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