txg.c revision 260763
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Portions Copyright 2011 Martin Matuska <mm@FreeBSD.org> 24 * Copyright (c) 2013 by Delphix. All rights reserved. 25 */ 26 27#include <sys/zfs_context.h> 28#include <sys/txg_impl.h> 29#include <sys/dmu_impl.h> 30#include <sys/dmu_tx.h> 31#include <sys/dsl_pool.h> 32#include <sys/dsl_scan.h> 33#include <sys/callb.h> 34 35/* 36 * ZFS Transaction Groups 37 * ---------------------- 38 * 39 * ZFS transaction groups are, as the name implies, groups of transactions 40 * that act on persistent state. ZFS asserts consistency at the granularity of 41 * these transaction groups. Each successive transaction group (txg) is 42 * assigned a 64-bit consecutive identifier. There are three active 43 * transaction group states: open, quiescing, or syncing. At any given time, 44 * there may be an active txg associated with each state; each active txg may 45 * either be processing, or blocked waiting to enter the next state. There may 46 * be up to three active txgs, and there is always a txg in the open state 47 * (though it may be blocked waiting to enter the quiescing state). In broad 48 * strokes, transactions -- operations that change in-memory structures -- are 49 * accepted into the txg in the open state, and are completed while the txg is 50 * in the open or quiescing states. The accumulated changes are written to 51 * disk in the syncing state. 52 * 53 * Open 54 * 55 * When a new txg becomes active, it first enters the open state. New 56 * transactions -- updates to in-memory structures -- are assigned to the 57 * currently open txg. There is always a txg in the open state so that ZFS can 58 * accept new changes (though the txg may refuse new changes if it has hit 59 * some limit). ZFS advances the open txg to the next state for a variety of 60 * reasons such as it hitting a time or size threshold, or the execution of an 61 * administrative action that must be completed in the syncing state. 62 * 63 * Quiescing 64 * 65 * After a txg exits the open state, it enters the quiescing state. The 66 * quiescing state is intended to provide a buffer between accepting new 67 * transactions in the open state and writing them out to stable storage in 68 * the syncing state. While quiescing, transactions can continue their 69 * operation without delaying either of the other states. Typically, a txg is 70 * in the quiescing state very briefly since the operations are bounded by 71 * software latencies rather than, say, slower I/O latencies. After all 72 * transactions complete, the txg is ready to enter the next state. 73 * 74 * Syncing 75 * 76 * In the syncing state, the in-memory state built up during the open and (to 77 * a lesser degree) the quiescing states is written to stable storage. The 78 * process of writing out modified data can, in turn modify more data. For 79 * example when we write new blocks, we need to allocate space for them; those 80 * allocations modify metadata (space maps)... which themselves must be 81 * written to stable storage. During the sync state, ZFS iterates, writing out 82 * data until it converges and all in-memory changes have been written out. 83 * The first such pass is the largest as it encompasses all the modified user 84 * data (as opposed to filesystem metadata). Subsequent passes typically have 85 * far less data to write as they consist exclusively of filesystem metadata. 86 * 87 * To ensure convergence, after a certain number of passes ZFS begins 88 * overwriting locations on stable storage that had been allocated earlier in 89 * the syncing state (and subsequently freed). ZFS usually allocates new 90 * blocks to optimize for large, continuous, writes. For the syncing state to 91 * converge however it must complete a pass where no new blocks are allocated 92 * since each allocation requires a modification of persistent metadata. 93 * Further, to hasten convergence, after a prescribed number of passes, ZFS 94 * also defers frees, and stops compressing. 95 * 96 * In addition to writing out user data, we must also execute synctasks during 97 * the syncing context. A synctask is the mechanism by which some 98 * administrative activities work such as creating and destroying snapshots or 99 * datasets. Note that when a synctask is initiated it enters the open txg, 100 * and ZFS then pushes that txg as quickly as possible to completion of the 101 * syncing state in order to reduce the latency of the administrative 102 * activity. To complete the syncing state, ZFS writes out a new uberblock, 103 * the root of the tree of blocks that comprise all state stored on the ZFS 104 * pool. Finally, if there is a quiesced txg waiting, we signal that it can 105 * now transition to the syncing state. 106 */ 107 108static void txg_sync_thread(void *arg); 109static void txg_quiesce_thread(void *arg); 110 111int zfs_txg_timeout = 5; /* max seconds worth of delta per txg */ 112 113SYSCTL_DECL(_vfs_zfs); 114SYSCTL_NODE(_vfs_zfs, OID_AUTO, txg, CTLFLAG_RW, 0, "ZFS TXG"); 115TUNABLE_INT("vfs.zfs.txg.timeout", &zfs_txg_timeout); 116SYSCTL_INT(_vfs_zfs_txg, OID_AUTO, timeout, CTLFLAG_RW, &zfs_txg_timeout, 0, 117 "Maximum seconds worth of delta per txg"); 118 119/* 120 * Prepare the txg subsystem. 121 */ 122void 123txg_init(dsl_pool_t *dp, uint64_t txg) 124{ 125 tx_state_t *tx = &dp->dp_tx; 126 int c; 127 bzero(tx, sizeof (tx_state_t)); 128 129 tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP); 130 131 for (c = 0; c < max_ncpus; c++) { 132 int i; 133 134 mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL); 135 mutex_init(&tx->tx_cpu[c].tc_open_lock, NULL, MUTEX_DEFAULT, 136 NULL); 137 for (i = 0; i < TXG_SIZE; i++) { 138 cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT, 139 NULL); 140 list_create(&tx->tx_cpu[c].tc_callbacks[i], 141 sizeof (dmu_tx_callback_t), 142 offsetof(dmu_tx_callback_t, dcb_node)); 143 } 144 } 145 146 mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL); 147 148 cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL); 149 cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL); 150 cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL); 151 cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL); 152 cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL); 153 154 tx->tx_open_txg = txg; 155} 156 157/* 158 * Close down the txg subsystem. 159 */ 160void 161txg_fini(dsl_pool_t *dp) 162{ 163 tx_state_t *tx = &dp->dp_tx; 164 int c; 165 166 ASSERT(tx->tx_threads == 0); 167 168 mutex_destroy(&tx->tx_sync_lock); 169 170 cv_destroy(&tx->tx_sync_more_cv); 171 cv_destroy(&tx->tx_sync_done_cv); 172 cv_destroy(&tx->tx_quiesce_more_cv); 173 cv_destroy(&tx->tx_quiesce_done_cv); 174 cv_destroy(&tx->tx_exit_cv); 175 176 for (c = 0; c < max_ncpus; c++) { 177 int i; 178 179 mutex_destroy(&tx->tx_cpu[c].tc_open_lock); 180 mutex_destroy(&tx->tx_cpu[c].tc_lock); 181 for (i = 0; i < TXG_SIZE; i++) { 182 cv_destroy(&tx->tx_cpu[c].tc_cv[i]); 183 list_destroy(&tx->tx_cpu[c].tc_callbacks[i]); 184 } 185 } 186 187 if (tx->tx_commit_cb_taskq != NULL) 188 taskq_destroy(tx->tx_commit_cb_taskq); 189 190 kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t)); 191 192 bzero(tx, sizeof (tx_state_t)); 193} 194 195/* 196 * Start syncing transaction groups. 197 */ 198void 199txg_sync_start(dsl_pool_t *dp) 200{ 201 tx_state_t *tx = &dp->dp_tx; 202 203 mutex_enter(&tx->tx_sync_lock); 204 205 dprintf("pool %p\n", dp); 206 207 ASSERT(tx->tx_threads == 0); 208 209 tx->tx_threads = 2; 210 211 tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread, 212 dp, 0, &p0, TS_RUN, minclsyspri); 213 214 /* 215 * The sync thread can need a larger-than-default stack size on 216 * 32-bit x86. This is due in part to nested pools and 217 * scrub_visitbp() recursion. 218 */ 219 tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread, 220 dp, 0, &p0, TS_RUN, minclsyspri); 221 222 mutex_exit(&tx->tx_sync_lock); 223} 224 225static void 226txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr) 227{ 228 CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG); 229 mutex_enter(&tx->tx_sync_lock); 230} 231 232static void 233txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp) 234{ 235 ASSERT(*tpp != NULL); 236 *tpp = NULL; 237 tx->tx_threads--; 238 cv_broadcast(&tx->tx_exit_cv); 239 CALLB_CPR_EXIT(cpr); /* drops &tx->tx_sync_lock */ 240 thread_exit(); 241} 242 243static void 244txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, clock_t time) 245{ 246 CALLB_CPR_SAFE_BEGIN(cpr); 247 248 if (time) 249 (void) cv_timedwait(cv, &tx->tx_sync_lock, time); 250 else 251 cv_wait(cv, &tx->tx_sync_lock); 252 253 CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock); 254} 255 256/* 257 * Stop syncing transaction groups. 258 */ 259void 260txg_sync_stop(dsl_pool_t *dp) 261{ 262 tx_state_t *tx = &dp->dp_tx; 263 264 dprintf("pool %p\n", dp); 265 /* 266 * Finish off any work in progress. 267 */ 268 ASSERT(tx->tx_threads == 2); 269 270 /* 271 * We need to ensure that we've vacated the deferred space_maps. 272 */ 273 txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE); 274 275 /* 276 * Wake all sync threads and wait for them to die. 277 */ 278 mutex_enter(&tx->tx_sync_lock); 279 280 ASSERT(tx->tx_threads == 2); 281 282 tx->tx_exiting = 1; 283 284 cv_broadcast(&tx->tx_quiesce_more_cv); 285 cv_broadcast(&tx->tx_quiesce_done_cv); 286 cv_broadcast(&tx->tx_sync_more_cv); 287 288 while (tx->tx_threads != 0) 289 cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock); 290 291 tx->tx_exiting = 0; 292 293 mutex_exit(&tx->tx_sync_lock); 294} 295 296uint64_t 297txg_hold_open(dsl_pool_t *dp, txg_handle_t *th) 298{ 299 tx_state_t *tx = &dp->dp_tx; 300 tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID]; 301 uint64_t txg; 302 303 mutex_enter(&tc->tc_open_lock); 304 txg = tx->tx_open_txg; 305 306 mutex_enter(&tc->tc_lock); 307 tc->tc_count[txg & TXG_MASK]++; 308 mutex_exit(&tc->tc_lock); 309 310 th->th_cpu = tc; 311 th->th_txg = txg; 312 313 return (txg); 314} 315 316void 317txg_rele_to_quiesce(txg_handle_t *th) 318{ 319 tx_cpu_t *tc = th->th_cpu; 320 321 ASSERT(!MUTEX_HELD(&tc->tc_lock)); 322 mutex_exit(&tc->tc_open_lock); 323} 324 325void 326txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks) 327{ 328 tx_cpu_t *tc = th->th_cpu; 329 int g = th->th_txg & TXG_MASK; 330 331 mutex_enter(&tc->tc_lock); 332 list_move_tail(&tc->tc_callbacks[g], tx_callbacks); 333 mutex_exit(&tc->tc_lock); 334} 335 336void 337txg_rele_to_sync(txg_handle_t *th) 338{ 339 tx_cpu_t *tc = th->th_cpu; 340 int g = th->th_txg & TXG_MASK; 341 342 mutex_enter(&tc->tc_lock); 343 ASSERT(tc->tc_count[g] != 0); 344 if (--tc->tc_count[g] == 0) 345 cv_broadcast(&tc->tc_cv[g]); 346 mutex_exit(&tc->tc_lock); 347 348 th->th_cpu = NULL; /* defensive */ 349} 350 351/* 352 * Blocks until all transactions in the group are committed. 353 * 354 * On return, the transaction group has reached a stable state in which it can 355 * then be passed off to the syncing context. 356 */ 357static void 358txg_quiesce(dsl_pool_t *dp, uint64_t txg) 359{ 360 tx_state_t *tx = &dp->dp_tx; 361 int g = txg & TXG_MASK; 362 int c; 363 364 /* 365 * Grab all tc_open_locks so nobody else can get into this txg. 366 */ 367 for (c = 0; c < max_ncpus; c++) 368 mutex_enter(&tx->tx_cpu[c].tc_open_lock); 369 370 ASSERT(txg == tx->tx_open_txg); 371 tx->tx_open_txg++; 372 tx->tx_open_time = gethrtime(); 373 374 DTRACE_PROBE2(txg__quiescing, dsl_pool_t *, dp, uint64_t, txg); 375 DTRACE_PROBE2(txg__opened, dsl_pool_t *, dp, uint64_t, tx->tx_open_txg); 376 377 /* 378 * Now that we've incremented tx_open_txg, we can let threads 379 * enter the next transaction group. 380 */ 381 for (c = 0; c < max_ncpus; c++) 382 mutex_exit(&tx->tx_cpu[c].tc_open_lock); 383 384 /* 385 * Quiesce the transaction group by waiting for everyone to txg_exit(). 386 */ 387 for (c = 0; c < max_ncpus; c++) { 388 tx_cpu_t *tc = &tx->tx_cpu[c]; 389 mutex_enter(&tc->tc_lock); 390 while (tc->tc_count[g] != 0) 391 cv_wait(&tc->tc_cv[g], &tc->tc_lock); 392 mutex_exit(&tc->tc_lock); 393 } 394} 395 396static void 397txg_do_callbacks(void *arg) 398{ 399 list_t *cb_list = arg; 400 401 dmu_tx_do_callbacks(cb_list, 0); 402 403 list_destroy(cb_list); 404 405 kmem_free(cb_list, sizeof (list_t)); 406} 407 408/* 409 * Dispatch the commit callbacks registered on this txg to worker threads. 410 * 411 * If no callbacks are registered for a given TXG, nothing happens. 412 * This function creates a taskq for the associated pool, if needed. 413 */ 414static void 415txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg) 416{ 417 int c; 418 tx_state_t *tx = &dp->dp_tx; 419 list_t *cb_list; 420 421 for (c = 0; c < max_ncpus; c++) { 422 tx_cpu_t *tc = &tx->tx_cpu[c]; 423 /* 424 * No need to lock tx_cpu_t at this point, since this can 425 * only be called once a txg has been synced. 426 */ 427 428 int g = txg & TXG_MASK; 429 430 if (list_is_empty(&tc->tc_callbacks[g])) 431 continue; 432 433 if (tx->tx_commit_cb_taskq == NULL) { 434 /* 435 * Commit callback taskq hasn't been created yet. 436 */ 437 tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb", 438 max_ncpus, minclsyspri, max_ncpus, max_ncpus * 2, 439 TASKQ_PREPOPULATE); 440 } 441 442 cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP); 443 list_create(cb_list, sizeof (dmu_tx_callback_t), 444 offsetof(dmu_tx_callback_t, dcb_node)); 445 446 list_move_tail(cb_list, &tc->tc_callbacks[g]); 447 448 (void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *) 449 txg_do_callbacks, cb_list, TQ_SLEEP); 450 } 451} 452 453static void 454txg_sync_thread(void *arg) 455{ 456 dsl_pool_t *dp = arg; 457 spa_t *spa = dp->dp_spa; 458 tx_state_t *tx = &dp->dp_tx; 459 callb_cpr_t cpr; 460 uint64_t start, delta; 461 462 txg_thread_enter(tx, &cpr); 463 464 start = delta = 0; 465 for (;;) { 466 uint64_t timeout = zfs_txg_timeout * hz; 467 uint64_t timer; 468 uint64_t txg; 469 470 /* 471 * We sync when we're scanning, there's someone waiting 472 * on us, or the quiesce thread has handed off a txg to 473 * us, or we have reached our timeout. 474 */ 475 timer = (delta >= timeout ? 0 : timeout - delta); 476 while (!dsl_scan_active(dp->dp_scan) && 477 !tx->tx_exiting && timer > 0 && 478 tx->tx_synced_txg >= tx->tx_sync_txg_waiting && 479 tx->tx_quiesced_txg == 0 && 480 dp->dp_dirty_total < zfs_dirty_data_sync) { 481 dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n", 482 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp); 483 txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer); 484 delta = ddi_get_lbolt() - start; 485 timer = (delta > timeout ? 0 : timeout - delta); 486 } 487 488 /* 489 * Wait until the quiesce thread hands off a txg to us, 490 * prompting it to do so if necessary. 491 */ 492 while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) { 493 if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1) 494 tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1; 495 cv_broadcast(&tx->tx_quiesce_more_cv); 496 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0); 497 } 498 499 if (tx->tx_exiting) 500 txg_thread_exit(tx, &cpr, &tx->tx_sync_thread); 501 502 /* 503 * Consume the quiesced txg which has been handed off to 504 * us. This may cause the quiescing thread to now be 505 * able to quiesce another txg, so we must signal it. 506 */ 507 txg = tx->tx_quiesced_txg; 508 tx->tx_quiesced_txg = 0; 509 tx->tx_syncing_txg = txg; 510 DTRACE_PROBE2(txg__syncing, dsl_pool_t *, dp, uint64_t, txg); 511 cv_broadcast(&tx->tx_quiesce_more_cv); 512 513 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n", 514 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting); 515 mutex_exit(&tx->tx_sync_lock); 516 517 start = ddi_get_lbolt(); 518 spa_sync(spa, txg); 519 delta = ddi_get_lbolt() - start; 520 521 mutex_enter(&tx->tx_sync_lock); 522 tx->tx_synced_txg = txg; 523 tx->tx_syncing_txg = 0; 524 DTRACE_PROBE2(txg__synced, dsl_pool_t *, dp, uint64_t, txg); 525 cv_broadcast(&tx->tx_sync_done_cv); 526 527 /* 528 * Dispatch commit callbacks to worker threads. 529 */ 530 txg_dispatch_callbacks(dp, txg); 531 } 532} 533 534static void 535txg_quiesce_thread(void *arg) 536{ 537 dsl_pool_t *dp = arg; 538 tx_state_t *tx = &dp->dp_tx; 539 callb_cpr_t cpr; 540 541 txg_thread_enter(tx, &cpr); 542 543 for (;;) { 544 uint64_t txg; 545 546 /* 547 * We quiesce when there's someone waiting on us. 548 * However, we can only have one txg in "quiescing" or 549 * "quiesced, waiting to sync" state. So we wait until 550 * the "quiesced, waiting to sync" txg has been consumed 551 * by the sync thread. 552 */ 553 while (!tx->tx_exiting && 554 (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting || 555 tx->tx_quiesced_txg != 0)) 556 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0); 557 558 if (tx->tx_exiting) 559 txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread); 560 561 txg = tx->tx_open_txg; 562 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n", 563 txg, tx->tx_quiesce_txg_waiting, 564 tx->tx_sync_txg_waiting); 565 mutex_exit(&tx->tx_sync_lock); 566 txg_quiesce(dp, txg); 567 mutex_enter(&tx->tx_sync_lock); 568 569 /* 570 * Hand this txg off to the sync thread. 571 */ 572 dprintf("quiesce done, handing off txg %llu\n", txg); 573 tx->tx_quiesced_txg = txg; 574 DTRACE_PROBE2(txg__quiesced, dsl_pool_t *, dp, uint64_t, txg); 575 cv_broadcast(&tx->tx_sync_more_cv); 576 cv_broadcast(&tx->tx_quiesce_done_cv); 577 } 578} 579 580/* 581 * Delay this thread by delay nanoseconds if we are still in the open 582 * transaction group and there is already a waiting txg quiesing or quiesced. 583 * Abort the delay if this txg stalls or enters the quiesing state. 584 */ 585void 586txg_delay(dsl_pool_t *dp, uint64_t txg, hrtime_t delay, hrtime_t resolution) 587{ 588 tx_state_t *tx = &dp->dp_tx; 589 hrtime_t start = gethrtime(); 590 591 /* don't delay if this txg could transition to quiescing immediately */ 592 if (tx->tx_open_txg > txg || 593 tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1) 594 return; 595 596 mutex_enter(&tx->tx_sync_lock); 597 if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) { 598 mutex_exit(&tx->tx_sync_lock); 599 return; 600 } 601 602 while (gethrtime() - start < delay && 603 tx->tx_syncing_txg < txg-1 && !txg_stalled(dp)) { 604 (void) cv_timedwait_hires(&tx->tx_quiesce_more_cv, 605 &tx->tx_sync_lock, delay, resolution, 0); 606 } 607 608 mutex_exit(&tx->tx_sync_lock); 609} 610 611void 612txg_wait_synced(dsl_pool_t *dp, uint64_t txg) 613{ 614 tx_state_t *tx = &dp->dp_tx; 615 616 ASSERT(!dsl_pool_config_held(dp)); 617 618 mutex_enter(&tx->tx_sync_lock); 619 ASSERT(tx->tx_threads == 2); 620 if (txg == 0) 621 txg = tx->tx_open_txg + TXG_DEFER_SIZE; 622 if (tx->tx_sync_txg_waiting < txg) 623 tx->tx_sync_txg_waiting = txg; 624 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n", 625 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting); 626 while (tx->tx_synced_txg < txg) { 627 dprintf("broadcasting sync more " 628 "tx_synced=%llu waiting=%llu dp=%p\n", 629 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp); 630 cv_broadcast(&tx->tx_sync_more_cv); 631 cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock); 632 } 633 mutex_exit(&tx->tx_sync_lock); 634} 635 636void 637txg_wait_open(dsl_pool_t *dp, uint64_t txg) 638{ 639 tx_state_t *tx = &dp->dp_tx; 640 641 ASSERT(!dsl_pool_config_held(dp)); 642 643 mutex_enter(&tx->tx_sync_lock); 644 ASSERT(tx->tx_threads == 2); 645 if (txg == 0) 646 txg = tx->tx_open_txg + 1; 647 if (tx->tx_quiesce_txg_waiting < txg) 648 tx->tx_quiesce_txg_waiting = txg; 649 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n", 650 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting); 651 while (tx->tx_open_txg < txg) { 652 cv_broadcast(&tx->tx_quiesce_more_cv); 653 cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock); 654 } 655 mutex_exit(&tx->tx_sync_lock); 656} 657 658/* 659 * If there isn't a txg syncing or in the pipeline, push another txg through 660 * the pipeline by queiscing the open txg. 661 */ 662void 663txg_kick(dsl_pool_t *dp) 664{ 665 tx_state_t *tx = &dp->dp_tx; 666 667 ASSERT(!dsl_pool_config_held(dp)); 668 669 mutex_enter(&tx->tx_sync_lock); 670 if (tx->tx_syncing_txg == 0 && 671 tx->tx_quiesce_txg_waiting <= tx->tx_open_txg && 672 tx->tx_sync_txg_waiting <= tx->tx_synced_txg && 673 tx->tx_quiesced_txg <= tx->tx_synced_txg) { 674 tx->tx_quiesce_txg_waiting = tx->tx_open_txg + 1; 675 cv_broadcast(&tx->tx_quiesce_more_cv); 676 } 677 mutex_exit(&tx->tx_sync_lock); 678} 679 680boolean_t 681txg_stalled(dsl_pool_t *dp) 682{ 683 tx_state_t *tx = &dp->dp_tx; 684 return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg); 685} 686 687boolean_t 688txg_sync_waiting(dsl_pool_t *dp) 689{ 690 tx_state_t *tx = &dp->dp_tx; 691 692 return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting || 693 tx->tx_quiesced_txg != 0); 694} 695 696/* 697 * Per-txg object lists. 698 */ 699void 700txg_list_create(txg_list_t *tl, size_t offset) 701{ 702 int t; 703 704 mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL); 705 706 tl->tl_offset = offset; 707 708 for (t = 0; t < TXG_SIZE; t++) 709 tl->tl_head[t] = NULL; 710} 711 712void 713txg_list_destroy(txg_list_t *tl) 714{ 715 int t; 716 717 for (t = 0; t < TXG_SIZE; t++) 718 ASSERT(txg_list_empty(tl, t)); 719 720 mutex_destroy(&tl->tl_lock); 721} 722 723boolean_t 724txg_list_empty(txg_list_t *tl, uint64_t txg) 725{ 726 return (tl->tl_head[txg & TXG_MASK] == NULL); 727} 728 729/* 730 * Add an entry to the list (unless it's already on the list). 731 * Returns B_TRUE if it was actually added. 732 */ 733boolean_t 734txg_list_add(txg_list_t *tl, void *p, uint64_t txg) 735{ 736 int t = txg & TXG_MASK; 737 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset); 738 boolean_t add; 739 740 mutex_enter(&tl->tl_lock); 741 add = (tn->tn_member[t] == 0); 742 if (add) { 743 tn->tn_member[t] = 1; 744 tn->tn_next[t] = tl->tl_head[t]; 745 tl->tl_head[t] = tn; 746 } 747 mutex_exit(&tl->tl_lock); 748 749 return (add); 750} 751 752/* 753 * Add an entry to the end of the list, unless it's already on the list. 754 * (walks list to find end) 755 * Returns B_TRUE if it was actually added. 756 */ 757boolean_t 758txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg) 759{ 760 int t = txg & TXG_MASK; 761 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset); 762 boolean_t add; 763 764 mutex_enter(&tl->tl_lock); 765 add = (tn->tn_member[t] == 0); 766 if (add) { 767 txg_node_t **tp; 768 769 for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t]) 770 continue; 771 772 tn->tn_member[t] = 1; 773 tn->tn_next[t] = NULL; 774 *tp = tn; 775 } 776 mutex_exit(&tl->tl_lock); 777 778 return (add); 779} 780 781/* 782 * Remove the head of the list and return it. 783 */ 784void * 785txg_list_remove(txg_list_t *tl, uint64_t txg) 786{ 787 int t = txg & TXG_MASK; 788 txg_node_t *tn; 789 void *p = NULL; 790 791 mutex_enter(&tl->tl_lock); 792 if ((tn = tl->tl_head[t]) != NULL) { 793 p = (char *)tn - tl->tl_offset; 794 tl->tl_head[t] = tn->tn_next[t]; 795 tn->tn_next[t] = NULL; 796 tn->tn_member[t] = 0; 797 } 798 mutex_exit(&tl->tl_lock); 799 800 return (p); 801} 802 803/* 804 * Remove a specific item from the list and return it. 805 */ 806void * 807txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg) 808{ 809 int t = txg & TXG_MASK; 810 txg_node_t *tn, **tp; 811 812 mutex_enter(&tl->tl_lock); 813 814 for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) { 815 if ((char *)tn - tl->tl_offset == p) { 816 *tp = tn->tn_next[t]; 817 tn->tn_next[t] = NULL; 818 tn->tn_member[t] = 0; 819 mutex_exit(&tl->tl_lock); 820 return (p); 821 } 822 } 823 824 mutex_exit(&tl->tl_lock); 825 826 return (NULL); 827} 828 829boolean_t 830txg_list_member(txg_list_t *tl, void *p, uint64_t txg) 831{ 832 int t = txg & TXG_MASK; 833 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset); 834 835 return (tn->tn_member[t] != 0); 836} 837 838/* 839 * Walk a txg list -- only safe if you know it's not changing. 840 */ 841void * 842txg_list_head(txg_list_t *tl, uint64_t txg) 843{ 844 int t = txg & TXG_MASK; 845 txg_node_t *tn = tl->tl_head[t]; 846 847 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset); 848} 849 850void * 851txg_list_next(txg_list_t *tl, void *p, uint64_t txg) 852{ 853 int t = txg & TXG_MASK; 854 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset); 855 856 tn = tn->tn_next[t]; 857 858 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset); 859} 860