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 uint64_t zfs_dirty_data_max = 0; 103 uint64_t 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 uint64_t 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 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000; 142 143 /* 144 * These tunables determine the behavior of how zil_itxg_clean() is 145 * called via zil_clean() in the context of spa_sync(). When an itxg 146 * list needs to be cleaned, TQ_NOSLEEP will be used when dispatching. 147 * If the dispatch fails, the call to zil_itxg_clean() will occur 148 * synchronously in the context of spa_sync(), which can negatively 149 * impact the performance of spa_sync() (e.g. in the case of the itxg 150 * list having a large number of itxs that needs to be cleaned). 151 * 152 * Thus, these tunables can be used to manipulate the behavior of the 153 * taskq used by zil_clean(); they determine the number of taskq entries 154 * that are pre-populated when the taskq is first created (via the 155 * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of 156 * taskq entries that are cached after an on-demand allocation (via the 157 * "zfs_zil_clean_taskq_maxalloc"). 158 * 159 * The idea being, we want to try reasonably hard to ensure there will 160 * already be a taskq entry pre-allocated by the time that it is needed 161 * by zil_clean(). This way, we can avoid the possibility of an 162 * on-demand allocation of a new taskq entry from failing, which would 163 * result in zil_itxg_clean() being called synchronously from zil_clean() 164 * (which can adversely affect performance of spa_sync()). 165 * 166 * Additionally, the number of threads used by the taskq can be 167 * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable. 168 */ 169 static int zfs_zil_clean_taskq_nthr_pct = 100; 170 static int zfs_zil_clean_taskq_minalloc = 1024; 171 static int zfs_zil_clean_taskq_maxalloc = 1024 * 1024; 172 173 int 174 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp) 175 { 176 uint64_t obj; 177 int err; 178 179 err = zap_lookup(dp->dp_meta_objset, 180 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj, 181 name, sizeof (obj), 1, &obj); 182 if (err) 183 return (err); 184 185 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp)); 186 } 187 188 static dsl_pool_t * 189 dsl_pool_open_impl(spa_t *spa, uint64_t txg) 190 { 191 dsl_pool_t *dp; 192 blkptr_t *bp = spa_get_rootblkptr(spa); 193 194 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP); 195 dp->dp_spa = spa; 196 dp->dp_meta_rootbp = *bp; 197 rrw_init(&dp->dp_config_rwlock, B_TRUE); 198 txg_init(dp, txg); 199 mmp_init(spa); 200 201 txg_list_create(&dp->dp_dirty_datasets, spa, 202 offsetof(dsl_dataset_t, ds_dirty_link)); 203 txg_list_create(&dp->dp_dirty_zilogs, spa, 204 offsetof(zilog_t, zl_dirty_link)); 205 txg_list_create(&dp->dp_dirty_dirs, spa, 206 offsetof(dsl_dir_t, dd_dirty_link)); 207 txg_list_create(&dp->dp_sync_tasks, spa, 208 offsetof(dsl_sync_task_t, dst_node)); 209 txg_list_create(&dp->dp_early_sync_tasks, spa, 210 offsetof(dsl_sync_task_t, dst_node)); 211 212 dp->dp_sync_taskq = spa_sync_tq_create(spa, "dp_sync_taskq"); 213 214 dp->dp_zil_clean_taskq = taskq_create("dp_zil_clean_taskq", 215 zfs_zil_clean_taskq_nthr_pct, minclsyspri, 216 zfs_zil_clean_taskq_minalloc, 217 zfs_zil_clean_taskq_maxalloc, 218 TASKQ_PREPOPULATE | TASKQ_THREADS_CPU_PCT); 219 220 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL); 221 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL); 222 223 aggsum_init(&dp->dp_wrlog_total, 0); 224 for (int i = 0; i < TXG_SIZE; i++) { 225 aggsum_init(&dp->dp_wrlog_pertxg[i], 0); 226 } 227 228 dp->dp_zrele_taskq = taskq_create("z_zrele", 100, defclsyspri, 229 boot_ncpus * 8, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC | 230 TASKQ_THREADS_CPU_PCT); 231 dp->dp_unlinked_drain_taskq = taskq_create("z_unlinked_drain", 232 100, defclsyspri, boot_ncpus, INT_MAX, 233 TASKQ_PREPOPULATE | TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT); 234 235 return (dp); 236 } 237 238 int 239 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp) 240 { 241 int err; 242 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg); 243 244 /* 245 * Initialize the caller's dsl_pool_t structure before we actually open 246 * the meta objset. This is done because a self-healing write zio may 247 * be issued as part of dmu_objset_open_impl() and the spa needs its 248 * dsl_pool_t initialized in order to handle the write. 249 */ 250 *dpp = dp; 251 252 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp, 253 &dp->dp_meta_objset); 254 if (err != 0) { 255 dsl_pool_close(dp); 256 *dpp = NULL; 257 } 258 259 return (err); 260 } 261 262 int 263 dsl_pool_open(dsl_pool_t *dp) 264 { 265 int err; 266 dsl_dir_t *dd; 267 dsl_dataset_t *ds; 268 uint64_t obj; 269 270 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 271 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 272 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1, 273 &dp->dp_root_dir_obj); 274 if (err) 275 goto out; 276 277 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, 278 NULL, dp, &dp->dp_root_dir); 279 if (err) 280 goto out; 281 282 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir); 283 if (err) 284 goto out; 285 286 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) { 287 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd); 288 if (err) 289 goto out; 290 err = dsl_dataset_hold_obj(dp, 291 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds); 292 if (err == 0) { 293 err = dsl_dataset_hold_obj(dp, 294 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp, 295 &dp->dp_origin_snap); 296 dsl_dataset_rele(ds, FTAG); 297 } 298 dsl_dir_rele(dd, dp); 299 if (err) 300 goto out; 301 } 302 303 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 304 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME, 305 &dp->dp_free_dir); 306 if (err) 307 goto out; 308 309 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 310 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj); 311 if (err) 312 goto out; 313 VERIFY0(bpobj_open(&dp->dp_free_bpobj, 314 dp->dp_meta_objset, obj)); 315 } 316 317 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS)) { 318 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 319 DMU_POOL_OBSOLETE_BPOBJ, sizeof (uint64_t), 1, &obj); 320 if (err == 0) { 321 VERIFY0(bpobj_open(&dp->dp_obsolete_bpobj, 322 dp->dp_meta_objset, obj)); 323 } else if (err == ENOENT) { 324 /* 325 * We might not have created the remap bpobj yet. 326 */ 327 } else { 328 goto out; 329 } 330 } 331 332 /* 333 * Note: errors ignored, because the these special dirs, used for 334 * space accounting, are only created on demand. 335 */ 336 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME, 337 &dp->dp_leak_dir); 338 339 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) { 340 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 341 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1, 342 &dp->dp_bptree_obj); 343 if (err != 0) 344 goto out; 345 } 346 347 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) { 348 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 349 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1, 350 &dp->dp_empty_bpobj); 351 if (err != 0) 352 goto out; 353 } 354 355 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 356 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1, 357 &dp->dp_tmp_userrefs_obj); 358 if (err == ENOENT) 359 err = 0; 360 if (err) 361 goto out; 362 363 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg); 364 365 out: 366 rrw_exit(&dp->dp_config_rwlock, FTAG); 367 return (err); 368 } 369 370 void 371 dsl_pool_close(dsl_pool_t *dp) 372 { 373 /* 374 * Drop our references from dsl_pool_open(). 375 * 376 * Since we held the origin_snap from "syncing" context (which 377 * includes pool-opening context), it actually only got a "ref" 378 * and not a hold, so just drop that here. 379 */ 380 if (dp->dp_origin_snap != NULL) 381 dsl_dataset_rele(dp->dp_origin_snap, dp); 382 if (dp->dp_mos_dir != NULL) 383 dsl_dir_rele(dp->dp_mos_dir, dp); 384 if (dp->dp_free_dir != NULL) 385 dsl_dir_rele(dp->dp_free_dir, dp); 386 if (dp->dp_leak_dir != NULL) 387 dsl_dir_rele(dp->dp_leak_dir, dp); 388 if (dp->dp_root_dir != NULL) 389 dsl_dir_rele(dp->dp_root_dir, dp); 390 391 bpobj_close(&dp->dp_free_bpobj); 392 bpobj_close(&dp->dp_obsolete_bpobj); 393 394 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */ 395 if (dp->dp_meta_objset != NULL) 396 dmu_objset_evict(dp->dp_meta_objset); 397 398 txg_list_destroy(&dp->dp_dirty_datasets); 399 txg_list_destroy(&dp->dp_dirty_zilogs); 400 txg_list_destroy(&dp->dp_sync_tasks); 401 txg_list_destroy(&dp->dp_early_sync_tasks); 402 txg_list_destroy(&dp->dp_dirty_dirs); 403 404 taskq_destroy(dp->dp_zil_clean_taskq); 405 spa_sync_tq_destroy(dp->dp_spa); 406 407 /* 408 * We can't set retry to TRUE since we're explicitly specifying 409 * a spa to flush. This is good enough; any missed buffers for 410 * this spa won't cause trouble, and they'll eventually fall 411 * out of the ARC just like any other unused buffer. 412 */ 413 arc_flush(dp->dp_spa, FALSE); 414 415 mmp_fini(dp->dp_spa); 416 txg_fini(dp); 417 dsl_scan_fini(dp); 418 dmu_buf_user_evict_wait(); 419 420 rrw_destroy(&dp->dp_config_rwlock); 421 mutex_destroy(&dp->dp_lock); 422 cv_destroy(&dp->dp_spaceavail_cv); 423 424 ASSERT0(aggsum_value(&dp->dp_wrlog_total)); 425 aggsum_fini(&dp->dp_wrlog_total); 426 for (int i = 0; i < TXG_SIZE; i++) { 427 ASSERT0(aggsum_value(&dp->dp_wrlog_pertxg[i])); 428 aggsum_fini(&dp->dp_wrlog_pertxg[i]); 429 } 430 431 taskq_destroy(dp->dp_unlinked_drain_taskq); 432 taskq_destroy(dp->dp_zrele_taskq); 433 if (dp->dp_blkstats != NULL) 434 vmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t)); 435 kmem_free(dp, sizeof (dsl_pool_t)); 436 } 437 438 void 439 dsl_pool_create_obsolete_bpobj(dsl_pool_t *dp, dmu_tx_t *tx) 440 { 441 uint64_t obj; 442 /* 443 * Currently, we only create the obsolete_bpobj where there are 444 * indirect vdevs with referenced mappings. 445 */ 446 ASSERT(spa_feature_is_active(dp->dp_spa, SPA_FEATURE_DEVICE_REMOVAL)); 447 /* create and open the obsolete_bpobj */ 448 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx); 449 VERIFY0(bpobj_open(&dp->dp_obsolete_bpobj, dp->dp_meta_objset, obj)); 450 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 451 DMU_POOL_OBSOLETE_BPOBJ, sizeof (uint64_t), 1, &obj, tx)); 452 spa_feature_incr(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); 453 } 454 455 void 456 dsl_pool_destroy_obsolete_bpobj(dsl_pool_t *dp, dmu_tx_t *tx) 457 { 458 spa_feature_decr(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); 459 VERIFY0(zap_remove(dp->dp_meta_objset, 460 DMU_POOL_DIRECTORY_OBJECT, 461 DMU_POOL_OBSOLETE_BPOBJ, tx)); 462 bpobj_free(dp->dp_meta_objset, 463 dp->dp_obsolete_bpobj.bpo_object, tx); 464 bpobj_close(&dp->dp_obsolete_bpobj); 465 } 466 467 dsl_pool_t * 468 dsl_pool_create(spa_t *spa, nvlist_t *zplprops __attribute__((unused)), 469 dsl_crypto_params_t *dcp, uint64_t txg) 470 { 471 int err; 472 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg); 473 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg); 474 #ifdef _KERNEL 475 objset_t *os; 476 #else 477 objset_t *os __attribute__((unused)); 478 #endif 479 dsl_dataset_t *ds; 480 uint64_t obj; 481 482 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 483 484 /* create and open the MOS (meta-objset) */ 485 dp->dp_meta_objset = dmu_objset_create_impl(spa, 486 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx); 487 spa->spa_meta_objset = dp->dp_meta_objset; 488 489 /* create the pool directory */ 490 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 491 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx); 492 ASSERT0(err); 493 494 /* Initialize scan structures */ 495 VERIFY0(dsl_scan_init(dp, txg)); 496 497 /* create and open the root dir */ 498 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx); 499 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, 500 NULL, dp, &dp->dp_root_dir)); 501 502 /* create and open the meta-objset dir */ 503 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx); 504 VERIFY0(dsl_pool_open_special_dir(dp, 505 MOS_DIR_NAME, &dp->dp_mos_dir)); 506 507 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) { 508 /* create and open the free dir */ 509 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, 510 FREE_DIR_NAME, tx); 511 VERIFY0(dsl_pool_open_special_dir(dp, 512 FREE_DIR_NAME, &dp->dp_free_dir)); 513 514 /* create and open the free_bplist */ 515 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx); 516 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 517 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0); 518 VERIFY0(bpobj_open(&dp->dp_free_bpobj, 519 dp->dp_meta_objset, obj)); 520 } 521 522 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) 523 dsl_pool_create_origin(dp, tx); 524 525 /* 526 * Some features may be needed when creating the root dataset, so we 527 * create the feature objects here. 528 */ 529 if (spa_version(spa) >= SPA_VERSION_FEATURES) 530 spa_feature_create_zap_objects(spa, tx); 531 532 if (dcp != NULL && dcp->cp_crypt != ZIO_CRYPT_OFF && 533 dcp->cp_crypt != ZIO_CRYPT_INHERIT) 534 spa_feature_enable(spa, SPA_FEATURE_ENCRYPTION, tx); 535 536 /* create the root dataset */ 537 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, dcp, 0, tx); 538 539 /* create the root objset */ 540 VERIFY0(dsl_dataset_hold_obj_flags(dp, obj, 541 DS_HOLD_FLAG_DECRYPT, FTAG, &ds)); 542 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG); 543 os = dmu_objset_create_impl(dp->dp_spa, ds, 544 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx); 545 rrw_exit(&ds->ds_bp_rwlock, FTAG); 546 #ifdef _KERNEL 547 zfs_create_fs(os, kcred, zplprops, tx); 548 #endif 549 dsl_dataset_rele_flags(ds, DS_HOLD_FLAG_DECRYPT, FTAG); 550 551 dmu_tx_commit(tx); 552 553 rrw_exit(&dp->dp_config_rwlock, FTAG); 554 555 return (dp); 556 } 557 558 /* 559 * Account for the meta-objset space in its placeholder dsl_dir. 560 */ 561 void 562 dsl_pool_mos_diduse_space(dsl_pool_t *dp, 563 int64_t used, int64_t comp, int64_t uncomp) 564 { 565 ASSERT3U(comp, ==, uncomp); /* it's all metadata */ 566 mutex_enter(&dp->dp_lock); 567 dp->dp_mos_used_delta += used; 568 dp->dp_mos_compressed_delta += comp; 569 dp->dp_mos_uncompressed_delta += uncomp; 570 mutex_exit(&dp->dp_lock); 571 } 572 573 static void 574 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx) 575 { 576 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 577 dmu_objset_sync(dp->dp_meta_objset, zio, tx); 578 VERIFY0(zio_wait(zio)); 579 dmu_objset_sync_done(dp->dp_meta_objset, tx); 580 taskq_wait(dp->dp_sync_taskq); 581 multilist_destroy(&dp->dp_meta_objset->os_synced_dnodes); 582 583 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", ""); 584 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp); 585 } 586 587 static void 588 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta) 589 { 590 ASSERT(MUTEX_HELD(&dp->dp_lock)); 591 592 if (delta < 0) 593 ASSERT3U(-delta, <=, dp->dp_dirty_total); 594 595 dp->dp_dirty_total += delta; 596 597 /* 598 * Note: we signal even when increasing dp_dirty_total. 599 * This ensures forward progress -- each thread wakes the next waiter. 600 */ 601 if (dp->dp_dirty_total < zfs_dirty_data_max) 602 cv_signal(&dp->dp_spaceavail_cv); 603 } 604 605 void 606 dsl_pool_wrlog_count(dsl_pool_t *dp, int64_t size, uint64_t txg) 607 { 608 ASSERT3S(size, >=, 0); 609 610 aggsum_add(&dp->dp_wrlog_pertxg[txg & TXG_MASK], size); 611 aggsum_add(&dp->dp_wrlog_total, size); 612 613 /* Choose a value slightly bigger than min dirty sync bytes */ 614 uint64_t sync_min = 615 zfs_wrlog_data_max * (zfs_dirty_data_sync_percent + 10) / 200; 616 if (aggsum_compare(&dp->dp_wrlog_pertxg[txg & TXG_MASK], sync_min) > 0) 617 txg_kick(dp, txg); 618 } 619 620 boolean_t 621 dsl_pool_need_wrlog_delay(dsl_pool_t *dp) 622 { 623 uint64_t delay_min_bytes = 624 zfs_wrlog_data_max * zfs_delay_min_dirty_percent / 100; 625 626 return (aggsum_compare(&dp->dp_wrlog_total, delay_min_bytes) > 0); 627 } 628 629 static void 630 dsl_pool_wrlog_clear(dsl_pool_t *dp, uint64_t txg) 631 { 632 int64_t delta; 633 delta = -(int64_t)aggsum_value(&dp->dp_wrlog_pertxg[txg & TXG_MASK]); 634 aggsum_add(&dp->dp_wrlog_pertxg[txg & TXG_MASK], delta); 635 aggsum_add(&dp->dp_wrlog_total, delta); 636 /* Compact per-CPU sums after the big change. */ 637 (void) aggsum_value(&dp->dp_wrlog_pertxg[txg & TXG_MASK]); 638 (void) aggsum_value(&dp->dp_wrlog_total); 639 } 640 641 #ifdef ZFS_DEBUG 642 static boolean_t 643 dsl_early_sync_task_verify(dsl_pool_t *dp, uint64_t txg) 644 { 645 spa_t *spa = dp->dp_spa; 646 vdev_t *rvd = spa->spa_root_vdev; 647 648 for (uint64_t c = 0; c < rvd->vdev_children; c++) { 649 vdev_t *vd = rvd->vdev_child[c]; 650 txg_list_t *tl = &vd->vdev_ms_list; 651 metaslab_t *ms; 652 653 for (ms = txg_list_head(tl, TXG_CLEAN(txg)); ms; 654 ms = txg_list_next(tl, ms, TXG_CLEAN(txg))) { 655 VERIFY(range_tree_is_empty(ms->ms_freeing)); 656 VERIFY(range_tree_is_empty(ms->ms_checkpointing)); 657 } 658 } 659 660 return (B_TRUE); 661 } 662 #else 663 #define dsl_early_sync_task_verify(dp, txg) \ 664 ((void) sizeof (dp), (void) sizeof (txg), B_TRUE) 665 #endif 666 667 void 668 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg) 669 { 670 zio_t *rio; /* root zio for all dirty dataset syncs */ 671 dmu_tx_t *tx; 672 dsl_dir_t *dd; 673 dsl_dataset_t *ds; 674 objset_t *mos = dp->dp_meta_objset; 675 list_t synced_datasets; 676 677 list_create(&synced_datasets, sizeof (dsl_dataset_t), 678 offsetof(dsl_dataset_t, ds_synced_link)); 679 680 tx = dmu_tx_create_assigned(dp, txg); 681 682 /* 683 * Run all early sync tasks before writing out any dirty blocks. 684 * For more info on early sync tasks see block comment in 685 * dsl_early_sync_task(). 686 */ 687 if (!txg_list_empty(&dp->dp_early_sync_tasks, txg)) { 688 dsl_sync_task_t *dst; 689 690 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1); 691 while ((dst = 692 txg_list_remove(&dp->dp_early_sync_tasks, txg)) != NULL) { 693 ASSERT(dsl_early_sync_task_verify(dp, txg)); 694 dsl_sync_task_sync(dst, tx); 695 } 696 ASSERT(dsl_early_sync_task_verify(dp, txg)); 697 } 698 699 /* 700 * Write out all dirty blocks of dirty datasets. Note, this could 701 * create a very large (+10k) zio tree. 702 */ 703 rio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 704 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) { 705 /* 706 * We must not sync any non-MOS datasets twice, because 707 * we may have taken a snapshot of them. However, we 708 * may sync newly-created datasets on pass 2. 709 */ 710 ASSERT(!list_link_active(&ds->ds_synced_link)); 711 list_insert_tail(&synced_datasets, ds); 712 dsl_dataset_sync(ds, rio, tx); 713 } 714 VERIFY0(zio_wait(rio)); 715 716 /* 717 * Update the long range free counter after 718 * we're done syncing user data 719 */ 720 mutex_enter(&dp->dp_lock); 721 ASSERT(spa_sync_pass(dp->dp_spa) == 1 || 722 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0); 723 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0; 724 mutex_exit(&dp->dp_lock); 725 726 /* 727 * After the data blocks have been written (ensured by the zio_wait() 728 * above), update the user/group/project space accounting. This happens 729 * in tasks dispatched to dp_sync_taskq, so wait for them before 730 * continuing. 731 */ 732 for (ds = list_head(&synced_datasets); ds != NULL; 733 ds = list_next(&synced_datasets, ds)) { 734 dmu_objset_sync_done(ds->ds_objset, tx); 735 } 736 taskq_wait(dp->dp_sync_taskq); 737 738 /* 739 * Sync the datasets again to push out the changes due to 740 * userspace updates. This must be done before we process the 741 * sync tasks, so that any snapshots will have the correct 742 * user accounting information (and we won't get confused 743 * about which blocks are part of the snapshot). 744 */ 745 rio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 746 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) { 747 objset_t *os = ds->ds_objset; 748 749 ASSERT(list_link_active(&ds->ds_synced_link)); 750 dmu_buf_rele(ds->ds_dbuf, ds); 751 dsl_dataset_sync(ds, rio, tx); 752 753 /* 754 * Release any key mappings created by calls to 755 * dsl_dataset_dirty() from the userquota accounting 756 * code paths. 757 */ 758 if (os->os_encrypted && !os->os_raw_receive && 759 !os->os_next_write_raw[txg & TXG_MASK]) { 760 ASSERT3P(ds->ds_key_mapping, !=, NULL); 761 key_mapping_rele(dp->dp_spa, ds->ds_key_mapping, ds); 762 } 763 } 764 VERIFY0(zio_wait(rio)); 765 766 /* 767 * Now that the datasets have been completely synced, we can 768 * clean up our in-memory structures accumulated while syncing: 769 * 770 * - move dead blocks from the pending deadlist and livelists 771 * to the on-disk versions 772 * - release hold from dsl_dataset_dirty() 773 * - release key mapping hold from dsl_dataset_dirty() 774 */ 775 while ((ds = list_remove_head(&synced_datasets)) != NULL) { 776 objset_t *os = ds->ds_objset; 777 778 if (os->os_encrypted && !os->os_raw_receive && 779 !os->os_next_write_raw[txg & TXG_MASK]) { 780 ASSERT3P(ds->ds_key_mapping, !=, NULL); 781 key_mapping_rele(dp->dp_spa, ds->ds_key_mapping, ds); 782 } 783 784 dsl_dataset_sync_done(ds, tx); 785 dmu_buf_rele(ds->ds_dbuf, ds); 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 uint64_t 951 dsl_pool_deferred_space(dsl_pool_t *dp) 952 { 953 return (metaslab_class_get_deferred(spa_normal_class(dp->dp_spa))); 954 } 955 956 boolean_t 957 dsl_pool_need_dirty_delay(dsl_pool_t *dp) 958 { 959 uint64_t delay_min_bytes = 960 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100; 961 962 /* 963 * We are not taking the dp_lock here and few other places, since torn 964 * reads are unlikely: on 64-bit systems due to register size and on 965 * 32-bit due to memory constraints. Pool-wide locks in hot path may 966 * be too expensive, while we do not need a precise result here. 967 */ 968 return (dp->dp_dirty_total > delay_min_bytes); 969 } 970 971 static boolean_t 972 dsl_pool_need_dirty_sync(dsl_pool_t *dp, uint64_t txg) 973 { 974 uint64_t dirty_min_bytes = 975 zfs_dirty_data_max * zfs_dirty_data_sync_percent / 100; 976 uint64_t dirty = dp->dp_dirty_pertxg[txg & TXG_MASK]; 977 978 return (dirty > dirty_min_bytes); 979 } 980 981 void 982 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx) 983 { 984 if (space > 0) { 985 mutex_enter(&dp->dp_lock); 986 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space; 987 dsl_pool_dirty_delta(dp, space); 988 boolean_t needsync = !dmu_tx_is_syncing(tx) && 989 dsl_pool_need_dirty_sync(dp, tx->tx_txg); 990 mutex_exit(&dp->dp_lock); 991 992 if (needsync) 993 txg_kick(dp, tx->tx_txg); 994 } 995 } 996 997 void 998 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg) 999 { 1000 ASSERT3S(space, >=, 0); 1001 if (space == 0) 1002 return; 1003 1004 mutex_enter(&dp->dp_lock); 1005 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) { 1006 /* XXX writing something we didn't dirty? */ 1007 space = dp->dp_dirty_pertxg[txg & TXG_MASK]; 1008 } 1009 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space); 1010 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space; 1011 ASSERT3U(dp->dp_dirty_total, >=, space); 1012 dsl_pool_dirty_delta(dp, -space); 1013 mutex_exit(&dp->dp_lock); 1014 } 1015 1016 static int 1017 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 1018 { 1019 dmu_tx_t *tx = arg; 1020 dsl_dataset_t *ds, *prev = NULL; 1021 int err; 1022 1023 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 1024 if (err) 1025 return (err); 1026 1027 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) { 1028 err = dsl_dataset_hold_obj(dp, 1029 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev); 1030 if (err) { 1031 dsl_dataset_rele(ds, FTAG); 1032 return (err); 1033 } 1034 1035 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) 1036 break; 1037 dsl_dataset_rele(ds, FTAG); 1038 ds = prev; 1039 prev = NULL; 1040 } 1041 1042 if (prev == NULL) { 1043 prev = dp->dp_origin_snap; 1044 1045 /* 1046 * The $ORIGIN can't have any data, or the accounting 1047 * will be wrong. 1048 */ 1049 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG); 1050 ASSERT0(BP_GET_LOGICAL_BIRTH(&dsl_dataset_phys(prev)->ds_bp)); 1051 rrw_exit(&ds->ds_bp_rwlock, FTAG); 1052 1053 /* The origin doesn't get attached to itself */ 1054 if (ds->ds_object == prev->ds_object) { 1055 dsl_dataset_rele(ds, FTAG); 1056 return (0); 1057 } 1058 1059 dmu_buf_will_dirty(ds->ds_dbuf, tx); 1060 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object; 1061 dsl_dataset_phys(ds)->ds_prev_snap_txg = 1062 dsl_dataset_phys(prev)->ds_creation_txg; 1063 1064 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx); 1065 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object; 1066 1067 dmu_buf_will_dirty(prev->ds_dbuf, tx); 1068 dsl_dataset_phys(prev)->ds_num_children++; 1069 1070 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) { 1071 ASSERT(ds->ds_prev == NULL); 1072 VERIFY0(dsl_dataset_hold_obj(dp, 1073 dsl_dataset_phys(ds)->ds_prev_snap_obj, 1074 ds, &ds->ds_prev)); 1075 } 1076 } 1077 1078 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object); 1079 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object); 1080 1081 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) { 1082 dmu_buf_will_dirty(prev->ds_dbuf, tx); 1083 dsl_dataset_phys(prev)->ds_next_clones_obj = 1084 zap_create(dp->dp_meta_objset, 1085 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx); 1086 } 1087 VERIFY0(zap_add_int(dp->dp_meta_objset, 1088 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx)); 1089 1090 dsl_dataset_rele(ds, FTAG); 1091 if (prev != dp->dp_origin_snap) 1092 dsl_dataset_rele(prev, FTAG); 1093 return (0); 1094 } 1095 1096 void 1097 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx) 1098 { 1099 ASSERT(dmu_tx_is_syncing(tx)); 1100 ASSERT(dp->dp_origin_snap != NULL); 1101 1102 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb, 1103 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE)); 1104 } 1105 1106 static int 1107 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg) 1108 { 1109 dmu_tx_t *tx = arg; 1110 objset_t *mos = dp->dp_meta_objset; 1111 1112 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) { 1113 dsl_dataset_t *origin; 1114 1115 VERIFY0(dsl_dataset_hold_obj(dp, 1116 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin)); 1117 1118 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) { 1119 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx); 1120 dsl_dir_phys(origin->ds_dir)->dd_clones = 1121 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE, 1122 0, tx); 1123 } 1124 1125 VERIFY0(zap_add_int(dp->dp_meta_objset, 1126 dsl_dir_phys(origin->ds_dir)->dd_clones, 1127 ds->ds_object, tx)); 1128 1129 dsl_dataset_rele(origin, FTAG); 1130 } 1131 return (0); 1132 } 1133 1134 void 1135 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx) 1136 { 1137 uint64_t obj; 1138 1139 ASSERT(dmu_tx_is_syncing(tx)); 1140 1141 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx); 1142 VERIFY0(dsl_pool_open_special_dir(dp, 1143 FREE_DIR_NAME, &dp->dp_free_dir)); 1144 1145 /* 1146 * We can't use bpobj_alloc(), because spa_version() still 1147 * returns the old version, and we need a new-version bpobj with 1148 * subobj support. So call dmu_object_alloc() directly. 1149 */ 1150 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ, 1151 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx); 1152 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 1153 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx)); 1154 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj)); 1155 1156 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 1157 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE)); 1158 } 1159 1160 void 1161 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx) 1162 { 1163 uint64_t dsobj; 1164 dsl_dataset_t *ds; 1165 1166 ASSERT(dmu_tx_is_syncing(tx)); 1167 ASSERT(dp->dp_origin_snap == NULL); 1168 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER)); 1169 1170 /* create the origin dir, ds, & snap-ds */ 1171 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME, 1172 NULL, 0, kcred, NULL, tx); 1173 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 1174 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx); 1175 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj, 1176 dp, &dp->dp_origin_snap)); 1177 dsl_dataset_rele(ds, FTAG); 1178 } 1179 1180 taskq_t * 1181 dsl_pool_zrele_taskq(dsl_pool_t *dp) 1182 { 1183 return (dp->dp_zrele_taskq); 1184 } 1185 1186 taskq_t * 1187 dsl_pool_unlinked_drain_taskq(dsl_pool_t *dp) 1188 { 1189 return (dp->dp_unlinked_drain_taskq); 1190 } 1191 1192 /* 1193 * Walk through the pool-wide zap object of temporary snapshot user holds 1194 * and release them. 1195 */ 1196 void 1197 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp) 1198 { 1199 zap_attribute_t za; 1200 zap_cursor_t zc; 1201 objset_t *mos = dp->dp_meta_objset; 1202 uint64_t zapobj = dp->dp_tmp_userrefs_obj; 1203 nvlist_t *holds; 1204 1205 if (zapobj == 0) 1206 return; 1207 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS); 1208 1209 holds = fnvlist_alloc(); 1210 1211 for (zap_cursor_init(&zc, mos, zapobj); 1212 zap_cursor_retrieve(&zc, &za) == 0; 1213 zap_cursor_advance(&zc)) { 1214 char *htag; 1215 nvlist_t *tags; 1216 1217 htag = strchr(za.za_name, '-'); 1218 *htag = '\0'; 1219 ++htag; 1220 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) { 1221 tags = fnvlist_alloc(); 1222 fnvlist_add_boolean(tags, htag); 1223 fnvlist_add_nvlist(holds, za.za_name, tags); 1224 fnvlist_free(tags); 1225 } else { 1226 fnvlist_add_boolean(tags, htag); 1227 } 1228 } 1229 dsl_dataset_user_release_tmp(dp, holds); 1230 fnvlist_free(holds); 1231 zap_cursor_fini(&zc); 1232 } 1233 1234 /* 1235 * Create the pool-wide zap object for storing temporary snapshot holds. 1236 */ 1237 static void 1238 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx) 1239 { 1240 objset_t *mos = dp->dp_meta_objset; 1241 1242 ASSERT(dp->dp_tmp_userrefs_obj == 0); 1243 ASSERT(dmu_tx_is_syncing(tx)); 1244 1245 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS, 1246 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx); 1247 } 1248 1249 static int 1250 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj, 1251 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding) 1252 { 1253 objset_t *mos = dp->dp_meta_objset; 1254 uint64_t zapobj = dp->dp_tmp_userrefs_obj; 1255 char *name; 1256 int error; 1257 1258 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS); 1259 ASSERT(dmu_tx_is_syncing(tx)); 1260 1261 /* 1262 * If the pool was created prior to SPA_VERSION_USERREFS, the 1263 * zap object for temporary holds might not exist yet. 1264 */ 1265 if (zapobj == 0) { 1266 if (holding) { 1267 dsl_pool_user_hold_create_obj(dp, tx); 1268 zapobj = dp->dp_tmp_userrefs_obj; 1269 } else { 1270 return (SET_ERROR(ENOENT)); 1271 } 1272 } 1273 1274 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag); 1275 if (holding) 1276 error = zap_add(mos, zapobj, name, 8, 1, &now, tx); 1277 else 1278 error = zap_remove(mos, zapobj, name, tx); 1279 kmem_strfree(name); 1280 1281 return (error); 1282 } 1283 1284 /* 1285 * Add a temporary hold for the given dataset object and tag. 1286 */ 1287 int 1288 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag, 1289 uint64_t now, dmu_tx_t *tx) 1290 { 1291 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE)); 1292 } 1293 1294 /* 1295 * Release a temporary hold for the given dataset object and tag. 1296 */ 1297 int 1298 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag, 1299 dmu_tx_t *tx) 1300 { 1301 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0, 1302 tx, B_FALSE)); 1303 } 1304 1305 /* 1306 * DSL Pool Configuration Lock 1307 * 1308 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset 1309 * creation / destruction / rename / property setting). It must be held for 1310 * read to hold a dataset or dsl_dir. I.e. you must call 1311 * dsl_pool_config_enter() or dsl_pool_hold() before calling 1312 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock 1313 * must be held continuously until all datasets and dsl_dirs are released. 1314 * 1315 * The only exception to this rule is that if a "long hold" is placed on 1316 * a dataset, then the dp_config_rwlock may be dropped while the dataset 1317 * is still held. The long hold will prevent the dataset from being 1318 * destroyed -- the destroy will fail with EBUSY. A long hold can be 1319 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset 1320 * (by calling dsl_{dataset,objset}_{try}own{_obj}). 1321 * 1322 * Legitimate long-holders (including owners) should be long-running, cancelable 1323 * tasks that should cause "zfs destroy" to fail. This includes DMU 1324 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open), 1325 * "zfs send", and "zfs diff". There are several other long-holders whose 1326 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()). 1327 * 1328 * The usual formula for long-holding would be: 1329 * dsl_pool_hold() 1330 * dsl_dataset_hold() 1331 * ... perform checks ... 1332 * dsl_dataset_long_hold() 1333 * dsl_pool_rele() 1334 * ... perform long-running task ... 1335 * dsl_dataset_long_rele() 1336 * dsl_dataset_rele() 1337 * 1338 * Note that when the long hold is released, the dataset is still held but 1339 * the pool is not held. The dataset may change arbitrarily during this time 1340 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the 1341 * dataset except release it. 1342 * 1343 * Operations generally fall somewhere into the following taxonomy: 1344 * 1345 * Read-Only Modifying 1346 * 1347 * Dataset Layer / MOS zfs get zfs destroy 1348 * 1349 * Individual Dataset read() write() 1350 * 1351 * 1352 * Dataset Layer Operations 1353 * 1354 * Modifying operations should generally use dsl_sync_task(). The synctask 1355 * infrastructure enforces proper locking strategy with respect to the 1356 * dp_config_rwlock. See the comment above dsl_sync_task() for details. 1357 * 1358 * Read-only operations will manually hold the pool, then the dataset, obtain 1359 * information from the dataset, then release the pool and dataset. 1360 * dmu_objset_{hold,rele}() are convenience routines that also do the pool 1361 * hold/rele. 1362 * 1363 * 1364 * Operations On Individual Datasets 1365 * 1366 * Objects _within_ an objset should only be modified by the current 'owner' 1367 * of the objset to prevent incorrect concurrent modification. Thus, use 1368 * {dmu_objset,dsl_dataset}_own to mark some entity as the current owner, 1369 * and fail with EBUSY if there is already an owner. The owner can then 1370 * implement its own locking strategy, independent of the dataset layer's 1371 * locking infrastructure. 1372 * (E.g., the ZPL has its own set of locks to control concurrency. A regular 1373 * vnop will not reach into the dataset layer). 1374 * 1375 * Ideally, objects would also only be read by the objset’s owner, so that we 1376 * don’t observe state mid-modification. 1377 * (E.g. the ZPL is creating a new object and linking it into a directory; if 1378 * you don’t coordinate with the ZPL to hold ZPL-level locks, you could see an 1379 * intermediate state. The ioctl level violates this but in pretty benign 1380 * ways, e.g. reading the zpl props object.) 1381 */ 1382 1383 int 1384 dsl_pool_hold(const char *name, const void *tag, dsl_pool_t **dp) 1385 { 1386 spa_t *spa; 1387 int error; 1388 1389 error = spa_open(name, &spa, tag); 1390 if (error == 0) { 1391 *dp = spa_get_dsl(spa); 1392 dsl_pool_config_enter(*dp, tag); 1393 } 1394 return (error); 1395 } 1396 1397 void 1398 dsl_pool_rele(dsl_pool_t *dp, const void *tag) 1399 { 1400 dsl_pool_config_exit(dp, tag); 1401 spa_close(dp->dp_spa, tag); 1402 } 1403 1404 void 1405 dsl_pool_config_enter(dsl_pool_t *dp, const void *tag) 1406 { 1407 /* 1408 * We use a "reentrant" reader-writer lock, but not reentrantly. 1409 * 1410 * The rrwlock can (with the track_all flag) track all reading threads, 1411 * which is very useful for debugging which code path failed to release 1412 * the lock, and for verifying that the *current* thread does hold 1413 * the lock. 1414 * 1415 * (Unlike a rwlock, which knows that N threads hold it for 1416 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE 1417 * if any thread holds it for read, even if this thread doesn't). 1418 */ 1419 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER)); 1420 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag); 1421 } 1422 1423 void 1424 dsl_pool_config_enter_prio(dsl_pool_t *dp, const void *tag) 1425 { 1426 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER)); 1427 rrw_enter_read_prio(&dp->dp_config_rwlock, tag); 1428 } 1429 1430 void 1431 dsl_pool_config_exit(dsl_pool_t *dp, const void *tag) 1432 { 1433 rrw_exit(&dp->dp_config_rwlock, tag); 1434 } 1435 1436 boolean_t 1437 dsl_pool_config_held(dsl_pool_t *dp) 1438 { 1439 return (RRW_LOCK_HELD(&dp->dp_config_rwlock)); 1440 } 1441 1442 boolean_t 1443 dsl_pool_config_held_writer(dsl_pool_t *dp) 1444 { 1445 return (RRW_WRITE_HELD(&dp->dp_config_rwlock)); 1446 } 1447 1448 EXPORT_SYMBOL(dsl_pool_config_enter); 1449 EXPORT_SYMBOL(dsl_pool_config_exit); 1450 1451 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */ 1452 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max_percent, UINT, ZMOD_RD, 1453 "Max percent of RAM allowed to be dirty"); 1454 1455 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */ 1456 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max_max_percent, UINT, ZMOD_RD, 1457 "zfs_dirty_data_max upper bound as % of RAM"); 1458 1459 ZFS_MODULE_PARAM(zfs, zfs_, delay_min_dirty_percent, UINT, ZMOD_RW, 1460 "Transaction delay threshold"); 1461 1462 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max, U64, ZMOD_RW, 1463 "Determines the dirty space limit"); 1464 1465 ZFS_MODULE_PARAM(zfs, zfs_, wrlog_data_max, U64, ZMOD_RW, 1466 "The size limit of write-transaction zil log data"); 1467 1468 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */ 1469 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max_max, U64, ZMOD_RD, 1470 "zfs_dirty_data_max upper bound in bytes"); 1471 1472 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_sync_percent, UINT, ZMOD_RW, 1473 "Dirty data txg sync threshold as a percentage of zfs_dirty_data_max"); 1474 1475 ZFS_MODULE_PARAM(zfs, zfs_, delay_scale, U64, ZMOD_RW, 1476 "How quickly delay approaches infinity"); 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