xref: /illumos-gate/usr/src/uts/common/fs/zfs/txg.c (revision ce636f8b38e8c9ff484e880d9abb27251a882860)
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) 2012 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  * Pool-wide transaction groups.
37  */
38 
39 static void txg_sync_thread(dsl_pool_t *dp);
40 static void txg_quiesce_thread(dsl_pool_t *dp);
41 
42 int zfs_txg_timeout = 5;	/* max seconds worth of delta per txg */
43 
44 /*
45  * Prepare the txg subsystem.
46  */
47 void
48 txg_init(dsl_pool_t *dp, uint64_t txg)
49 {
50 	tx_state_t *tx = &dp->dp_tx;
51 	int c;
52 	bzero(tx, sizeof (tx_state_t));
53 
54 	tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
55 
56 	for (c = 0; c < max_ncpus; c++) {
57 		int i;
58 
59 		mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
60 		for (i = 0; i < TXG_SIZE; i++) {
61 			cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
62 			    NULL);
63 			list_create(&tx->tx_cpu[c].tc_callbacks[i],
64 			    sizeof (dmu_tx_callback_t),
65 			    offsetof(dmu_tx_callback_t, dcb_node));
66 		}
67 	}
68 
69 	mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
70 
71 	cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
72 	cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
73 	cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
74 	cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
75 	cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
76 
77 	tx->tx_open_txg = txg;
78 }
79 
80 /*
81  * Close down the txg subsystem.
82  */
83 void
84 txg_fini(dsl_pool_t *dp)
85 {
86 	tx_state_t *tx = &dp->dp_tx;
87 	int c;
88 
89 	ASSERT(tx->tx_threads == 0);
90 
91 	mutex_destroy(&tx->tx_sync_lock);
92 
93 	cv_destroy(&tx->tx_sync_more_cv);
94 	cv_destroy(&tx->tx_sync_done_cv);
95 	cv_destroy(&tx->tx_quiesce_more_cv);
96 	cv_destroy(&tx->tx_quiesce_done_cv);
97 	cv_destroy(&tx->tx_exit_cv);
98 
99 	for (c = 0; c < max_ncpus; c++) {
100 		int i;
101 
102 		mutex_destroy(&tx->tx_cpu[c].tc_lock);
103 		for (i = 0; i < TXG_SIZE; i++) {
104 			cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
105 			list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
106 		}
107 	}
108 
109 	if (tx->tx_commit_cb_taskq != NULL)
110 		taskq_destroy(tx->tx_commit_cb_taskq);
111 
112 	kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
113 
114 	bzero(tx, sizeof (tx_state_t));
115 }
116 
117 /*
118  * Start syncing transaction groups.
119  */
120 void
121 txg_sync_start(dsl_pool_t *dp)
122 {
123 	tx_state_t *tx = &dp->dp_tx;
124 
125 	mutex_enter(&tx->tx_sync_lock);
126 
127 	dprintf("pool %p\n", dp);
128 
129 	ASSERT(tx->tx_threads == 0);
130 
131 	tx->tx_threads = 2;
132 
133 	tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
134 	    dp, 0, &p0, TS_RUN, minclsyspri);
135 
136 	/*
137 	 * The sync thread can need a larger-than-default stack size on
138 	 * 32-bit x86.  This is due in part to nested pools and
139 	 * scrub_visitbp() recursion.
140 	 */
141 	tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread,
142 	    dp, 0, &p0, TS_RUN, minclsyspri);
143 
144 	mutex_exit(&tx->tx_sync_lock);
145 }
146 
147 static void
148 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
149 {
150 	CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
151 	mutex_enter(&tx->tx_sync_lock);
152 }
153 
154 static void
155 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
156 {
157 	ASSERT(*tpp != NULL);
158 	*tpp = NULL;
159 	tx->tx_threads--;
160 	cv_broadcast(&tx->tx_exit_cv);
161 	CALLB_CPR_EXIT(cpr);		/* drops &tx->tx_sync_lock */
162 	thread_exit();
163 }
164 
165 static void
166 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, uint64_t time)
167 {
168 	CALLB_CPR_SAFE_BEGIN(cpr);
169 
170 	if (time)
171 		(void) cv_timedwait(cv, &tx->tx_sync_lock,
172 		    ddi_get_lbolt() + time);
173 	else
174 		cv_wait(cv, &tx->tx_sync_lock);
175 
176 	CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
177 }
178 
179 /*
180  * Stop syncing transaction groups.
181  */
182 void
183 txg_sync_stop(dsl_pool_t *dp)
184 {
185 	tx_state_t *tx = &dp->dp_tx;
186 
187 	dprintf("pool %p\n", dp);
188 	/*
189 	 * Finish off any work in progress.
190 	 */
191 	ASSERT(tx->tx_threads == 2);
192 
193 	/*
194 	 * We need to ensure that we've vacated the deferred space_maps.
195 	 */
196 	txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
197 
198 	/*
199 	 * Wake all sync threads and wait for them to die.
200 	 */
201 	mutex_enter(&tx->tx_sync_lock);
202 
203 	ASSERT(tx->tx_threads == 2);
204 
205 	tx->tx_exiting = 1;
206 
207 	cv_broadcast(&tx->tx_quiesce_more_cv);
208 	cv_broadcast(&tx->tx_quiesce_done_cv);
209 	cv_broadcast(&tx->tx_sync_more_cv);
210 
211 	while (tx->tx_threads != 0)
212 		cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
213 
214 	tx->tx_exiting = 0;
215 
216 	mutex_exit(&tx->tx_sync_lock);
217 }
218 
219 uint64_t
220 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
221 {
222 	tx_state_t *tx = &dp->dp_tx;
223 	tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
224 	uint64_t txg;
225 
226 	mutex_enter(&tc->tc_lock);
227 
228 	txg = tx->tx_open_txg;
229 	tc->tc_count[txg & TXG_MASK]++;
230 
231 	th->th_cpu = tc;
232 	th->th_txg = txg;
233 
234 	return (txg);
235 }
236 
237 void
238 txg_rele_to_quiesce(txg_handle_t *th)
239 {
240 	tx_cpu_t *tc = th->th_cpu;
241 
242 	mutex_exit(&tc->tc_lock);
243 }
244 
245 void
246 txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
247 {
248 	tx_cpu_t *tc = th->th_cpu;
249 	int g = th->th_txg & TXG_MASK;
250 
251 	mutex_enter(&tc->tc_lock);
252 	list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
253 	mutex_exit(&tc->tc_lock);
254 }
255 
256 void
257 txg_rele_to_sync(txg_handle_t *th)
258 {
259 	tx_cpu_t *tc = th->th_cpu;
260 	int g = th->th_txg & TXG_MASK;
261 
262 	mutex_enter(&tc->tc_lock);
263 	ASSERT(tc->tc_count[g] != 0);
264 	if (--tc->tc_count[g] == 0)
265 		cv_broadcast(&tc->tc_cv[g]);
266 	mutex_exit(&tc->tc_lock);
267 
268 	th->th_cpu = NULL;	/* defensive */
269 }
270 
271 static void
272 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
273 {
274 	tx_state_t *tx = &dp->dp_tx;
275 	int g = txg & TXG_MASK;
276 	int c;
277 
278 	/*
279 	 * Grab all tx_cpu locks so nobody else can get into this txg.
280 	 */
281 	for (c = 0; c < max_ncpus; c++)
282 		mutex_enter(&tx->tx_cpu[c].tc_lock);
283 
284 	ASSERT(txg == tx->tx_open_txg);
285 	tx->tx_open_txg++;
286 
287 	/*
288 	 * Now that we've incremented tx_open_txg, we can let threads
289 	 * enter the next transaction group.
290 	 */
291 	for (c = 0; c < max_ncpus; c++)
292 		mutex_exit(&tx->tx_cpu[c].tc_lock);
293 
294 	/*
295 	 * Quiesce the transaction group by waiting for everyone to txg_exit().
296 	 */
297 	for (c = 0; c < max_ncpus; c++) {
298 		tx_cpu_t *tc = &tx->tx_cpu[c];
299 		mutex_enter(&tc->tc_lock);
300 		while (tc->tc_count[g] != 0)
301 			cv_wait(&tc->tc_cv[g], &tc->tc_lock);
302 		mutex_exit(&tc->tc_lock);
303 	}
304 }
305 
306 static void
307 txg_do_callbacks(list_t *cb_list)
308 {
309 	dmu_tx_do_callbacks(cb_list, 0);
310 
311 	list_destroy(cb_list);
312 
313 	kmem_free(cb_list, sizeof (list_t));
314 }
315 
316 /*
317  * Dispatch the commit callbacks registered on this txg to worker threads.
318  */
319 static void
320 txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg)
321 {
322 	int c;
323 	tx_state_t *tx = &dp->dp_tx;
324 	list_t *cb_list;
325 
326 	for (c = 0; c < max_ncpus; c++) {
327 		tx_cpu_t *tc = &tx->tx_cpu[c];
328 		/* No need to lock tx_cpu_t at this point */
329 
330 		int g = txg & TXG_MASK;
331 
332 		if (list_is_empty(&tc->tc_callbacks[g]))
333 			continue;
334 
335 		if (tx->tx_commit_cb_taskq == NULL) {
336 			/*
337 			 * Commit callback taskq hasn't been created yet.
338 			 */
339 			tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
340 			    max_ncpus, minclsyspri, max_ncpus, max_ncpus * 2,
341 			    TASKQ_PREPOPULATE);
342 		}
343 
344 		cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
345 		list_create(cb_list, sizeof (dmu_tx_callback_t),
346 		    offsetof(dmu_tx_callback_t, dcb_node));
347 
348 		list_move_tail(&tc->tc_callbacks[g], cb_list);
349 
350 		(void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
351 		    txg_do_callbacks, cb_list, TQ_SLEEP);
352 	}
353 }
354 
355 static void
356 txg_sync_thread(dsl_pool_t *dp)
357 {
358 	spa_t *spa = dp->dp_spa;
359 	tx_state_t *tx = &dp->dp_tx;
360 	callb_cpr_t cpr;
361 	uint64_t start, delta;
362 
363 	txg_thread_enter(tx, &cpr);
364 
365 	start = delta = 0;
366 	for (;;) {
367 		uint64_t timer, timeout = zfs_txg_timeout * hz;
368 		uint64_t txg;
369 
370 		/*
371 		 * We sync when we're scanning, there's someone waiting
372 		 * on us, or the quiesce thread has handed off a txg to
373 		 * us, or we have reached our timeout.
374 		 */
375 		timer = (delta >= timeout ? 0 : timeout - delta);
376 		while (!dsl_scan_active(dp->dp_scan) &&
377 		    !tx->tx_exiting && timer > 0 &&
378 		    tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
379 		    tx->tx_quiesced_txg == 0) {
380 			dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
381 			    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
382 			txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
383 			delta = ddi_get_lbolt() - start;
384 			timer = (delta > timeout ? 0 : timeout - delta);
385 		}
386 
387 		/*
388 		 * Wait until the quiesce thread hands off a txg to us,
389 		 * prompting it to do so if necessary.
390 		 */
391 		while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
392 			if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
393 				tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
394 			cv_broadcast(&tx->tx_quiesce_more_cv);
395 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
396 		}
397 
398 		if (tx->tx_exiting)
399 			txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
400 
401 		/*
402 		 * Consume the quiesced txg which has been handed off to
403 		 * us.  This may cause the quiescing thread to now be
404 		 * able to quiesce another txg, so we must signal it.
405 		 */
406 		txg = tx->tx_quiesced_txg;
407 		tx->tx_quiesced_txg = 0;
408 		tx->tx_syncing_txg = txg;
409 		cv_broadcast(&tx->tx_quiesce_more_cv);
410 
411 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
412 		    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
413 		mutex_exit(&tx->tx_sync_lock);
414 
415 		start = ddi_get_lbolt();
416 		spa_sync(spa, txg);
417 		delta = ddi_get_lbolt() - start;
418 
419 		mutex_enter(&tx->tx_sync_lock);
420 		tx->tx_synced_txg = txg;
421 		tx->tx_syncing_txg = 0;
422 		cv_broadcast(&tx->tx_sync_done_cv);
423 
424 		/*
425 		 * Dispatch commit callbacks to worker threads.
426 		 */
427 		txg_dispatch_callbacks(dp, txg);
428 	}
429 }
430 
431 static void
432 txg_quiesce_thread(dsl_pool_t *dp)
433 {
434 	tx_state_t *tx = &dp->dp_tx;
435 	callb_cpr_t cpr;
436 
437 	txg_thread_enter(tx, &cpr);
438 
439 	for (;;) {
440 		uint64_t txg;
441 
442 		/*
443 		 * We quiesce when there's someone waiting on us.
444 		 * However, we can only have one txg in "quiescing" or
445 		 * "quiesced, waiting to sync" state.  So we wait until
446 		 * the "quiesced, waiting to sync" txg has been consumed
447 		 * by the sync thread.
448 		 */
449 		while (!tx->tx_exiting &&
450 		    (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
451 		    tx->tx_quiesced_txg != 0))
452 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
453 
454 		if (tx->tx_exiting)
455 			txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
456 
457 		txg = tx->tx_open_txg;
458 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
459 		    txg, tx->tx_quiesce_txg_waiting,
460 		    tx->tx_sync_txg_waiting);
461 		mutex_exit(&tx->tx_sync_lock);
462 		txg_quiesce(dp, txg);
463 		mutex_enter(&tx->tx_sync_lock);
464 
465 		/*
466 		 * Hand this txg off to the sync thread.
467 		 */
468 		dprintf("quiesce done, handing off txg %llu\n", txg);
469 		tx->tx_quiesced_txg = txg;
470 		cv_broadcast(&tx->tx_sync_more_cv);
471 		cv_broadcast(&tx->tx_quiesce_done_cv);
472 	}
473 }
474 
475 /*
476  * Delay this thread by 'ticks' if we are still in the open transaction
477  * group and there is already a waiting txg quiesing or quiesced.  Abort
478  * the delay if this txg stalls or enters the quiesing state.
479  */
480 void
481 txg_delay(dsl_pool_t *dp, uint64_t txg, int ticks)
482 {
483 	tx_state_t *tx = &dp->dp_tx;
484 	clock_t timeout = ddi_get_lbolt() + ticks;
485 
486 	/* don't delay if this txg could transition to quiesing immediately */
487 	if (tx->tx_open_txg > txg ||
488 	    tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
489 		return;
490 
491 	mutex_enter(&tx->tx_sync_lock);
492 	if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
493 		mutex_exit(&tx->tx_sync_lock);
494 		return;
495 	}
496 
497 	while (ddi_get_lbolt() < timeout &&
498 	    tx->tx_syncing_txg < txg-1 && !txg_stalled(dp))
499 		(void) cv_timedwait(&tx->tx_quiesce_more_cv, &tx->tx_sync_lock,
500 		    timeout);
501 
502 	mutex_exit(&tx->tx_sync_lock);
503 }
504 
505 void
506 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
507 {
508 	tx_state_t *tx = &dp->dp_tx;
509 
510 	mutex_enter(&tx->tx_sync_lock);
511 	ASSERT(tx->tx_threads == 2);
512 	if (txg == 0)
513 		txg = tx->tx_open_txg + TXG_DEFER_SIZE;
514 	if (tx->tx_sync_txg_waiting < txg)
515 		tx->tx_sync_txg_waiting = txg;
516 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
517 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
518 	while (tx->tx_synced_txg < txg) {
519 		dprintf("broadcasting sync more "
520 		    "tx_synced=%llu waiting=%llu dp=%p\n",
521 		    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
522 		cv_broadcast(&tx->tx_sync_more_cv);
523 		cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
524 	}
525 	mutex_exit(&tx->tx_sync_lock);
526 }
527 
528 void
529 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
530 {
531 	tx_state_t *tx = &dp->dp_tx;
532 
533 	mutex_enter(&tx->tx_sync_lock);
534 	ASSERT(tx->tx_threads == 2);
535 	if (txg == 0)
536 		txg = tx->tx_open_txg + 1;
537 	if (tx->tx_quiesce_txg_waiting < txg)
538 		tx->tx_quiesce_txg_waiting = txg;
539 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
540 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
541 	while (tx->tx_open_txg < txg) {
542 		cv_broadcast(&tx->tx_quiesce_more_cv);
543 		cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
544 	}
545 	mutex_exit(&tx->tx_sync_lock);
546 }
547 
548 boolean_t
549 txg_stalled(dsl_pool_t *dp)
550 {
551 	tx_state_t *tx = &dp->dp_tx;
552 	return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
553 }
554 
555 boolean_t
556 txg_sync_waiting(dsl_pool_t *dp)
557 {
558 	tx_state_t *tx = &dp->dp_tx;
559 
560 	return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
561 	    tx->tx_quiesced_txg != 0);
562 }
563 
564 /*
565  * Per-txg object lists.
566  */
567 void
568 txg_list_create(txg_list_t *tl, size_t offset)
569 {
570 	int t;
571 
572 	mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
573 
574 	tl->tl_offset = offset;
575 
576 	for (t = 0; t < TXG_SIZE; t++)
577 		tl->tl_head[t] = NULL;
578 }
579 
580 void
581 txg_list_destroy(txg_list_t *tl)
582 {
583 	int t;
584 
585 	for (t = 0; t < TXG_SIZE; t++)
586 		ASSERT(txg_list_empty(tl, t));
587 
588 	mutex_destroy(&tl->tl_lock);
589 }
590 
591 boolean_t
592 txg_list_empty(txg_list_t *tl, uint64_t txg)
593 {
594 	return (tl->tl_head[txg & TXG_MASK] == NULL);
595 }
596 
597 /*
598  * Add an entry to the list.
599  * Returns 0 if it's a new entry, 1 if it's already there.
600  */
601 int
602 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
603 {
604 	int t = txg & TXG_MASK;
605 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
606 	int already_on_list;
607 
608 	mutex_enter(&tl->tl_lock);
609 	already_on_list = tn->tn_member[t];
610 	if (!already_on_list) {
611 		tn->tn_member[t] = 1;
612 		tn->tn_next[t] = tl->tl_head[t];
613 		tl->tl_head[t] = tn;
614 	}
615 	mutex_exit(&tl->tl_lock);
616 
617 	return (already_on_list);
618 }
619 
620 /*
621  * Add an entry to the end of the list (walks list to find end).
622  * Returns 0 if it's a new entry, 1 if it's already there.
623  */
624 int
625 txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
626 {
627 	int t = txg & TXG_MASK;
628 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
629 	int already_on_list;
630 
631 	mutex_enter(&tl->tl_lock);
632 	already_on_list = tn->tn_member[t];
633 	if (!already_on_list) {
634 		txg_node_t **tp;
635 
636 		for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
637 			continue;
638 
639 		tn->tn_member[t] = 1;
640 		tn->tn_next[t] = NULL;
641 		*tp = tn;
642 	}
643 	mutex_exit(&tl->tl_lock);
644 
645 	return (already_on_list);
646 }
647 
648 /*
649  * Remove the head of the list and return it.
650  */
651 void *
652 txg_list_remove(txg_list_t *tl, uint64_t txg)
653 {
654 	int t = txg & TXG_MASK;
655 	txg_node_t *tn;
656 	void *p = NULL;
657 
658 	mutex_enter(&tl->tl_lock);
659 	if ((tn = tl->tl_head[t]) != NULL) {
660 		p = (char *)tn - tl->tl_offset;
661 		tl->tl_head[t] = tn->tn_next[t];
662 		tn->tn_next[t] = NULL;
663 		tn->tn_member[t] = 0;
664 	}
665 	mutex_exit(&tl->tl_lock);
666 
667 	return (p);
668 }
669 
670 /*
671  * Remove a specific item from the list and return it.
672  */
673 void *
674 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
675 {
676 	int t = txg & TXG_MASK;
677 	txg_node_t *tn, **tp;
678 
679 	mutex_enter(&tl->tl_lock);
680 
681 	for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
682 		if ((char *)tn - tl->tl_offset == p) {
683 			*tp = tn->tn_next[t];
684 			tn->tn_next[t] = NULL;
685 			tn->tn_member[t] = 0;
686 			mutex_exit(&tl->tl_lock);
687 			return (p);
688 		}
689 	}
690 
691 	mutex_exit(&tl->tl_lock);
692 
693 	return (NULL);
694 }
695 
696 int
697 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
698 {
699 	int t = txg & TXG_MASK;
700 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
701 
702 	return (tn->tn_member[t]);
703 }
704 
705 /*
706  * Walk a txg list -- only safe if you know it's not changing.
707  */
708 void *
709 txg_list_head(txg_list_t *tl, uint64_t txg)
710 {
711 	int t = txg & TXG_MASK;
712 	txg_node_t *tn = tl->tl_head[t];
713 
714 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
715 }
716 
717 void *
718 txg_list_next(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 
723 	tn = tn->tn_next[t];
724 
725 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
726 }
727