xref: /titanic_41/usr/src/uts/common/fs/zfs/txg.c (revision 52ccf843e173e2a4a657360b0a22853fd413905f)
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 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/zfs_context.h>
29 #include <sys/txg_impl.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/callb.h>
33 
34 /*
35  * Pool-wide transaction groups.
36  */
37 
38 static void txg_sync_thread(dsl_pool_t *dp);
39 static void txg_quiesce_thread(dsl_pool_t *dp);
40 
41 int zfs_txg_timeout = 30;	/* max seconds worth of delta per txg */
42 int zfs_txg_synctime = 5;	/* target seconds to sync a txg */
43 
44 int zfs_write_limit_shift = 3;	/* 1/8th of physical memory */
45 
46 uint64_t zfs_write_limit_min = 32 << 20; /* min write limit is 32MB */
47 uint64_t zfs_write_limit_max = 0; /* max data payload per txg */
48 uint64_t zfs_write_limit_inflated = 0;
49 
50 /*
51  * Prepare the txg subsystem.
52  */
53 void
54 txg_init(dsl_pool_t *dp, uint64_t txg)
55 {
56 	tx_state_t *tx = &dp->dp_tx;
57 	int c;
58 	bzero(tx, sizeof (tx_state_t));
59 
60 	tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
61 
62 	for (c = 0; c < max_ncpus; c++) {
63 		int i;
64 
65 		mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
66 		for (i = 0; i < TXG_SIZE; i++) {
67 			cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
68 			    NULL);
69 		}
70 	}
71 
72 	rw_init(&tx->tx_suspend, NULL, RW_DEFAULT, NULL);
73 	mutex_init(&tx->tx_sync_lock, NULL, MUTEX_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 	rw_destroy(&tx->tx_suspend);
90 	mutex_destroy(&tx->tx_sync_lock);
91 
92 	for (c = 0; c < max_ncpus; c++) {
93 		int i;
94 
95 		mutex_destroy(&tx->tx_cpu[c].tc_lock);
96 		for (i = 0; i < TXG_SIZE; i++)
97 			cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
98 	}
99 
100 	kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
101 
102 	bzero(tx, sizeof (tx_state_t));
103 }
104 
105 /*
106  * Start syncing transaction groups.
107  */
108 void
109 txg_sync_start(dsl_pool_t *dp)
110 {
111 	tx_state_t *tx = &dp->dp_tx;
112 
113 	mutex_enter(&tx->tx_sync_lock);
114 
115 	dprintf("pool %p\n", dp);
116 
117 	ASSERT(tx->tx_threads == 0);
118 
119 	tx->tx_threads = 2;
120 
121 	tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
122 	    dp, 0, &p0, TS_RUN, minclsyspri);
123 
124 	tx->tx_sync_thread = thread_create(NULL, 0, txg_sync_thread,
125 	    dp, 0, &p0, TS_RUN, minclsyspri);
126 
127 	mutex_exit(&tx->tx_sync_lock);
128 }
129 
130 static void
131 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
132 {
133 	CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
134 	mutex_enter(&tx->tx_sync_lock);
135 }
136 
137 static void
138 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
139 {
140 	ASSERT(*tpp != NULL);
141 	*tpp = NULL;
142 	tx->tx_threads--;
143 	cv_broadcast(&tx->tx_exit_cv);
144 	CALLB_CPR_EXIT(cpr);		/* drops &tx->tx_sync_lock */
145 	thread_exit();
146 }
147 
148 static void
149 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, uint64_t time)
150 {
151 	CALLB_CPR_SAFE_BEGIN(cpr);
152 
153 	if (time)
154 		(void) cv_timedwait(cv, &tx->tx_sync_lock, lbolt + time);
155 	else
156 		cv_wait(cv, &tx->tx_sync_lock);
157 
158 	CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
159 }
160 
161 /*
162  * Stop syncing transaction groups.
163  */
164 void
165 txg_sync_stop(dsl_pool_t *dp)
166 {
167 	tx_state_t *tx = &dp->dp_tx;
168 
169 	dprintf("pool %p\n", dp);
170 	/*
171 	 * Finish off any work in progress.
172 	 */
173 	ASSERT(tx->tx_threads == 2);
174 	txg_wait_synced(dp, 0);
175 
176 	/*
177 	 * Wake all sync threads and wait for them to die.
178 	 */
179 	mutex_enter(&tx->tx_sync_lock);
180 
181 	ASSERT(tx->tx_threads == 2);
182 
183 	tx->tx_exiting = 1;
184 
185 	cv_broadcast(&tx->tx_quiesce_more_cv);
186 	cv_broadcast(&tx->tx_quiesce_done_cv);
187 	cv_broadcast(&tx->tx_sync_more_cv);
188 
189 	while (tx->tx_threads != 0)
190 		cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
191 
192 	tx->tx_exiting = 0;
193 
194 	mutex_exit(&tx->tx_sync_lock);
195 }
196 
197 uint64_t
198 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
199 {
200 	tx_state_t *tx = &dp->dp_tx;
201 	tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
202 	uint64_t txg;
203 
204 	mutex_enter(&tc->tc_lock);
205 
206 	txg = tx->tx_open_txg;
207 	tc->tc_count[txg & TXG_MASK]++;
208 
209 	th->th_cpu = tc;
210 	th->th_txg = txg;
211 
212 	return (txg);
213 }
214 
215 void
216 txg_rele_to_quiesce(txg_handle_t *th)
217 {
218 	tx_cpu_t *tc = th->th_cpu;
219 
220 	mutex_exit(&tc->tc_lock);
221 }
222 
223 void
224 txg_rele_to_sync(txg_handle_t *th)
225 {
226 	tx_cpu_t *tc = th->th_cpu;
227 	int g = th->th_txg & TXG_MASK;
228 
229 	mutex_enter(&tc->tc_lock);
230 	ASSERT(tc->tc_count[g] != 0);
231 	if (--tc->tc_count[g] == 0)
232 		cv_broadcast(&tc->tc_cv[g]);
233 	mutex_exit(&tc->tc_lock);
234 
235 	th->th_cpu = NULL;	/* defensive */
236 }
237 
238 static void
239 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
240 {
241 	tx_state_t *tx = &dp->dp_tx;
242 	int g = txg & TXG_MASK;
243 	int c;
244 
245 	/*
246 	 * Grab all tx_cpu locks so nobody else can get into this txg.
247 	 */
248 	for (c = 0; c < max_ncpus; c++)
249 		mutex_enter(&tx->tx_cpu[c].tc_lock);
250 
251 	ASSERT(txg == tx->tx_open_txg);
252 	tx->tx_open_txg++;
253 
254 	/*
255 	 * Now that we've incremented tx_open_txg, we can let threads
256 	 * enter the next transaction group.
257 	 */
258 	for (c = 0; c < max_ncpus; c++)
259 		mutex_exit(&tx->tx_cpu[c].tc_lock);
260 
261 	/*
262 	 * Quiesce the transaction group by waiting for everyone to txg_exit().
263 	 */
264 	for (c = 0; c < max_ncpus; c++) {
265 		tx_cpu_t *tc = &tx->tx_cpu[c];
266 		mutex_enter(&tc->tc_lock);
267 		while (tc->tc_count[g] != 0)
268 			cv_wait(&tc->tc_cv[g], &tc->tc_lock);
269 		mutex_exit(&tc->tc_lock);
270 	}
271 }
272 
273 static void
274 txg_sync_thread(dsl_pool_t *dp)
275 {
276 	tx_state_t *tx = &dp->dp_tx;
277 	callb_cpr_t cpr;
278 	uint64_t timeout, start, delta, timer;
279 	int target;
280 
281 	txg_thread_enter(tx, &cpr);
282 
283 	start = delta = 0;
284 	timeout = zfs_txg_timeout * hz;
285 	for (;;) {
286 		uint64_t txg, written;
287 
288 		/*
289 		 * We sync when there's someone waiting on us, or the
290 		 * quiesce thread has handed off a txg to us, or we have
291 		 * reached our timeout.
292 		 */
293 		timer = (delta >= timeout ? 0 : timeout - delta);
294 		while (!tx->tx_exiting && timer > 0 &&
295 		    tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
296 		    tx->tx_quiesced_txg == 0) {
297 			dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
298 			    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
299 			txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
300 			delta = lbolt - start;
301 			timer = (delta > timeout ? 0 : timeout - delta);
302 		}
303 
304 		/*
305 		 * Wait until the quiesce thread hands off a txg to us,
306 		 * prompting it to do so if necessary.
307 		 */
308 		while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
309 			if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
310 				tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
311 			cv_broadcast(&tx->tx_quiesce_more_cv);
312 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
313 		}
314 
315 		if (tx->tx_exiting)
316 			txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
317 
318 		rw_enter(&tx->tx_suspend, RW_WRITER);
319 
320 		/*
321 		 * Consume the quiesced txg which has been handed off to
322 		 * us.  This may cause the quiescing thread to now be
323 		 * able to quiesce another txg, so we must signal it.
324 		 */
325 		txg = tx->tx_quiesced_txg;
326 		tx->tx_quiesced_txg = 0;
327 		tx->tx_syncing_txg = txg;
328 		cv_broadcast(&tx->tx_quiesce_more_cv);
329 		rw_exit(&tx->tx_suspend);
330 
331 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
332 		    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
333 		mutex_exit(&tx->tx_sync_lock);
334 		start = lbolt;
335 		spa_sync(dp->dp_spa, txg);
336 		delta = (lbolt - start) + 1;
337 
338 		written = dp->dp_space_towrite[txg & TXG_MASK];
339 		dp->dp_space_towrite[txg & TXG_MASK] = 0;
340 		ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
341 
342 		/*
343 		 * If the write limit max has not been explicitly set, set it
344 		 * to a fraction of available phisical memory (default 1/8th).
345 		 * Note that we must inflate the limit because the spa
346 		 * inflates write sizes to account for data replication.
347 		 * Check this each sync phase to catch changing memory size.
348 		 */
349 		if (zfs_write_limit_inflated == 0 ||
350 		    (zfs_write_limit_shift && zfs_write_limit_max !=
351 		    physmem * PAGESIZE >> zfs_write_limit_shift)) {
352 			zfs_write_limit_max =
353 			    physmem * PAGESIZE >> zfs_write_limit_shift;
354 			zfs_write_limit_inflated =
355 			    spa_get_asize(dp->dp_spa, zfs_write_limit_max);
356 			if (zfs_write_limit_min > zfs_write_limit_inflated)
357 				zfs_write_limit_inflated = zfs_write_limit_min;
358 		}
359 
360 		/*
361 		 * Attempt to keep the sync time consistant by adjusting the
362 		 * amount of write traffic allowed into each transaction group.
363 		 */
364 		target = zfs_txg_synctime * hz;
365 		if (delta > target) {
366 			uint64_t old = MIN(dp->dp_write_limit, written);
367 
368 			dp->dp_write_limit = MAX(zfs_write_limit_min,
369 			    old * target / delta);
370 		} else if (written >= dp->dp_write_limit &&
371 		    delta >> 3 < target >> 3) {
372 			uint64_t rescale =
373 			    MIN((100 * target) / delta, 200);
374 
375 			dp->dp_write_limit = MIN(zfs_write_limit_inflated,
376 			    written * rescale / 100);
377 		}
378 
379 		mutex_enter(&tx->tx_sync_lock);
380 		rw_enter(&tx->tx_suspend, RW_WRITER);
381 		tx->tx_synced_txg = txg;
382 		tx->tx_syncing_txg = 0;
383 		rw_exit(&tx->tx_suspend);
384 		cv_broadcast(&tx->tx_sync_done_cv);
385 	}
386 }
387 
388 static void
389 txg_quiesce_thread(dsl_pool_t *dp)
390 {
391 	tx_state_t *tx = &dp->dp_tx;
392 	callb_cpr_t cpr;
393 
394 	txg_thread_enter(tx, &cpr);
395 
396 	for (;;) {
397 		uint64_t txg;
398 
399 		/*
400 		 * We quiesce when there's someone waiting on us.
401 		 * However, we can only have one txg in "quiescing" or
402 		 * "quiesced, waiting to sync" state.  So we wait until
403 		 * the "quiesced, waiting to sync" txg has been consumed
404 		 * by the sync thread.
405 		 */
406 		while (!tx->tx_exiting &&
407 		    (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
408 		    tx->tx_quiesced_txg != 0))
409 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
410 
411 		if (tx->tx_exiting)
412 			txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
413 
414 		txg = tx->tx_open_txg;
415 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
416 		    txg, tx->tx_quiesce_txg_waiting,
417 		    tx->tx_sync_txg_waiting);
418 		mutex_exit(&tx->tx_sync_lock);
419 		txg_quiesce(dp, txg);
420 		mutex_enter(&tx->tx_sync_lock);
421 
422 		/*
423 		 * Hand this txg off to the sync thread.
424 		 */
425 		dprintf("quiesce done, handing off txg %llu\n", txg);
426 		tx->tx_quiesced_txg = txg;
427 		cv_broadcast(&tx->tx_sync_more_cv);
428 		cv_broadcast(&tx->tx_quiesce_done_cv);
429 	}
430 }
431 
432 /*
433  * Delay this thread by 'ticks' if we are still in the open transaction
434  * group and there is already a waiting txg quiesing or quiesced.  Abort
435  * the delay if this txg stalls or enters the quiesing state.
436  */
437 void
438 txg_delay(dsl_pool_t *dp, uint64_t txg, int ticks)
439 {
440 	tx_state_t *tx = &dp->dp_tx;
441 	int timeout = lbolt + ticks;
442 
443 	/* don't delay if this txg could transition to quiesing immediately */
444 	if (tx->tx_open_txg > txg ||
445 	    tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
446 		return;
447 
448 	mutex_enter(&tx->tx_sync_lock);
449 	if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
450 		mutex_exit(&tx->tx_sync_lock);
451 		return;
452 	}
453 
454 	while (lbolt < timeout &&
455 	    tx->tx_syncing_txg < txg-1 && !txg_stalled(dp))
456 		(void) cv_timedwait(&tx->tx_quiesce_more_cv, &tx->tx_sync_lock,
457 		    timeout);
458 
459 	mutex_exit(&tx->tx_sync_lock);
460 }
461 
462 void
463 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
464 {
465 	tx_state_t *tx = &dp->dp_tx;
466 
467 	mutex_enter(&tx->tx_sync_lock);
468 	ASSERT(tx->tx_threads == 2);
469 	if (txg == 0)
470 		txg = tx->tx_open_txg;
471 	if (tx->tx_sync_txg_waiting < txg)
472 		tx->tx_sync_txg_waiting = txg;
473 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
474 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
475 	while (tx->tx_synced_txg < txg) {
476 		dprintf("broadcasting sync more "
477 		    "tx_synced=%llu waiting=%llu dp=%p\n",
478 		    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
479 		cv_broadcast(&tx->tx_sync_more_cv);
480 		cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
481 	}
482 	mutex_exit(&tx->tx_sync_lock);
483 }
484 
485 void
486 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
487 {
488 	tx_state_t *tx = &dp->dp_tx;
489 
490 	mutex_enter(&tx->tx_sync_lock);
491 	ASSERT(tx->tx_threads == 2);
492 	if (txg == 0)
493 		txg = tx->tx_open_txg + 1;
494 	if (tx->tx_quiesce_txg_waiting < txg)
495 		tx->tx_quiesce_txg_waiting = txg;
496 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
497 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
498 	while (tx->tx_open_txg < txg) {
499 		cv_broadcast(&tx->tx_quiesce_more_cv);
500 		cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
501 	}
502 	mutex_exit(&tx->tx_sync_lock);
503 }
504 
505 int
506 txg_stalled(dsl_pool_t *dp)
507 {
508 	tx_state_t *tx = &dp->dp_tx;
509 	return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
510 }
511 
512 void
513 txg_suspend(dsl_pool_t *dp)
514 {
515 	tx_state_t *tx = &dp->dp_tx;
516 	/* XXX some code paths suspend when they are already suspended! */
517 	rw_enter(&tx->tx_suspend, RW_READER);
518 }
519 
520 void
521 txg_resume(dsl_pool_t *dp)
522 {
523 	tx_state_t *tx = &dp->dp_tx;
524 	rw_exit(&tx->tx_suspend);
525 }
526 
527 /*
528  * Per-txg object lists.
529  */
530 void
531 txg_list_create(txg_list_t *tl, size_t offset)
532 {
533 	int t;
534 
535 	mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
536 
537 	tl->tl_offset = offset;
538 
539 	for (t = 0; t < TXG_SIZE; t++)
540 		tl->tl_head[t] = NULL;
541 }
542 
543 void
544 txg_list_destroy(txg_list_t *tl)
545 {
546 	int t;
547 
548 	for (t = 0; t < TXG_SIZE; t++)
549 		ASSERT(txg_list_empty(tl, t));
550 
551 	mutex_destroy(&tl->tl_lock);
552 }
553 
554 int
555 txg_list_empty(txg_list_t *tl, uint64_t txg)
556 {
557 	return (tl->tl_head[txg & TXG_MASK] == NULL);
558 }
559 
560 /*
561  * Add an entry to the list.
562  * Returns 0 if it's a new entry, 1 if it's already there.
563  */
564 int
565 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
566 {
567 	int t = txg & TXG_MASK;
568 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
569 	int already_on_list;
570 
571 	mutex_enter(&tl->tl_lock);
572 	already_on_list = tn->tn_member[t];
573 	if (!already_on_list) {
574 		tn->tn_member[t] = 1;
575 		tn->tn_next[t] = tl->tl_head[t];
576 		tl->tl_head[t] = tn;
577 	}
578 	mutex_exit(&tl->tl_lock);
579 
580 	return (already_on_list);
581 }
582 
583 /*
584  * Remove the head of the list and return it.
585  */
586 void *
587 txg_list_remove(txg_list_t *tl, uint64_t txg)
588 {
589 	int t = txg & TXG_MASK;
590 	txg_node_t *tn;
591 	void *p = NULL;
592 
593 	mutex_enter(&tl->tl_lock);
594 	if ((tn = tl->tl_head[t]) != NULL) {
595 		p = (char *)tn - tl->tl_offset;
596 		tl->tl_head[t] = tn->tn_next[t];
597 		tn->tn_next[t] = NULL;
598 		tn->tn_member[t] = 0;
599 	}
600 	mutex_exit(&tl->tl_lock);
601 
602 	return (p);
603 }
604 
605 /*
606  * Remove a specific item from the list and return it.
607  */
608 void *
609 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
610 {
611 	int t = txg & TXG_MASK;
612 	txg_node_t *tn, **tp;
613 
614 	mutex_enter(&tl->tl_lock);
615 
616 	for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
617 		if ((char *)tn - tl->tl_offset == p) {
618 			*tp = tn->tn_next[t];
619 			tn->tn_next[t] = NULL;
620 			tn->tn_member[t] = 0;
621 			mutex_exit(&tl->tl_lock);
622 			return (p);
623 		}
624 	}
625 
626 	mutex_exit(&tl->tl_lock);
627 
628 	return (NULL);
629 }
630 
631 int
632 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
633 {
634 	int t = txg & TXG_MASK;
635 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
636 
637 	return (tn->tn_member[t]);
638 }
639 
640 /*
641  * Walk a txg list -- only safe if you know it's not changing.
642  */
643 void *
644 txg_list_head(txg_list_t *tl, uint64_t txg)
645 {
646 	int t = txg & TXG_MASK;
647 	txg_node_t *tn = tl->tl_head[t];
648 
649 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
650 }
651 
652 void *
653 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
654 {
655 	int t = txg & TXG_MASK;
656 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
657 
658 	tn = tn->tn_next[t];
659 
660 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
661 }
662