Print this page
3741 zfs needs better comments
Submitted by: Will Andrews <willa@spectralogic.com>
Submitted by: Justin Gibbs <justing@spectralogic.com>
Submitted by: Alan Somers <alans@spectralogic.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Eric Schrock <eric.schrock@delphix.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/txg.c
+++ new/usr/src/uts/common/fs/zfs/txg.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Portions Copyright 2011 Martin Matuska
24 24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 25 */
26 26
27 27 #include <sys/zfs_context.h>
28 28 #include <sys/txg_impl.h>
29 29 #include <sys/dmu_impl.h>
30 30 #include <sys/dmu_tx.h>
31 31 #include <sys/dsl_pool.h>
32 32 #include <sys/dsl_scan.h>
33 33 #include <sys/callb.h>
34 34
35 35 /*
36 36 * ZFS Transaction Groups
37 37 * ----------------------
38 38 *
39 39 * ZFS transaction groups are, as the name implies, groups of transactions
40 40 * that act on persistent state. ZFS asserts consistency at the granularity of
41 41 * these transaction groups. Each successive transaction group (txg) is
42 42 * assigned a 64-bit consecutive identifier. There are three active
43 43 * transaction group states: open, quiescing, or syncing. At any given time,
44 44 * there may be an active txg associated with each state; each active txg may
45 45 * either be processing, or blocked waiting to enter the next state. There may
46 46 * be up to three active txgs, and there is always a txg in the open state
47 47 * (though it may be blocked waiting to enter the quiescing state). In broad
48 48 * strokes, transactions — operations that change in-memory structures — are
49 49 * accepted into the txg in the open state, and are completed while the txg is
50 50 * in the open or quiescing states. The accumulated changes are written to
51 51 * disk in the syncing state.
52 52 *
53 53 * Open
54 54 *
55 55 * When a new txg becomes active, it first enters the open state. New
56 56 * transactions — updates to in-memory structures — are assigned to the
57 57 * currently open txg. There is always a txg in the open state so that ZFS can
58 58 * accept new changes (though the txg may refuse new changes if it has hit
59 59 * some limit). ZFS advances the open txg to the next state for a variety of
60 60 * reasons such as it hitting a time or size threshold, or the execution of an
61 61 * administrative action that must be completed in the syncing state.
62 62 *
63 63 * Quiescing
64 64 *
65 65 * After a txg exits the open state, it enters the quiescing state. The
66 66 * quiescing state is intended to provide a buffer between accepting new
67 67 * transactions in the open state and writing them out to stable storage in
68 68 * the syncing state. While quiescing, transactions can continue their
69 69 * operation without delaying either of the other states. Typically, a txg is
70 70 * in the quiescing state very briefly since the operations are bounded by
71 71 * software latencies rather than, say, slower I/O latencies. After all
72 72 * transactions complete, the txg is ready to enter the next state.
73 73 *
74 74 * Syncing
75 75 *
76 76 * In the syncing state, the in-memory state built up during the open and (to
77 77 * a lesser degree) the quiescing states is written to stable storage. The
78 78 * process of writing out modified data can, in turn modify more data. For
79 79 * example when we write new blocks, we need to allocate space for them; those
80 80 * allocations modify metadata (space maps)... which themselves must be
81 81 * written to stable storage. During the sync state, ZFS iterates, writing out
82 82 * data until it converges and all in-memory changes have been written out.
83 83 * The first such pass is the largest as it encompasses all the modified user
84 84 * data (as opposed to filesystem metadata). Subsequent passes typically have
85 85 * far less data to write as they consist exclusively of filesystem metadata.
86 86 *
87 87 * To ensure convergence, after a certain number of passes ZFS begins
88 88 * overwriting locations on stable storage that had been allocated earlier in
89 89 * the syncing state (and subsequently freed). ZFS usually allocates new
90 90 * blocks to optimize for large, continuous, writes. For the syncing state to
91 91 * converge however it must complete a pass where no new blocks are allocated
92 92 * since each allocation requires a modification of persistent metadata.
93 93 * Further, to hasten convergence, after a prescribed number of passes, ZFS
94 94 * also defers frees, and stops compressing.
95 95 *
96 96 * In addition to writing out user data, we must also execute synctasks during
97 97 * the syncing context. A synctask is the mechanism by which some
98 98 * administrative activities work such as creating and destroying snapshots or
99 99 * datasets. Note that when a synctask is initiated it enters the open txg,
100 100 * and ZFS then pushes that txg as quickly as possible to completion of the
101 101 * syncing state in order to reduce the latency of the administrative
102 102 * activity. To complete the syncing state, ZFS writes out a new uberblock,
103 103 * the root of the tree of blocks that comprise all state stored on the ZFS
104 104 * pool. Finally, if there is a quiesced txg waiting, we signal that it can
105 105 * now transition to the syncing state.
106 106 */
107 107
108 108 static void txg_sync_thread(dsl_pool_t *dp);
109 109 static void txg_quiesce_thread(dsl_pool_t *dp);
110 110
111 111 int zfs_txg_timeout = 5; /* max seconds worth of delta per txg */
112 112
113 113 /*
114 114 * Prepare the txg subsystem.
115 115 */
116 116 void
117 117 txg_init(dsl_pool_t *dp, uint64_t txg)
118 118 {
119 119 tx_state_t *tx = &dp->dp_tx;
120 120 int c;
121 121 bzero(tx, sizeof (tx_state_t));
122 122
123 123 tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
124 124
125 125 for (c = 0; c < max_ncpus; c++) {
126 126 int i;
127 127
128 128 mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
129 129 for (i = 0; i < TXG_SIZE; i++) {
130 130 cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
131 131 NULL);
132 132 list_create(&tx->tx_cpu[c].tc_callbacks[i],
133 133 sizeof (dmu_tx_callback_t),
134 134 offsetof(dmu_tx_callback_t, dcb_node));
135 135 }
136 136 }
137 137
138 138 mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
139 139
140 140 cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
141 141 cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
142 142 cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
143 143 cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
144 144 cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
145 145
146 146 tx->tx_open_txg = txg;
147 147 }
148 148
149 149 /*
150 150 * Close down the txg subsystem.
151 151 */
152 152 void
153 153 txg_fini(dsl_pool_t *dp)
154 154 {
155 155 tx_state_t *tx = &dp->dp_tx;
156 156 int c;
157 157
158 158 ASSERT(tx->tx_threads == 0);
159 159
160 160 mutex_destroy(&tx->tx_sync_lock);
161 161
162 162 cv_destroy(&tx->tx_sync_more_cv);
163 163 cv_destroy(&tx->tx_sync_done_cv);
164 164 cv_destroy(&tx->tx_quiesce_more_cv);
165 165 cv_destroy(&tx->tx_quiesce_done_cv);
166 166 cv_destroy(&tx->tx_exit_cv);
167 167
168 168 for (c = 0; c < max_ncpus; c++) {
169 169 int i;
170 170
171 171 mutex_destroy(&tx->tx_cpu[c].tc_lock);
172 172 for (i = 0; i < TXG_SIZE; i++) {
173 173 cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
174 174 list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
175 175 }
176 176 }
177 177
178 178 if (tx->tx_commit_cb_taskq != NULL)
179 179 taskq_destroy(tx->tx_commit_cb_taskq);
180 180
181 181 kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
182 182
183 183 bzero(tx, sizeof (tx_state_t));
184 184 }
185 185
186 186 /*
187 187 * Start syncing transaction groups.
188 188 */
189 189 void
190 190 txg_sync_start(dsl_pool_t *dp)
191 191 {
192 192 tx_state_t *tx = &dp->dp_tx;
193 193
194 194 mutex_enter(&tx->tx_sync_lock);
195 195
196 196 dprintf("pool %p\n", dp);
197 197
198 198 ASSERT(tx->tx_threads == 0);
199 199
200 200 tx->tx_threads = 2;
201 201
202 202 tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
203 203 dp, 0, &p0, TS_RUN, minclsyspri);
204 204
205 205 /*
206 206 * The sync thread can need a larger-than-default stack size on
207 207 * 32-bit x86. This is due in part to nested pools and
208 208 * scrub_visitbp() recursion.
209 209 */
210 210 tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread,
211 211 dp, 0, &p0, TS_RUN, minclsyspri);
212 212
213 213 mutex_exit(&tx->tx_sync_lock);
214 214 }
215 215
216 216 static void
217 217 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
218 218 {
219 219 CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
220 220 mutex_enter(&tx->tx_sync_lock);
221 221 }
222 222
223 223 static void
224 224 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
225 225 {
226 226 ASSERT(*tpp != NULL);
227 227 *tpp = NULL;
228 228 tx->tx_threads--;
229 229 cv_broadcast(&tx->tx_exit_cv);
230 230 CALLB_CPR_EXIT(cpr); /* drops &tx->tx_sync_lock */
231 231 thread_exit();
232 232 }
233 233
234 234 static void
235 235 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, clock_t time)
236 236 {
237 237 CALLB_CPR_SAFE_BEGIN(cpr);
238 238
239 239 if (time)
240 240 (void) cv_timedwait(cv, &tx->tx_sync_lock,
241 241 ddi_get_lbolt() + time);
242 242 else
243 243 cv_wait(cv, &tx->tx_sync_lock);
244 244
245 245 CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
246 246 }
247 247
248 248 /*
249 249 * Stop syncing transaction groups.
250 250 */
251 251 void
252 252 txg_sync_stop(dsl_pool_t *dp)
253 253 {
254 254 tx_state_t *tx = &dp->dp_tx;
255 255
256 256 dprintf("pool %p\n", dp);
257 257 /*
258 258 * Finish off any work in progress.
259 259 */
260 260 ASSERT(tx->tx_threads == 2);
261 261
262 262 /*
263 263 * We need to ensure that we've vacated the deferred space_maps.
264 264 */
265 265 txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
266 266
267 267 /*
268 268 * Wake all sync threads and wait for them to die.
269 269 */
270 270 mutex_enter(&tx->tx_sync_lock);
271 271
272 272 ASSERT(tx->tx_threads == 2);
273 273
274 274 tx->tx_exiting = 1;
275 275
276 276 cv_broadcast(&tx->tx_quiesce_more_cv);
277 277 cv_broadcast(&tx->tx_quiesce_done_cv);
278 278 cv_broadcast(&tx->tx_sync_more_cv);
279 279
280 280 while (tx->tx_threads != 0)
281 281 cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
282 282
283 283 tx->tx_exiting = 0;
284 284
285 285 mutex_exit(&tx->tx_sync_lock);
286 286 }
287 287
288 288 uint64_t
289 289 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
290 290 {
291 291 tx_state_t *tx = &dp->dp_tx;
292 292 tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
293 293 uint64_t txg;
294 294
295 295 mutex_enter(&tc->tc_lock);
296 296
297 297 txg = tx->tx_open_txg;
298 298 tc->tc_count[txg & TXG_MASK]++;
299 299
300 300 th->th_cpu = tc;
301 301 th->th_txg = txg;
302 302
303 303 return (txg);
304 304 }
305 305
306 306 void
307 307 txg_rele_to_quiesce(txg_handle_t *th)
308 308 {
309 309 tx_cpu_t *tc = th->th_cpu;
310 310
311 311 mutex_exit(&tc->tc_lock);
312 312 }
313 313
314 314 void
315 315 txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
316 316 {
317 317 tx_cpu_t *tc = th->th_cpu;
318 318 int g = th->th_txg & TXG_MASK;
319 319
320 320 mutex_enter(&tc->tc_lock);
321 321 list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
322 322 mutex_exit(&tc->tc_lock);
323 323 }
324 324
325 325 void
326 326 txg_rele_to_sync(txg_handle_t *th)
327 327 {
328 328 tx_cpu_t *tc = th->th_cpu;
329 329 int g = th->th_txg & TXG_MASK;
|
↓ open down ↓ |
329 lines elided |
↑ open up ↑ |
330 330
331 331 mutex_enter(&tc->tc_lock);
332 332 ASSERT(tc->tc_count[g] != 0);
333 333 if (--tc->tc_count[g] == 0)
334 334 cv_broadcast(&tc->tc_cv[g]);
335 335 mutex_exit(&tc->tc_lock);
336 336
337 337 th->th_cpu = NULL; /* defensive */
338 338 }
339 339
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 + */
340 346 static void
341 347 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
342 348 {
343 349 tx_state_t *tx = &dp->dp_tx;
344 350 int g = txg & TXG_MASK;
345 351 int c;
346 352
347 353 /*
348 354 * Grab all tx_cpu locks so nobody else can get into this txg.
349 355 */
350 356 for (c = 0; c < max_ncpus; c++)
351 357 mutex_enter(&tx->tx_cpu[c].tc_lock);
352 358
353 359 ASSERT(txg == tx->tx_open_txg);
354 360 tx->tx_open_txg++;
355 361
356 362 DTRACE_PROBE2(txg__quiescing, dsl_pool_t *, dp, uint64_t, txg);
357 363 DTRACE_PROBE2(txg__opened, dsl_pool_t *, dp, uint64_t, tx->tx_open_txg);
358 364
359 365 /*
360 366 * Now that we've incremented tx_open_txg, we can let threads
361 367 * enter the next transaction group.
362 368 */
363 369 for (c = 0; c < max_ncpus; c++)
364 370 mutex_exit(&tx->tx_cpu[c].tc_lock);
365 371
366 372 /*
367 373 * Quiesce the transaction group by waiting for everyone to txg_exit().
368 374 */
369 375 for (c = 0; c < max_ncpus; c++) {
370 376 tx_cpu_t *tc = &tx->tx_cpu[c];
371 377 mutex_enter(&tc->tc_lock);
372 378 while (tc->tc_count[g] != 0)
373 379 cv_wait(&tc->tc_cv[g], &tc->tc_lock);
374 380 mutex_exit(&tc->tc_lock);
375 381 }
376 382 }
377 383
378 384 static void
379 385 txg_do_callbacks(list_t *cb_list)
|
↓ open down ↓ |
30 lines elided |
↑ open up ↑ |
380 386 {
381 387 dmu_tx_do_callbacks(cb_list, 0);
382 388
383 389 list_destroy(cb_list);
384 390
385 391 kmem_free(cb_list, sizeof (list_t));
386 392 }
387 393
388 394 /*
389 395 * 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.
390 399 */
391 400 static void
392 401 txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg)
393 402 {
394 403 int c;
395 404 tx_state_t *tx = &dp->dp_tx;
396 405 list_t *cb_list;
397 406
398 407 for (c = 0; c < max_ncpus; c++) {
399 408 tx_cpu_t *tc = &tx->tx_cpu[c];
400 - /* No need to lock tx_cpu_t at this point */
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 + */
401 413
402 414 int g = txg & TXG_MASK;
403 415
404 416 if (list_is_empty(&tc->tc_callbacks[g]))
405 417 continue;
406 418
407 419 if (tx->tx_commit_cb_taskq == NULL) {
408 420 /*
409 421 * Commit callback taskq hasn't been created yet.
410 422 */
411 423 tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
412 424 max_ncpus, minclsyspri, max_ncpus, max_ncpus * 2,
413 425 TASKQ_PREPOPULATE);
414 426 }
415 427
416 428 cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
417 429 list_create(cb_list, sizeof (dmu_tx_callback_t),
418 430 offsetof(dmu_tx_callback_t, dcb_node));
419 431
420 432 list_move_tail(&tc->tc_callbacks[g], cb_list);
421 433
422 434 (void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
423 435 txg_do_callbacks, cb_list, TQ_SLEEP);
424 436 }
425 437 }
426 438
427 439 static void
428 440 txg_sync_thread(dsl_pool_t *dp)
429 441 {
430 442 spa_t *spa = dp->dp_spa;
431 443 tx_state_t *tx = &dp->dp_tx;
432 444 callb_cpr_t cpr;
433 445 uint64_t start, delta;
434 446
435 447 txg_thread_enter(tx, &cpr);
436 448
437 449 start = delta = 0;
438 450 for (;;) {
439 451 uint64_t timer, timeout = zfs_txg_timeout * hz;
440 452 uint64_t txg;
441 453
442 454 /*
443 455 * We sync when we're scanning, there's someone waiting
444 456 * on us, or the quiesce thread has handed off a txg to
445 457 * us, or we have reached our timeout.
446 458 */
447 459 timer = (delta >= timeout ? 0 : timeout - delta);
448 460 while (!dsl_scan_active(dp->dp_scan) &&
449 461 !tx->tx_exiting && timer > 0 &&
450 462 tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
451 463 tx->tx_quiesced_txg == 0) {
452 464 dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
453 465 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
454 466 txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
455 467 delta = ddi_get_lbolt() - start;
456 468 timer = (delta > timeout ? 0 : timeout - delta);
457 469 }
458 470
459 471 /*
460 472 * Wait until the quiesce thread hands off a txg to us,
461 473 * prompting it to do so if necessary.
462 474 */
463 475 while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
464 476 if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
465 477 tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
466 478 cv_broadcast(&tx->tx_quiesce_more_cv);
467 479 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
468 480 }
469 481
470 482 if (tx->tx_exiting)
471 483 txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
472 484
473 485 /*
474 486 * Consume the quiesced txg which has been handed off to
475 487 * us. This may cause the quiescing thread to now be
476 488 * able to quiesce another txg, so we must signal it.
477 489 */
478 490 txg = tx->tx_quiesced_txg;
479 491 tx->tx_quiesced_txg = 0;
480 492 tx->tx_syncing_txg = txg;
481 493 DTRACE_PROBE2(txg__syncing, dsl_pool_t *, dp, uint64_t, txg);
482 494 cv_broadcast(&tx->tx_quiesce_more_cv);
483 495
484 496 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
485 497 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
486 498 mutex_exit(&tx->tx_sync_lock);
487 499
488 500 start = ddi_get_lbolt();
489 501 spa_sync(spa, txg);
490 502 delta = ddi_get_lbolt() - start;
491 503
492 504 mutex_enter(&tx->tx_sync_lock);
493 505 tx->tx_synced_txg = txg;
494 506 tx->tx_syncing_txg = 0;
495 507 DTRACE_PROBE2(txg__synced, dsl_pool_t *, dp, uint64_t, txg);
496 508 cv_broadcast(&tx->tx_sync_done_cv);
497 509
498 510 /*
499 511 * Dispatch commit callbacks to worker threads.
500 512 */
501 513 txg_dispatch_callbacks(dp, txg);
502 514 }
503 515 }
504 516
505 517 static void
506 518 txg_quiesce_thread(dsl_pool_t *dp)
507 519 {
508 520 tx_state_t *tx = &dp->dp_tx;
509 521 callb_cpr_t cpr;
510 522
511 523 txg_thread_enter(tx, &cpr);
512 524
513 525 for (;;) {
514 526 uint64_t txg;
515 527
516 528 /*
517 529 * We quiesce when there's someone waiting on us.
518 530 * However, we can only have one txg in "quiescing" or
519 531 * "quiesced, waiting to sync" state. So we wait until
520 532 * the "quiesced, waiting to sync" txg has been consumed
521 533 * by the sync thread.
522 534 */
523 535 while (!tx->tx_exiting &&
524 536 (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
525 537 tx->tx_quiesced_txg != 0))
526 538 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
527 539
528 540 if (tx->tx_exiting)
529 541 txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
530 542
531 543 txg = tx->tx_open_txg;
532 544 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
533 545 txg, tx->tx_quiesce_txg_waiting,
534 546 tx->tx_sync_txg_waiting);
535 547 mutex_exit(&tx->tx_sync_lock);
536 548 txg_quiesce(dp, txg);
537 549 mutex_enter(&tx->tx_sync_lock);
538 550
539 551 /*
540 552 * Hand this txg off to the sync thread.
541 553 */
542 554 dprintf("quiesce done, handing off txg %llu\n", txg);
543 555 tx->tx_quiesced_txg = txg;
544 556 DTRACE_PROBE2(txg__quiesced, dsl_pool_t *, dp, uint64_t, txg);
545 557 cv_broadcast(&tx->tx_sync_more_cv);
546 558 cv_broadcast(&tx->tx_quiesce_done_cv);
547 559 }
548 560 }
549 561
550 562 /*
551 563 * Delay this thread by delay nanoseconds if we are still in the open
552 564 * transaction group and there is already a waiting txg quiesing or quiesced.
553 565 * Abort the delay if this txg stalls or enters the quiesing state.
554 566 */
555 567 void
556 568 txg_delay(dsl_pool_t *dp, uint64_t txg, hrtime_t delay, hrtime_t resolution)
557 569 {
558 570 tx_state_t *tx = &dp->dp_tx;
559 571 hrtime_t start = gethrtime();
560 572
561 573 /* don't delay if this txg could transition to quiesing immediately */
562 574 if (tx->tx_open_txg > txg ||
563 575 tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
564 576 return;
565 577
566 578 mutex_enter(&tx->tx_sync_lock);
567 579 if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
568 580 mutex_exit(&tx->tx_sync_lock);
569 581 return;
570 582 }
571 583
572 584 while (gethrtime() - start < delay &&
573 585 tx->tx_syncing_txg < txg-1 && !txg_stalled(dp)) {
574 586 (void) cv_timedwait_hires(&tx->tx_quiesce_more_cv,
575 587 &tx->tx_sync_lock, delay, resolution, 0);
576 588 }
577 589
578 590 mutex_exit(&tx->tx_sync_lock);
579 591 }
580 592
581 593 void
582 594 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
583 595 {
584 596 tx_state_t *tx = &dp->dp_tx;
585 597
586 598 ASSERT(!dsl_pool_config_held(dp));
587 599
588 600 mutex_enter(&tx->tx_sync_lock);
589 601 ASSERT(tx->tx_threads == 2);
590 602 if (txg == 0)
591 603 txg = tx->tx_open_txg + TXG_DEFER_SIZE;
592 604 if (tx->tx_sync_txg_waiting < txg)
593 605 tx->tx_sync_txg_waiting = txg;
594 606 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
595 607 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
596 608 while (tx->tx_synced_txg < txg) {
597 609 dprintf("broadcasting sync more "
598 610 "tx_synced=%llu waiting=%llu dp=%p\n",
599 611 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
600 612 cv_broadcast(&tx->tx_sync_more_cv);
601 613 cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
602 614 }
603 615 mutex_exit(&tx->tx_sync_lock);
604 616 }
605 617
606 618 void
607 619 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
608 620 {
609 621 tx_state_t *tx = &dp->dp_tx;
610 622
611 623 ASSERT(!dsl_pool_config_held(dp));
612 624
613 625 mutex_enter(&tx->tx_sync_lock);
614 626 ASSERT(tx->tx_threads == 2);
615 627 if (txg == 0)
616 628 txg = tx->tx_open_txg + 1;
617 629 if (tx->tx_quiesce_txg_waiting < txg)
618 630 tx->tx_quiesce_txg_waiting = txg;
619 631 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
620 632 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
621 633 while (tx->tx_open_txg < txg) {
622 634 cv_broadcast(&tx->tx_quiesce_more_cv);
623 635 cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
624 636 }
625 637 mutex_exit(&tx->tx_sync_lock);
626 638 }
627 639
628 640 boolean_t
629 641 txg_stalled(dsl_pool_t *dp)
630 642 {
631 643 tx_state_t *tx = &dp->dp_tx;
632 644 return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
633 645 }
634 646
635 647 boolean_t
636 648 txg_sync_waiting(dsl_pool_t *dp)
637 649 {
638 650 tx_state_t *tx = &dp->dp_tx;
639 651
640 652 return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
641 653 tx->tx_quiesced_txg != 0);
642 654 }
643 655
644 656 /*
645 657 * Per-txg object lists.
646 658 */
647 659 void
648 660 txg_list_create(txg_list_t *tl, size_t offset)
649 661 {
650 662 int t;
651 663
652 664 mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
653 665
654 666 tl->tl_offset = offset;
655 667
656 668 for (t = 0; t < TXG_SIZE; t++)
657 669 tl->tl_head[t] = NULL;
658 670 }
659 671
660 672 void
661 673 txg_list_destroy(txg_list_t *tl)
662 674 {
663 675 int t;
664 676
665 677 for (t = 0; t < TXG_SIZE; t++)
666 678 ASSERT(txg_list_empty(tl, t));
667 679
668 680 mutex_destroy(&tl->tl_lock);
669 681 }
670 682
671 683 boolean_t
672 684 txg_list_empty(txg_list_t *tl, uint64_t txg)
673 685 {
674 686 return (tl->tl_head[txg & TXG_MASK] == NULL);
675 687 }
676 688
677 689 /*
678 690 * Add an entry to the list (unless it's already on the list).
679 691 * Returns B_TRUE if it was actually added.
680 692 */
681 693 boolean_t
682 694 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
683 695 {
684 696 int t = txg & TXG_MASK;
685 697 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
686 698 boolean_t add;
687 699
688 700 mutex_enter(&tl->tl_lock);
689 701 add = (tn->tn_member[t] == 0);
690 702 if (add) {
691 703 tn->tn_member[t] = 1;
692 704 tn->tn_next[t] = tl->tl_head[t];
693 705 tl->tl_head[t] = tn;
694 706 }
695 707 mutex_exit(&tl->tl_lock);
696 708
697 709 return (add);
698 710 }
699 711
700 712 /*
701 713 * Add an entry to the end of the list, unless it's already on the list.
702 714 * (walks list to find end)
703 715 * Returns B_TRUE if it was actually added.
704 716 */
705 717 boolean_t
706 718 txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
707 719 {
708 720 int t = txg & TXG_MASK;
709 721 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
710 722 boolean_t add;
711 723
712 724 mutex_enter(&tl->tl_lock);
713 725 add = (tn->tn_member[t] == 0);
714 726 if (add) {
715 727 txg_node_t **tp;
716 728
717 729 for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
718 730 continue;
719 731
720 732 tn->tn_member[t] = 1;
721 733 tn->tn_next[t] = NULL;
722 734 *tp = tn;
723 735 }
724 736 mutex_exit(&tl->tl_lock);
725 737
726 738 return (add);
727 739 }
728 740
729 741 /*
730 742 * Remove the head of the list and return it.
731 743 */
732 744 void *
733 745 txg_list_remove(txg_list_t *tl, uint64_t txg)
734 746 {
735 747 int t = txg & TXG_MASK;
736 748 txg_node_t *tn;
737 749 void *p = NULL;
738 750
739 751 mutex_enter(&tl->tl_lock);
740 752 if ((tn = tl->tl_head[t]) != NULL) {
741 753 p = (char *)tn - tl->tl_offset;
742 754 tl->tl_head[t] = tn->tn_next[t];
743 755 tn->tn_next[t] = NULL;
744 756 tn->tn_member[t] = 0;
745 757 }
746 758 mutex_exit(&tl->tl_lock);
747 759
748 760 return (p);
749 761 }
750 762
751 763 /*
752 764 * Remove a specific item from the list and return it.
753 765 */
754 766 void *
755 767 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
756 768 {
757 769 int t = txg & TXG_MASK;
758 770 txg_node_t *tn, **tp;
759 771
760 772 mutex_enter(&tl->tl_lock);
761 773
762 774 for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
763 775 if ((char *)tn - tl->tl_offset == p) {
764 776 *tp = tn->tn_next[t];
765 777 tn->tn_next[t] = NULL;
766 778 tn->tn_member[t] = 0;
767 779 mutex_exit(&tl->tl_lock);
768 780 return (p);
769 781 }
770 782 }
771 783
772 784 mutex_exit(&tl->tl_lock);
773 785
774 786 return (NULL);
775 787 }
776 788
777 789 boolean_t
778 790 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
779 791 {
780 792 int t = txg & TXG_MASK;
781 793 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
782 794
783 795 return (tn->tn_member[t] != 0);
784 796 }
785 797
786 798 /*
787 799 * Walk a txg list -- only safe if you know it's not changing.
788 800 */
789 801 void *
790 802 txg_list_head(txg_list_t *tl, uint64_t txg)
791 803 {
792 804 int t = txg & TXG_MASK;
793 805 txg_node_t *tn = tl->tl_head[t];
794 806
795 807 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
796 808 }
797 809
798 810 void *
799 811 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
800 812 {
801 813 int t = txg & TXG_MASK;
802 814 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
803 815
804 816 tn = tn->tn_next[t];
805 817
806 818 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
807 819 }
|
↓ open down ↓ |
397 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX