dsl_pool.c revision 260763
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2013 by Delphix. All rights reserved. 24 * Copyright (c) 2013 Steven Hartland. All rights reserved. 25 */ 26 27#include <sys/dsl_pool.h> 28#include <sys/dsl_dataset.h> 29#include <sys/dsl_prop.h> 30#include <sys/dsl_dir.h> 31#include <sys/dsl_synctask.h> 32#include <sys/dsl_scan.h> 33#include <sys/dnode.h> 34#include <sys/dmu_tx.h> 35#include <sys/dmu_objset.h> 36#include <sys/arc.h> 37#include <sys/zap.h> 38#include <sys/zio.h> 39#include <sys/zfs_context.h> 40#include <sys/fs/zfs.h> 41#include <sys/zfs_znode.h> 42#include <sys/spa_impl.h> 43#include <sys/dsl_deadlist.h> 44#include <sys/bptree.h> 45#include <sys/zfeature.h> 46#include <sys/zil_impl.h> 47#include <sys/dsl_userhold.h> 48 49/* 50 * ZFS Write Throttle 51 * ------------------ 52 * 53 * ZFS must limit the rate of incoming writes to the rate at which it is able 54 * to sync data modifications to the backend storage. Throttling by too much 55 * creates an artificial limit; throttling by too little can only be sustained 56 * for short periods and would lead to highly lumpy performance. On a per-pool 57 * basis, ZFS tracks the amount of modified (dirty) data. As operations change 58 * data, the amount of dirty data increases; as ZFS syncs out data, the amount 59 * of dirty data decreases. When the amount of dirty data exceeds a 60 * predetermined threshold further modifications are blocked until the amount 61 * of dirty data decreases (as data is synced out). 62 * 63 * The limit on dirty data is tunable, and should be adjusted according to 64 * both the IO capacity and available memory of the system. The larger the 65 * window, the more ZFS is able to aggregate and amortize metadata (and data) 66 * changes. However, memory is a limited resource, and allowing for more dirty 67 * data comes at the cost of keeping other useful data in memory (for example 68 * ZFS data cached by the ARC). 69 * 70 * Implementation 71 * 72 * As buffers are modified dsl_pool_willuse_space() increments both the per- 73 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of 74 * dirty space used; dsl_pool_dirty_space() decrements those values as data 75 * is synced out from dsl_pool_sync(). While only the poolwide value is 76 * relevant, the per-txg value is useful for debugging. The tunable 77 * zfs_dirty_data_max determines the dirty space limit. Once that value is 78 * exceeded, new writes are halted until space frees up. 79 * 80 * The zfs_dirty_data_sync tunable dictates the threshold at which we 81 * ensure that there is a txg syncing (see the comment in txg.c for a full 82 * description of transaction group stages). 83 * 84 * The IO scheduler uses both the dirty space limit and current amount of 85 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS 86 * issues. See the comment in vdev_queue.c for details of the IO scheduler. 87 * 88 * The delay is also calculated based on the amount of dirty data. See the 89 * comment above dmu_tx_delay() for details. 90 */ 91 92/* 93 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory, 94 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system. 95 */ 96uint64_t zfs_dirty_data_max; 97uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024; 98int zfs_dirty_data_max_percent = 10; 99 100/* 101 * If there is at least this much dirty data, push out a txg. 102 */ 103uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024; 104 105/* 106 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in 107 * and delay each transaction. 108 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent. 109 */ 110int zfs_delay_min_dirty_percent = 60; 111 112/* 113 * This controls how quickly the delay approaches infinity. 114 * Larger values cause it to delay less for a given amount of dirty data. 115 * Therefore larger values will cause there to be more dirty data for a 116 * given throughput. 117 * 118 * For the smoothest delay, this value should be about 1 billion divided 119 * by the maximum number of operations per second. This will smoothly 120 * handle between 10x and 1/10th this number. 121 * 122 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the 123 * multiply in dmu_tx_delay(). 124 */ 125uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000; 126 127 128/* 129 * XXX someday maybe turn these into #defines, and you have to tune it on a 130 * per-pool basis using zfs.conf. 131 */ 132 133 134SYSCTL_DECL(_vfs_zfs); 135#if 0 136TUNABLE_INT("vfs.zfs.no_write_throttle", &zfs_no_write_throttle); 137SYSCTL_INT(_vfs_zfs, OID_AUTO, no_write_throttle, CTLFLAG_RDTUN, 138 &zfs_no_write_throttle, 0, ""); 139TUNABLE_INT("vfs.zfs.write_limit_shift", &zfs_write_limit_shift); 140SYSCTL_INT(_vfs_zfs, OID_AUTO, write_limit_shift, CTLFLAG_RDTUN, 141 &zfs_write_limit_shift, 0, "2^N of physical memory"); 142SYSCTL_DECL(_vfs_zfs_txg); 143TUNABLE_INT("vfs.zfs.txg.synctime_ms", &zfs_txg_synctime_ms); 144SYSCTL_INT(_vfs_zfs_txg, OID_AUTO, synctime_ms, CTLFLAG_RDTUN, 145 &zfs_txg_synctime_ms, 0, "Target milliseconds to sync a txg"); 146 147TUNABLE_QUAD("vfs.zfs.write_limit_min", &zfs_write_limit_min); 148SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_min, CTLFLAG_RDTUN, 149 &zfs_write_limit_min, 0, "Minimum write limit"); 150TUNABLE_QUAD("vfs.zfs.write_limit_max", &zfs_write_limit_max); 151SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_max, CTLFLAG_RDTUN, 152 &zfs_write_limit_max, 0, "Maximum data payload per txg"); 153TUNABLE_QUAD("vfs.zfs.write_limit_inflated", &zfs_write_limit_inflated); 154SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_inflated, CTLFLAG_RDTUN, 155 &zfs_write_limit_inflated, 0, "Maximum size of the dynamic write limit"); 156TUNABLE_QUAD("vfs.zfs.write_limit_override", &zfs_write_limit_override); 157SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_override, CTLFLAG_RDTUN, 158 &zfs_write_limit_override, 0, 159 "Force a txg if dirty buffers exceed this value (bytes)"); 160#endif 161 162hrtime_t zfs_throttle_delay = MSEC2NSEC(10); 163hrtime_t zfs_throttle_resolution = MSEC2NSEC(10); 164 165int 166dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp) 167{ 168 uint64_t obj; 169 int err; 170 171 err = zap_lookup(dp->dp_meta_objset, 172 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj, 173 name, sizeof (obj), 1, &obj); 174 if (err) 175 return (err); 176 177 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp)); 178} 179 180static dsl_pool_t * 181dsl_pool_open_impl(spa_t *spa, uint64_t txg) 182{ 183 dsl_pool_t *dp; 184 blkptr_t *bp = spa_get_rootblkptr(spa); 185 186 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP); 187 dp->dp_spa = spa; 188 dp->dp_meta_rootbp = *bp; 189 rrw_init(&dp->dp_config_rwlock, B_TRUE); 190 txg_init(dp, txg); 191 192 txg_list_create(&dp->dp_dirty_datasets, 193 offsetof(dsl_dataset_t, ds_dirty_link)); 194 txg_list_create(&dp->dp_dirty_zilogs, 195 offsetof(zilog_t, zl_dirty_link)); 196 txg_list_create(&dp->dp_dirty_dirs, 197 offsetof(dsl_dir_t, dd_dirty_link)); 198 txg_list_create(&dp->dp_sync_tasks, 199 offsetof(dsl_sync_task_t, dst_node)); 200 201 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL); 202 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL); 203 204 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri, 205 1, 4, 0); 206 207 return (dp); 208} 209 210int 211dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp) 212{ 213 int err; 214 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg); 215 216 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp, 217 &dp->dp_meta_objset); 218 if (err != 0) 219 dsl_pool_close(dp); 220 else 221 *dpp = dp; 222 223 return (err); 224} 225 226int 227dsl_pool_open(dsl_pool_t *dp) 228{ 229 int err; 230 dsl_dir_t *dd; 231 dsl_dataset_t *ds; 232 uint64_t obj; 233 234 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 235 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 236 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1, 237 &dp->dp_root_dir_obj); 238 if (err) 239 goto out; 240 241 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, 242 NULL, dp, &dp->dp_root_dir); 243 if (err) 244 goto out; 245 246 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir); 247 if (err) 248 goto out; 249 250 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) { 251 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd); 252 if (err) 253 goto out; 254 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj, 255 FTAG, &ds); 256 if (err == 0) { 257 err = dsl_dataset_hold_obj(dp, 258 ds->ds_phys->ds_prev_snap_obj, dp, 259 &dp->dp_origin_snap); 260 dsl_dataset_rele(ds, FTAG); 261 } 262 dsl_dir_rele(dd, dp); 263 if (err) 264 goto out; 265 } 266 267 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 268 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME, 269 &dp->dp_free_dir); 270 if (err) 271 goto out; 272 273 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 274 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj); 275 if (err) 276 goto out; 277 VERIFY0(bpobj_open(&dp->dp_free_bpobj, 278 dp->dp_meta_objset, obj)); 279 } 280 281 if (spa_feature_is_active(dp->dp_spa, 282 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) { 283 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 284 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1, 285 &dp->dp_bptree_obj); 286 if (err != 0) 287 goto out; 288 } 289 290 if (spa_feature_is_active(dp->dp_spa, 291 &spa_feature_table[SPA_FEATURE_EMPTY_BPOBJ])) { 292 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 293 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1, 294 &dp->dp_empty_bpobj); 295 if (err != 0) 296 goto out; 297 } 298 299 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 300 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1, 301 &dp->dp_tmp_userrefs_obj); 302 if (err == ENOENT) 303 err = 0; 304 if (err) 305 goto out; 306 307 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg); 308 309out: 310 rrw_exit(&dp->dp_config_rwlock, FTAG); 311 return (err); 312} 313 314void 315dsl_pool_close(dsl_pool_t *dp) 316{ 317 /* 318 * Drop our references from dsl_pool_open(). 319 * 320 * Since we held the origin_snap from "syncing" context (which 321 * includes pool-opening context), it actually only got a "ref" 322 * and not a hold, so just drop that here. 323 */ 324 if (dp->dp_origin_snap) 325 dsl_dataset_rele(dp->dp_origin_snap, dp); 326 if (dp->dp_mos_dir) 327 dsl_dir_rele(dp->dp_mos_dir, dp); 328 if (dp->dp_free_dir) 329 dsl_dir_rele(dp->dp_free_dir, dp); 330 if (dp->dp_root_dir) 331 dsl_dir_rele(dp->dp_root_dir, dp); 332 333 bpobj_close(&dp->dp_free_bpobj); 334 335 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */ 336 if (dp->dp_meta_objset) 337 dmu_objset_evict(dp->dp_meta_objset); 338 339 txg_list_destroy(&dp->dp_dirty_datasets); 340 txg_list_destroy(&dp->dp_dirty_zilogs); 341 txg_list_destroy(&dp->dp_sync_tasks); 342 txg_list_destroy(&dp->dp_dirty_dirs); 343 344 arc_flush(dp->dp_spa); 345 txg_fini(dp); 346 dsl_scan_fini(dp); 347 rrw_destroy(&dp->dp_config_rwlock); 348 mutex_destroy(&dp->dp_lock); 349 taskq_destroy(dp->dp_vnrele_taskq); 350 if (dp->dp_blkstats) 351 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t)); 352 kmem_free(dp, sizeof (dsl_pool_t)); 353} 354 355dsl_pool_t * 356dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg) 357{ 358 int err; 359 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg); 360 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg); 361 objset_t *os; 362 dsl_dataset_t *ds; 363 uint64_t obj; 364 365 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 366 367 /* create and open the MOS (meta-objset) */ 368 dp->dp_meta_objset = dmu_objset_create_impl(spa, 369 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx); 370 371 /* create the pool directory */ 372 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 373 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx); 374 ASSERT0(err); 375 376 /* Initialize scan structures */ 377 VERIFY0(dsl_scan_init(dp, txg)); 378 379 /* create and open the root dir */ 380 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx); 381 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, 382 NULL, dp, &dp->dp_root_dir)); 383 384 /* create and open the meta-objset dir */ 385 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx); 386 VERIFY0(dsl_pool_open_special_dir(dp, 387 MOS_DIR_NAME, &dp->dp_mos_dir)); 388 389 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) { 390 /* create and open the free dir */ 391 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, 392 FREE_DIR_NAME, tx); 393 VERIFY0(dsl_pool_open_special_dir(dp, 394 FREE_DIR_NAME, &dp->dp_free_dir)); 395 396 /* create and open the free_bplist */ 397 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx); 398 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 399 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0); 400 VERIFY0(bpobj_open(&dp->dp_free_bpobj, 401 dp->dp_meta_objset, obj)); 402 } 403 404 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) 405 dsl_pool_create_origin(dp, tx); 406 407 /* create the root dataset */ 408 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx); 409 410 /* create the root objset */ 411 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds)); 412 os = dmu_objset_create_impl(dp->dp_spa, ds, 413 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx); 414#ifdef _KERNEL 415 zfs_create_fs(os, kcred, zplprops, tx); 416#endif 417 dsl_dataset_rele(ds, FTAG); 418 419 dmu_tx_commit(tx); 420 421 rrw_exit(&dp->dp_config_rwlock, FTAG); 422 423 return (dp); 424} 425 426/* 427 * Account for the meta-objset space in its placeholder dsl_dir. 428 */ 429void 430dsl_pool_mos_diduse_space(dsl_pool_t *dp, 431 int64_t used, int64_t comp, int64_t uncomp) 432{ 433 ASSERT3U(comp, ==, uncomp); /* it's all metadata */ 434 mutex_enter(&dp->dp_lock); 435 dp->dp_mos_used_delta += used; 436 dp->dp_mos_compressed_delta += comp; 437 dp->dp_mos_uncompressed_delta += uncomp; 438 mutex_exit(&dp->dp_lock); 439} 440 441static int 442deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 443{ 444 dsl_deadlist_t *dl = arg; 445 dsl_deadlist_insert(dl, bp, tx); 446 return (0); 447} 448 449static void 450dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx) 451{ 452 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 453 dmu_objset_sync(dp->dp_meta_objset, zio, tx); 454 VERIFY0(zio_wait(zio)); 455 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", ""); 456 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp); 457} 458 459static void 460dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta) 461{ 462 ASSERT(MUTEX_HELD(&dp->dp_lock)); 463 464 if (delta < 0) 465 ASSERT3U(-delta, <=, dp->dp_dirty_total); 466 467 dp->dp_dirty_total += delta; 468 469 /* 470 * Note: we signal even when increasing dp_dirty_total. 471 * This ensures forward progress -- each thread wakes the next waiter. 472 */ 473 if (dp->dp_dirty_total <= zfs_dirty_data_max) 474 cv_signal(&dp->dp_spaceavail_cv); 475} 476 477void 478dsl_pool_sync(dsl_pool_t *dp, uint64_t txg) 479{ 480 zio_t *zio; 481 dmu_tx_t *tx; 482 dsl_dir_t *dd; 483 dsl_dataset_t *ds; 484 objset_t *mos = dp->dp_meta_objset; 485 list_t synced_datasets; 486 487 list_create(&synced_datasets, sizeof (dsl_dataset_t), 488 offsetof(dsl_dataset_t, ds_synced_link)); 489 490 tx = dmu_tx_create_assigned(dp, txg); 491 492 /* 493 * Write out all dirty blocks of dirty datasets. 494 */ 495 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 496 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) { 497 /* 498 * We must not sync any non-MOS datasets twice, because 499 * we may have taken a snapshot of them. However, we 500 * may sync newly-created datasets on pass 2. 501 */ 502 ASSERT(!list_link_active(&ds->ds_synced_link)); 503 list_insert_tail(&synced_datasets, ds); 504 dsl_dataset_sync(ds, zio, tx); 505 } 506 VERIFY0(zio_wait(zio)); 507 508 /* 509 * We have written all of the accounted dirty data, so our 510 * dp_space_towrite should now be zero. However, some seldom-used 511 * code paths do not adhere to this (e.g. dbuf_undirty(), also 512 * rounding error in dbuf_write_physdone). 513 * Shore up the accounting of any dirtied space now. 514 */ 515 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg); 516 517 /* 518 * After the data blocks have been written (ensured by the zio_wait() 519 * above), update the user/group space accounting. 520 */ 521 for (ds = list_head(&synced_datasets); ds != NULL; 522 ds = list_next(&synced_datasets, ds)) { 523 dmu_objset_do_userquota_updates(ds->ds_objset, tx); 524 } 525 526 /* 527 * Sync the datasets again to push out the changes due to 528 * userspace updates. This must be done before we process the 529 * sync tasks, so that any snapshots will have the correct 530 * user accounting information (and we won't get confused 531 * about which blocks are part of the snapshot). 532 */ 533 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 534 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) { 535 ASSERT(list_link_active(&ds->ds_synced_link)); 536 dmu_buf_rele(ds->ds_dbuf, ds); 537 dsl_dataset_sync(ds, zio, tx); 538 } 539 VERIFY0(zio_wait(zio)); 540 541 /* 542 * Now that the datasets have been completely synced, we can 543 * clean up our in-memory structures accumulated while syncing: 544 * 545 * - move dead blocks from the pending deadlist to the on-disk deadlist 546 * - release hold from dsl_dataset_dirty() 547 */ 548 while ((ds = list_remove_head(&synced_datasets)) != NULL) { 549 objset_t *os = ds->ds_objset; 550 bplist_iterate(&ds->ds_pending_deadlist, 551 deadlist_enqueue_cb, &ds->ds_deadlist, tx); 552 ASSERT(!dmu_objset_is_dirty(os, txg)); 553 dmu_buf_rele(ds->ds_dbuf, ds); 554 } 555 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) { 556 dsl_dir_sync(dd, tx); 557 } 558 559 /* 560 * The MOS's space is accounted for in the pool/$MOS 561 * (dp_mos_dir). We can't modify the mos while we're syncing 562 * it, so we remember the deltas and apply them here. 563 */ 564 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 || 565 dp->dp_mos_uncompressed_delta != 0) { 566 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD, 567 dp->dp_mos_used_delta, 568 dp->dp_mos_compressed_delta, 569 dp->dp_mos_uncompressed_delta, tx); 570 dp->dp_mos_used_delta = 0; 571 dp->dp_mos_compressed_delta = 0; 572 dp->dp_mos_uncompressed_delta = 0; 573 } 574 575 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL || 576 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) { 577 dsl_pool_sync_mos(dp, tx); 578 } 579 580 /* 581 * If we modify a dataset in the same txg that we want to destroy it, 582 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it. 583 * dsl_dir_destroy_check() will fail if there are unexpected holds. 584 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf 585 * and clearing the hold on it) before we process the sync_tasks. 586 * The MOS data dirtied by the sync_tasks will be synced on the next 587 * pass. 588 */ 589 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) { 590 dsl_sync_task_t *dst; 591 /* 592 * No more sync tasks should have been added while we 593 * were syncing. 594 */ 595 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1); 596 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL) 597 dsl_sync_task_sync(dst, tx); 598 } 599 600 dmu_tx_commit(tx); 601 602 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg); 603} 604 605void 606dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg) 607{ 608 zilog_t *zilog; 609 610 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) { 611 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 612 zil_clean(zilog, txg); 613 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg)); 614 dmu_buf_rele(ds->ds_dbuf, zilog); 615 } 616 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg)); 617} 618 619/* 620 * TRUE if the current thread is the tx_sync_thread or if we 621 * are being called from SPA context during pool initialization. 622 */ 623int 624dsl_pool_sync_context(dsl_pool_t *dp) 625{ 626 return (curthread == dp->dp_tx.tx_sync_thread || 627 spa_is_initializing(dp->dp_spa)); 628} 629 630uint64_t 631dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree) 632{ 633 uint64_t space, resv; 634 635 /* 636 * Reserve about 1.6% (1/64), or at least 32MB, for allocation 637 * efficiency. 638 * XXX The intent log is not accounted for, so it must fit 639 * within this slop. 640 * 641 * If we're trying to assess whether it's OK to do a free, 642 * cut the reservation in half to allow forward progress 643 * (e.g. make it possible to rm(1) files from a full pool). 644 */ 645 space = spa_get_dspace(dp->dp_spa); 646 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1); 647 if (netfree) 648 resv >>= 1; 649 650 return (space - resv); 651} 652 653boolean_t 654dsl_pool_need_dirty_delay(dsl_pool_t *dp) 655{ 656 uint64_t delay_min_bytes = 657 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100; 658 boolean_t rv; 659 660 mutex_enter(&dp->dp_lock); 661 if (dp->dp_dirty_total > zfs_dirty_data_sync) 662 txg_kick(dp); 663 rv = (dp->dp_dirty_total > delay_min_bytes); 664 mutex_exit(&dp->dp_lock); 665 return (rv); 666} 667 668void 669dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx) 670{ 671 if (space > 0) { 672 mutex_enter(&dp->dp_lock); 673 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space; 674 dsl_pool_dirty_delta(dp, space); 675 mutex_exit(&dp->dp_lock); 676 } 677} 678 679void 680dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg) 681{ 682 ASSERT3S(space, >=, 0); 683 if (space == 0) 684 return; 685 mutex_enter(&dp->dp_lock); 686 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) { 687 /* XXX writing something we didn't dirty? */ 688 space = dp->dp_dirty_pertxg[txg & TXG_MASK]; 689 } 690 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space); 691 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space; 692 ASSERT3U(dp->dp_dirty_total, >=, space); 693 dsl_pool_dirty_delta(dp, -space); 694 mutex_exit(&dp->dp_lock); 695} 696 697/* ARGSUSED */ 698static int 699upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 700{ 701 dmu_tx_t *tx = arg; 702 dsl_dataset_t *ds, *prev = NULL; 703 int err; 704 705 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 706 if (err) 707 return (err); 708 709 while (ds->ds_phys->ds_prev_snap_obj != 0) { 710 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj, 711 FTAG, &prev); 712 if (err) { 713 dsl_dataset_rele(ds, FTAG); 714 return (err); 715 } 716 717 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) 718 break; 719 dsl_dataset_rele(ds, FTAG); 720 ds = prev; 721 prev = NULL; 722 } 723 724 if (prev == NULL) { 725 prev = dp->dp_origin_snap; 726 727 /* 728 * The $ORIGIN can't have any data, or the accounting 729 * will be wrong. 730 */ 731 ASSERT0(prev->ds_phys->ds_bp.blk_birth); 732 733 /* The origin doesn't get attached to itself */ 734 if (ds->ds_object == prev->ds_object) { 735 dsl_dataset_rele(ds, FTAG); 736 return (0); 737 } 738 739 dmu_buf_will_dirty(ds->ds_dbuf, tx); 740 ds->ds_phys->ds_prev_snap_obj = prev->ds_object; 741 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg; 742 743 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx); 744 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object; 745 746 dmu_buf_will_dirty(prev->ds_dbuf, tx); 747 prev->ds_phys->ds_num_children++; 748 749 if (ds->ds_phys->ds_next_snap_obj == 0) { 750 ASSERT(ds->ds_prev == NULL); 751 VERIFY0(dsl_dataset_hold_obj(dp, 752 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev)); 753 } 754 } 755 756 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object); 757 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object); 758 759 if (prev->ds_phys->ds_next_clones_obj == 0) { 760 dmu_buf_will_dirty(prev->ds_dbuf, tx); 761 prev->ds_phys->ds_next_clones_obj = 762 zap_create(dp->dp_meta_objset, 763 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx); 764 } 765 VERIFY0(zap_add_int(dp->dp_meta_objset, 766 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx)); 767 768 dsl_dataset_rele(ds, FTAG); 769 if (prev != dp->dp_origin_snap) 770 dsl_dataset_rele(prev, FTAG); 771 return (0); 772} 773 774void 775dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx) 776{ 777 ASSERT(dmu_tx_is_syncing(tx)); 778 ASSERT(dp->dp_origin_snap != NULL); 779 780 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb, 781 tx, DS_FIND_CHILDREN)); 782} 783 784/* ARGSUSED */ 785static int 786upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg) 787{ 788 dmu_tx_t *tx = arg; 789 objset_t *mos = dp->dp_meta_objset; 790 791 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) { 792 dsl_dataset_t *origin; 793 794 VERIFY0(dsl_dataset_hold_obj(dp, 795 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin)); 796 797 if (origin->ds_dir->dd_phys->dd_clones == 0) { 798 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx); 799 origin->ds_dir->dd_phys->dd_clones = zap_create(mos, 800 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx); 801 } 802 803 VERIFY0(zap_add_int(dp->dp_meta_objset, 804 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx)); 805 806 dsl_dataset_rele(origin, FTAG); 807 } 808 return (0); 809} 810 811void 812dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx) 813{ 814 ASSERT(dmu_tx_is_syncing(tx)); 815 uint64_t obj; 816 817 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx); 818 VERIFY0(dsl_pool_open_special_dir(dp, 819 FREE_DIR_NAME, &dp->dp_free_dir)); 820 821 /* 822 * We can't use bpobj_alloc(), because spa_version() still 823 * returns the old version, and we need a new-version bpobj with 824 * subobj support. So call dmu_object_alloc() directly. 825 */ 826 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ, 827 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx); 828 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 829 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx)); 830 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj)); 831 832 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 833 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN)); 834} 835 836void 837dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx) 838{ 839 uint64_t dsobj; 840 dsl_dataset_t *ds; 841 842 ASSERT(dmu_tx_is_syncing(tx)); 843 ASSERT(dp->dp_origin_snap == NULL); 844 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER)); 845 846 /* create the origin dir, ds, & snap-ds */ 847 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME, 848 NULL, 0, kcred, tx); 849 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 850 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx); 851 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj, 852 dp, &dp->dp_origin_snap)); 853 dsl_dataset_rele(ds, FTAG); 854} 855 856taskq_t * 857dsl_pool_vnrele_taskq(dsl_pool_t *dp) 858{ 859 return (dp->dp_vnrele_taskq); 860} 861 862/* 863 * Walk through the pool-wide zap object of temporary snapshot user holds 864 * and release them. 865 */ 866void 867dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp) 868{ 869 zap_attribute_t za; 870 zap_cursor_t zc; 871 objset_t *mos = dp->dp_meta_objset; 872 uint64_t zapobj = dp->dp_tmp_userrefs_obj; 873 nvlist_t *holds; 874 875 if (zapobj == 0) 876 return; 877 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS); 878 879 holds = fnvlist_alloc(); 880 881 for (zap_cursor_init(&zc, mos, zapobj); 882 zap_cursor_retrieve(&zc, &za) == 0; 883 zap_cursor_advance(&zc)) { 884 char *htag; 885 nvlist_t *tags; 886 887 htag = strchr(za.za_name, '-'); 888 *htag = '\0'; 889 ++htag; 890 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) { 891 tags = fnvlist_alloc(); 892 fnvlist_add_boolean(tags, htag); 893 fnvlist_add_nvlist(holds, za.za_name, tags); 894 fnvlist_free(tags); 895 } else { 896 fnvlist_add_boolean(tags, htag); 897 } 898 } 899 dsl_dataset_user_release_tmp(dp, holds); 900 fnvlist_free(holds); 901 zap_cursor_fini(&zc); 902} 903 904/* 905 * Create the pool-wide zap object for storing temporary snapshot holds. 906 */ 907void 908dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx) 909{ 910 objset_t *mos = dp->dp_meta_objset; 911 912 ASSERT(dp->dp_tmp_userrefs_obj == 0); 913 ASSERT(dmu_tx_is_syncing(tx)); 914 915 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS, 916 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx); 917} 918 919static int 920dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj, 921 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding) 922{ 923 objset_t *mos = dp->dp_meta_objset; 924 uint64_t zapobj = dp->dp_tmp_userrefs_obj; 925 char *name; 926 int error; 927 928 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS); 929 ASSERT(dmu_tx_is_syncing(tx)); 930 931 /* 932 * If the pool was created prior to SPA_VERSION_USERREFS, the 933 * zap object for temporary holds might not exist yet. 934 */ 935 if (zapobj == 0) { 936 if (holding) { 937 dsl_pool_user_hold_create_obj(dp, tx); 938 zapobj = dp->dp_tmp_userrefs_obj; 939 } else { 940 return (SET_ERROR(ENOENT)); 941 } 942 } 943 944 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag); 945 if (holding) 946 error = zap_add(mos, zapobj, name, 8, 1, &now, tx); 947 else 948 error = zap_remove(mos, zapobj, name, tx); 949 strfree(name); 950 951 return (error); 952} 953 954/* 955 * Add a temporary hold for the given dataset object and tag. 956 */ 957int 958dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag, 959 uint64_t now, dmu_tx_t *tx) 960{ 961 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE)); 962} 963 964/* 965 * Release a temporary hold for the given dataset object and tag. 966 */ 967int 968dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag, 969 dmu_tx_t *tx) 970{ 971 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0, 972 tx, B_FALSE)); 973} 974 975/* 976 * DSL Pool Configuration Lock 977 * 978 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset 979 * creation / destruction / rename / property setting). It must be held for 980 * read to hold a dataset or dsl_dir. I.e. you must call 981 * dsl_pool_config_enter() or dsl_pool_hold() before calling 982 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock 983 * must be held continuously until all datasets and dsl_dirs are released. 984 * 985 * The only exception to this rule is that if a "long hold" is placed on 986 * a dataset, then the dp_config_rwlock may be dropped while the dataset 987 * is still held. The long hold will prevent the dataset from being 988 * destroyed -- the destroy will fail with EBUSY. A long hold can be 989 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset 990 * (by calling dsl_{dataset,objset}_{try}own{_obj}). 991 * 992 * Legitimate long-holders (including owners) should be long-running, cancelable 993 * tasks that should cause "zfs destroy" to fail. This includes DMU 994 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open), 995 * "zfs send", and "zfs diff". There are several other long-holders whose 996 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()). 997 * 998 * The usual formula for long-holding would be: 999 * dsl_pool_hold() 1000 * dsl_dataset_hold() 1001 * ... perform checks ... 1002 * dsl_dataset_long_hold() 1003 * dsl_pool_rele() 1004 * ... perform long-running task ... 1005 * dsl_dataset_long_rele() 1006 * dsl_dataset_rele() 1007 * 1008 * Note that when the long hold is released, the dataset is still held but 1009 * the pool is not held. The dataset may change arbitrarily during this time 1010 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the 1011 * dataset except release it. 1012 * 1013 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only 1014 * or modifying operations. 1015 * 1016 * Modifying operations should generally use dsl_sync_task(). The synctask 1017 * infrastructure enforces proper locking strategy with respect to the 1018 * dp_config_rwlock. See the comment above dsl_sync_task() for details. 1019 * 1020 * Read-only operations will manually hold the pool, then the dataset, obtain 1021 * information from the dataset, then release the pool and dataset. 1022 * dmu_objset_{hold,rele}() are convenience routines that also do the pool 1023 * hold/rele. 1024 */ 1025 1026int 1027dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp) 1028{ 1029 spa_t *spa; 1030 int error; 1031 1032 error = spa_open(name, &spa, tag); 1033 if (error == 0) { 1034 *dp = spa_get_dsl(spa); 1035 dsl_pool_config_enter(*dp, tag); 1036 } 1037 return (error); 1038} 1039 1040void 1041dsl_pool_rele(dsl_pool_t *dp, void *tag) 1042{ 1043 dsl_pool_config_exit(dp, tag); 1044 spa_close(dp->dp_spa, tag); 1045} 1046 1047void 1048dsl_pool_config_enter(dsl_pool_t *dp, void *tag) 1049{ 1050 /* 1051 * We use a "reentrant" reader-writer lock, but not reentrantly. 1052 * 1053 * The rrwlock can (with the track_all flag) track all reading threads, 1054 * which is very useful for debugging which code path failed to release 1055 * the lock, and for verifying that the *current* thread does hold 1056 * the lock. 1057 * 1058 * (Unlike a rwlock, which knows that N threads hold it for 1059 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE 1060 * if any thread holds it for read, even if this thread doesn't). 1061 */ 1062 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER)); 1063 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag); 1064} 1065 1066void 1067dsl_pool_config_exit(dsl_pool_t *dp, void *tag) 1068{ 1069 rrw_exit(&dp->dp_config_rwlock, tag); 1070} 1071 1072boolean_t 1073dsl_pool_config_held(dsl_pool_t *dp) 1074{ 1075 return (RRW_LOCK_HELD(&dp->dp_config_rwlock)); 1076} 1077