating first_offset based on
687 * a prior finalization when the first finalization of the next flush
688 * cycle occurs, removing any undo info from the prior finalization
689 * from consideration.
690 *
9f5097dc 691 * The volume header will be flushed out synchronously.
059819e3 692 */
9f5097dc
MD
693 dundomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
694 cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
09ac686b 695
9f5097dc 696 if (dundomap->first_offset != cundomap->first_offset ||
cdb6e4e6 697 dundomap->next_offset != cundomap->next_offset) {
0729c8c8 698 hammer_modify_volume(NULL, root_volume, NULL, 0);
9f5097dc
MD
699 dundomap->first_offset = cundomap->first_offset;
700 dundomap->next_offset = cundomap->next_offset;
701 hammer_crc_set_blockmap(dundomap);
0729c8c8
MD
702 hammer_modify_volume_done(root_volume);
703 }
c9b9e29d 704
4889cbd4
MD
705 /*
706 * vol0_next_tid is used for TID selection and is updated without
707 * an UNDO so we do not reuse a TID that may have been rolled-back.
708 *
709 * vol0_last_tid is the highest fully-synchronized TID. It is
710 * set-up when the UNDO fifo is fully synced, later on (not here).
711 */
19619882 712 if (root_volume->io.modified) {
adf01747
MD
713 hammer_modify_volume(NULL, root_volume, NULL, 0);
714 if (root_volume->ondisk->vol0_next_tid < trans->tid)
715 root_volume->ondisk->vol0_next_tid = trans->tid;
716 hammer_crc_set_volume(root_volume->ondisk);
717 hammer_modify_volume_done(root_volume);
710733a6 718 hammer_io_flush(&root_volume->io, 0);
19619882 719 }
059819e3
MD
720
721 /*
02428fb6
MD
722 * Wait for I/Os to complete.
723 *
724 * For HAMMER VERSION 4+ filesystems we do not have to wait for
725 * the I/O to complete as the new UNDO FIFO entries are recognized
726 * even without the volume header update. This allows the volume
727 * header to flushed along with meta-data, significantly reducing
728 * flush overheads.
059819e3 729 */
a99b9ea2 730 hammer_flusher_clean_loose_ios(hmp);
02428fb6
MD
731 if (hmp->version < HAMMER_VOL_VERSION_FOUR)
732 hammer_io_wait_all(hmp, "hmrfl2");
059819e3 733
cdb6e4e6
MD
734 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
735 goto failed;
736
059819e3 737 /*
e8599db1 738 * Flush meta-data. The meta-data will be undone if we crash
02428fb6
MD
739 * so we can safely flush it asynchronously. There is no need
740 * to wait for I/O to complete (or issue a synchronous disk flush).
9f5097dc 741 *
02428fb6
MD
742 * In fact, even if we did wait the meta-data will still be undone
743 * by a crash up until the next flush cycle due to the first_offset
744 * in the volume header for the UNDO FIFO not being adjusted until
745 * the following flush cycle.
059819e3 746 */
c9b9e29d 747 count = 0;
10a5d1ba 748 while ((io = TAILQ_FIRST(&hmp->meta_list)) != NULL) {
cdb6e4e6
MD
749 if (io->ioerror)
750 break;
10a5d1ba 751 KKASSERT(io->modify_refs == 0);
a99b9ea2
MD
752 if (io->lock.refs == 0)
753 ++hammer_count_refedbufs;
10a5d1ba
MD
754 hammer_ref(&io->lock);
755 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
710733a6 756 hammer_io_flush(io, 0);
09ac686b 757 hammer_rel_buffer((hammer_buffer_t)io, 0);
c9b9e29d 758 ++count;
059819e3 759 }
9f5097dc
MD
760
761 /*
762 * If this is the final finalization for the flush group set
763 * up for the next sequence by setting a new first_offset in
06ad81ff
MD
764 * our cached blockmap and clearing the undo history.
765 *
766 * Even though we have updated our cached first_offset, the on-disk
767 * first_offset still governs available-undo-space calculations.
9f5097dc
MD
768 */
769 if (final) {
770 cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
1b0ab2c3
MD
771 if (cundomap->first_offset == cundomap->next_offset) {
772 hmp->hflags &= ~HMNT_UNDO_DIRTY;
773 } else {
774 cundomap->first_offset = cundomap->next_offset;
775 hmp->hflags |= HMNT_UNDO_DIRTY;
776 }
9f5097dc 777 hammer_clear_undo_history(hmp);
4889cbd4
MD
778
779 /*
780 * Flush tid sequencing. flush_tid1 is fully synchronized,
781 * meaning a crash will not roll it back. flush_tid2 has
782 * been written out asynchronously and a crash will roll
783 * it back. flush_tid1 is used for all mirroring masters.
784 */
785 if (hmp->flush_tid1 != hmp->flush_tid2) {
786 hmp->flush_tid1 = hmp->flush_tid2;
787 wakeup(&hmp->flush_tid1);
788 }
789 hmp->flush_tid2 = trans->tid;
9f5097dc
MD
790 }
791
cdb6e4e6
MD
792 /*
793 * Cleanup. Report any critical errors.
794 */
795failed:
2f85fa4d 796 hammer_sync_unlock(trans);
6c1f89f4 797
cdb6e4e6
MD
798 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) {
799 kprintf("HAMMER(%s): Critical write error during flush, "
800 "refusing to sync UNDO FIFO\n",
801 root_volume->ondisk->vol_name);
802 }
803
6c1f89f4
MD
804done:
805 hammer_unlock(&hmp->flusher.finalize_lock);
4889cbd4 806
6c1f89f4
MD
807 if (--hmp->flusher.finalize_want == 0)
808 wakeup(&hmp->flusher.finalize_want);
ce0138a6 809 hammer_stats_commits += final;
059819e3
MD
810}
811
06ad81ff
MD
812/*
813 * Return non-zero if too many dirty meta-data buffers have built up.
814 *
815 * Since we cannot allow such buffers to flush until we have dealt with
816 * the UNDOs, we risk deadlocking the kernel's buffer cache.
817 */
818int
819hammer_flusher_meta_limit(hammer_mount_t hmp)
820{
f5a07a7a
MD
821 if (hmp->locked_dirty_space + hmp->io_running_space >
822 hammer_limit_dirtybufspace) {
06ad81ff
MD
823 return(1);
824 }
825 return(0);
826}
827
1b0ab2c3
MD
828/*
829 * Return non-zero if too many dirty meta-data buffers have built up.
830 *
831 * This version is used by background operations (mirror, prune, reblock)
832 * to leave room for foreground operations.
833 */
93291532
MD
834int
835hammer_flusher_meta_halflimit(hammer_mount_t hmp)
836{
837 if (hmp->locked_dirty_space + hmp->io_running_space >
838 hammer_limit_dirtybufspace / 2) {
839 return(1);
840 }
841 return(0);
842}
843
1b0ab2c3
MD
844/*
845 * Return non-zero if the flusher still has something to flush.
846 */
847int
848hammer_flusher_haswork(hammer_mount_t hmp)
849{
cdb6e4e6
MD
850 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
851 return(0);
1b0ab2c3
MD
852 if (TAILQ_FIRST(&hmp->flush_group_list) || /* dirty inodes */
853 TAILQ_FIRST(&hmp->volu_list) || /* dirty bufffers */
854 TAILQ_FIRST(&hmp->undo_list) ||
855 TAILQ_FIRST(&hmp->data_list) ||
856 TAILQ_FIRST(&hmp->meta_list) ||
857 (hmp->hflags & HMNT_UNDO_DIRTY) /* UNDO FIFO sync */
858 ) {
859 return(1);
860 }
861 return(0);
862}
863