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 * Copyright (c) 2011, 2020 by Delphix. All rights reserved. 24 * Copyright (c) 2013 Steven Hartland. All rights reserved. 25 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 26 * Copyright 2016 Nexenta Systems, Inc. All rights reserved. 27 */ 28 29 #include <sys/dsl_pool.h> 30 #include <sys/dsl_dataset.h> 31 #include <sys/dsl_prop.h> 32 #include <sys/dsl_dir.h> 33 #include <sys/dsl_synctask.h> 34 #include <sys/dsl_scan.h> 35 #include <sys/dnode.h> 36 #include <sys/dmu_tx.h> 37 #include <sys/dmu_objset.h> 38 #include <sys/arc.h> 39 #include <sys/zap.h> 40 #include <sys/zio.h> 41 #include <sys/zfs_context.h> 42 #include <sys/fs/zfs.h> 43 #include <sys/zfs_znode.h> 44 #include <sys/spa_impl.h> 45 #include <sys/vdev_impl.h> 46 #include <sys/metaslab_impl.h> 47 #include <sys/bptree.h> 48 #include <sys/zfeature.h> 49 #include <sys/zil_impl.h> 50 #include <sys/dsl_userhold.h> 51 #include <sys/trace_zfs.h> 52 #include <sys/mmp.h> 53 54 /* 55 * ZFS Write Throttle 56 * ------------------ 57 * 58 * ZFS must limit the rate of incoming writes to the rate at which it is able 59 * to sync data modifications to the backend storage. Throttling by too much 60 * creates an artificial limit; throttling by too little can only be sustained 61 * for short periods and would lead to highly lumpy performance. On a per-pool 62 * basis, ZFS tracks the amount of modified (dirty) data. As operations change 63 * data, the amount of dirty data increases; as ZFS syncs out data, the amount 64 * of dirty data decreases. When the amount of dirty data exceeds a 65 * predetermined threshold further modifications are blocked until the amount 66 * of dirty data decreases (as data is synced out). 67 * 68 * The limit on dirty data is tunable, and should be adjusted according to 69 * both the IO capacity and available memory of the system. The larger the 70 * window, the more ZFS is able to aggregate and amortize metadata (and data) 71 * changes. However, memory is a limited resource, and allowing for more dirty 72 * data comes at the cost of keeping other useful data in memory (for example 73 * ZFS data cached by the ARC). 74 * 75 * Implementation 76 * 77 * As buffers are modified dsl_pool_willuse_space() increments both the per- 78 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of 79 * dirty space used; dsl_pool_dirty_space() decrements those values as data 80 * is synced out from dsl_pool_sync(). While only the poolwide value is 81 * relevant, the per-txg value is useful for debugging. The tunable 82 * zfs_dirty_data_max determines the dirty space limit. Once that value is 83 * exceeded, new writes are halted until space frees up. 84 * 85 * The zfs_dirty_data_sync_percent tunable dictates the threshold at which we 86 * ensure that there is a txg syncing (see the comment in txg.c for a full 87 * description of transaction group stages). 88 * 89 * The IO scheduler uses both the dirty space limit and current amount of 90 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS 91 * issues. See the comment in vdev_queue.c for details of the IO scheduler. 92 * 93 * The delay is also calculated based on the amount of dirty data. See the 94 * comment above dmu_tx_delay() for details. 95 */ 96 97 /* 98 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory, 99 * capped at zfs_dirty_data_max_max. It can also be overridden with a module 100 * parameter. 101 */ 102 unsigned long zfs_dirty_data_max = 0; 103 unsigned long zfs_dirty_data_max_max = 0; 104 int zfs_dirty_data_max_percent = 10; 105 int zfs_dirty_data_max_max_percent = 25; 106 107 /* 108 * zfs_wrlog_data_max, the upper limit of TX_WRITE log data. 109 * Once it is reached, write operation is blocked, 110 * until log data is cleared out after txg sync. 111 * It only counts TX_WRITE log with WR_COPIED or WR_NEED_COPY. 112 */ 113 unsigned long zfs_wrlog_data_max = 0; 114 115 /* 116 * If there's at least this much dirty data (as a percentage of 117 * zfs_dirty_data_max), push out a txg. This should be less than 118 * zfs_vdev_async_write_active_min_dirty_percent. 119 */ 120 int zfs_dirty_data_sync_percent = 20; 121 122 /* 123 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in 124 * and delay each transaction. 125 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent. 126 */ 127 int zfs_delay_min_dirty_percent = 60; 128 129 /* 130 * This controls how quickly the delay approaches infinity. 131 * Larger values cause it to delay more for a given amount of dirty data. 132 * Therefore larger values will cause there to be less dirty data for a 133 * given throughput. 134 * 135 * For the smoothest delay, this value should be about 1 billion divided 136 * by the maximum number of operations per second. This will smoothly 137 * handle between 10x and 1/10th this number. 138 * 139 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the 140 * multiply in dmu_tx_delay(). 141 */ 142 unsigned long zfs_delay_scale = 1000 * 1000 * 1000 / 2000; 143 144 /* 145 * This determines the number of threads used by the dp_sync_taskq. 146 */ 147 int zfs_sync_taskq_batch_pct = 75; 148 149 /* 150 * These tunables determine the behavior of how zil_itxg_clean() is 151 * called via zil_clean() in the context of spa_sync(). When an itxg 152 * list needs to be cleaned, TQ_NOSLEEP will be used when dispatching. 153 * If the dispatch fails, the call to zil_itxg_clean() will occur 154 * synchronously in the context of spa_sync(), which can negatively 155 * impact the performance of spa_sync() (e.g. in the case of the itxg 156 * list having a large number of itxs that needs to be cleaned). 157 * 158 * Thus, these tunables can be used to manipulate the behavior of the 159 * taskq used by zil_clean(); they determine the number of taskq entries 160 * that are pre-populated when the taskq is first created (via the 161 * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of 162 * taskq entries that are cached after an on-demand allocation (via the 163 * "zfs_zil_clean_taskq_maxalloc"). 164 * 165 * The idea being, we want to try reasonably hard to ensure there will 166 * already be a taskq entry pre-allocated by the time that it is needed 167 * by zil_clean(). This way, we can avoid the possibility of an 168 * on-demand allocation of a new taskq entry from failing, which would 169 * result in zil_itxg_clean() being called synchronously from zil_clean() 170 * (which can adversely affect performance of spa_sync()). 171 * 172 * Additionally, the number of threads used by the taskq can be 173 * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable. 174 */ 175 int zfs_zil_clean_taskq_nthr_pct = 100; 176 int zfs_zil_clean_taskq_minalloc = 1024; 177 int zfs_zil_clean_taskq_maxalloc = 1024 * 1024; 178 179 int 180 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp) 181 { 182 uint64_t obj; 183 int err; 184 185 err = zap_lookup(dp->dp_meta_objset, 186 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj, 187 name, sizeof (obj), 1, &obj); 188 if (err) 189 return (err); 190 191 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp)); 192 } 193 194 static dsl_pool_t * 195 dsl_pool_open_impl(spa_t *spa, uint64_t txg) 196 { 197 dsl_pool_t *dp; 198 blkptr_t *bp = spa_get_rootblkptr(spa); 199 200 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP); 201 dp->dp_spa = spa; 202 dp->dp_meta_rootbp = *bp; 203 rrw_init(&dp->dp_config_rwlock, B_TRUE); 204 txg_init(dp, txg); 205 mmp_init(spa); 206 207 txg_list_create(&dp->dp_dirty_datasets, spa, 208 offsetof(dsl_dataset_t, ds_dirty_link)); 209 txg_list_create(&dp->dp_dirty_zilogs, spa, 210 offsetof(zilog_t, zl_dirty_link)); 211 txg_list_create(&dp->dp_dirty_dirs, spa, 212 offsetof(dsl_dir_t, dd_dirty_link)); 213 txg_list_create(&dp->dp_sync_tasks, spa, 214 offsetof(dsl_sync_task_t, dst_node)); 215 txg_list_create(&dp->dp_early_sync_tasks, spa, 216 offsetof(dsl_sync_task_t, dst_node)); 217 218 dp->dp_sync_taskq = taskq_create("dp_sync_taskq", 219 zfs_sync_taskq_batch_pct, minclsyspri, 1, INT_MAX, 220 TASKQ_THREADS_CPU_PCT); 221 222 dp->dp_zil_clean_taskq = taskq_create("dp_zil_clean_taskq", 223 zfs_zil_clean_taskq_nthr_pct, minclsyspri, 224 zfs_zil_clean_taskq_minalloc, 225 zfs_zil_clean_taskq_maxalloc, 226 TASKQ_PREPOPULATE | TASKQ_THREADS_CPU_PCT); 227 228 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL); 229 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL); 230 231 aggsum_init(&dp->dp_wrlog_total, 0); 232 for (int i = 0; i < TXG_SIZE; i++) { 233 aggsum_init(&dp->dp_wrlog_pertxg[i], 0); 234 } 235 236 dp->dp_zrele_taskq = taskq_create("z_zrele", 100, defclsyspri, 237 boot_ncpus * 8, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC | 238 TASKQ_THREADS_CPU_PCT); 239 dp->dp_unlinked_drain_taskq = taskq_create("z_unlinked_drain", 240 100, defclsyspri, boot_ncpus, INT_MAX, 241 TASKQ_PREPOPULATE | TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT); 242 243 return (dp); 244 } 245 246 int 247 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp) 248 { 249 int err; 250 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg); 251 252 /* 253 * Initialize the caller's dsl_pool_t structure before we actually open 254 * the meta objset. This is done because a self-healing write zio may 255 * be issued as part of dmu_objset_open_impl() and the spa needs its 256 * dsl_pool_t initialized in order to handle the write. 257 */ 258 *dpp = dp; 259 260 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp, 261 &dp->dp_meta_objset); 262 if (err != 0) { 263 dsl_pool_close(dp); 264 *dpp = NULL; 265 } 266 267 return (err); 268 } 269 270 int 271 dsl_pool_open(dsl_pool_t *dp) 272 { 273 int err; 274 dsl_dir_t *dd; 275 dsl_dataset_t *ds; 276 uint64_t obj; 277 278 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 279 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 280 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1, 281 &dp->dp_root_dir_obj); 282 if (err) 283 goto out; 284 285 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, 286 NULL, dp, &dp->dp_root_dir); 287 if (err) 288 goto out; 289 290 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir); 291 if (err) 292 goto out; 293 294 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) { 295 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd); 296 if (err) 297 goto out; 298 err = dsl_dataset_hold_obj(dp, 299 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds); 300 if (err == 0) { 301 err = dsl_dataset_hold_obj(dp, 302 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp, 303 &dp->dp_origin_snap); 304 dsl_dataset_rele(ds, FTAG); 305 } 306 dsl_dir_rele(dd, dp); 307 if (err) 308 goto out; 309 } 310 311 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 312 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME, 313 &dp->dp_free_dir); 314 if (err) 315 goto out; 316 317 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 318 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj); 319 if (err) 320 goto out; 321 VERIFY0(bpobj_open(&dp->dp_free_bpobj, 322 dp->dp_meta_objset, obj)); 323 } 324 325 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS)) { 326 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 327 DMU_POOL_OBSOLETE_BPOBJ, sizeof (uint64_t), 1, &obj); 328 if (err == 0) { 329 VERIFY0(bpobj_open(&dp->dp_obsolete_bpobj, 330 dp->dp_meta_objset, obj)); 331 } else if (err == ENOENT) { 332 /* 333 * We might not have created the remap bpobj yet. 334 */ 335 err = 0; 336 } else { 337 goto out; 338 } 339 } 340 341 /* 342 * Note: errors ignored, because the these special dirs, used for 343 * space accounting, are only created on demand. 344 */ 345 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME, 346 &dp->dp_leak_dir); 347 348 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) { 349 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 350 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1, 351 &dp->dp_bptree_obj); 352 if (err != 0) 353 goto out; 354 } 355 356 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) { 357 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 358 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1, 359 &dp->dp_empty_bpobj); 360 if (err != 0) 361 goto out; 362 } 363 364 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 365 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1, 366 &dp->dp_tmp_userrefs_obj); 367 if (err == ENOENT) 368 err = 0; 369 if (err) 370 goto out; 371 372 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg); 373 374 out: 375 rrw_exit(&dp->dp_config_rwlock, FTAG); 376 return (err); 377 } 378 379 void 380 dsl_pool_close(dsl_pool_t *dp) 381 { 382 /* 383 * Drop our references from dsl_pool_open(). 384 * 385 * Since we held the origin_snap from "syncing" context (which 386 * includes pool-opening context), it actually only got a "ref" 387 * and not a hold, so just drop that here. 388 */ 389 if (dp->dp_origin_snap != NULL) 390 dsl_dataset_rele(dp->dp_origin_snap, dp); 391 if (dp->dp_mos_dir != NULL) 392 dsl_dir_rele(dp->dp_mos_dir, dp); 393 if (dp->dp_free_dir != NULL) 394 dsl_dir_rele(dp->dp_free_dir, dp); 395 if (dp->dp_leak_dir != NULL) 396 dsl_dir_rele(dp->dp_leak_dir, dp); 397 if (dp->dp_root_dir != NULL) 398 dsl_dir_rele(dp->dp_root_dir, dp); 399 400 bpobj_close(&dp->dp_free_bpobj); 401 bpobj_close(&dp->dp_obsolete_bpobj); 402 403 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */ 404 if (dp->dp_meta_objset != NULL) 405 dmu_objset_evict(dp->dp_meta_objset); 406 407 txg_list_destroy(&dp->dp_dirty_datasets); 408 txg_list_destroy(&dp->dp_dirty_zilogs); 409 txg_list_destroy(&dp->dp_sync_tasks); 410 txg_list_destroy(&dp->dp_early_sync_tasks); 411 txg_list_destroy(&dp->dp_dirty_dirs); 412 413 taskq_destroy(dp->dp_zil_clean_taskq); 414 taskq_destroy(dp->dp_sync_taskq); 415 416 /* 417 * We can't set retry to TRUE since we're explicitly specifying 418 * a spa to flush. This is good enough; any missed buffers for 419 * this spa won't cause trouble, and they'll eventually fall 420 * out of the ARC just like any other unused buffer. 421 */ 422 arc_flush(dp->dp_spa, FALSE); 423 424 mmp_fini(dp->dp_spa); 425 txg_fini(dp); 426 dsl_scan_fini(dp); 427 dmu_buf_user_evict_wait(); 428 429 rrw_destroy(&dp->dp_config_rwlock); 430 mutex_destroy(&dp->dp_lock); 431 cv_destroy(&dp->dp_spaceavail_cv); 432 433 ASSERT0(aggsum_value(&dp->dp_wrlog_total)); 434 aggsum_fini(&dp->dp_wrlog_total); 435 for (int i = 0; i < TXG_SIZE; i++) { 436 ASSERT0(aggsum_value(&dp->dp_wrlog_pertxg[i])); 437 aggsum_fini(&dp->dp_wrlog_pertxg[i]); 438 } 439 440 taskq_destroy(dp->dp_unlinked_drain_taskq); 441 taskq_destroy(dp->dp_zrele_taskq); 442 if (dp->dp_blkstats != NULL) { 443 mutex_destroy(&dp->dp_blkstats->zab_lock); 444 vmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t)); 445 } 446 kmem_free(dp, sizeof (dsl_pool_t)); 447 } 448 449 void 450 dsl_pool_create_obsolete_bpobj(dsl_pool_t *dp, dmu_tx_t *tx) 451 { 452 uint64_t obj; 453 /* 454 * Currently, we only create the obsolete_bpobj where there are 455 * indirect vdevs with referenced mappings. 456 */ 457 ASSERT(spa_feature_is_active(dp->dp_spa, SPA_FEATURE_DEVICE_REMOVAL)); 458 /* create and open the obsolete_bpobj */ 459 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx); 460 VERIFY0(bpobj_open(&dp->dp_obsolete_bpobj, dp->dp_meta_objset, obj)); 461 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 462 DMU_POOL_OBSOLETE_BPOBJ, sizeof (uint64_t), 1, &obj, tx)); 463 spa_feature_incr(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); 464 } 465 466 void 467 dsl_pool_destroy_obsolete_bpobj(dsl_pool_t *dp, dmu_tx_t *tx) 468 { 469 spa_feature_decr(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); 470 VERIFY0(zap_remove(dp->dp_meta_objset, 471 DMU_POOL_DIRECTORY_OBJECT, 472 DMU_POOL_OBSOLETE_BPOBJ, tx)); 473 bpobj_free(dp->dp_meta_objset, 474 dp->dp_obsolete_bpobj.bpo_object, tx); 475 bpobj_close(&dp->dp_obsolete_bpobj); 476 } 477 478 dsl_pool_t * 479 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, dsl_crypto_params_t *dcp, 480 uint64_t txg) 481 { 482 int err; 483 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg); 484 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg); 485 #ifdef _KERNEL 486 objset_t *os; 487 #else 488 objset_t *os __attribute__((unused)); 489 #endif 490 dsl_dataset_t *ds; 491 uint64_t obj; 492 493 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 494 495 /* create and open the MOS (meta-objset) */ 496 dp->dp_meta_objset = dmu_objset_create_impl(spa, 497 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx); 498 spa->spa_meta_objset = dp->dp_meta_objset; 499 500 /* create the pool directory */ 501 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 502 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx); 503 ASSERT0(err); 504 505 /* Initialize scan structures */ 506 VERIFY0(dsl_scan_init(dp, txg)); 507 508 /* create and open the root dir */ 509 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx); 510 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, 511 NULL, dp, &dp->dp_root_dir)); 512 513 /* create and open the meta-objset dir */ 514 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx); 515 VERIFY0(dsl_pool_open_special_dir(dp, 516 MOS_DIR_NAME, &dp->dp_mos_dir)); 517 518 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) { 519 /* create and open the free dir */ 520 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, 521 FREE_DIR_NAME, tx); 522 VERIFY0(dsl_pool_open_special_dir(dp, 523 FREE_DIR_NAME, &dp->dp_free_dir)); 524 525 /* create and open the free_bplist */ 526 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx); 527 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 528 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0); 529 VERIFY0(bpobj_open(&dp->dp_free_bpobj, 530 dp->dp_meta_objset, obj)); 531 } 532 533 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) 534 dsl_pool_create_origin(dp, tx); 535 536 /* 537 * Some features may be needed when creating the root dataset, so we 538 * create the feature objects here. 539 */ 540 if (spa_version(spa) >= SPA_VERSION_FEATURES) 541 spa_feature_create_zap_objects(spa, tx); 542 543 if (dcp != NULL && dcp->cp_crypt != ZIO_CRYPT_OFF && 544 dcp->cp_crypt != ZIO_CRYPT_INHERIT) 545 spa_feature_enable(spa, SPA_FEATURE_ENCRYPTION, tx); 546 547 /* create the root dataset */ 548 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, dcp, 0, tx); 549 550 /* create the root objset */ 551 VERIFY0(dsl_dataset_hold_obj_flags(dp, obj, 552 DS_HOLD_FLAG_DECRYPT, FTAG, &ds)); 553 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG); 554 os = dmu_objset_create_impl(dp->dp_spa, ds, 555 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx); 556 rrw_exit(&ds->ds_bp_rwlock, FTAG); 557 #ifdef _KERNEL 558 zfs_create_fs(os, kcred, zplprops, tx); 559 #endif 560 dsl_dataset_rele_flags(ds, DS_HOLD_FLAG_DECRYPT, FTAG); 561 562 dmu_tx_commit(tx); 563 564 rrw_exit(&dp->dp_config_rwlock, FTAG); 565 566 return (dp); 567 } 568 569 /* 570 * Account for the meta-objset space in its placeholder dsl_dir. 571 */ 572 void 573 dsl_pool_mos_diduse_space(dsl_pool_t *dp, 574 int64_t used, int64_t comp, int64_t uncomp) 575 { 576 ASSERT3U(comp, ==, uncomp); /* it's all metadata */ 577 mutex_enter(&dp->dp_lock); 578 dp->dp_mos_used_delta += used; 579 dp->dp_mos_compressed_delta += comp; 580 dp->dp_mos_uncompressed_delta += uncomp; 581 mutex_exit(&dp->dp_lock); 582 } 583 584 static void 585 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx) 586 { 587 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 588 dmu_objset_sync(dp->dp_meta_objset, zio, tx); 589 VERIFY0(zio_wait(zio)); 590 dmu_objset_sync_done(dp->dp_meta_objset, tx); 591 taskq_wait(dp->dp_sync_taskq); 592 multilist_destroy(&dp->dp_meta_objset->os_synced_dnodes); 593 594 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", ""); 595 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp); 596 } 597 598 static void 599 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta) 600 { 601 ASSERT(MUTEX_HELD(&dp->dp_lock)); 602 603 if (delta < 0) 604 ASSERT3U(-delta, <=, dp->dp_dirty_total); 605 606 dp->dp_dirty_total += delta; 607 608 /* 609 * Note: we signal even when increasing dp_dirty_total. 610 * This ensures forward progress -- each thread wakes the next waiter. 611 */ 612 if (dp->dp_dirty_total < zfs_dirty_data_max) 613 cv_signal(&dp->dp_spaceavail_cv); 614 } 615 616 void 617 dsl_pool_wrlog_count(dsl_pool_t *dp, int64_t size, uint64_t txg) 618 { 619 ASSERT3S(size, >=, 0); 620 621 aggsum_add(&dp->dp_wrlog_pertxg[txg & TXG_MASK], size); 622 aggsum_add(&dp->dp_wrlog_total, size); 623 624 /* Choose a value slightly bigger than min dirty sync bytes */ 625 uint64_t sync_min = 626 zfs_dirty_data_max * (zfs_dirty_data_sync_percent + 10) / 100; 627 if (aggsum_compare(&dp->dp_wrlog_pertxg[txg & TXG_MASK], sync_min) > 0) 628 txg_kick(dp, txg); 629 } 630 631 boolean_t 632 dsl_pool_wrlog_over_max(dsl_pool_t *dp) 633 { 634 return (aggsum_compare(&dp->dp_wrlog_total, zfs_wrlog_data_max) > 0); 635 } 636 637 static void 638 dsl_pool_wrlog_clear(dsl_pool_t *dp, uint64_t txg) 639 { 640 int64_t delta; 641 delta = -(int64_t)aggsum_value(&dp->dp_wrlog_pertxg[txg & TXG_MASK]); 642 aggsum_add(&dp->dp_wrlog_pertxg[txg & TXG_MASK], delta); 643 aggsum_add(&dp->dp_wrlog_total, delta); 644 } 645 646 #ifdef ZFS_DEBUG 647 static boolean_t 648 dsl_early_sync_task_verify(dsl_pool_t *dp, uint64_t txg) 649 { 650 spa_t *spa = dp->dp_spa; 651 vdev_t *rvd = spa->spa_root_vdev; 652 653 for (uint64_t c = 0; c < rvd->vdev_children; c++) { 654 vdev_t *vd = rvd->vdev_child[c]; 655 txg_list_t *tl = &vd->vdev_ms_list; 656 metaslab_t *ms; 657 658 for (ms = txg_list_head(tl, TXG_CLEAN(txg)); ms; 659 ms = txg_list_next(tl, ms, TXG_CLEAN(txg))) { 660 VERIFY(range_tree_is_empty(ms->ms_freeing)); 661 VERIFY(range_tree_is_empty(ms->ms_checkpointing)); 662 } 663 } 664 665 return (B_TRUE); 666 } 667 #endif 668 669 void 670 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg) 671 { 672 zio_t *zio; 673 dmu_tx_t *tx; 674 dsl_dir_t *dd; 675 dsl_dataset_t *ds; 676 objset_t *mos = dp->dp_meta_objset; 677 list_t synced_datasets; 678 679 list_create(&synced_datasets, sizeof (dsl_dataset_t), 680 offsetof(dsl_dataset_t, ds_synced_link)); 681 682 tx = dmu_tx_create_assigned(dp, txg); 683 684 /* 685 * Run all early sync tasks before writing out any dirty blocks. 686 * For more info on early sync tasks see block comment in 687 * dsl_early_sync_task(). 688 */ 689 if (!txg_list_empty(&dp->dp_early_sync_tasks, txg)) { 690 dsl_sync_task_t *dst; 691 692 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1); 693 while ((dst = 694 txg_list_remove(&dp->dp_early_sync_tasks, txg)) != NULL) { 695 ASSERT(dsl_early_sync_task_verify(dp, txg)); 696 dsl_sync_task_sync(dst, tx); 697 } 698 ASSERT(dsl_early_sync_task_verify(dp, txg)); 699 } 700 701 /* 702 * Write out all dirty blocks of dirty datasets. 703 */ 704 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 705 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) { 706 /* 707 * We must not sync any non-MOS datasets twice, because 708 * we may have taken a snapshot of them. However, we 709 * may sync newly-created datasets on pass 2. 710 */ 711 ASSERT(!list_link_active(&ds->ds_synced_link)); 712 list_insert_tail(&synced_datasets, ds); 713 dsl_dataset_sync(ds, zio, tx); 714 } 715 VERIFY0(zio_wait(zio)); 716 717 /* 718 * Update the long range free counter after 719 * we're done syncing user data 720 */ 721 mutex_enter(&dp->dp_lock); 722 ASSERT(spa_sync_pass(dp->dp_spa) == 1 || 723 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0); 724 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0; 725 mutex_exit(&dp->dp_lock); 726 727 /* 728 * After the data blocks have been written (ensured by the zio_wait() 729 * above), update the user/group/project space accounting. This happens 730 * in tasks dispatched to dp_sync_taskq, so wait for them before 731 * continuing. 732 */ 733 for (ds = list_head(&synced_datasets); ds != NULL; 734 ds = list_next(&synced_datasets, ds)) { 735 dmu_objset_sync_done(ds->ds_objset, tx); 736 } 737 taskq_wait(dp->dp_sync_taskq); 738 739 /* 740 * Sync the datasets again to push out the changes due to 741 * userspace updates. This must be done before we process the 742 * sync tasks, so that any snapshots will have the correct 743 * user accounting information (and we won't get confused 744 * about which blocks are part of the snapshot). 745 */ 746 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 747 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) { 748 objset_t *os = ds->ds_objset; 749 750 ASSERT(list_link_active(&ds->ds_synced_link)); 751 dmu_buf_rele(ds->ds_dbuf, ds); 752 dsl_dataset_sync(ds, zio, tx); 753 754 /* 755 * Release any key mappings created by calls to 756 * dsl_dataset_dirty() from the userquota accounting 757 * code paths. 758 */ 759 if (os->os_encrypted && !os->os_raw_receive && 760 !os->os_next_write_raw[txg & TXG_MASK]) { 761 ASSERT3P(ds->ds_key_mapping, !=, NULL); 762 key_mapping_rele(dp->dp_spa, ds->ds_key_mapping, ds); 763 } 764 } 765 VERIFY0(zio_wait(zio)); 766 767 /* 768 * Now that the datasets have been completely synced, we can 769 * clean up our in-memory structures accumulated while syncing: 770 * 771 * - move dead blocks from the pending deadlist and livelists 772 * to the on-disk versions 773 * - release hold from dsl_dataset_dirty() 774 * - release key mapping hold from dsl_dataset_dirty() 775 */ 776 while ((ds = list_remove_head(&synced_datasets)) != NULL) { 777 objset_t *os = ds->ds_objset; 778 779 if (os->os_encrypted && !os->os_raw_receive && 780 !os->os_next_write_raw[txg & TXG_MASK]) { 781 ASSERT3P(ds->ds_key_mapping, !=, NULL); 782 key_mapping_rele(dp->dp_spa, ds->ds_key_mapping, ds); 783 } 784 785 dsl_dataset_sync_done(ds, tx); 786 } 787 788 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) { 789 dsl_dir_sync(dd, tx); 790 } 791 792 /* 793 * The MOS's space is accounted for in the pool/$MOS 794 * (dp_mos_dir). We can't modify the mos while we're syncing 795 * it, so we remember the deltas and apply them here. 796 */ 797 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 || 798 dp->dp_mos_uncompressed_delta != 0) { 799 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD, 800 dp->dp_mos_used_delta, 801 dp->dp_mos_compressed_delta, 802 dp->dp_mos_uncompressed_delta, tx); 803 dp->dp_mos_used_delta = 0; 804 dp->dp_mos_compressed_delta = 0; 805 dp->dp_mos_uncompressed_delta = 0; 806 } 807 808 if (dmu_objset_is_dirty(mos, txg)) { 809 dsl_pool_sync_mos(dp, tx); 810 } 811 812 /* 813 * We have written all of the accounted dirty data, so our 814 * dp_space_towrite should now be zero. However, some seldom-used 815 * code paths do not adhere to this (e.g. dbuf_undirty()). Shore up 816 * the accounting of any dirtied space now. 817 * 818 * Note that, besides any dirty data from datasets, the amount of 819 * dirty data in the MOS is also accounted by the pool. Therefore, 820 * we want to do this cleanup after dsl_pool_sync_mos() so we don't 821 * attempt to update the accounting for the same dirty data twice. 822 * (i.e. at this point we only update the accounting for the space 823 * that we know that we "leaked"). 824 */ 825 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg); 826 827 /* 828 * If we modify a dataset in the same txg that we want to destroy it, 829 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it. 830 * dsl_dir_destroy_check() will fail if there are unexpected holds. 831 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf 832 * and clearing the hold on it) before we process the sync_tasks. 833 * The MOS data dirtied by the sync_tasks will be synced on the next 834 * pass. 835 */ 836 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) { 837 dsl_sync_task_t *dst; 838 /* 839 * No more sync tasks should have been added while we 840 * were syncing. 841 */ 842 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1); 843 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL) 844 dsl_sync_task_sync(dst, tx); 845 } 846 847 dmu_tx_commit(tx); 848 849 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg); 850 } 851 852 void 853 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg) 854 { 855 zilog_t *zilog; 856 857 while ((zilog = txg_list_head(&dp->dp_dirty_zilogs, txg))) { 858 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 859 /* 860 * We don't remove the zilog from the dp_dirty_zilogs 861 * list until after we've cleaned it. This ensures that 862 * callers of zilog_is_dirty() receive an accurate 863 * answer when they are racing with the spa sync thread. 864 */ 865 zil_clean(zilog, txg); 866 (void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg); 867 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg)); 868 dmu_buf_rele(ds->ds_dbuf, zilog); 869 } 870 871 dsl_pool_wrlog_clear(dp, txg); 872 873 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg)); 874 } 875 876 /* 877 * TRUE if the current thread is the tx_sync_thread or if we 878 * are being called from SPA context during pool initialization. 879 */ 880 int 881 dsl_pool_sync_context(dsl_pool_t *dp) 882 { 883 return (curthread == dp->dp_tx.tx_sync_thread || 884 spa_is_initializing(dp->dp_spa) || 885 taskq_member(dp->dp_sync_taskq, curthread)); 886 } 887 888 /* 889 * This function returns the amount of allocatable space in the pool 890 * minus whatever space is currently reserved by ZFS for specific 891 * purposes. Specifically: 892 * 893 * 1] Any reserved SLOP space 894 * 2] Any space used by the checkpoint 895 * 3] Any space used for deferred frees 896 * 897 * The latter 2 are especially important because they are needed to 898 * rectify the SPA's and DMU's different understanding of how much space 899 * is used. Now the DMU is aware of that extra space tracked by the SPA 900 * without having to maintain a separate special dir (e.g similar to 901 * $MOS, $FREEING, and $LEAKED). 902 * 903 * Note: By deferred frees here, we mean the frees that were deferred 904 * in spa_sync() after sync pass 1 (spa_deferred_bpobj), and not the 905 * segments placed in ms_defer trees during metaslab_sync_done(). 906 */ 907 uint64_t 908 dsl_pool_adjustedsize(dsl_pool_t *dp, zfs_space_check_t slop_policy) 909 { 910 spa_t *spa = dp->dp_spa; 911 uint64_t space, resv, adjustedsize; 912 uint64_t spa_deferred_frees = 913 spa->spa_deferred_bpobj.bpo_phys->bpo_bytes; 914 915 space = spa_get_dspace(spa) 916 - spa_get_checkpoint_space(spa) - spa_deferred_frees; 917 resv = spa_get_slop_space(spa); 918 919 switch (slop_policy) { 920 case ZFS_SPACE_CHECK_NORMAL: 921 break; 922 case ZFS_SPACE_CHECK_RESERVED: 923 resv >>= 1; 924 break; 925 case ZFS_SPACE_CHECK_EXTRA_RESERVED: 926 resv >>= 2; 927 break; 928 case ZFS_SPACE_CHECK_NONE: 929 resv = 0; 930 break; 931 default: 932 panic("invalid slop policy value: %d", slop_policy); 933 break; 934 } 935 adjustedsize = (space >= resv) ? (space - resv) : 0; 936 937 return (adjustedsize); 938 } 939 940 uint64_t 941 dsl_pool_unreserved_space(dsl_pool_t *dp, zfs_space_check_t slop_policy) 942 { 943 uint64_t poolsize = dsl_pool_adjustedsize(dp, slop_policy); 944 uint64_t deferred = 945 metaslab_class_get_deferred(spa_normal_class(dp->dp_spa)); 946 uint64_t quota = (poolsize >= deferred) ? (poolsize - deferred) : 0; 947 return (quota); 948 } 949 950 boolean_t 951 dsl_pool_need_dirty_delay(dsl_pool_t *dp) 952 { 953 uint64_t delay_min_bytes = 954 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100; 955 956 mutex_enter(&dp->dp_lock); 957 uint64_t dirty = dp->dp_dirty_total; 958 mutex_exit(&dp->dp_lock); 959 960 return (dirty > delay_min_bytes); 961 } 962 963 static boolean_t 964 dsl_pool_need_dirty_sync(dsl_pool_t *dp, uint64_t txg) 965 { 966 ASSERT(MUTEX_HELD(&dp->dp_lock)); 967 968 uint64_t dirty_min_bytes = 969 zfs_dirty_data_max * zfs_dirty_data_sync_percent / 100; 970 uint64_t dirty = dp->dp_dirty_pertxg[txg & TXG_MASK]; 971 972 return (dirty > dirty_min_bytes); 973 } 974 975 void 976 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx) 977 { 978 if (space > 0) { 979 mutex_enter(&dp->dp_lock); 980 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space; 981 dsl_pool_dirty_delta(dp, space); 982 boolean_t needsync = !dmu_tx_is_syncing(tx) && 983 dsl_pool_need_dirty_sync(dp, tx->tx_txg); 984 mutex_exit(&dp->dp_lock); 985 986 if (needsync) 987 txg_kick(dp, tx->tx_txg); 988 } 989 } 990 991 void 992 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg) 993 { 994 ASSERT3S(space, >=, 0); 995 if (space == 0) 996 return; 997 998 mutex_enter(&dp->dp_lock); 999 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) { 1000 /* XXX writing something we didn't dirty? */ 1001 space = dp->dp_dirty_pertxg[txg & TXG_MASK]; 1002 } 1003 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space); 1004 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space; 1005 ASSERT3U(dp->dp_dirty_total, >=, space); 1006 dsl_pool_dirty_delta(dp, -space); 1007 mutex_exit(&dp->dp_lock); 1008 } 1009 1010 /* ARGSUSED */ 1011 static int 1012 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 1013 { 1014 dmu_tx_t *tx = arg; 1015 dsl_dataset_t *ds, *prev = NULL; 1016 int err; 1017 1018 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 1019 if (err) 1020 return (err); 1021 1022 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) { 1023 err = dsl_dataset_hold_obj(dp, 1024 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev); 1025 if (err) { 1026 dsl_dataset_rele(ds, FTAG); 1027 return (err); 1028 } 1029 1030 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) 1031 break; 1032 dsl_dataset_rele(ds, FTAG); 1033 ds = prev; 1034 prev = NULL; 1035 } 1036 1037 if (prev == NULL) { 1038 prev = dp->dp_origin_snap; 1039 1040 /* 1041 * The $ORIGIN can't have any data, or the accounting 1042 * will be wrong. 1043 */ 1044 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG); 1045 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth); 1046 rrw_exit(&ds->ds_bp_rwlock, FTAG); 1047 1048 /* The origin doesn't get attached to itself */ 1049 if (ds->ds_object == prev->ds_object) { 1050 dsl_dataset_rele(ds, FTAG); 1051 return (0); 1052 } 1053 1054 dmu_buf_will_dirty(ds->ds_dbuf, tx); 1055 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object; 1056 dsl_dataset_phys(ds)->ds_prev_snap_txg = 1057 dsl_dataset_phys(prev)->ds_creation_txg; 1058 1059 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx); 1060 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object; 1061 1062 dmu_buf_will_dirty(prev->ds_dbuf, tx); 1063 dsl_dataset_phys(prev)->ds_num_children++; 1064 1065 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) { 1066 ASSERT(ds->ds_prev == NULL); 1067 VERIFY0(dsl_dataset_hold_obj(dp, 1068 dsl_dataset_phys(ds)->ds_prev_snap_obj, 1069 ds, &ds->ds_prev)); 1070 } 1071 } 1072 1073 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object); 1074 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object); 1075 1076 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) { 1077 dmu_buf_will_dirty(prev->ds_dbuf, tx); 1078 dsl_dataset_phys(prev)->ds_next_clones_obj = 1079 zap_create(dp->dp_meta_objset, 1080 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx); 1081 } 1082 VERIFY0(zap_add_int(dp->dp_meta_objset, 1083 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx)); 1084 1085 dsl_dataset_rele(ds, FTAG); 1086 if (prev != dp->dp_origin_snap) 1087 dsl_dataset_rele(prev, FTAG); 1088 return (0); 1089 } 1090 1091 void 1092 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx) 1093 { 1094 ASSERT(dmu_tx_is_syncing(tx)); 1095 ASSERT(dp->dp_origin_snap != NULL); 1096 1097 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb, 1098 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE)); 1099 } 1100 1101 /* ARGSUSED */ 1102 static int 1103 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg) 1104 { 1105 dmu_tx_t *tx = arg; 1106 objset_t *mos = dp->dp_meta_objset; 1107 1108 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) { 1109 dsl_dataset_t *origin; 1110 1111 VERIFY0(dsl_dataset_hold_obj(dp, 1112 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin)); 1113 1114 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) { 1115 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx); 1116 dsl_dir_phys(origin->ds_dir)->dd_clones = 1117 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE, 1118 0, tx); 1119 } 1120 1121 VERIFY0(zap_add_int(dp->dp_meta_objset, 1122 dsl_dir_phys(origin->ds_dir)->dd_clones, 1123 ds->ds_object, tx)); 1124 1125 dsl_dataset_rele(origin, FTAG); 1126 } 1127 return (0); 1128 } 1129 1130 void 1131 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx) 1132 { 1133 uint64_t obj; 1134 1135 ASSERT(dmu_tx_is_syncing(tx)); 1136 1137 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx); 1138 VERIFY0(dsl_pool_open_special_dir(dp, 1139 FREE_DIR_NAME, &dp->dp_free_dir)); 1140 1141 /* 1142 * We can't use bpobj_alloc(), because spa_version() still 1143 * returns the old version, and we need a new-version bpobj with 1144 * subobj support. So call dmu_object_alloc() directly. 1145 */ 1146 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ, 1147 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx); 1148 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 1149 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx)); 1150 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj)); 1151 1152 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 1153 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE)); 1154 } 1155 1156 void 1157 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx) 1158 { 1159 uint64_t dsobj; 1160 dsl_dataset_t *ds; 1161 1162 ASSERT(dmu_tx_is_syncing(tx)); 1163 ASSERT(dp->dp_origin_snap == NULL); 1164 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER)); 1165 1166 /* create the origin dir, ds, & snap-ds */ 1167 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME, 1168 NULL, 0, kcred, NULL, tx); 1169 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 1170 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx); 1171 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj, 1172 dp, &dp->dp_origin_snap)); 1173 dsl_dataset_rele(ds, FTAG); 1174 } 1175 1176 taskq_t * 1177 dsl_pool_zrele_taskq(dsl_pool_t *dp) 1178 { 1179 return (dp->dp_zrele_taskq); 1180 } 1181 1182 taskq_t * 1183 dsl_pool_unlinked_drain_taskq(dsl_pool_t *dp) 1184 { 1185 return (dp->dp_unlinked_drain_taskq); 1186 } 1187 1188 /* 1189 * Walk through the pool-wide zap object of temporary snapshot user holds 1190 * and release them. 1191 */ 1192 void 1193 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp) 1194 { 1195 zap_attribute_t za; 1196 zap_cursor_t zc; 1197 objset_t *mos = dp->dp_meta_objset; 1198 uint64_t zapobj = dp->dp_tmp_userrefs_obj; 1199 nvlist_t *holds; 1200 1201 if (zapobj == 0) 1202 return; 1203 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS); 1204 1205 holds = fnvlist_alloc(); 1206 1207 for (zap_cursor_init(&zc, mos, zapobj); 1208 zap_cursor_retrieve(&zc, &za) == 0; 1209 zap_cursor_advance(&zc)) { 1210 char *htag; 1211 nvlist_t *tags; 1212 1213 htag = strchr(za.za_name, '-'); 1214 *htag = '\0'; 1215 ++htag; 1216 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) { 1217 tags = fnvlist_alloc(); 1218 fnvlist_add_boolean(tags, htag); 1219 fnvlist_add_nvlist(holds, za.za_name, tags); 1220 fnvlist_free(tags); 1221 } else { 1222 fnvlist_add_boolean(tags, htag); 1223 } 1224 } 1225 dsl_dataset_user_release_tmp(dp, holds); 1226 fnvlist_free(holds); 1227 zap_cursor_fini(&zc); 1228 } 1229 1230 /* 1231 * Create the pool-wide zap object for storing temporary snapshot holds. 1232 */ 1233 static void 1234 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx) 1235 { 1236 objset_t *mos = dp->dp_meta_objset; 1237 1238 ASSERT(dp->dp_tmp_userrefs_obj == 0); 1239 ASSERT(dmu_tx_is_syncing(tx)); 1240 1241 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS, 1242 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx); 1243 } 1244 1245 static int 1246 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj, 1247 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding) 1248 { 1249 objset_t *mos = dp->dp_meta_objset; 1250 uint64_t zapobj = dp->dp_tmp_userrefs_obj; 1251 char *name; 1252 int error; 1253 1254 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS); 1255 ASSERT(dmu_tx_is_syncing(tx)); 1256 1257 /* 1258 * If the pool was created prior to SPA_VERSION_USERREFS, the 1259 * zap object for temporary holds might not exist yet. 1260 */ 1261 if (zapobj == 0) { 1262 if (holding) { 1263 dsl_pool_user_hold_create_obj(dp, tx); 1264 zapobj = dp->dp_tmp_userrefs_obj; 1265 } else { 1266 return (SET_ERROR(ENOENT)); 1267 } 1268 } 1269 1270 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag); 1271 if (holding) 1272 error = zap_add(mos, zapobj, name, 8, 1, &now, tx); 1273 else 1274 error = zap_remove(mos, zapobj, name, tx); 1275 kmem_strfree(name); 1276 1277 return (error); 1278 } 1279 1280 /* 1281 * Add a temporary hold for the given dataset object and tag. 1282 */ 1283 int 1284 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag, 1285 uint64_t now, dmu_tx_t *tx) 1286 { 1287 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE)); 1288 } 1289 1290 /* 1291 * Release a temporary hold for the given dataset object and tag. 1292 */ 1293 int 1294 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag, 1295 dmu_tx_t *tx) 1296 { 1297 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0, 1298 tx, B_FALSE)); 1299 } 1300 1301 /* 1302 * DSL Pool Configuration Lock 1303 * 1304 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset 1305 * creation / destruction / rename / property setting). It must be held for 1306 * read to hold a dataset or dsl_dir. I.e. you must call 1307 * dsl_pool_config_enter() or dsl_pool_hold() before calling 1308 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock 1309 * must be held continuously until all datasets and dsl_dirs are released. 1310 * 1311 * The only exception to this rule is that if a "long hold" is placed on 1312 * a dataset, then the dp_config_rwlock may be dropped while the dataset 1313 * is still held. The long hold will prevent the dataset from being 1314 * destroyed -- the destroy will fail with EBUSY. A long hold can be 1315 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset 1316 * (by calling dsl_{dataset,objset}_{try}own{_obj}). 1317 * 1318 * Legitimate long-holders (including owners) should be long-running, cancelable 1319 * tasks that should cause "zfs destroy" to fail. This includes DMU 1320 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open), 1321 * "zfs send", and "zfs diff". There are several other long-holders whose 1322 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()). 1323 * 1324 * The usual formula for long-holding would be: 1325 * dsl_pool_hold() 1326 * dsl_dataset_hold() 1327 * ... perform checks ... 1328 * dsl_dataset_long_hold() 1329 * dsl_pool_rele() 1330 * ... perform long-running task ... 1331 * dsl_dataset_long_rele() 1332 * dsl_dataset_rele() 1333 * 1334 * Note that when the long hold is released, the dataset is still held but 1335 * the pool is not held. The dataset may change arbitrarily during this time 1336 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the 1337 * dataset except release it. 1338 * 1339 * Operations generally fall somewhere into the following taxonomy: 1340 * 1341 * Read-Only Modifying 1342 * 1343 * Dataset Layer / MOS zfs get zfs destroy 1344 * 1345 * Individual Dataset read() write() 1346 * 1347 * 1348 * Dataset Layer Operations 1349 * 1350 * Modifying operations should generally use dsl_sync_task(). The synctask 1351 * infrastructure enforces proper locking strategy with respect to the 1352 * dp_config_rwlock. See the comment above dsl_sync_task() for details. 1353 * 1354 * Read-only operations will manually hold the pool, then the dataset, obtain 1355 * information from the dataset, then release the pool and dataset. 1356 * dmu_objset_{hold,rele}() are convenience routines that also do the pool 1357 * hold/rele. 1358 * 1359 * 1360 * Operations On Individual Datasets 1361 * 1362 * Objects _within_ an objset should only be modified by the current 'owner' 1363 * of the objset to prevent incorrect concurrent modification. Thus, use 1364 * {dmu_objset,dsl_dataset}_own to mark some entity as the current owner, 1365 * and fail with EBUSY if there is already an owner. The owner can then 1366 * implement its own locking strategy, independent of the dataset layer's 1367 * locking infrastructure. 1368 * (E.g., the ZPL has its own set of locks to control concurrency. A regular 1369 * vnop will not reach into the dataset layer). 1370 * 1371 * Ideally, objects would also only be read by the objset’s owner, so that we 1372 * don’t observe state mid-modification. 1373 * (E.g. the ZPL is creating a new object and linking it into a directory; if 1374 * you don’t coordinate with the ZPL to hold ZPL-level locks, you could see an 1375 * intermediate state. The ioctl level violates this but in pretty benign 1376 * ways, e.g. reading the zpl props object.) 1377 */ 1378 1379 int 1380 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp) 1381 { 1382 spa_t *spa; 1383 int error; 1384 1385 error = spa_open(name, &spa, tag); 1386 if (error == 0) { 1387 *dp = spa_get_dsl(spa); 1388 dsl_pool_config_enter(*dp, tag); 1389 } 1390 return (error); 1391 } 1392 1393 void 1394 dsl_pool_rele(dsl_pool_t *dp, void *tag) 1395 { 1396 dsl_pool_config_exit(dp, tag); 1397 spa_close(dp->dp_spa, tag); 1398 } 1399 1400 void 1401 dsl_pool_config_enter(dsl_pool_t *dp, void *tag) 1402 { 1403 /* 1404 * We use a "reentrant" reader-writer lock, but not reentrantly. 1405 * 1406 * The rrwlock can (with the track_all flag) track all reading threads, 1407 * which is very useful for debugging which code path failed to release 1408 * the lock, and for verifying that the *current* thread does hold 1409 * the lock. 1410 * 1411 * (Unlike a rwlock, which knows that N threads hold it for 1412 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE 1413 * if any thread holds it for read, even if this thread doesn't). 1414 */ 1415 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER)); 1416 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag); 1417 } 1418 1419 void 1420 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag) 1421 { 1422 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER)); 1423 rrw_enter_read_prio(&dp->dp_config_rwlock, tag); 1424 } 1425 1426 void 1427 dsl_pool_config_exit(dsl_pool_t *dp, void *tag) 1428 { 1429 rrw_exit(&dp->dp_config_rwlock, tag); 1430 } 1431 1432 boolean_t 1433 dsl_pool_config_held(dsl_pool_t *dp) 1434 { 1435 return (RRW_LOCK_HELD(&dp->dp_config_rwlock)); 1436 } 1437 1438 boolean_t 1439 dsl_pool_config_held_writer(dsl_pool_t *dp) 1440 { 1441 return (RRW_WRITE_HELD(&dp->dp_config_rwlock)); 1442 } 1443 1444 EXPORT_SYMBOL(dsl_pool_config_enter); 1445 EXPORT_SYMBOL(dsl_pool_config_exit); 1446 1447 /* BEGIN CSTYLED */ 1448 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */ 1449 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max_percent, INT, ZMOD_RD, 1450 "Max percent of RAM allowed to be dirty"); 1451 1452 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */ 1453 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max_max_percent, INT, ZMOD_RD, 1454 "zfs_dirty_data_max upper bound as % of RAM"); 1455 1456 ZFS_MODULE_PARAM(zfs, zfs_, delay_min_dirty_percent, INT, ZMOD_RW, 1457 "Transaction delay threshold"); 1458 1459 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max, ULONG, ZMOD_RW, 1460 "Determines the dirty space limit"); 1461 1462 ZFS_MODULE_PARAM(zfs, zfs_, wrlog_data_max, ULONG, ZMOD_RW, 1463 "The size limit of write-transaction zil log data"); 1464 1465 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */ 1466 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max_max, ULONG, ZMOD_RD, 1467 "zfs_dirty_data_max upper bound in bytes"); 1468 1469 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_sync_percent, INT, ZMOD_RW, 1470 "Dirty data txg sync threshold as a percentage of zfs_dirty_data_max"); 1471 1472 ZFS_MODULE_PARAM(zfs, zfs_, delay_scale, ULONG, ZMOD_RW, 1473 "How quickly delay approaches infinity"); 1474 1475 ZFS_MODULE_PARAM(zfs, zfs_, sync_taskq_batch_pct, INT, ZMOD_RW, 1476 "Max percent of CPUs that are used to sync dirty data"); 1477 1478 ZFS_MODULE_PARAM(zfs_zil, zfs_zil_, clean_taskq_nthr_pct, INT, ZMOD_RW, 1479 "Max percent of CPUs that are used per dp_sync_taskq"); 1480 1481 ZFS_MODULE_PARAM(zfs_zil, zfs_zil_, clean_taskq_minalloc, INT, ZMOD_RW, 1482 "Number of taskq entries that are pre-populated"); 1483 1484 ZFS_MODULE_PARAM(zfs_zil, zfs_zil_, clean_taskq_maxalloc, INT, ZMOD_RW, 1485 "Max number of taskq entries that are cached"); 1486 /* END CSTYLED */ 1487