xref: /illumos-gate/usr/src/uts/common/fs/zfs/txg.c (revision 3d393ee6c37fa10ac512ed6d36109ad616dc7c1a)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/txg_impl.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dsl_pool.h>
30 #include <sys/callb.h>
31 
32 /*
33  * Pool-wide transaction groups.
34  */
35 
36 static void txg_sync_thread(dsl_pool_t *dp);
37 static void txg_quiesce_thread(dsl_pool_t *dp);
38 
39 int zfs_txg_timeout = 30;	/* max seconds worth of delta per txg */
40 
41 /*
42  * Prepare the txg subsystem.
43  */
44 void
45 txg_init(dsl_pool_t *dp, uint64_t txg)
46 {
47 	tx_state_t *tx = &dp->dp_tx;
48 	int c;
49 	bzero(tx, sizeof (tx_state_t));
50 
51 	tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
52 
53 	for (c = 0; c < max_ncpus; c++) {
54 		int i;
55 
56 		mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
57 		for (i = 0; i < TXG_SIZE; i++) {
58 			cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
59 			    NULL);
60 		}
61 	}
62 
63 	rw_init(&tx->tx_suspend, NULL, RW_DEFAULT, NULL);
64 	mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
65 
66 	cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
67 	cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
68 	cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
69 	cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
70 	cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
71 
72 	tx->tx_open_txg = txg;
73 }
74 
75 /*
76  * Close down the txg subsystem.
77  */
78 void
79 txg_fini(dsl_pool_t *dp)
80 {
81 	tx_state_t *tx = &dp->dp_tx;
82 	int c;
83 
84 	ASSERT(tx->tx_threads == 0);
85 
86 	rw_destroy(&tx->tx_suspend);
87 	mutex_destroy(&tx->tx_sync_lock);
88 
89 	cv_destroy(&tx->tx_sync_more_cv);
90 	cv_destroy(&tx->tx_sync_done_cv);
91 	cv_destroy(&tx->tx_quiesce_more_cv);
92 	cv_destroy(&tx->tx_quiesce_done_cv);
93 	cv_destroy(&tx->tx_exit_cv);
94 
95 	for (c = 0; c < max_ncpus; c++) {
96 		int i;
97 
98 		mutex_destroy(&tx->tx_cpu[c].tc_lock);
99 		for (i = 0; i < TXG_SIZE; i++)
100 			cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
101 	}
102 
103 	kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
104 
105 	bzero(tx, sizeof (tx_state_t));
106 }
107 
108 /*
109  * Start syncing transaction groups.
110  */
111 void
112 txg_sync_start(dsl_pool_t *dp)
113 {
114 	tx_state_t *tx = &dp->dp_tx;
115 
116 	mutex_enter(&tx->tx_sync_lock);
117 
118 	dprintf("pool %p\n", dp);
119 
120 	ASSERT(tx->tx_threads == 0);
121 
122 	tx->tx_threads = 2;
123 
124 	tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
125 	    dp, 0, &p0, TS_RUN, minclsyspri);
126 
127 	/*
128 	 * The sync thread can need a larger-than-default stack size on
129 	 * 32-bit x86.  This is due in part to nested pools and
130 	 * scrub_visitbp() recursion.
131 	 */
132 	tx->tx_sync_thread = thread_create(NULL, 12<<10, txg_sync_thread,
133 	    dp, 0, &p0, TS_RUN, minclsyspri);
134 
135 	mutex_exit(&tx->tx_sync_lock);
136 }
137 
138 static void
139 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
140 {
141 	CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
142 	mutex_enter(&tx->tx_sync_lock);
143 }
144 
145 static void
146 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
147 {
148 	ASSERT(*tpp != NULL);
149 	*tpp = NULL;
150 	tx->tx_threads--;
151 	cv_broadcast(&tx->tx_exit_cv);
152 	CALLB_CPR_EXIT(cpr);		/* drops &tx->tx_sync_lock */
153 	thread_exit();
154 }
155 
156 static void
157 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, uint64_t time)
158 {
159 	CALLB_CPR_SAFE_BEGIN(cpr);
160 
161 	if (time)
162 		(void) cv_timedwait(cv, &tx->tx_sync_lock, lbolt + time);
163 	else
164 		cv_wait(cv, &tx->tx_sync_lock);
165 
166 	CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
167 }
168 
169 /*
170  * Stop syncing transaction groups.
171  */
172 void
173 txg_sync_stop(dsl_pool_t *dp)
174 {
175 	tx_state_t *tx = &dp->dp_tx;
176 
177 	dprintf("pool %p\n", dp);
178 	/*
179 	 * Finish off any work in progress.
180 	 */
181 	ASSERT(tx->tx_threads == 2);
182 	txg_wait_synced(dp, 0);
183 
184 	/*
185 	 * Wake all sync threads and wait for them to die.
186 	 */
187 	mutex_enter(&tx->tx_sync_lock);
188 
189 	ASSERT(tx->tx_threads == 2);
190 
191 	tx->tx_exiting = 1;
192 
193 	cv_broadcast(&tx->tx_quiesce_more_cv);
194 	cv_broadcast(&tx->tx_quiesce_done_cv);
195 	cv_broadcast(&tx->tx_sync_more_cv);
196 
197 	while (tx->tx_threads != 0)
198 		cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
199 
200 	tx->tx_exiting = 0;
201 
202 	mutex_exit(&tx->tx_sync_lock);
203 }
204 
205 uint64_t
206 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
207 {
208 	tx_state_t *tx = &dp->dp_tx;
209 	tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
210 	uint64_t txg;
211 
212 	mutex_enter(&tc->tc_lock);
213 
214 	txg = tx->tx_open_txg;
215 	tc->tc_count[txg & TXG_MASK]++;
216 
217 	th->th_cpu = tc;
218 	th->th_txg = txg;
219 
220 	return (txg);
221 }
222 
223 void
224 txg_rele_to_quiesce(txg_handle_t *th)
225 {
226 	tx_cpu_t *tc = th->th_cpu;
227 
228 	mutex_exit(&tc->tc_lock);
229 }
230 
231 void
232 txg_rele_to_sync(txg_handle_t *th)
233 {
234 	tx_cpu_t *tc = th->th_cpu;
235 	int g = th->th_txg & TXG_MASK;
236 
237 	mutex_enter(&tc->tc_lock);
238 	ASSERT(tc->tc_count[g] != 0);
239 	if (--tc->tc_count[g] == 0)
240 		cv_broadcast(&tc->tc_cv[g]);
241 	mutex_exit(&tc->tc_lock);
242 
243 	th->th_cpu = NULL;	/* defensive */
244 }
245 
246 static void
247 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
248 {
249 	tx_state_t *tx = &dp->dp_tx;
250 	int g = txg & TXG_MASK;
251 	int c;
252 
253 	/*
254 	 * Grab all tx_cpu locks so nobody else can get into this txg.
255 	 */
256 	for (c = 0; c < max_ncpus; c++)
257 		mutex_enter(&tx->tx_cpu[c].tc_lock);
258 
259 	ASSERT(txg == tx->tx_open_txg);
260 	tx->tx_open_txg++;
261 
262 	/*
263 	 * Now that we've incremented tx_open_txg, we can let threads
264 	 * enter the next transaction group.
265 	 */
266 	for (c = 0; c < max_ncpus; c++)
267 		mutex_exit(&tx->tx_cpu[c].tc_lock);
268 
269 	/*
270 	 * Quiesce the transaction group by waiting for everyone to txg_exit().
271 	 */
272 	for (c = 0; c < max_ncpus; c++) {
273 		tx_cpu_t *tc = &tx->tx_cpu[c];
274 		mutex_enter(&tc->tc_lock);
275 		while (tc->tc_count[g] != 0)
276 			cv_wait(&tc->tc_cv[g], &tc->tc_lock);
277 		mutex_exit(&tc->tc_lock);
278 	}
279 }
280 
281 static void
282 txg_sync_thread(dsl_pool_t *dp)
283 {
284 	tx_state_t *tx = &dp->dp_tx;
285 	callb_cpr_t cpr;
286 	uint64_t start, delta;
287 
288 	txg_thread_enter(tx, &cpr);
289 
290 	start = delta = 0;
291 	for (;;) {
292 		uint64_t timer, timeout = zfs_txg_timeout * hz;
293 		uint64_t txg;
294 
295 		/*
296 		 * We sync when we're scrubbing, there's someone waiting
297 		 * on us, or the quiesce thread has handed off a txg to
298 		 * us, or we have reached our timeout.
299 		 */
300 		timer = (delta >= timeout ? 0 : timeout - delta);
301 		while ((dp->dp_scrub_func == SCRUB_FUNC_NONE ||
302 		    spa_shutting_down(dp->dp_spa)) &&
303 		    !tx->tx_exiting && timer > 0 &&
304 		    tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
305 		    tx->tx_quiesced_txg == 0) {
306 			dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
307 			    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
308 			txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
309 			delta = lbolt - start;
310 			timer = (delta > timeout ? 0 : timeout - delta);
311 		}
312 
313 		/*
314 		 * Wait until the quiesce thread hands off a txg to us,
315 		 * prompting it to do so if necessary.
316 		 */
317 		while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
318 			if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
319 				tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
320 			cv_broadcast(&tx->tx_quiesce_more_cv);
321 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
322 		}
323 
324 		if (tx->tx_exiting)
325 			txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
326 
327 		rw_enter(&tx->tx_suspend, RW_WRITER);
328 
329 		/*
330 		 * Consume the quiesced txg which has been handed off to
331 		 * us.  This may cause the quiescing thread to now be
332 		 * able to quiesce another txg, so we must signal it.
333 		 */
334 		txg = tx->tx_quiesced_txg;
335 		tx->tx_quiesced_txg = 0;
336 		tx->tx_syncing_txg = txg;
337 		cv_broadcast(&tx->tx_quiesce_more_cv);
338 		rw_exit(&tx->tx_suspend);
339 
340 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
341 		    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
342 		mutex_exit(&tx->tx_sync_lock);
343 
344 		start = lbolt;
345 		spa_sync(dp->dp_spa, txg);
346 		delta = lbolt - start;
347 
348 		mutex_enter(&tx->tx_sync_lock);
349 		rw_enter(&tx->tx_suspend, RW_WRITER);
350 		tx->tx_synced_txg = txg;
351 		tx->tx_syncing_txg = 0;
352 		rw_exit(&tx->tx_suspend);
353 		cv_broadcast(&tx->tx_sync_done_cv);
354 	}
355 }
356 
357 static void
358 txg_quiesce_thread(dsl_pool_t *dp)
359 {
360 	tx_state_t *tx = &dp->dp_tx;
361 	callb_cpr_t cpr;
362 
363 	txg_thread_enter(tx, &cpr);
364 
365 	for (;;) {
366 		uint64_t txg;
367 
368 		/*
369 		 * We quiesce when there's someone waiting on us.
370 		 * However, we can only have one txg in "quiescing" or
371 		 * "quiesced, waiting to sync" state.  So we wait until
372 		 * the "quiesced, waiting to sync" txg has been consumed
373 		 * by the sync thread.
374 		 */
375 		while (!tx->tx_exiting &&
376 		    (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
377 		    tx->tx_quiesced_txg != 0))
378 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
379 
380 		if (tx->tx_exiting)
381 			txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
382 
383 		txg = tx->tx_open_txg;
384 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
385 		    txg, tx->tx_quiesce_txg_waiting,
386 		    tx->tx_sync_txg_waiting);
387 		mutex_exit(&tx->tx_sync_lock);
388 		txg_quiesce(dp, txg);
389 		mutex_enter(&tx->tx_sync_lock);
390 
391 		/*
392 		 * Hand this txg off to the sync thread.
393 		 */
394 		dprintf("quiesce done, handing off txg %llu\n", txg);
395 		tx->tx_quiesced_txg = txg;
396 		cv_broadcast(&tx->tx_sync_more_cv);
397 		cv_broadcast(&tx->tx_quiesce_done_cv);
398 	}
399 }
400 
401 /*
402  * Delay this thread by 'ticks' if we are still in the open transaction
403  * group and there is already a waiting txg quiesing or quiesced.  Abort
404  * the delay if this txg stalls or enters the quiesing state.
405  */
406 void
407 txg_delay(dsl_pool_t *dp, uint64_t txg, int ticks)
408 {
409 	tx_state_t *tx = &dp->dp_tx;
410 	int timeout = lbolt + ticks;
411 
412 	/* don't delay if this txg could transition to quiesing immediately */
413 	if (tx->tx_open_txg > txg ||
414 	    tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
415 		return;
416 
417 	mutex_enter(&tx->tx_sync_lock);
418 	if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
419 		mutex_exit(&tx->tx_sync_lock);
420 		return;
421 	}
422 
423 	while (lbolt < timeout &&
424 	    tx->tx_syncing_txg < txg-1 && !txg_stalled(dp))
425 		(void) cv_timedwait(&tx->tx_quiesce_more_cv, &tx->tx_sync_lock,
426 		    timeout);
427 
428 	mutex_exit(&tx->tx_sync_lock);
429 }
430 
431 void
432 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
433 {
434 	tx_state_t *tx = &dp->dp_tx;
435 
436 	mutex_enter(&tx->tx_sync_lock);
437 	ASSERT(tx->tx_threads == 2);
438 	if (txg == 0)
439 		txg = tx->tx_open_txg;
440 	if (tx->tx_sync_txg_waiting < txg)
441 		tx->tx_sync_txg_waiting = txg;
442 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
443 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
444 	while (tx->tx_synced_txg < txg) {
445 		dprintf("broadcasting sync more "
446 		    "tx_synced=%llu waiting=%llu dp=%p\n",
447 		    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
448 		cv_broadcast(&tx->tx_sync_more_cv);
449 		cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
450 	}
451 	mutex_exit(&tx->tx_sync_lock);
452 }
453 
454 void
455 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
456 {
457 	tx_state_t *tx = &dp->dp_tx;
458 
459 	mutex_enter(&tx->tx_sync_lock);
460 	ASSERT(tx->tx_threads == 2);
461 	if (txg == 0)
462 		txg = tx->tx_open_txg + 1;
463 	if (tx->tx_quiesce_txg_waiting < txg)
464 		tx->tx_quiesce_txg_waiting = txg;
465 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
466 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
467 	while (tx->tx_open_txg < txg) {
468 		cv_broadcast(&tx->tx_quiesce_more_cv);
469 		cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
470 	}
471 	mutex_exit(&tx->tx_sync_lock);
472 }
473 
474 boolean_t
475 txg_stalled(dsl_pool_t *dp)
476 {
477 	tx_state_t *tx = &dp->dp_tx;
478 	return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
479 }
480 
481 boolean_t
482 txg_sync_waiting(dsl_pool_t *dp)
483 {
484 	tx_state_t *tx = &dp->dp_tx;
485 
486 	return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
487 	    tx->tx_quiesced_txg != 0);
488 }
489 
490 void
491 txg_suspend(dsl_pool_t *dp)
492 {
493 	tx_state_t *tx = &dp->dp_tx;
494 	/* XXX some code paths suspend when they are already suspended! */
495 	rw_enter(&tx->tx_suspend, RW_READER);
496 }
497 
498 void
499 txg_resume(dsl_pool_t *dp)
500 {
501 	tx_state_t *tx = &dp->dp_tx;
502 	rw_exit(&tx->tx_suspend);
503 }
504 
505 /*
506  * Per-txg object lists.
507  */
508 void
509 txg_list_create(txg_list_t *tl, size_t offset)
510 {
511 	int t;
512 
513 	mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
514 
515 	tl->tl_offset = offset;
516 
517 	for (t = 0; t < TXG_SIZE; t++)
518 		tl->tl_head[t] = NULL;
519 }
520 
521 void
522 txg_list_destroy(txg_list_t *tl)
523 {
524 	int t;
525 
526 	for (t = 0; t < TXG_SIZE; t++)
527 		ASSERT(txg_list_empty(tl, t));
528 
529 	mutex_destroy(&tl->tl_lock);
530 }
531 
532 int
533 txg_list_empty(txg_list_t *tl, uint64_t txg)
534 {
535 	return (tl->tl_head[txg & TXG_MASK] == NULL);
536 }
537 
538 /*
539  * Add an entry to the list.
540  * Returns 0 if it's a new entry, 1 if it's already there.
541  */
542 int
543 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
544 {
545 	int t = txg & TXG_MASK;
546 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
547 	int already_on_list;
548 
549 	mutex_enter(&tl->tl_lock);
550 	already_on_list = tn->tn_member[t];
551 	if (!already_on_list) {
552 		tn->tn_member[t] = 1;
553 		tn->tn_next[t] = tl->tl_head[t];
554 		tl->tl_head[t] = tn;
555 	}
556 	mutex_exit(&tl->tl_lock);
557 
558 	return (already_on_list);
559 }
560 
561 /*
562  * Remove the head of the list and return it.
563  */
564 void *
565 txg_list_remove(txg_list_t *tl, uint64_t txg)
566 {
567 	int t = txg & TXG_MASK;
568 	txg_node_t *tn;
569 	void *p = NULL;
570 
571 	mutex_enter(&tl->tl_lock);
572 	if ((tn = tl->tl_head[t]) != NULL) {
573 		p = (char *)tn - tl->tl_offset;
574 		tl->tl_head[t] = tn->tn_next[t];
575 		tn->tn_next[t] = NULL;
576 		tn->tn_member[t] = 0;
577 	}
578 	mutex_exit(&tl->tl_lock);
579 
580 	return (p);
581 }
582 
583 /*
584  * Remove a specific item from the list and return it.
585  */
586 void *
587 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
588 {
589 	int t = txg & TXG_MASK;
590 	txg_node_t *tn, **tp;
591 
592 	mutex_enter(&tl->tl_lock);
593 
594 	for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
595 		if ((char *)tn - tl->tl_offset == p) {
596 			*tp = tn->tn_next[t];
597 			tn->tn_next[t] = NULL;
598 			tn->tn_member[t] = 0;
599 			mutex_exit(&tl->tl_lock);
600 			return (p);
601 		}
602 	}
603 
604 	mutex_exit(&tl->tl_lock);
605 
606 	return (NULL);
607 }
608 
609 int
610 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
611 {
612 	int t = txg & TXG_MASK;
613 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
614 
615 	return (tn->tn_member[t]);
616 }
617 
618 /*
619  * Walk a txg list -- only safe if you know it's not changing.
620  */
621 void *
622 txg_list_head(txg_list_t *tl, uint64_t txg)
623 {
624 	int t = txg & TXG_MASK;
625 	txg_node_t *tn = tl->tl_head[t];
626 
627 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
628 }
629 
630 void *
631 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
632 {
633 	int t = txg & TXG_MASK;
634 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
635 
636 	tn = tn->tn_next[t];
637 
638 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
639 }
640