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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Portions Copyright 2011 Martin Matuska
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/txg_impl.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dsl_scan.h>
33 #include <sys/callb.h>
34
35 /*
36 * ZFS Transaction Groups
37 * ----------------------
38 *
39 * ZFS transaction groups are, as the name implies, groups of transactions
40 * that act on persistent state. ZFS asserts consistency at the granularity of
41 * these transaction groups. Each successive transaction group (txg) is
42 * assigned a 64-bit consecutive identifier. There are three active
43 * transaction group states: open, quiescing, or syncing. At any given time,
44 * there may be an active txg associated with each state; each active txg may
45 * either be processing, or blocked waiting to enter the next state. There may
46 * be up to three active txgs, and there is always a txg in the open state
47 * (though it may be blocked waiting to enter the quiescing state). In broad
48 * strokes, transactions — operations that change in-memory structures — are
49 * accepted into the txg in the open state, and are completed while the txg is
50 * in the open or quiescing states. The accumulated changes are written to
51 * disk in the syncing state.
52 *
53 * Open
54 *
55 * When a new txg becomes active, it first enters the open state. New
56 * transactions — updates to in-memory structures — are assigned to the
57 * currently open txg. There is always a txg in the open state so that ZFS can
58 * accept new changes (though the txg may refuse new changes if it has hit
59 * some limit). ZFS advances the open txg to the next state for a variety of
60 * reasons such as it hitting a time or size threshold, or the execution of an
61 * administrative action that must be completed in the syncing state.
62 *
63 * Quiescing
64 *
65 * After a txg exits the open state, it enters the quiescing state. The
66 * quiescing state is intended to provide a buffer between accepting new
67 * transactions in the open state and writing them out to stable storage in
68 * the syncing state. While quiescing, transactions can continue their
69 * operation without delaying either of the other states. Typically, a txg is
70 * in the quiescing state very briefly since the operations are bounded by
71 * software latencies rather than, say, slower I/O latencies. After all
72 * transactions complete, the txg is ready to enter the next state.
73 *
74 * Syncing
75 *
76 * In the syncing state, the in-memory state built up during the open and (to
77 * a lesser degree) the quiescing states is written to stable storage. The
78 * process of writing out modified data can, in turn modify more data. For
79 * example when we write new blocks, we need to allocate space for them; those
80 * allocations modify metadata (space maps)... which themselves must be
81 * written to stable storage. During the sync state, ZFS iterates, writing out
82 * data until it converges and all in-memory changes have been written out.
83 * The first such pass is the largest as it encompasses all the modified user
84 * data (as opposed to filesystem metadata). Subsequent passes typically have
85 * far less data to write as they consist exclusively of filesystem metadata.
86 *
87 * To ensure convergence, after a certain number of passes ZFS begins
88 * overwriting locations on stable storage that had been allocated earlier in
89 * the syncing state (and subsequently freed). ZFS usually allocates new
90 * blocks to optimize for large, continuous, writes. For the syncing state to
91 * converge however it must complete a pass where no new blocks are allocated
92 * since each allocation requires a modification of persistent metadata.
93 * Further, to hasten convergence, after a prescribed number of passes, ZFS
94 * also defers frees, and stops compressing.
95 *
96 * In addition to writing out user data, we must also execute synctasks during
97 * the syncing context. A synctask is the mechanism by which some
98 * administrative activities work such as creating and destroying snapshots or
99 * datasets. Note that when a synctask is initiated it enters the open txg,
100 * and ZFS then pushes that txg as quickly as possible to completion of the
101 * syncing state in order to reduce the latency of the administrative
102 * activity. To complete the syncing state, ZFS writes out a new uberblock,
103 * the root of the tree of blocks that comprise all state stored on the ZFS
104 * pool. Finally, if there is a quiesced txg waiting, we signal that it can
105 * now transition to the syncing state.
106 */
107
108 static void txg_sync_thread(dsl_pool_t *dp);
109 static void txg_quiesce_thread(dsl_pool_t *dp);
110
111 int zfs_txg_timeout = 5; /* max seconds worth of delta per txg */
112
113 /*
114 * Prepare the txg subsystem.
115 */
116 void
117 txg_init(dsl_pool_t *dp, uint64_t txg)
118 {
119 tx_state_t *tx = &dp->dp_tx;
120 int c;
121 bzero(tx, sizeof (tx_state_t));
122
123 tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
124
125 for (c = 0; c < max_ncpus; c++) {
126 int i;
127
128 mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
129 for (i = 0; i < TXG_SIZE; i++) {
130 cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
131 NULL);
132 list_create(&tx->tx_cpu[c].tc_callbacks[i],
133 sizeof (dmu_tx_callback_t),
134 offsetof(dmu_tx_callback_t, dcb_node));
135 }
136 }
137
138 mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
139
140 cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
141 cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
142 cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
143 cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
144 cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
145
146 tx->tx_open_txg = txg;
147 }
148
149 /*
150 * Close down the txg subsystem.
151 */
152 void
153 txg_fini(dsl_pool_t *dp)
154 {
155 tx_state_t *tx = &dp->dp_tx;
156 int c;
157
158 ASSERT(tx->tx_threads == 0);
159
160 mutex_destroy(&tx->tx_sync_lock);
161
162 cv_destroy(&tx->tx_sync_more_cv);
163 cv_destroy(&tx->tx_sync_done_cv);
164 cv_destroy(&tx->tx_quiesce_more_cv);
165 cv_destroy(&tx->tx_quiesce_done_cv);
166 cv_destroy(&tx->tx_exit_cv);
167
168 for (c = 0; c < max_ncpus; c++) {
169 int i;
170
171 mutex_destroy(&tx->tx_cpu[c].tc_lock);
172 for (i = 0; i < TXG_SIZE; i++) {
173 cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
174 list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
175 }
176 }
177
178 if (tx->tx_commit_cb_taskq != NULL)
179 taskq_destroy(tx->tx_commit_cb_taskq);
180
181 kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
182
183 bzero(tx, sizeof (tx_state_t));
184 }
185
186 /*
187 * Start syncing transaction groups.
188 */
189 void
190 txg_sync_start(dsl_pool_t *dp)
191 {
192 tx_state_t *tx = &dp->dp_tx;
193
194 mutex_enter(&tx->tx_sync_lock);
195
196 dprintf("pool %p\n", dp);
197
198 ASSERT(tx->tx_threads == 0);
199
200 tx->tx_threads = 2;
201
202 tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
203 dp, 0, &p0, TS_RUN, minclsyspri);
204
205 /*
206 * The sync thread can need a larger-than-default stack size on
207 * 32-bit x86. This is due in part to nested pools and
208 * scrub_visitbp() recursion.
209 */
210 tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread,
211 dp, 0, &p0, TS_RUN, minclsyspri);
212
213 mutex_exit(&tx->tx_sync_lock);
214 }
215
216 static void
217 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
218 {
219 CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
220 mutex_enter(&tx->tx_sync_lock);
221 }
222
223 static void
224 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
225 {
226 ASSERT(*tpp != NULL);
227 *tpp = NULL;
228 tx->tx_threads--;
229 cv_broadcast(&tx->tx_exit_cv);
230 CALLB_CPR_EXIT(cpr); /* drops &tx->tx_sync_lock */
231 thread_exit();
232 }
233
234 static void
235 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, clock_t time)
236 {
237 CALLB_CPR_SAFE_BEGIN(cpr);
238
239 if (time)
240 (void) cv_timedwait(cv, &tx->tx_sync_lock,
241 ddi_get_lbolt() + time);
242 else
243 cv_wait(cv, &tx->tx_sync_lock);
244
245 CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
246 }
247
248 /*
249 * Stop syncing transaction groups.
250 */
251 void
252 txg_sync_stop(dsl_pool_t *dp)
253 {
254 tx_state_t *tx = &dp->dp_tx;
255
256 dprintf("pool %p\n", dp);
257 /*
258 * Finish off any work in progress.
259 */
260 ASSERT(tx->tx_threads == 2);
261
262 /*
263 * We need to ensure that we've vacated the deferred space_maps.
264 */
265 txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
266
267 /*
268 * Wake all sync threads and wait for them to die.
269 */
270 mutex_enter(&tx->tx_sync_lock);
271
272 ASSERT(tx->tx_threads == 2);
273
274 tx->tx_exiting = 1;
275
276 cv_broadcast(&tx->tx_quiesce_more_cv);
277 cv_broadcast(&tx->tx_quiesce_done_cv);
278 cv_broadcast(&tx->tx_sync_more_cv);
279
280 while (tx->tx_threads != 0)
281 cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
282
283 tx->tx_exiting = 0;
284
285 mutex_exit(&tx->tx_sync_lock);
286 }
287
288 uint64_t
289 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
290 {
291 tx_state_t *tx = &dp->dp_tx;
292 tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
293 uint64_t txg;
294
295 mutex_enter(&tc->tc_lock);
296
297 txg = tx->tx_open_txg;
298 tc->tc_count[txg & TXG_MASK]++;
299
300 th->th_cpu = tc;
301 th->th_txg = txg;
302
303 return (txg);
304 }
305
306 void
307 txg_rele_to_quiesce(txg_handle_t *th)
308 {
309 tx_cpu_t *tc = th->th_cpu;
310
311 mutex_exit(&tc->tc_lock);
312 }
313
314 void
315 txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
316 {
317 tx_cpu_t *tc = th->th_cpu;
318 int g = th->th_txg & TXG_MASK;
319
320 mutex_enter(&tc->tc_lock);
321 list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
322 mutex_exit(&tc->tc_lock);
323 }
324
325 void
326 txg_rele_to_sync(txg_handle_t *th)
327 {
328 tx_cpu_t *tc = th->th_cpu;
329 int g = th->th_txg & TXG_MASK;
330
331 mutex_enter(&tc->tc_lock);
332 ASSERT(tc->tc_count[g] != 0);
333 if (--tc->tc_count[g] == 0)
334 cv_broadcast(&tc->tc_cv[g]);
335 mutex_exit(&tc->tc_lock);
336
337 th->th_cpu = NULL; /* defensive */
338 }
339
340 /*
341 * Blocks until all transactions in the group are committed.
342 *
343 * On return, the transaction group has reached a stable state in which it can
344 * then be passed off to the syncing context.
345 */
346 static void
347 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
348 {
349 tx_state_t *tx = &dp->dp_tx;
350 int g = txg & TXG_MASK;
351 int c;
352
353 /*
354 * Grab all tx_cpu locks so nobody else can get into this txg.
355 */
356 for (c = 0; c < max_ncpus; c++)
357 mutex_enter(&tx->tx_cpu[c].tc_lock);
358
359 ASSERT(txg == tx->tx_open_txg);
360 tx->tx_open_txg++;
361
362 DTRACE_PROBE2(txg__quiescing, dsl_pool_t *, dp, uint64_t, txg);
363 DTRACE_PROBE2(txg__opened, dsl_pool_t *, dp, uint64_t, tx->tx_open_txg);
364
365 /*
366 * Now that we've incremented tx_open_txg, we can let threads
367 * enter the next transaction group.
368 */
369 for (c = 0; c < max_ncpus; c++)
370 mutex_exit(&tx->tx_cpu[c].tc_lock);
371
372 /*
373 * Quiesce the transaction group by waiting for everyone to txg_exit().
374 */
375 for (c = 0; c < max_ncpus; c++) {
376 tx_cpu_t *tc = &tx->tx_cpu[c];
377 mutex_enter(&tc->tc_lock);
378 while (tc->tc_count[g] != 0)
379 cv_wait(&tc->tc_cv[g], &tc->tc_lock);
380 mutex_exit(&tc->tc_lock);
381 }
382 }
383
384 static void
385 txg_do_callbacks(list_t *cb_list)
386 {
387 dmu_tx_do_callbacks(cb_list, 0);
388
389 list_destroy(cb_list);
390
391 kmem_free(cb_list, sizeof (list_t));
392 }
393
394 /*
395 * Dispatch the commit callbacks registered on this txg to worker threads.
396 *
397 * If no callbacks are registered for a given TXG, nothing happens.
398 * This function creates a taskq for the associated pool, if needed.
399 */
400 static void
401 txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg)
402 {
403 int c;
404 tx_state_t *tx = &dp->dp_tx;
405 list_t *cb_list;
406
407 for (c = 0; c < max_ncpus; c++) {
408 tx_cpu_t *tc = &tx->tx_cpu[c];
409 /*
410 * No need to lock tx_cpu_t at this point, since this can
411 * only be called once a txg has been synced.
412 */
413
414 int g = txg & TXG_MASK;
415
416 if (list_is_empty(&tc->tc_callbacks[g]))
417 continue;
418
419 if (tx->tx_commit_cb_taskq == NULL) {
420 /*
421 * Commit callback taskq hasn't been created yet.
422 */
423 tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
424 max_ncpus, minclsyspri, max_ncpus, max_ncpus * 2,
425 TASKQ_PREPOPULATE);
426 }
427
428 cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
429 list_create(cb_list, sizeof (dmu_tx_callback_t),
430 offsetof(dmu_tx_callback_t, dcb_node));
431
432 list_move_tail(&tc->tc_callbacks[g], cb_list);
433
434 (void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
435 txg_do_callbacks, cb_list, TQ_SLEEP);
436 }
437 }
438
439 static void
440 txg_sync_thread(dsl_pool_t *dp)
441 {
442 spa_t *spa = dp->dp_spa;
443 tx_state_t *tx = &dp->dp_tx;
444 callb_cpr_t cpr;
445 uint64_t start, delta;
446
447 txg_thread_enter(tx, &cpr);
448
449 start = delta = 0;
450 for (;;) {
451 uint64_t timer, timeout = zfs_txg_timeout * hz;
452 uint64_t txg;
453
454 /*
455 * We sync when we're scanning, there's someone waiting
456 * on us, or the quiesce thread has handed off a txg to
457 * us, or we have reached our timeout.
458 */
459 timer = (delta >= timeout ? 0 : timeout - delta);
460 while (!dsl_scan_active(dp->dp_scan) &&
461 !tx->tx_exiting && timer > 0 &&
462 tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
463 tx->tx_quiesced_txg == 0) {
464 dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
465 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
466 txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
467 delta = ddi_get_lbolt() - start;
468 timer = (delta > timeout ? 0 : timeout - delta);
469 }
470
471 /*
472 * Wait until the quiesce thread hands off a txg to us,
473 * prompting it to do so if necessary.
474 */
475 while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
476 if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
477 tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
478 cv_broadcast(&tx->tx_quiesce_more_cv);
479 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
480 }
481
482 if (tx->tx_exiting)
483 txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
484
485 /*
486 * Consume the quiesced txg which has been handed off to
487 * us. This may cause the quiescing thread to now be
488 * able to quiesce another txg, so we must signal it.
489 */
490 txg = tx->tx_quiesced_txg;
491 tx->tx_quiesced_txg = 0;
492 tx->tx_syncing_txg = txg;
493 DTRACE_PROBE2(txg__syncing, dsl_pool_t *, dp, uint64_t, txg);
494 cv_broadcast(&tx->tx_quiesce_more_cv);
495
496 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
497 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
498 mutex_exit(&tx->tx_sync_lock);
499
500 start = ddi_get_lbolt();
501 spa_sync(spa, txg);
502 delta = ddi_get_lbolt() - start;
503
504 mutex_enter(&tx->tx_sync_lock);
505 tx->tx_synced_txg = txg;
506 tx->tx_syncing_txg = 0;
507 DTRACE_PROBE2(txg__synced, dsl_pool_t *, dp, uint64_t, txg);
508 cv_broadcast(&tx->tx_sync_done_cv);
509
510 /*
511 * Dispatch commit callbacks to worker threads.
512 */
513 txg_dispatch_callbacks(dp, txg);
514 }
515 }
516
517 static void
518 txg_quiesce_thread(dsl_pool_t *dp)
519 {
520 tx_state_t *tx = &dp->dp_tx;
521 callb_cpr_t cpr;
522
523 txg_thread_enter(tx, &cpr);
524
525 for (;;) {
526 uint64_t txg;
527
528 /*
529 * We quiesce when there's someone waiting on us.
530 * However, we can only have one txg in "quiescing" or
531 * "quiesced, waiting to sync" state. So we wait until
532 * the "quiesced, waiting to sync" txg has been consumed
533 * by the sync thread.
534 */
535 while (!tx->tx_exiting &&
536 (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
537 tx->tx_quiesced_txg != 0))
538 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
539
540 if (tx->tx_exiting)
541 txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
542
543 txg = tx->tx_open_txg;
544 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
545 txg, tx->tx_quiesce_txg_waiting,
546 tx->tx_sync_txg_waiting);
547 mutex_exit(&tx->tx_sync_lock);
548 txg_quiesce(dp, txg);
549 mutex_enter(&tx->tx_sync_lock);
550
551 /*
552 * Hand this txg off to the sync thread.
553 */
554 dprintf("quiesce done, handing off txg %llu\n", txg);
555 tx->tx_quiesced_txg = txg;
556 DTRACE_PROBE2(txg__quiesced, dsl_pool_t *, dp, uint64_t, txg);
557 cv_broadcast(&tx->tx_sync_more_cv);
558 cv_broadcast(&tx->tx_quiesce_done_cv);
559 }
560 }
561
562 /*
563 * Delay this thread by delay nanoseconds if we are still in the open
564 * transaction group and there is already a waiting txg quiescing or quiesced.
565 * Abort the delay if this txg stalls or enters the quiescing state.
566 */
567 void
568 txg_delay(dsl_pool_t *dp, uint64_t txg, hrtime_t delay, hrtime_t resolution)
569 {
570 tx_state_t *tx = &dp->dp_tx;
571 hrtime_t start = gethrtime();
572
573 /* don't delay if this txg could transition to quiescing immediately */
574 if (tx->tx_open_txg > txg ||
575 tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
576 return;
577
578 mutex_enter(&tx->tx_sync_lock);
579 if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
580 mutex_exit(&tx->tx_sync_lock);
581 return;
582 }
583
584 while (gethrtime() - start < delay &&
585 tx->tx_syncing_txg < txg-1 && !txg_stalled(dp)) {
586 (void) cv_timedwait_hires(&tx->tx_quiesce_more_cv,
587 &tx->tx_sync_lock, delay, resolution, 0);
588 }
589
590 mutex_exit(&tx->tx_sync_lock);
591 }
592
593 void
594 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
595 {
596 tx_state_t *tx = &dp->dp_tx;
597
598 ASSERT(!dsl_pool_config_held(dp));
599
600 mutex_enter(&tx->tx_sync_lock);
601 ASSERT(tx->tx_threads == 2);
602 if (txg == 0)
603 txg = tx->tx_open_txg + TXG_DEFER_SIZE;
604 if (tx->tx_sync_txg_waiting < txg)
605 tx->tx_sync_txg_waiting = txg;
606 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
607 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
608 while (tx->tx_synced_txg < txg) {
609 dprintf("broadcasting sync more "
610 "tx_synced=%llu waiting=%llu dp=%p\n",
611 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
612 cv_broadcast(&tx->tx_sync_more_cv);
613 cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
614 }
615 mutex_exit(&tx->tx_sync_lock);
616 }
617
618 void
619 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
620 {
621 tx_state_t *tx = &dp->dp_tx;
622
623 ASSERT(!dsl_pool_config_held(dp));
624
625 mutex_enter(&tx->tx_sync_lock);
626 ASSERT(tx->tx_threads == 2);
627 if (txg == 0)
628 txg = tx->tx_open_txg + 1;
629 if (tx->tx_quiesce_txg_waiting < txg)
630 tx->tx_quiesce_txg_waiting = txg;
631 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
632 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
633 while (tx->tx_open_txg < txg) {
634 cv_broadcast(&tx->tx_quiesce_more_cv);
635 cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
636 }
637 mutex_exit(&tx->tx_sync_lock);
638 }
639
640 boolean_t
641 txg_stalled(dsl_pool_t *dp)
642 {
643 tx_state_t *tx = &dp->dp_tx;
644 return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
645 }
646
647 boolean_t
648 txg_sync_waiting(dsl_pool_t *dp)
649 {
650 tx_state_t *tx = &dp->dp_tx;
651
652 return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
653 tx->tx_quiesced_txg != 0);
654 }
655
656 /*
657 * Per-txg object lists.
658 */
659 void
660 txg_list_create(txg_list_t *tl, size_t offset)
661 {
662 int t;
663
664 mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
665
666 tl->tl_offset = offset;
667
668 for (t = 0; t < TXG_SIZE; t++)
669 tl->tl_head[t] = NULL;
670 }
671
672 void
673 txg_list_destroy(txg_list_t *tl)
674 {
675 int t;
676
677 for (t = 0; t < TXG_SIZE; t++)
678 ASSERT(txg_list_empty(tl, t));
679
680 mutex_destroy(&tl->tl_lock);
681 }
682
683 boolean_t
684 txg_list_empty(txg_list_t *tl, uint64_t txg)
685 {
686 return (tl->tl_head[txg & TXG_MASK] == NULL);
687 }
688
689 /*
690 * Add an entry to the list (unless it's already on the list).
691 * Returns B_TRUE if it was actually added.
692 */
693 boolean_t
694 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
695 {
696 int t = txg & TXG_MASK;
697 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
698 boolean_t add;
699
700 mutex_enter(&tl->tl_lock);
701 add = (tn->tn_member[t] == 0);
702 if (add) {
703 tn->tn_member[t] = 1;
704 tn->tn_next[t] = tl->tl_head[t];
705 tl->tl_head[t] = tn;
706 }
707 mutex_exit(&tl->tl_lock);
708
709 return (add);
710 }
711
712 /*
713 * Add an entry to the end of the list, unless it's already on the list.
714 * (walks list to find end)
715 * Returns B_TRUE if it was actually added.
716 */
717 boolean_t
718 txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
719 {
720 int t = txg & TXG_MASK;
721 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
722 boolean_t add;
723
724 mutex_enter(&tl->tl_lock);
725 add = (tn->tn_member[t] == 0);
726 if (add) {
727 txg_node_t **tp;
728
729 for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
730 continue;
731
732 tn->tn_member[t] = 1;
733 tn->tn_next[t] = NULL;
734 *tp = tn;
735 }
736 mutex_exit(&tl->tl_lock);
737
738 return (add);
739 }
740
741 /*
742 * Remove the head of the list and return it.
743 */
744 void *
745 txg_list_remove(txg_list_t *tl, uint64_t txg)
746 {
747 int t = txg & TXG_MASK;
748 txg_node_t *tn;
749 void *p = NULL;
750
751 mutex_enter(&tl->tl_lock);
752 if ((tn = tl->tl_head[t]) != NULL) {
753 p = (char *)tn - tl->tl_offset;
754 tl->tl_head[t] = tn->tn_next[t];
755 tn->tn_next[t] = NULL;
756 tn->tn_member[t] = 0;
757 }
758 mutex_exit(&tl->tl_lock);
759
760 return (p);
761 }
762
763 /*
764 * Remove a specific item from the list and return it.
765 */
766 void *
767 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
768 {
769 int t = txg & TXG_MASK;
770 txg_node_t *tn, **tp;
771
772 mutex_enter(&tl->tl_lock);
773
774 for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
775 if ((char *)tn - tl->tl_offset == p) {
776 *tp = tn->tn_next[t];
777 tn->tn_next[t] = NULL;
778 tn->tn_member[t] = 0;
779 mutex_exit(&tl->tl_lock);
780 return (p);
781 }
782 }
783
784 mutex_exit(&tl->tl_lock);
785
786 return (NULL);
787 }
788
789 boolean_t
790 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
791 {
792 int t = txg & TXG_MASK;
793 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
794
795 return (tn->tn_member[t] != 0);
796 }
797
798 /*
799 * Walk a txg list -- only safe if you know it's not changing.
800 */
801 void *
802 txg_list_head(txg_list_t *tl, uint64_t txg)
803 {
804 int t = txg & TXG_MASK;
805 txg_node_t *tn = tl->tl_head[t];
806
807 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
808 }
809
810 void *
811 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
812 {
813 int t = txg & TXG_MASK;
814 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
815
816 tn = tn->tn_next[t];
817
818 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
819 }