xref: /illumos-gate/usr/src/uts/common/fs/zfs/txg.c (revision 89b42a211fa7d3527b9615260f495d22e430c5c5)
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  */
24 
25 #include <sys/zfs_context.h>
26 #include <sys/txg_impl.h>
27 #include <sys/dmu_impl.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/dsl_pool.h>
30 #include <sys/dsl_scan.h>
31 #include <sys/callb.h>
32 
33 /*
34  * Pool-wide transaction groups.
35  */
36 
37 static void txg_sync_thread(dsl_pool_t *dp);
38 static void txg_quiesce_thread(dsl_pool_t *dp);
39 
40 int zfs_txg_timeout = 5;	/* max seconds worth of delta per txg */
41 
42 /*
43  * Prepare the txg subsystem.
44  */
45 void
46 txg_init(dsl_pool_t *dp, uint64_t txg)
47 {
48 	tx_state_t *tx = &dp->dp_tx;
49 	int c;
50 	bzero(tx, sizeof (tx_state_t));
51 
52 	tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
53 
54 	for (c = 0; c < max_ncpus; c++) {
55 		int i;
56 
57 		mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
58 		for (i = 0; i < TXG_SIZE; i++) {
59 			cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
60 			    NULL);
61 			list_create(&tx->tx_cpu[c].tc_callbacks[i],
62 			    sizeof (dmu_tx_callback_t),
63 			    offsetof(dmu_tx_callback_t, dcb_node));
64 		}
65 	}
66 
67 	mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
68 
69 	cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
70 	cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
71 	cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
72 	cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
73 	cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
74 
75 	tx->tx_open_txg = txg;
76 }
77 
78 /*
79  * Close down the txg subsystem.
80  */
81 void
82 txg_fini(dsl_pool_t *dp)
83 {
84 	tx_state_t *tx = &dp->dp_tx;
85 	int c;
86 
87 	ASSERT(tx->tx_threads == 0);
88 
89 	mutex_destroy(&tx->tx_sync_lock);
90 
91 	cv_destroy(&tx->tx_sync_more_cv);
92 	cv_destroy(&tx->tx_sync_done_cv);
93 	cv_destroy(&tx->tx_quiesce_more_cv);
94 	cv_destroy(&tx->tx_quiesce_done_cv);
95 	cv_destroy(&tx->tx_exit_cv);
96 
97 	for (c = 0; c < max_ncpus; c++) {
98 		int i;
99 
100 		mutex_destroy(&tx->tx_cpu[c].tc_lock);
101 		for (i = 0; i < TXG_SIZE; i++) {
102 			cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
103 			list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
104 		}
105 	}
106 
107 	if (tx->tx_commit_cb_taskq != NULL)
108 		taskq_destroy(tx->tx_commit_cb_taskq);
109 
110 	kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
111 
112 	bzero(tx, sizeof (tx_state_t));
113 }
114 
115 /*
116  * Start syncing transaction groups.
117  */
118 void
119 txg_sync_start(dsl_pool_t *dp)
120 {
121 	tx_state_t *tx = &dp->dp_tx;
122 
123 	mutex_enter(&tx->tx_sync_lock);
124 
125 	dprintf("pool %p\n", dp);
126 
127 	ASSERT(tx->tx_threads == 0);
128 
129 	tx->tx_threads = 2;
130 
131 	tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
132 	    dp, 0, &p0, TS_RUN, minclsyspri);
133 
134 	/*
135 	 * The sync thread can need a larger-than-default stack size on
136 	 * 32-bit x86.  This is due in part to nested pools and
137 	 * scrub_visitbp() recursion.
138 	 */
139 	tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread,
140 	    dp, 0, &p0, TS_RUN, minclsyspri);
141 
142 	mutex_exit(&tx->tx_sync_lock);
143 }
144 
145 static void
146 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
147 {
148 	CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
149 	mutex_enter(&tx->tx_sync_lock);
150 }
151 
152 static void
153 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
154 {
155 	ASSERT(*tpp != NULL);
156 	*tpp = NULL;
157 	tx->tx_threads--;
158 	cv_broadcast(&tx->tx_exit_cv);
159 	CALLB_CPR_EXIT(cpr);		/* drops &tx->tx_sync_lock */
160 	thread_exit();
161 }
162 
163 static void
164 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, uint64_t time)
165 {
166 	CALLB_CPR_SAFE_BEGIN(cpr);
167 
168 	if (time)
169 		(void) cv_timedwait(cv, &tx->tx_sync_lock,
170 		    ddi_get_lbolt() + time);
171 	else
172 		cv_wait(cv, &tx->tx_sync_lock);
173 
174 	CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
175 }
176 
177 /*
178  * Stop syncing transaction groups.
179  */
180 void
181 txg_sync_stop(dsl_pool_t *dp)
182 {
183 	tx_state_t *tx = &dp->dp_tx;
184 
185 	dprintf("pool %p\n", dp);
186 	/*
187 	 * Finish off any work in progress.
188 	 */
189 	ASSERT(tx->tx_threads == 2);
190 
191 	/*
192 	 * We need to ensure that we've vacated the deferred space_maps.
193 	 */
194 	txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
195 
196 	/*
197 	 * Wake all sync threads and wait for them to die.
198 	 */
199 	mutex_enter(&tx->tx_sync_lock);
200 
201 	ASSERT(tx->tx_threads == 2);
202 
203 	tx->tx_exiting = 1;
204 
205 	cv_broadcast(&tx->tx_quiesce_more_cv);
206 	cv_broadcast(&tx->tx_quiesce_done_cv);
207 	cv_broadcast(&tx->tx_sync_more_cv);
208 
209 	while (tx->tx_threads != 0)
210 		cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
211 
212 	tx->tx_exiting = 0;
213 
214 	mutex_exit(&tx->tx_sync_lock);
215 }
216 
217 uint64_t
218 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
219 {
220 	tx_state_t *tx = &dp->dp_tx;
221 	tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
222 	uint64_t txg;
223 
224 	mutex_enter(&tc->tc_lock);
225 
226 	txg = tx->tx_open_txg;
227 	tc->tc_count[txg & TXG_MASK]++;
228 
229 	th->th_cpu = tc;
230 	th->th_txg = txg;
231 
232 	return (txg);
233 }
234 
235 void
236 txg_rele_to_quiesce(txg_handle_t *th)
237 {
238 	tx_cpu_t *tc = th->th_cpu;
239 
240 	mutex_exit(&tc->tc_lock);
241 }
242 
243 void
244 txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
245 {
246 	tx_cpu_t *tc = th->th_cpu;
247 	int g = th->th_txg & TXG_MASK;
248 
249 	mutex_enter(&tc->tc_lock);
250 	list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
251 	mutex_exit(&tc->tc_lock);
252 }
253 
254 void
255 txg_rele_to_sync(txg_handle_t *th)
256 {
257 	tx_cpu_t *tc = th->th_cpu;
258 	int g = th->th_txg & TXG_MASK;
259 
260 	mutex_enter(&tc->tc_lock);
261 	ASSERT(tc->tc_count[g] != 0);
262 	if (--tc->tc_count[g] == 0)
263 		cv_broadcast(&tc->tc_cv[g]);
264 	mutex_exit(&tc->tc_lock);
265 
266 	th->th_cpu = NULL;	/* defensive */
267 }
268 
269 static void
270 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
271 {
272 	tx_state_t *tx = &dp->dp_tx;
273 	int g = txg & TXG_MASK;
274 	int c;
275 
276 	/*
277 	 * Grab all tx_cpu locks so nobody else can get into this txg.
278 	 */
279 	for (c = 0; c < max_ncpus; c++)
280 		mutex_enter(&tx->tx_cpu[c].tc_lock);
281 
282 	ASSERT(txg == tx->tx_open_txg);
283 	tx->tx_open_txg++;
284 
285 	/*
286 	 * Now that we've incremented tx_open_txg, we can let threads
287 	 * enter the next transaction group.
288 	 */
289 	for (c = 0; c < max_ncpus; c++)
290 		mutex_exit(&tx->tx_cpu[c].tc_lock);
291 
292 	/*
293 	 * Quiesce the transaction group by waiting for everyone to txg_exit().
294 	 */
295 	for (c = 0; c < max_ncpus; c++) {
296 		tx_cpu_t *tc = &tx->tx_cpu[c];
297 		mutex_enter(&tc->tc_lock);
298 		while (tc->tc_count[g] != 0)
299 			cv_wait(&tc->tc_cv[g], &tc->tc_lock);
300 		mutex_exit(&tc->tc_lock);
301 	}
302 }
303 
304 static void
305 txg_do_callbacks(list_t *cb_list)
306 {
307 	dmu_tx_do_callbacks(cb_list, 0);
308 
309 	list_destroy(cb_list);
310 
311 	kmem_free(cb_list, sizeof (list_t));
312 }
313 
314 /*
315  * Dispatch the commit callbacks registered on this txg to worker threads.
316  */
317 static void
318 txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg)
319 {
320 	int c;
321 	tx_state_t *tx = &dp->dp_tx;
322 	list_t *cb_list;
323 
324 	for (c = 0; c < max_ncpus; c++) {
325 		tx_cpu_t *tc = &tx->tx_cpu[c];
326 		/* No need to lock tx_cpu_t at this point */
327 
328 		int g = txg & TXG_MASK;
329 
330 		if (list_is_empty(&tc->tc_callbacks[g]))
331 			continue;
332 
333 		if (tx->tx_commit_cb_taskq == NULL) {
334 			/*
335 			 * Commit callback taskq hasn't been created yet.
336 			 */
337 			tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
338 			    max_ncpus, minclsyspri, max_ncpus, max_ncpus * 2,
339 			    TASKQ_PREPOPULATE);
340 		}
341 
342 		cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
343 		list_create(cb_list, sizeof (dmu_tx_callback_t),
344 		    offsetof(dmu_tx_callback_t, dcb_node));
345 
346 		list_move_tail(&tc->tc_callbacks[g], cb_list);
347 
348 		(void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
349 		    txg_do_callbacks, cb_list, TQ_SLEEP);
350 	}
351 }
352 
353 static void
354 txg_sync_thread(dsl_pool_t *dp)
355 {
356 	spa_t *spa = dp->dp_spa;
357 	tx_state_t *tx = &dp->dp_tx;
358 	callb_cpr_t cpr;
359 	uint64_t start, delta;
360 
361 	txg_thread_enter(tx, &cpr);
362 
363 	start = delta = 0;
364 	for (;;) {
365 		uint64_t timer, timeout = zfs_txg_timeout * hz;
366 		uint64_t txg;
367 
368 		/*
369 		 * We sync when we're scanning, there's someone waiting
370 		 * on us, or the quiesce thread has handed off a txg to
371 		 * us, or we have reached our timeout.
372 		 */
373 		timer = (delta >= timeout ? 0 : timeout - delta);
374 		while (!dsl_scan_active(dp->dp_scan) &&
375 		    !tx->tx_exiting && timer > 0 &&
376 		    tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
377 		    tx->tx_quiesced_txg == 0) {
378 			dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
379 			    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
380 			txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
381 			delta = ddi_get_lbolt() - start;
382 			timer = (delta > timeout ? 0 : timeout - delta);
383 		}
384 
385 		/*
386 		 * Wait until the quiesce thread hands off a txg to us,
387 		 * prompting it to do so if necessary.
388 		 */
389 		while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
390 			if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
391 				tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
392 			cv_broadcast(&tx->tx_quiesce_more_cv);
393 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
394 		}
395 
396 		if (tx->tx_exiting)
397 			txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
398 
399 		/*
400 		 * Consume the quiesced txg which has been handed off to
401 		 * us.  This may cause the quiescing thread to now be
402 		 * able to quiesce another txg, so we must signal it.
403 		 */
404 		txg = tx->tx_quiesced_txg;
405 		tx->tx_quiesced_txg = 0;
406 		tx->tx_syncing_txg = txg;
407 		cv_broadcast(&tx->tx_quiesce_more_cv);
408 
409 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
410 		    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
411 		mutex_exit(&tx->tx_sync_lock);
412 
413 		start = ddi_get_lbolt();
414 		spa_sync(spa, txg);
415 		delta = ddi_get_lbolt() - start;
416 
417 		mutex_enter(&tx->tx_sync_lock);
418 		tx->tx_synced_txg = txg;
419 		tx->tx_syncing_txg = 0;
420 		cv_broadcast(&tx->tx_sync_done_cv);
421 
422 		/*
423 		 * Dispatch commit callbacks to worker threads.
424 		 */
425 		txg_dispatch_callbacks(dp, txg);
426 	}
427 }
428 
429 static void
430 txg_quiesce_thread(dsl_pool_t *dp)
431 {
432 	tx_state_t *tx = &dp->dp_tx;
433 	callb_cpr_t cpr;
434 
435 	txg_thread_enter(tx, &cpr);
436 
437 	for (;;) {
438 		uint64_t txg;
439 
440 		/*
441 		 * We quiesce when there's someone waiting on us.
442 		 * However, we can only have one txg in "quiescing" or
443 		 * "quiesced, waiting to sync" state.  So we wait until
444 		 * the "quiesced, waiting to sync" txg has been consumed
445 		 * by the sync thread.
446 		 */
447 		while (!tx->tx_exiting &&
448 		    (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
449 		    tx->tx_quiesced_txg != 0))
450 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
451 
452 		if (tx->tx_exiting)
453 			txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
454 
455 		txg = tx->tx_open_txg;
456 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
457 		    txg, tx->tx_quiesce_txg_waiting,
458 		    tx->tx_sync_txg_waiting);
459 		mutex_exit(&tx->tx_sync_lock);
460 		txg_quiesce(dp, txg);
461 		mutex_enter(&tx->tx_sync_lock);
462 
463 		/*
464 		 * Hand this txg off to the sync thread.
465 		 */
466 		dprintf("quiesce done, handing off txg %llu\n", txg);
467 		tx->tx_quiesced_txg = txg;
468 		cv_broadcast(&tx->tx_sync_more_cv);
469 		cv_broadcast(&tx->tx_quiesce_done_cv);
470 	}
471 }
472 
473 /*
474  * Delay this thread by 'ticks' if we are still in the open transaction
475  * group and there is already a waiting txg quiesing or quiesced.  Abort
476  * the delay if this txg stalls or enters the quiesing state.
477  */
478 void
479 txg_delay(dsl_pool_t *dp, uint64_t txg, int ticks)
480 {
481 	tx_state_t *tx = &dp->dp_tx;
482 	int timeout = ddi_get_lbolt() + ticks;
483 
484 	/* don't delay if this txg could transition to quiesing immediately */
485 	if (tx->tx_open_txg > txg ||
486 	    tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
487 		return;
488 
489 	mutex_enter(&tx->tx_sync_lock);
490 	if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
491 		mutex_exit(&tx->tx_sync_lock);
492 		return;
493 	}
494 
495 	while (ddi_get_lbolt() < timeout &&
496 	    tx->tx_syncing_txg < txg-1 && !txg_stalled(dp))
497 		(void) cv_timedwait(&tx->tx_quiesce_more_cv, &tx->tx_sync_lock,
498 		    timeout);
499 
500 	mutex_exit(&tx->tx_sync_lock);
501 }
502 
503 void
504 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
505 {
506 	tx_state_t *tx = &dp->dp_tx;
507 
508 	mutex_enter(&tx->tx_sync_lock);
509 	ASSERT(tx->tx_threads == 2);
510 	if (txg == 0)
511 		txg = tx->tx_open_txg + TXG_DEFER_SIZE;
512 	if (tx->tx_sync_txg_waiting < txg)
513 		tx->tx_sync_txg_waiting = txg;
514 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
515 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
516 	while (tx->tx_synced_txg < txg) {
517 		dprintf("broadcasting sync more "
518 		    "tx_synced=%llu waiting=%llu dp=%p\n",
519 		    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
520 		cv_broadcast(&tx->tx_sync_more_cv);
521 		cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
522 	}
523 	mutex_exit(&tx->tx_sync_lock);
524 }
525 
526 void
527 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
528 {
529 	tx_state_t *tx = &dp->dp_tx;
530 
531 	mutex_enter(&tx->tx_sync_lock);
532 	ASSERT(tx->tx_threads == 2);
533 	if (txg == 0)
534 		txg = tx->tx_open_txg + 1;
535 	if (tx->tx_quiesce_txg_waiting < txg)
536 		tx->tx_quiesce_txg_waiting = txg;
537 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
538 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
539 	while (tx->tx_open_txg < txg) {
540 		cv_broadcast(&tx->tx_quiesce_more_cv);
541 		cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
542 	}
543 	mutex_exit(&tx->tx_sync_lock);
544 }
545 
546 boolean_t
547 txg_stalled(dsl_pool_t *dp)
548 {
549 	tx_state_t *tx = &dp->dp_tx;
550 	return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
551 }
552 
553 boolean_t
554 txg_sync_waiting(dsl_pool_t *dp)
555 {
556 	tx_state_t *tx = &dp->dp_tx;
557 
558 	return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
559 	    tx->tx_quiesced_txg != 0);
560 }
561 
562 /*
563  * Per-txg object lists.
564  */
565 void
566 txg_list_create(txg_list_t *tl, size_t offset)
567 {
568 	int t;
569 
570 	mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
571 
572 	tl->tl_offset = offset;
573 
574 	for (t = 0; t < TXG_SIZE; t++)
575 		tl->tl_head[t] = NULL;
576 }
577 
578 void
579 txg_list_destroy(txg_list_t *tl)
580 {
581 	int t;
582 
583 	for (t = 0; t < TXG_SIZE; t++)
584 		ASSERT(txg_list_empty(tl, t));
585 
586 	mutex_destroy(&tl->tl_lock);
587 }
588 
589 int
590 txg_list_empty(txg_list_t *tl, uint64_t txg)
591 {
592 	return (tl->tl_head[txg & TXG_MASK] == NULL);
593 }
594 
595 /*
596  * Add an entry to the list.
597  * Returns 0 if it's a new entry, 1 if it's already there.
598  */
599 int
600 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
601 {
602 	int t = txg & TXG_MASK;
603 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
604 	int already_on_list;
605 
606 	mutex_enter(&tl->tl_lock);
607 	already_on_list = tn->tn_member[t];
608 	if (!already_on_list) {
609 		tn->tn_member[t] = 1;
610 		tn->tn_next[t] = tl->tl_head[t];
611 		tl->tl_head[t] = tn;
612 	}
613 	mutex_exit(&tl->tl_lock);
614 
615 	return (already_on_list);
616 }
617 
618 /*
619  * Add an entry to the end of the list (walks list to find end).
620  * Returns 0 if it's a new entry, 1 if it's already there.
621  */
622 int
623 txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
624 {
625 	int t = txg & TXG_MASK;
626 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
627 	int already_on_list;
628 
629 	mutex_enter(&tl->tl_lock);
630 	already_on_list = tn->tn_member[t];
631 	if (!already_on_list) {
632 		txg_node_t **tp;
633 
634 		for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
635 			continue;
636 
637 		tn->tn_member[t] = 1;
638 		tn->tn_next[t] = NULL;
639 		*tp = tn;
640 	}
641 	mutex_exit(&tl->tl_lock);
642 
643 	return (already_on_list);
644 }
645 
646 /*
647  * Remove the head of the list and return it.
648  */
649 void *
650 txg_list_remove(txg_list_t *tl, uint64_t txg)
651 {
652 	int t = txg & TXG_MASK;
653 	txg_node_t *tn;
654 	void *p = NULL;
655 
656 	mutex_enter(&tl->tl_lock);
657 	if ((tn = tl->tl_head[t]) != NULL) {
658 		p = (char *)tn - tl->tl_offset;
659 		tl->tl_head[t] = tn->tn_next[t];
660 		tn->tn_next[t] = NULL;
661 		tn->tn_member[t] = 0;
662 	}
663 	mutex_exit(&tl->tl_lock);
664 
665 	return (p);
666 }
667 
668 /*
669  * Remove a specific item from the list and return it.
670  */
671 void *
672 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
673 {
674 	int t = txg & TXG_MASK;
675 	txg_node_t *tn, **tp;
676 
677 	mutex_enter(&tl->tl_lock);
678 
679 	for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
680 		if ((char *)tn - tl->tl_offset == p) {
681 			*tp = tn->tn_next[t];
682 			tn->tn_next[t] = NULL;
683 			tn->tn_member[t] = 0;
684 			mutex_exit(&tl->tl_lock);
685 			return (p);
686 		}
687 	}
688 
689 	mutex_exit(&tl->tl_lock);
690 
691 	return (NULL);
692 }
693 
694 int
695 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
696 {
697 	int t = txg & TXG_MASK;
698 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
699 
700 	return (tn->tn_member[t]);
701 }
702 
703 /*
704  * Walk a txg list -- only safe if you know it's not changing.
705  */
706 void *
707 txg_list_head(txg_list_t *tl, uint64_t txg)
708 {
709 	int t = txg & TXG_MASK;
710 	txg_node_t *tn = tl->tl_head[t];
711 
712 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
713 }
714 
715 void *
716 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
717 {
718 	int t = txg & TXG_MASK;
719 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
720 
721 	tn = tn->tn_next[t];
722 
723 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
724 }
725