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