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 https://opensource.org/licenses/CDDL-1.0. 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 uint_t zfs_dirty_data_max_percent = 10; 105 uint_t zfs_dirty_data_max_max_percent = 25; 106 107 /* 108 * The upper limit of TX_WRITE log data. Write operations are throttled 109 * when approaching the limit until log data is cleared out after txg sync. 110 * It only counts TX_WRITE log with WR_COPIED or WR_NEED_COPY. 111 */ 112 unsigned long zfs_wrlog_data_max = 0; 113 114 /* 115 * If there's at least this much dirty data (as a percentage of 116 * zfs_dirty_data_max), push out a txg. This should be less than 117 * zfs_vdev_async_write_active_min_dirty_percent. 118 */ 119 static uint_t zfs_dirty_data_sync_percent = 20; 120 121 /* 122 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in 123 * and delay each transaction. 124 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent. 125 */ 126 uint_t zfs_delay_min_dirty_percent = 60; 127 128 /* 129 * This controls how quickly the delay approaches infinity. 130 * Larger values cause it to delay more for a given amount of dirty data. 131 * Therefore larger values will cause there to be less dirty data for a 132 * given throughput. 133 * 134 * For the smoothest delay, this value should be about 1 billion divided 135 * by the maximum number of operations per second. This will smoothly 136 * handle between 10x and 1/10th this number. 137 * 138 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the 139 * multiply in dmu_tx_delay(). 140 */ 141 unsigned long zfs_delay_scale = 1000 * 1000 * 1000 / 2000; 142 143 /* 144 * This determines the number of threads used by the dp_sync_taskq. 145 */ 146 static int zfs_sync_taskq_batch_pct = 75; 147 148 /* 149 * These tunables determine the behavior of how zil_itxg_clean() is 150 * called via zil_clean() in the context of spa_sync(). When an itxg 151 * list needs to be cleaned, TQ_NOSLEEP will be used when dispatching. 152 * If the dispatch fails, the call to zil_itxg_clean() will occur 153 * synchronously in the context of spa_sync(), which can negatively 154 * impact the performance of spa_sync() (e.g. in the case of the itxg 155 * list having a large number of itxs that needs to be cleaned). 156 * 157 * Thus, these tunables can be used to manipulate the behavior of the 158 * taskq used by zil_clean(); they determine the number of taskq entries 159 * that are pre-populated when the taskq is first created (via the 160 * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of 161 * taskq entries that are cached after an on-demand allocation (via the 162 * "zfs_zil_clean_taskq_maxalloc"). 163 * 164 * The idea being, we want to try reasonably hard to ensure there will 165 * already be a taskq entry pre-allocated by the time that it is needed 166 * by zil_clean(). This way, we can avoid the possibility of an 167 * on-demand allocation of a new taskq entry from failing, which would 168 * result in zil_itxg_clean() being called synchronously from zil_clean() 169 * (which can adversely affect performance of spa_sync()). 170 * 171 * Additionally, the number of threads used by the taskq can be 172 * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable. 173 */ 174 static int zfs_zil_clean_taskq_nthr_pct = 100; 175 static int zfs_zil_clean_taskq_minalloc = 1024; 176 static int zfs_zil_clean_taskq_maxalloc = 1024 * 1024; 177 178 int 179 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp) 180 { 181 uint64_t obj; 182 int err; 183 184 err = zap_lookup(dp->dp_meta_objset, 185 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj, 186 name, sizeof (obj), 1, &obj); 187 if (err) 188 return (err); 189 190 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp)); 191 } 192 193 static dsl_pool_t * 194 dsl_pool_open_impl(spa_t *spa, uint64_t txg) 195 { 196 dsl_pool_t *dp; 197 blkptr_t *bp = spa_get_rootblkptr(spa); 198 199 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP); 200 dp->dp_spa = spa; 201 dp->dp_meta_rootbp = *bp; 202 rrw_init(&dp->dp_config_rwlock, B_TRUE); 203 txg_init(dp, txg); 204 mmp_init(spa); 205 206 txg_list_create(&dp->dp_dirty_datasets, spa, 207 offsetof(dsl_dataset_t, ds_dirty_link)); 208 txg_list_create(&dp->dp_dirty_zilogs, spa, 209 offsetof(zilog_t, zl_dirty_link)); 210 txg_list_create(&dp->dp_dirty_dirs, spa, 211 offsetof(dsl_dir_t, dd_dirty_link)); 212 txg_list_create(&dp->dp_sync_tasks, spa, 213 offsetof(dsl_sync_task_t, dst_node)); 214 txg_list_create(&dp->dp_early_sync_tasks, spa, 215 offsetof(dsl_sync_task_t, dst_node)); 216 217 dp->dp_sync_taskq = taskq_create("dp_sync_taskq", 218 zfs_sync_taskq_batch_pct, minclsyspri, 1, INT_MAX, 219 TASKQ_THREADS_CPU_PCT); 220 221 dp->dp_zil_clean_taskq = taskq_create("dp_zil_clean_taskq", 222 zfs_zil_clean_taskq_nthr_pct, minclsyspri, 223 zfs_zil_clean_taskq_minalloc, 224 zfs_zil_clean_taskq_maxalloc, 225 TASKQ_PREPOPULATE | TASKQ_THREADS_CPU_PCT); 226 227 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL); 228 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL); 229 230 aggsum_init(&dp->dp_wrlog_total, 0); 231 for (int i = 0; i < TXG_SIZE; i++) { 232 aggsum_init(&dp->dp_wrlog_pertxg[i], 0); 233 } 234 235 dp->dp_zrele_taskq = taskq_create("z_zrele", 100, defclsyspri, 236 boot_ncpus * 8, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC | 237 TASKQ_THREADS_CPU_PCT); 238 dp->dp_unlinked_drain_taskq = taskq_create("z_unlinked_drain", 239 100, defclsyspri, boot_ncpus, INT_MAX, 240 TASKQ_PREPOPULATE | TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT); 241 242 return (dp); 243 } 244 245 int 246 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp) 247 { 248 int err; 249 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg); 250 251 /* 252 * Initialize the caller's dsl_pool_t structure before we actually open 253 * the meta objset. This is done because a self-healing write zio may 254 * be issued as part of dmu_objset_open_impl() and the spa needs its 255 * dsl_pool_t initialized in order to handle the write. 256 */ 257 *dpp = dp; 258 259 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp, 260 &dp->dp_meta_objset); 261 if (err != 0) { 262 dsl_pool_close(dp); 263 *dpp = NULL; 264 } 265 266 return (err); 267 } 268 269 int 270 dsl_pool_open(dsl_pool_t *dp) 271 { 272 int err; 273 dsl_dir_t *dd; 274 dsl_dataset_t *ds; 275 uint64_t obj; 276 277 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 278 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 279 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1, 280 &dp->dp_root_dir_obj); 281 if (err) 282 goto out; 283 284 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, 285 NULL, dp, &dp->dp_root_dir); 286 if (err) 287 goto out; 288 289 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir); 290 if (err) 291 goto out; 292 293 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) { 294 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd); 295 if (err) 296 goto out; 297 err = dsl_dataset_hold_obj(dp, 298 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds); 299 if (err == 0) { 300 err = dsl_dataset_hold_obj(dp, 301 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp, 302 &dp->dp_origin_snap); 303 dsl_dataset_rele(ds, FTAG); 304 } 305 dsl_dir_rele(dd, dp); 306 if (err) 307 goto out; 308 } 309 310 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 311 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME, 312 &dp->dp_free_dir); 313 if (err) 314 goto out; 315 316 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 317 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj); 318 if (err) 319 goto out; 320 VERIFY0(bpobj_open(&dp->dp_free_bpobj, 321 dp->dp_meta_objset, obj)); 322 } 323 324 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS)) { 325 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 326 DMU_POOL_OBSOLETE_BPOBJ, sizeof (uint64_t), 1, &obj); 327 if (err == 0) { 328 VERIFY0(bpobj_open(&dp->dp_obsolete_bpobj, 329 dp->dp_meta_objset, obj)); 330 } else if (err == ENOENT) { 331 /* 332 * We might not have created the remap bpobj yet. 333 */ 334 err = 0; 335 } else { 336 goto out; 337 } 338 } 339 340 /* 341 * Note: errors ignored, because the these special dirs, used for 342 * space accounting, are only created on demand. 343 */ 344 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME, 345 &dp->dp_leak_dir); 346 347 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) { 348 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 349 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1, 350 &dp->dp_bptree_obj); 351 if (err != 0) 352 goto out; 353 } 354 355 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) { 356 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 357 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1, 358 &dp->dp_empty_bpobj); 359 if (err != 0) 360 goto out; 361 } 362 363 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 364 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1, 365 &dp->dp_tmp_userrefs_obj); 366 if (err == ENOENT) 367 err = 0; 368 if (err) 369 goto out; 370 371 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg); 372 373 out: 374 rrw_exit(&dp->dp_config_rwlock, FTAG); 375 return (err); 376 } 377 378 void 379 dsl_pool_close(dsl_pool_t *dp) 380 { 381 /* 382 * Drop our references from dsl_pool_open(). 383 * 384 * Since we held the origin_snap from "syncing" context (which 385 * includes pool-opening context), it actually only got a "ref" 386 * and not a hold, so just drop that here. 387 */ 388 if (dp->dp_origin_snap != NULL) 389 dsl_dataset_rele(dp->dp_origin_snap, dp); 390 if (dp->dp_mos_dir != NULL) 391 dsl_dir_rele(dp->dp_mos_dir, dp); 392 if (dp->dp_free_dir != NULL) 393 dsl_dir_rele(dp->dp_free_dir, dp); 394 if (dp->dp_leak_dir != NULL) 395 dsl_dir_rele(dp->dp_leak_dir, dp); 396 if (dp->dp_root_dir != NULL) 397 dsl_dir_rele(dp->dp_root_dir, dp); 398 399 bpobj_close(&dp->dp_free_bpobj); 400 bpobj_close(&dp->dp_obsolete_bpobj); 401 402 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */ 403 if (dp->dp_meta_objset != NULL) 404 dmu_objset_evict(dp->dp_meta_objset); 405 406 txg_list_destroy(&dp->dp_dirty_datasets); 407 txg_list_destroy(&dp->dp_dirty_zilogs); 408 txg_list_destroy(&dp->dp_sync_tasks); 409 txg_list_destroy(&dp->dp_early_sync_tasks); 410 txg_list_destroy(&dp->dp_dirty_dirs); 411 412 taskq_destroy(dp->dp_zil_clean_taskq); 413 taskq_destroy(dp->dp_sync_taskq); 414 415 /* 416 * We can't set retry to TRUE since we're explicitly specifying 417 * a spa to flush. This is good enough; any missed buffers for 418 * this spa won't cause trouble, and they'll eventually fall 419 * out of the ARC just like any other unused buffer. 420 */ 421 arc_flush(dp->dp_spa, FALSE); 422 423 mmp_fini(dp->dp_spa); 424 txg_fini(dp); 425 dsl_scan_fini(dp); 426 dmu_buf_user_evict_wait(); 427 428 rrw_destroy(&dp->dp_config_rwlock); 429 mutex_destroy(&dp->dp_lock); 430 cv_destroy(&dp->dp_spaceavail_cv); 431 432 ASSERT0(aggsum_value(&dp->dp_wrlog_total)); 433 aggsum_fini(&dp->dp_wrlog_total); 434 for (int i = 0; i < TXG_SIZE; i++) { 435 ASSERT0(aggsum_value(&dp->dp_wrlog_pertxg[i])); 436 aggsum_fini(&dp->dp_wrlog_pertxg[i]); 437 } 438 439 taskq_destroy(dp->dp_unlinked_drain_taskq); 440 taskq_destroy(dp->dp_zrele_taskq); 441 if (dp->dp_blkstats != NULL) 442 vmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t)); 443 kmem_free(dp, sizeof (dsl_pool_t)); 444 } 445 446 void 447 dsl_pool_create_obsolete_bpobj(dsl_pool_t *dp, dmu_tx_t *tx) 448 { 449 uint64_t obj; 450 /* 451 * Currently, we only create the obsolete_bpobj where there are 452 * indirect vdevs with referenced mappings. 453 */ 454 ASSERT(spa_feature_is_active(dp->dp_spa, SPA_FEATURE_DEVICE_REMOVAL)); 455 /* create and open the obsolete_bpobj */ 456 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx); 457 VERIFY0(bpobj_open(&dp->dp_obsolete_bpobj, dp->dp_meta_objset, obj)); 458 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 459 DMU_POOL_OBSOLETE_BPOBJ, sizeof (uint64_t), 1, &obj, tx)); 460 spa_feature_incr(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); 461 } 462 463 void 464 dsl_pool_destroy_obsolete_bpobj(dsl_pool_t *dp, dmu_tx_t *tx) 465 { 466 spa_feature_decr(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); 467 VERIFY0(zap_remove(dp->dp_meta_objset, 468 DMU_POOL_DIRECTORY_OBJECT, 469 DMU_POOL_OBSOLETE_BPOBJ, tx)); 470 bpobj_free(dp->dp_meta_objset, 471 dp->dp_obsolete_bpobj.bpo_object, tx); 472 bpobj_close(&dp->dp_obsolete_bpobj); 473 } 474 475 dsl_pool_t * 476 dsl_pool_create(spa_t *spa, nvlist_t *zplprops __attribute__((unused)), 477 dsl_crypto_params_t *dcp, uint64_t txg) 478 { 479 int err; 480 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg); 481 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg); 482 #ifdef _KERNEL 483 objset_t *os; 484 #else 485 objset_t *os __attribute__((unused)); 486 #endif 487 dsl_dataset_t *ds; 488 uint64_t obj; 489 490 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 491 492 /* create and open the MOS (meta-objset) */ 493 dp->dp_meta_objset = dmu_objset_create_impl(spa, 494 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx); 495 spa->spa_meta_objset = dp->dp_meta_objset; 496 497 /* create the pool directory */ 498 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 499 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx); 500 ASSERT0(err); 501 502 /* Initialize scan structures */ 503 VERIFY0(dsl_scan_init(dp, txg)); 504 505 /* create and open the root dir */ 506 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx); 507 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, 508 NULL, dp, &dp->dp_root_dir)); 509 510 /* create and open the meta-objset dir */ 511 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx); 512 VERIFY0(dsl_pool_open_special_dir(dp, 513 MOS_DIR_NAME, &dp->dp_mos_dir)); 514 515 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) { 516 /* create and open the free dir */ 517 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, 518 FREE_DIR_NAME, tx); 519 VERIFY0(dsl_pool_open_special_dir(dp, 520 FREE_DIR_NAME, &dp->dp_free_dir)); 521 522 /* create and open the free_bplist */ 523 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx); 524 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 525 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0); 526 VERIFY0(bpobj_open(&dp->dp_free_bpobj, 527 dp->dp_meta_objset, obj)); 528 } 529 530 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) 531 dsl_pool_create_origin(dp, tx); 532 533 /* 534 * Some features may be needed when creating the root dataset, so we 535 * create the feature objects here. 536 */ 537 if (spa_version(spa) >= SPA_VERSION_FEATURES) 538 spa_feature_create_zap_objects(spa, tx); 539 540 if (dcp != NULL && dcp->cp_crypt != ZIO_CRYPT_OFF && 541 dcp->cp_crypt != ZIO_CRYPT_INHERIT) 542 spa_feature_enable(spa, SPA_FEATURE_ENCRYPTION, tx); 543 544 /* create the root dataset */ 545 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, dcp, 0, tx); 546 547 /* create the root objset */ 548 VERIFY0(dsl_dataset_hold_obj_flags(dp, obj, 549 DS_HOLD_FLAG_DECRYPT, FTAG, &ds)); 550 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG); 551 os = dmu_objset_create_impl(dp->dp_spa, ds, 552 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx); 553 rrw_exit(&ds->ds_bp_rwlock, FTAG); 554 #ifdef _KERNEL 555 zfs_create_fs(os, kcred, zplprops, tx); 556 #endif 557 dsl_dataset_rele_flags(ds, DS_HOLD_FLAG_DECRYPT, FTAG); 558 559 dmu_tx_commit(tx); 560 561 rrw_exit(&dp->dp_config_rwlock, FTAG); 562 563 return (dp); 564 } 565 566 /* 567 * Account for the meta-objset space in its placeholder dsl_dir. 568 */ 569 void 570 dsl_pool_mos_diduse_space(dsl_pool_t *dp, 571 int64_t used, int64_t comp, int64_t uncomp) 572 { 573 ASSERT3U(comp, ==, uncomp); /* it's all metadata */ 574 mutex_enter(&dp->dp_lock); 575 dp->dp_mos_used_delta += used; 576 dp->dp_mos_compressed_delta += comp; 577 dp->dp_mos_uncompressed_delta += uncomp; 578 mutex_exit(&dp->dp_lock); 579 } 580 581 static void 582 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx) 583 { 584 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 585 dmu_objset_sync(dp->dp_meta_objset, zio, tx); 586 VERIFY0(zio_wait(zio)); 587 dmu_objset_sync_done(dp->dp_meta_objset, tx); 588 taskq_wait(dp->dp_sync_taskq); 589 multilist_destroy(&dp->dp_meta_objset->os_synced_dnodes); 590 591 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", ""); 592 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp); 593 } 594 595 static void 596 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta) 597 { 598 ASSERT(MUTEX_HELD(&dp->dp_lock)); 599 600 if (delta < 0) 601 ASSERT3U(-delta, <=, dp->dp_dirty_total); 602 603 dp->dp_dirty_total += delta; 604 605 /* 606 * Note: we signal even when increasing dp_dirty_total. 607 * This ensures forward progress -- each thread wakes the next waiter. 608 */ 609 if (dp->dp_dirty_total < zfs_dirty_data_max) 610 cv_signal(&dp->dp_spaceavail_cv); 611 } 612 613 void 614 dsl_pool_wrlog_count(dsl_pool_t *dp, int64_t size, uint64_t txg) 615 { 616 ASSERT3S(size, >=, 0); 617 618 aggsum_add(&dp->dp_wrlog_pertxg[txg & TXG_MASK], size); 619 aggsum_add(&dp->dp_wrlog_total, size); 620 621 /* Choose a value slightly bigger than min dirty sync bytes */ 622 uint64_t sync_min = 623 zfs_wrlog_data_max * (zfs_dirty_data_sync_percent + 10) / 200; 624 if (aggsum_compare(&dp->dp_wrlog_pertxg[txg & TXG_MASK], sync_min) > 0) 625 txg_kick(dp, txg); 626 } 627 628 boolean_t 629 dsl_pool_need_wrlog_delay(dsl_pool_t *dp) 630 { 631 uint64_t delay_min_bytes = 632 zfs_wrlog_data_max * zfs_delay_min_dirty_percent / 100; 633 634 return (aggsum_compare(&dp->dp_wrlog_total, delay_min_bytes) > 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 /* Compact per-CPU sums after the big change. */ 645 (void) aggsum_value(&dp->dp_wrlog_pertxg[txg & TXG_MASK]); 646 (void) aggsum_value(&dp->dp_wrlog_total); 647 } 648 649 #ifdef ZFS_DEBUG 650 static boolean_t 651 dsl_early_sync_task_verify(dsl_pool_t *dp, uint64_t txg) 652 { 653 spa_t *spa = dp->dp_spa; 654 vdev_t *rvd = spa->spa_root_vdev; 655 656 for (uint64_t c = 0; c < rvd->vdev_children; c++) { 657 vdev_t *vd = rvd->vdev_child[c]; 658 txg_list_t *tl = &vd->vdev_ms_list; 659 metaslab_t *ms; 660 661 for (ms = txg_list_head(tl, TXG_CLEAN(txg)); ms; 662 ms = txg_list_next(tl, ms, TXG_CLEAN(txg))) { 663 VERIFY(range_tree_is_empty(ms->ms_freeing)); 664 VERIFY(range_tree_is_empty(ms->ms_checkpointing)); 665 } 666 } 667 668 return (B_TRUE); 669 } 670 #else 671 #define dsl_early_sync_task_verify(dp, txg) \ 672 ((void) sizeof (dp), (void) sizeof (txg), B_TRUE) 673 #endif 674 675 void 676 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg) 677 { 678 zio_t *zio; 679 dmu_tx_t *tx; 680 dsl_dir_t *dd; 681 dsl_dataset_t *ds; 682 objset_t *mos = dp->dp_meta_objset; 683 list_t synced_datasets; 684 685 list_create(&synced_datasets, sizeof (dsl_dataset_t), 686 offsetof(dsl_dataset_t, ds_synced_link)); 687 688 tx = dmu_tx_create_assigned(dp, txg); 689 690 /* 691 * Run all early sync tasks before writing out any dirty blocks. 692 * For more info on early sync tasks see block comment in 693 * dsl_early_sync_task(). 694 */ 695 if (!txg_list_empty(&dp->dp_early_sync_tasks, txg)) { 696 dsl_sync_task_t *dst; 697 698 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1); 699 while ((dst = 700 txg_list_remove(&dp->dp_early_sync_tasks, txg)) != NULL) { 701 ASSERT(dsl_early_sync_task_verify(dp, txg)); 702 dsl_sync_task_sync(dst, tx); 703 } 704 ASSERT(dsl_early_sync_task_verify(dp, txg)); 705 } 706 707 /* 708 * Write out all dirty blocks of dirty datasets. 709 */ 710 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 711 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) { 712 /* 713 * We must not sync any non-MOS datasets twice, because 714 * we may have taken a snapshot of them. However, we 715 * may sync newly-created datasets on pass 2. 716 */ 717 ASSERT(!list_link_active(&ds->ds_synced_link)); 718 list_insert_tail(&synced_datasets, ds); 719 dsl_dataset_sync(ds, zio, tx); 720 } 721 VERIFY0(zio_wait(zio)); 722 723 /* 724 * Update the long range free counter after 725 * we're done syncing user data 726 */ 727 mutex_enter(&dp->dp_lock); 728 ASSERT(spa_sync_pass(dp->dp_spa) == 1 || 729 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0); 730 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0; 731 mutex_exit(&dp->dp_lock); 732 733 /* 734 * After the data blocks have been written (ensured by the zio_wait() 735 * above), update the user/group/project space accounting. This happens 736 * in tasks dispatched to dp_sync_taskq, so wait for them before 737 * continuing. 738 */ 739 for (ds = list_head(&synced_datasets); ds != NULL; 740 ds = list_next(&synced_datasets, ds)) { 741 dmu_objset_sync_done(ds->ds_objset, tx); 742 } 743 taskq_wait(dp->dp_sync_taskq); 744 745 /* 746 * Sync the datasets again to push out the changes due to 747 * userspace updates. This must be done before we process the 748 * sync tasks, so that any snapshots will have the correct 749 * user accounting information (and we won't get confused 750 * about which blocks are part of the snapshot). 751 */ 752 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 753 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) { 754 objset_t *os = ds->ds_objset; 755 756 ASSERT(list_link_active(&ds->ds_synced_link)); 757 dmu_buf_rele(ds->ds_dbuf, ds); 758 dsl_dataset_sync(ds, zio, tx); 759 760 /* 761 * Release any key mappings created by calls to 762 * dsl_dataset_dirty() from the userquota accounting 763 * code paths. 764 */ 765 if (os->os_encrypted && !os->os_raw_receive && 766 !os->os_next_write_raw[txg & TXG_MASK]) { 767 ASSERT3P(ds->ds_key_mapping, !=, NULL); 768 key_mapping_rele(dp->dp_spa, ds->ds_key_mapping, ds); 769 } 770 } 771 VERIFY0(zio_wait(zio)); 772 773 /* 774 * Now that the datasets have been completely synced, we can 775 * clean up our in-memory structures accumulated while syncing: 776 * 777 * - move dead blocks from the pending deadlist and livelists 778 * to the on-disk versions 779 * - release hold from dsl_dataset_dirty() 780 * - release key mapping hold from dsl_dataset_dirty() 781 */ 782 while ((ds = list_remove_head(&synced_datasets)) != NULL) { 783 objset_t *os = ds->ds_objset; 784 785 if (os->os_encrypted && !os->os_raw_receive && 786 !os->os_next_write_raw[txg & TXG_MASK]) { 787 ASSERT3P(ds->ds_key_mapping, !=, NULL); 788 key_mapping_rele(dp->dp_spa, ds->ds_key_mapping, ds); 789 } 790 791 dsl_dataset_sync_done(ds, tx); 792 } 793 794 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) { 795 dsl_dir_sync(dd, tx); 796 } 797 798 /* 799 * The MOS's space is accounted for in the pool/$MOS 800 * (dp_mos_dir). We can't modify the mos while we're syncing 801 * it, so we remember the deltas and apply them here. 802 */ 803 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 || 804 dp->dp_mos_uncompressed_delta != 0) { 805 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD, 806 dp->dp_mos_used_delta, 807 dp->dp_mos_compressed_delta, 808 dp->dp_mos_uncompressed_delta, tx); 809 dp->dp_mos_used_delta = 0; 810 dp->dp_mos_compressed_delta = 0; 811 dp->dp_mos_uncompressed_delta = 0; 812 } 813 814 if (dmu_objset_is_dirty(mos, txg)) { 815 dsl_pool_sync_mos(dp, tx); 816 } 817 818 /* 819 * We have written all of the accounted dirty data, so our 820 * dp_space_towrite should now be zero. However, some seldom-used 821 * code paths do not adhere to this (e.g. dbuf_undirty()). Shore up 822 * the accounting of any dirtied space now. 823 * 824 * Note that, besides any dirty data from datasets, the amount of 825 * dirty data in the MOS is also accounted by the pool. Therefore, 826 * we want to do this cleanup after dsl_pool_sync_mos() so we don't 827 * attempt to update the accounting for the same dirty data twice. 828 * (i.e. at this point we only update the accounting for the space 829 * that we know that we "leaked"). 830 */ 831 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg); 832 833 /* 834 * If we modify a dataset in the same txg that we want to destroy it, 835 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it. 836 * dsl_dir_destroy_check() will fail if there are unexpected holds. 837 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf 838 * and clearing the hold on it) before we process the sync_tasks. 839 * The MOS data dirtied by the sync_tasks will be synced on the next 840 * pass. 841 */ 842 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) { 843 dsl_sync_task_t *dst; 844 /* 845 * No more sync tasks should have been added while we 846 * were syncing. 847 */ 848 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1); 849 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL) 850 dsl_sync_task_sync(dst, tx); 851 } 852 853 dmu_tx_commit(tx); 854 855 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg); 856 } 857 858 void 859 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg) 860 { 861 zilog_t *zilog; 862 863 while ((zilog = txg_list_head(&dp->dp_dirty_zilogs, txg))) { 864 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 865 /* 866 * We don't remove the zilog from the dp_dirty_zilogs 867 * list until after we've cleaned it. This ensures that 868 * callers of zilog_is_dirty() receive an accurate 869 * answer when they are racing with the spa sync thread. 870 */ 871 zil_clean(zilog, txg); 872 (void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg); 873 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg)); 874 dmu_buf_rele(ds->ds_dbuf, zilog); 875 } 876 877 dsl_pool_wrlog_clear(dp, txg); 878 879 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg)); 880 } 881 882 /* 883 * TRUE if the current thread is the tx_sync_thread or if we 884 * are being called from SPA context during pool initialization. 885 */ 886 int 887 dsl_pool_sync_context(dsl_pool_t *dp) 888 { 889 return (curthread == dp->dp_tx.tx_sync_thread || 890 spa_is_initializing(dp->dp_spa) || 891 taskq_member(dp->dp_sync_taskq, curthread)); 892 } 893 894 /* 895 * This function returns the amount of allocatable space in the pool 896 * minus whatever space is currently reserved by ZFS for specific 897 * purposes. Specifically: 898 * 899 * 1] Any reserved SLOP space 900 * 2] Any space used by the checkpoint 901 * 3] Any space used for deferred frees 902 * 903 * The latter 2 are especially important because they are needed to 904 * rectify the SPA's and DMU's different understanding of how much space 905 * is used. Now the DMU is aware of that extra space tracked by the SPA 906 * without having to maintain a separate special dir (e.g similar to 907 * $MOS, $FREEING, and $LEAKED). 908 * 909 * Note: By deferred frees here, we mean the frees that were deferred 910 * in spa_sync() after sync pass 1 (spa_deferred_bpobj), and not the 911 * segments placed in ms_defer trees during metaslab_sync_done(). 912 */ 913 uint64_t 914 dsl_pool_adjustedsize(dsl_pool_t *dp, zfs_space_check_t slop_policy) 915 { 916 spa_t *spa = dp->dp_spa; 917 uint64_t space, resv, adjustedsize; 918 uint64_t spa_deferred_frees = 919 spa->spa_deferred_bpobj.bpo_phys->bpo_bytes; 920 921 space = spa_get_dspace(spa) 922 - spa_get_checkpoint_space(spa) - spa_deferred_frees; 923 resv = spa_get_slop_space(spa); 924 925 switch (slop_policy) { 926 case ZFS_SPACE_CHECK_NORMAL: 927 break; 928 case ZFS_SPACE_CHECK_RESERVED: 929 resv >>= 1; 930 break; 931 case ZFS_SPACE_CHECK_EXTRA_RESERVED: 932 resv >>= 2; 933 break; 934 case ZFS_SPACE_CHECK_NONE: 935 resv = 0; 936 break; 937 default: 938 panic("invalid slop policy value: %d", slop_policy); 939 break; 940 } 941 adjustedsize = (space >= resv) ? (space - resv) : 0; 942 943 return (adjustedsize); 944 } 945 946 uint64_t 947 dsl_pool_unreserved_space(dsl_pool_t *dp, zfs_space_check_t slop_policy) 948 { 949 uint64_t poolsize = dsl_pool_adjustedsize(dp, slop_policy); 950 uint64_t deferred = 951 metaslab_class_get_deferred(spa_normal_class(dp->dp_spa)); 952 uint64_t quota = (poolsize >= deferred) ? (poolsize - deferred) : 0; 953 return (quota); 954 } 955 956 uint64_t 957 dsl_pool_deferred_space(dsl_pool_t *dp) 958 { 959 return (metaslab_class_get_deferred(spa_normal_class(dp->dp_spa))); 960 } 961 962 boolean_t 963 dsl_pool_need_dirty_delay(dsl_pool_t *dp) 964 { 965 uint64_t delay_min_bytes = 966 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100; 967 968 mutex_enter(&dp->dp_lock); 969 uint64_t dirty = dp->dp_dirty_total; 970 mutex_exit(&dp->dp_lock); 971 972 return (dirty > delay_min_bytes); 973 } 974 975 static boolean_t 976 dsl_pool_need_dirty_sync(dsl_pool_t *dp, uint64_t txg) 977 { 978 ASSERT(MUTEX_HELD(&dp->dp_lock)); 979 980 uint64_t dirty_min_bytes = 981 zfs_dirty_data_max * zfs_dirty_data_sync_percent / 100; 982 uint64_t dirty = dp->dp_dirty_pertxg[txg & TXG_MASK]; 983 984 return (dirty > dirty_min_bytes); 985 } 986 987 void 988 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx) 989 { 990 if (space > 0) { 991 mutex_enter(&dp->dp_lock); 992 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space; 993 dsl_pool_dirty_delta(dp, space); 994 boolean_t needsync = !dmu_tx_is_syncing(tx) && 995 dsl_pool_need_dirty_sync(dp, tx->tx_txg); 996 mutex_exit(&dp->dp_lock); 997 998 if (needsync) 999 txg_kick(dp, tx->tx_txg); 1000 } 1001 } 1002 1003 void 1004 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg) 1005 { 1006 ASSERT3S(space, >=, 0); 1007 if (space == 0) 1008 return; 1009 1010 mutex_enter(&dp->dp_lock); 1011 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) { 1012 /* XXX writing something we didn't dirty? */ 1013 space = dp->dp_dirty_pertxg[txg & TXG_MASK]; 1014 } 1015 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space); 1016 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space; 1017 ASSERT3U(dp->dp_dirty_total, >=, space); 1018 dsl_pool_dirty_delta(dp, -space); 1019 mutex_exit(&dp->dp_lock); 1020 } 1021 1022 static int 1023 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 1024 { 1025 dmu_tx_t *tx = arg; 1026 dsl_dataset_t *ds, *prev = NULL; 1027 int err; 1028 1029 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 1030 if (err) 1031 return (err); 1032 1033 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) { 1034 err = dsl_dataset_hold_obj(dp, 1035 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev); 1036 if (err) { 1037 dsl_dataset_rele(ds, FTAG); 1038 return (err); 1039 } 1040 1041 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) 1042 break; 1043 dsl_dataset_rele(ds, FTAG); 1044 ds = prev; 1045 prev = NULL; 1046 } 1047 1048 if (prev == NULL) { 1049 prev = dp->dp_origin_snap; 1050 1051 /* 1052 * The $ORIGIN can't have any data, or the accounting 1053 * will be wrong. 1054 */ 1055 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG); 1056 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth); 1057 rrw_exit(&ds->ds_bp_rwlock, FTAG); 1058 1059 /* The origin doesn't get attached to itself */ 1060 if (ds->ds_object == prev->ds_object) { 1061 dsl_dataset_rele(ds, FTAG); 1062 return (0); 1063 } 1064 1065 dmu_buf_will_dirty(ds->ds_dbuf, tx); 1066 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object; 1067 dsl_dataset_phys(ds)->ds_prev_snap_txg = 1068 dsl_dataset_phys(prev)->ds_creation_txg; 1069 1070 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx); 1071 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object; 1072 1073 dmu_buf_will_dirty(prev->ds_dbuf, tx); 1074 dsl_dataset_phys(prev)->ds_num_children++; 1075 1076 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) { 1077 ASSERT(ds->ds_prev == NULL); 1078 VERIFY0(dsl_dataset_hold_obj(dp, 1079 dsl_dataset_phys(ds)->ds_prev_snap_obj, 1080 ds, &ds->ds_prev)); 1081 } 1082 } 1083 1084 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object); 1085 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object); 1086 1087 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) { 1088 dmu_buf_will_dirty(prev->ds_dbuf, tx); 1089 dsl_dataset_phys(prev)->ds_next_clones_obj = 1090 zap_create(dp->dp_meta_objset, 1091 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx); 1092 } 1093 VERIFY0(zap_add_int(dp->dp_meta_objset, 1094 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx)); 1095 1096 dsl_dataset_rele(ds, FTAG); 1097 if (prev != dp->dp_origin_snap) 1098 dsl_dataset_rele(prev, FTAG); 1099 return (0); 1100 } 1101 1102 void 1103 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx) 1104 { 1105 ASSERT(dmu_tx_is_syncing(tx)); 1106 ASSERT(dp->dp_origin_snap != NULL); 1107 1108 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb, 1109 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE)); 1110 } 1111 1112 static int 1113 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg) 1114 { 1115 dmu_tx_t *tx = arg; 1116 objset_t *mos = dp->dp_meta_objset; 1117 1118 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) { 1119 dsl_dataset_t *origin; 1120 1121 VERIFY0(dsl_dataset_hold_obj(dp, 1122 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin)); 1123 1124 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) { 1125 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx); 1126 dsl_dir_phys(origin->ds_dir)->dd_clones = 1127 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE, 1128 0, tx); 1129 } 1130 1131 VERIFY0(zap_add_int(dp->dp_meta_objset, 1132 dsl_dir_phys(origin->ds_dir)->dd_clones, 1133 ds->ds_object, tx)); 1134 1135 dsl_dataset_rele(origin, FTAG); 1136 } 1137 return (0); 1138 } 1139 1140 void 1141 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx) 1142 { 1143 uint64_t obj; 1144 1145 ASSERT(dmu_tx_is_syncing(tx)); 1146 1147 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx); 1148 VERIFY0(dsl_pool_open_special_dir(dp, 1149 FREE_DIR_NAME, &dp->dp_free_dir)); 1150 1151 /* 1152 * We can't use bpobj_alloc(), because spa_version() still 1153 * returns the old version, and we need a new-version bpobj with 1154 * subobj support. So call dmu_object_alloc() directly. 1155 */ 1156 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ, 1157 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx); 1158 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 1159 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx)); 1160 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj)); 1161 1162 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 1163 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE)); 1164 } 1165 1166 void 1167 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx) 1168 { 1169 uint64_t dsobj; 1170 dsl_dataset_t *ds; 1171 1172 ASSERT(dmu_tx_is_syncing(tx)); 1173 ASSERT(dp->dp_origin_snap == NULL); 1174 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER)); 1175 1176 /* create the origin dir, ds, & snap-ds */ 1177 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME, 1178 NULL, 0, kcred, NULL, tx); 1179 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 1180 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx); 1181 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj, 1182 dp, &dp->dp_origin_snap)); 1183 dsl_dataset_rele(ds, FTAG); 1184 } 1185 1186 taskq_t * 1187 dsl_pool_zrele_taskq(dsl_pool_t *dp) 1188 { 1189 return (dp->dp_zrele_taskq); 1190 } 1191 1192 taskq_t * 1193 dsl_pool_unlinked_drain_taskq(dsl_pool_t *dp) 1194 { 1195 return (dp->dp_unlinked_drain_taskq); 1196 } 1197 1198 /* 1199 * Walk through the pool-wide zap object of temporary snapshot user holds 1200 * and release them. 1201 */ 1202 void 1203 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp) 1204 { 1205 zap_attribute_t za; 1206 zap_cursor_t zc; 1207 objset_t *mos = dp->dp_meta_objset; 1208 uint64_t zapobj = dp->dp_tmp_userrefs_obj; 1209 nvlist_t *holds; 1210 1211 if (zapobj == 0) 1212 return; 1213 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS); 1214 1215 holds = fnvlist_alloc(); 1216 1217 for (zap_cursor_init(&zc, mos, zapobj); 1218 zap_cursor_retrieve(&zc, &za) == 0; 1219 zap_cursor_advance(&zc)) { 1220 char *htag; 1221 nvlist_t *tags; 1222 1223 htag = strchr(za.za_name, '-'); 1224 *htag = '\0'; 1225 ++htag; 1226 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) { 1227 tags = fnvlist_alloc(); 1228 fnvlist_add_boolean(tags, htag); 1229 fnvlist_add_nvlist(holds, za.za_name, tags); 1230 fnvlist_free(tags); 1231 } else { 1232 fnvlist_add_boolean(tags, htag); 1233 } 1234 } 1235 dsl_dataset_user_release_tmp(dp, holds); 1236 fnvlist_free(holds); 1237 zap_cursor_fini(&zc); 1238 } 1239 1240 /* 1241 * Create the pool-wide zap object for storing temporary snapshot holds. 1242 */ 1243 static void 1244 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx) 1245 { 1246 objset_t *mos = dp->dp_meta_objset; 1247 1248 ASSERT(dp->dp_tmp_userrefs_obj == 0); 1249 ASSERT(dmu_tx_is_syncing(tx)); 1250 1251 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS, 1252 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx); 1253 } 1254 1255 static int 1256 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj, 1257 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding) 1258 { 1259 objset_t *mos = dp->dp_meta_objset; 1260 uint64_t zapobj = dp->dp_tmp_userrefs_obj; 1261 char *name; 1262 int error; 1263 1264 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS); 1265 ASSERT(dmu_tx_is_syncing(tx)); 1266 1267 /* 1268 * If the pool was created prior to SPA_VERSION_USERREFS, the 1269 * zap object for temporary holds might not exist yet. 1270 */ 1271 if (zapobj == 0) { 1272 if (holding) { 1273 dsl_pool_user_hold_create_obj(dp, tx); 1274 zapobj = dp->dp_tmp_userrefs_obj; 1275 } else { 1276 return (SET_ERROR(ENOENT)); 1277 } 1278 } 1279 1280 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag); 1281 if (holding) 1282 error = zap_add(mos, zapobj, name, 8, 1, &now, tx); 1283 else 1284 error = zap_remove(mos, zapobj, name, tx); 1285 kmem_strfree(name); 1286 1287 return (error); 1288 } 1289 1290 /* 1291 * Add a temporary hold for the given dataset object and tag. 1292 */ 1293 int 1294 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag, 1295 uint64_t now, dmu_tx_t *tx) 1296 { 1297 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE)); 1298 } 1299 1300 /* 1301 * Release a temporary hold for the given dataset object and tag. 1302 */ 1303 int 1304 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag, 1305 dmu_tx_t *tx) 1306 { 1307 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0, 1308 tx, B_FALSE)); 1309 } 1310 1311 /* 1312 * DSL Pool Configuration Lock 1313 * 1314 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset 1315 * creation / destruction / rename / property setting). It must be held for 1316 * read to hold a dataset or dsl_dir. I.e. you must call 1317 * dsl_pool_config_enter() or dsl_pool_hold() before calling 1318 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock 1319 * must be held continuously until all datasets and dsl_dirs are released. 1320 * 1321 * The only exception to this rule is that if a "long hold" is placed on 1322 * a dataset, then the dp_config_rwlock may be dropped while the dataset 1323 * is still held. The long hold will prevent the dataset from being 1324 * destroyed -- the destroy will fail with EBUSY. A long hold can be 1325 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset 1326 * (by calling dsl_{dataset,objset}_{try}own{_obj}). 1327 * 1328 * Legitimate long-holders (including owners) should be long-running, cancelable 1329 * tasks that should cause "zfs destroy" to fail. This includes DMU 1330 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open), 1331 * "zfs send", and "zfs diff". There are several other long-holders whose 1332 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()). 1333 * 1334 * The usual formula for long-holding would be: 1335 * dsl_pool_hold() 1336 * dsl_dataset_hold() 1337 * ... perform checks ... 1338 * dsl_dataset_long_hold() 1339 * dsl_pool_rele() 1340 * ... perform long-running task ... 1341 * dsl_dataset_long_rele() 1342 * dsl_dataset_rele() 1343 * 1344 * Note that when the long hold is released, the dataset is still held but 1345 * the pool is not held. The dataset may change arbitrarily during this time 1346 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the 1347 * dataset except release it. 1348 * 1349 * Operations generally fall somewhere into the following taxonomy: 1350 * 1351 * Read-Only Modifying 1352 * 1353 * Dataset Layer / MOS zfs get zfs destroy 1354 * 1355 * Individual Dataset read() write() 1356 * 1357 * 1358 * Dataset Layer Operations 1359 * 1360 * Modifying operations should generally use dsl_sync_task(). The synctask 1361 * infrastructure enforces proper locking strategy with respect to the 1362 * dp_config_rwlock. See the comment above dsl_sync_task() for details. 1363 * 1364 * Read-only operations will manually hold the pool, then the dataset, obtain 1365 * information from the dataset, then release the pool and dataset. 1366 * dmu_objset_{hold,rele}() are convenience routines that also do the pool 1367 * hold/rele. 1368 * 1369 * 1370 * Operations On Individual Datasets 1371 * 1372 * Objects _within_ an objset should only be modified by the current 'owner' 1373 * of the objset to prevent incorrect concurrent modification. Thus, use 1374 * {dmu_objset,dsl_dataset}_own to mark some entity as the current owner, 1375 * and fail with EBUSY if there is already an owner. The owner can then 1376 * implement its own locking strategy, independent of the dataset layer's 1377 * locking infrastructure. 1378 * (E.g., the ZPL has its own set of locks to control concurrency. A regular 1379 * vnop will not reach into the dataset layer). 1380 * 1381 * Ideally, objects would also only be read by the objset’s owner, so that we 1382 * don’t observe state mid-modification. 1383 * (E.g. the ZPL is creating a new object and linking it into a directory; if 1384 * you don’t coordinate with the ZPL to hold ZPL-level locks, you could see an 1385 * intermediate state. The ioctl level violates this but in pretty benign 1386 * ways, e.g. reading the zpl props object.) 1387 */ 1388 1389 int 1390 dsl_pool_hold(const char *name, const void *tag, dsl_pool_t **dp) 1391 { 1392 spa_t *spa; 1393 int error; 1394 1395 error = spa_open(name, &spa, tag); 1396 if (error == 0) { 1397 *dp = spa_get_dsl(spa); 1398 dsl_pool_config_enter(*dp, tag); 1399 } 1400 return (error); 1401 } 1402 1403 void 1404 dsl_pool_rele(dsl_pool_t *dp, const void *tag) 1405 { 1406 dsl_pool_config_exit(dp, tag); 1407 spa_close(dp->dp_spa, tag); 1408 } 1409 1410 void 1411 dsl_pool_config_enter(dsl_pool_t *dp, const void *tag) 1412 { 1413 /* 1414 * We use a "reentrant" reader-writer lock, but not reentrantly. 1415 * 1416 * The rrwlock can (with the track_all flag) track all reading threads, 1417 * which is very useful for debugging which code path failed to release 1418 * the lock, and for verifying that the *current* thread does hold 1419 * the lock. 1420 * 1421 * (Unlike a rwlock, which knows that N threads hold it for 1422 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE 1423 * if any thread holds it for read, even if this thread doesn't). 1424 */ 1425 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER)); 1426 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag); 1427 } 1428 1429 void 1430 dsl_pool_config_enter_prio(dsl_pool_t *dp, const void *tag) 1431 { 1432 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER)); 1433 rrw_enter_read_prio(&dp->dp_config_rwlock, tag); 1434 } 1435 1436 void 1437 dsl_pool_config_exit(dsl_pool_t *dp, const void *tag) 1438 { 1439 rrw_exit(&dp->dp_config_rwlock, tag); 1440 } 1441 1442 boolean_t 1443 dsl_pool_config_held(dsl_pool_t *dp) 1444 { 1445 return (RRW_LOCK_HELD(&dp->dp_config_rwlock)); 1446 } 1447 1448 boolean_t 1449 dsl_pool_config_held_writer(dsl_pool_t *dp) 1450 { 1451 return (RRW_WRITE_HELD(&dp->dp_config_rwlock)); 1452 } 1453 1454 EXPORT_SYMBOL(dsl_pool_config_enter); 1455 EXPORT_SYMBOL(dsl_pool_config_exit); 1456 1457 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */ 1458 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max_percent, UINT, ZMOD_RD, 1459 "Max percent of RAM allowed to be dirty"); 1460 1461 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */ 1462 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max_max_percent, UINT, ZMOD_RD, 1463 "zfs_dirty_data_max upper bound as % of RAM"); 1464 1465 ZFS_MODULE_PARAM(zfs, zfs_, delay_min_dirty_percent, UINT, ZMOD_RW, 1466 "Transaction delay threshold"); 1467 1468 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max, ULONG, ZMOD_RW, 1469 "Determines the dirty space limit"); 1470 1471 ZFS_MODULE_PARAM(zfs, zfs_, wrlog_data_max, ULONG, ZMOD_RW, 1472 "The size limit of write-transaction zil log data"); 1473 1474 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */ 1475 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max_max, ULONG, ZMOD_RD, 1476 "zfs_dirty_data_max upper bound in bytes"); 1477 1478 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_sync_percent, UINT, ZMOD_RW, 1479 "Dirty data txg sync threshold as a percentage of zfs_dirty_data_max"); 1480 1481 ZFS_MODULE_PARAM(zfs, zfs_, delay_scale, ULONG, ZMOD_RW, 1482 "How quickly delay approaches infinity"); 1483 1484 ZFS_MODULE_PARAM(zfs, zfs_, sync_taskq_batch_pct, INT, ZMOD_RW, 1485 "Max percent of CPUs that are used to sync dirty data"); 1486 1487 ZFS_MODULE_PARAM(zfs_zil, zfs_zil_, clean_taskq_nthr_pct, INT, ZMOD_RW, 1488 "Max percent of CPUs that are used per dp_sync_taskq"); 1489 1490 ZFS_MODULE_PARAM(zfs_zil, zfs_zil_, clean_taskq_minalloc, INT, ZMOD_RW, 1491 "Number of taskq entries that are pre-populated"); 1492 1493 ZFS_MODULE_PARAM(zfs_zil, zfs_zil_, clean_taskq_maxalloc, INT, ZMOD_RW, 1494 "Max number of taskq entries that are cached"); 1495