dsl_pool.c revision 269416
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 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 /* 282 * Note: errors ignored, because the leak dir will not exist if we 283 * have not encountered a leak yet. 284 */ 285 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME, 286 &dp->dp_leak_dir); 287 288 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) { 289 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 290 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1, 291 &dp->dp_bptree_obj); 292 if (err != 0) 293 goto out; 294 } 295 296 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) { 297 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 298 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1, 299 &dp->dp_empty_bpobj); 300 if (err != 0) 301 goto out; 302 } 303 304 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 305 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1, 306 &dp->dp_tmp_userrefs_obj); 307 if (err == ENOENT) 308 err = 0; 309 if (err) 310 goto out; 311 312 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg); 313 314out: 315 rrw_exit(&dp->dp_config_rwlock, FTAG); 316 return (err); 317} 318 319void 320dsl_pool_close(dsl_pool_t *dp) 321{ 322 /* 323 * Drop our references from dsl_pool_open(). 324 * 325 * Since we held the origin_snap from "syncing" context (which 326 * includes pool-opening context), it actually only got a "ref" 327 * and not a hold, so just drop that here. 328 */ 329 if (dp->dp_origin_snap) 330 dsl_dataset_rele(dp->dp_origin_snap, dp); 331 if (dp->dp_mos_dir) 332 dsl_dir_rele(dp->dp_mos_dir, dp); 333 if (dp->dp_free_dir) 334 dsl_dir_rele(dp->dp_free_dir, dp); 335 if (dp->dp_leak_dir) 336 dsl_dir_rele(dp->dp_leak_dir, dp); 337 if (dp->dp_root_dir) 338 dsl_dir_rele(dp->dp_root_dir, dp); 339 340 bpobj_close(&dp->dp_free_bpobj); 341 342 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */ 343 if (dp->dp_meta_objset) 344 dmu_objset_evict(dp->dp_meta_objset); 345 346 txg_list_destroy(&dp->dp_dirty_datasets); 347 txg_list_destroy(&dp->dp_dirty_zilogs); 348 txg_list_destroy(&dp->dp_sync_tasks); 349 txg_list_destroy(&dp->dp_dirty_dirs); 350 351 arc_flush(dp->dp_spa); 352 txg_fini(dp); 353 dsl_scan_fini(dp); 354 rrw_destroy(&dp->dp_config_rwlock); 355 mutex_destroy(&dp->dp_lock); 356 taskq_destroy(dp->dp_vnrele_taskq); 357 if (dp->dp_blkstats) 358 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t)); 359 kmem_free(dp, sizeof (dsl_pool_t)); 360} 361 362dsl_pool_t * 363dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg) 364{ 365 int err; 366 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg); 367 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg); 368 objset_t *os; 369 dsl_dataset_t *ds; 370 uint64_t obj; 371 372 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 373 374 /* create and open the MOS (meta-objset) */ 375 dp->dp_meta_objset = dmu_objset_create_impl(spa, 376 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx); 377 378 /* create the pool directory */ 379 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 380 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx); 381 ASSERT0(err); 382 383 /* Initialize scan structures */ 384 VERIFY0(dsl_scan_init(dp, txg)); 385 386 /* create and open the root dir */ 387 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx); 388 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, 389 NULL, dp, &dp->dp_root_dir)); 390 391 /* create and open the meta-objset dir */ 392 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx); 393 VERIFY0(dsl_pool_open_special_dir(dp, 394 MOS_DIR_NAME, &dp->dp_mos_dir)); 395 396 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) { 397 /* create and open the free dir */ 398 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, 399 FREE_DIR_NAME, tx); 400 VERIFY0(dsl_pool_open_special_dir(dp, 401 FREE_DIR_NAME, &dp->dp_free_dir)); 402 403 /* create and open the free_bplist */ 404 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx); 405 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 406 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0); 407 VERIFY0(bpobj_open(&dp->dp_free_bpobj, 408 dp->dp_meta_objset, obj)); 409 } 410 411 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) 412 dsl_pool_create_origin(dp, tx); 413 414 /* create the root dataset */ 415 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx); 416 417 /* create the root objset */ 418 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds)); 419 os = dmu_objset_create_impl(dp->dp_spa, ds, 420 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx); 421#ifdef _KERNEL 422 zfs_create_fs(os, kcred, zplprops, tx); 423#endif 424 dsl_dataset_rele(ds, FTAG); 425 426 dmu_tx_commit(tx); 427 428 rrw_exit(&dp->dp_config_rwlock, FTAG); 429 430 return (dp); 431} 432 433/* 434 * Account for the meta-objset space in its placeholder dsl_dir. 435 */ 436void 437dsl_pool_mos_diduse_space(dsl_pool_t *dp, 438 int64_t used, int64_t comp, int64_t uncomp) 439{ 440 ASSERT3U(comp, ==, uncomp); /* it's all metadata */ 441 mutex_enter(&dp->dp_lock); 442 dp->dp_mos_used_delta += used; 443 dp->dp_mos_compressed_delta += comp; 444 dp->dp_mos_uncompressed_delta += uncomp; 445 mutex_exit(&dp->dp_lock); 446} 447 448static int 449deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 450{ 451 dsl_deadlist_t *dl = arg; 452 dsl_deadlist_insert(dl, bp, tx); 453 return (0); 454} 455 456static void 457dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx) 458{ 459 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 460 dmu_objset_sync(dp->dp_meta_objset, zio, tx); 461 VERIFY0(zio_wait(zio)); 462 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", ""); 463 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp); 464} 465 466static void 467dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta) 468{ 469 ASSERT(MUTEX_HELD(&dp->dp_lock)); 470 471 if (delta < 0) 472 ASSERT3U(-delta, <=, dp->dp_dirty_total); 473 474 dp->dp_dirty_total += delta; 475 476 /* 477 * Note: we signal even when increasing dp_dirty_total. 478 * This ensures forward progress -- each thread wakes the next waiter. 479 */ 480 if (dp->dp_dirty_total <= zfs_dirty_data_max) 481 cv_signal(&dp->dp_spaceavail_cv); 482} 483 484void 485dsl_pool_sync(dsl_pool_t *dp, uint64_t txg) 486{ 487 zio_t *zio; 488 dmu_tx_t *tx; 489 dsl_dir_t *dd; 490 dsl_dataset_t *ds; 491 objset_t *mos = dp->dp_meta_objset; 492 list_t synced_datasets; 493 494 list_create(&synced_datasets, sizeof (dsl_dataset_t), 495 offsetof(dsl_dataset_t, ds_synced_link)); 496 497 tx = dmu_tx_create_assigned(dp, txg); 498 499 /* 500 * Write out all dirty blocks of dirty datasets. 501 */ 502 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 503 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) { 504 /* 505 * We must not sync any non-MOS datasets twice, because 506 * we may have taken a snapshot of them. However, we 507 * may sync newly-created datasets on pass 2. 508 */ 509 ASSERT(!list_link_active(&ds->ds_synced_link)); 510 list_insert_tail(&synced_datasets, ds); 511 dsl_dataset_sync(ds, zio, tx); 512 } 513 VERIFY0(zio_wait(zio)); 514 515 /* 516 * We have written all of the accounted dirty data, so our 517 * dp_space_towrite should now be zero. However, some seldom-used 518 * code paths do not adhere to this (e.g. dbuf_undirty(), also 519 * rounding error in dbuf_write_physdone). 520 * Shore up the accounting of any dirtied space now. 521 */ 522 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg); 523 524 /* 525 * After the data blocks have been written (ensured by the zio_wait() 526 * above), update the user/group space accounting. 527 */ 528 for (ds = list_head(&synced_datasets); ds != NULL; 529 ds = list_next(&synced_datasets, ds)) { 530 dmu_objset_do_userquota_updates(ds->ds_objset, tx); 531 } 532 533 /* 534 * Sync the datasets again to push out the changes due to 535 * userspace updates. This must be done before we process the 536 * sync tasks, so that any snapshots will have the correct 537 * user accounting information (and we won't get confused 538 * about which blocks are part of the snapshot). 539 */ 540 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED); 541 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) { 542 ASSERT(list_link_active(&ds->ds_synced_link)); 543 dmu_buf_rele(ds->ds_dbuf, ds); 544 dsl_dataset_sync(ds, zio, tx); 545 } 546 VERIFY0(zio_wait(zio)); 547 548 /* 549 * Now that the datasets have been completely synced, we can 550 * clean up our in-memory structures accumulated while syncing: 551 * 552 * - move dead blocks from the pending deadlist to the on-disk deadlist 553 * - release hold from dsl_dataset_dirty() 554 */ 555 while ((ds = list_remove_head(&synced_datasets)) != NULL) { 556 objset_t *os = ds->ds_objset; 557 bplist_iterate(&ds->ds_pending_deadlist, 558 deadlist_enqueue_cb, &ds->ds_deadlist, tx); 559 ASSERT(!dmu_objset_is_dirty(os, txg)); 560 dmu_buf_rele(ds->ds_dbuf, ds); 561 } 562 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) { 563 dsl_dir_sync(dd, tx); 564 } 565 566 /* 567 * The MOS's space is accounted for in the pool/$MOS 568 * (dp_mos_dir). We can't modify the mos while we're syncing 569 * it, so we remember the deltas and apply them here. 570 */ 571 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 || 572 dp->dp_mos_uncompressed_delta != 0) { 573 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD, 574 dp->dp_mos_used_delta, 575 dp->dp_mos_compressed_delta, 576 dp->dp_mos_uncompressed_delta, tx); 577 dp->dp_mos_used_delta = 0; 578 dp->dp_mos_compressed_delta = 0; 579 dp->dp_mos_uncompressed_delta = 0; 580 } 581 582 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL || 583 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) { 584 dsl_pool_sync_mos(dp, tx); 585 } 586 587 /* 588 * If we modify a dataset in the same txg that we want to destroy it, 589 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it. 590 * dsl_dir_destroy_check() will fail if there are unexpected holds. 591 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf 592 * and clearing the hold on it) before we process the sync_tasks. 593 * The MOS data dirtied by the sync_tasks will be synced on the next 594 * pass. 595 */ 596 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) { 597 dsl_sync_task_t *dst; 598 /* 599 * No more sync tasks should have been added while we 600 * were syncing. 601 */ 602 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1); 603 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL) 604 dsl_sync_task_sync(dst, tx); 605 } 606 607 dmu_tx_commit(tx); 608 609 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg); 610} 611 612void 613dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg) 614{ 615 zilog_t *zilog; 616 617 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) { 618 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 619 zil_clean(zilog, txg); 620 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg)); 621 dmu_buf_rele(ds->ds_dbuf, zilog); 622 } 623 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg)); 624} 625 626/* 627 * TRUE if the current thread is the tx_sync_thread or if we 628 * are being called from SPA context during pool initialization. 629 */ 630int 631dsl_pool_sync_context(dsl_pool_t *dp) 632{ 633 return (curthread == dp->dp_tx.tx_sync_thread || 634 spa_is_initializing(dp->dp_spa)); 635} 636 637uint64_t 638dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree) 639{ 640 uint64_t space, resv; 641 642 /* 643 * If we're trying to assess whether it's OK to do a free, 644 * cut the reservation in half to allow forward progress 645 * (e.g. make it possible to rm(1) files from a full pool). 646 */ 647 space = spa_get_dspace(dp->dp_spa); 648 resv = spa_get_slop_space(dp->dp_spa); 649 if (netfree) 650 resv >>= 1; 651 652 return (space - resv); 653} 654 655boolean_t 656dsl_pool_need_dirty_delay(dsl_pool_t *dp) 657{ 658 uint64_t delay_min_bytes = 659 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100; 660 boolean_t rv; 661 662 mutex_enter(&dp->dp_lock); 663 if (dp->dp_dirty_total > zfs_dirty_data_sync) 664 txg_kick(dp); 665 rv = (dp->dp_dirty_total > delay_min_bytes); 666 mutex_exit(&dp->dp_lock); 667 return (rv); 668} 669 670void 671dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx) 672{ 673 if (space > 0) { 674 mutex_enter(&dp->dp_lock); 675 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space; 676 dsl_pool_dirty_delta(dp, space); 677 mutex_exit(&dp->dp_lock); 678 } 679} 680 681void 682dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg) 683{ 684 ASSERT3S(space, >=, 0); 685 if (space == 0) 686 return; 687 mutex_enter(&dp->dp_lock); 688 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) { 689 /* XXX writing something we didn't dirty? */ 690 space = dp->dp_dirty_pertxg[txg & TXG_MASK]; 691 } 692 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space); 693 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space; 694 ASSERT3U(dp->dp_dirty_total, >=, space); 695 dsl_pool_dirty_delta(dp, -space); 696 mutex_exit(&dp->dp_lock); 697} 698 699/* ARGSUSED */ 700static int 701upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 702{ 703 dmu_tx_t *tx = arg; 704 dsl_dataset_t *ds, *prev = NULL; 705 int err; 706 707 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 708 if (err) 709 return (err); 710 711 while (ds->ds_phys->ds_prev_snap_obj != 0) { 712 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj, 713 FTAG, &prev); 714 if (err) { 715 dsl_dataset_rele(ds, FTAG); 716 return (err); 717 } 718 719 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) 720 break; 721 dsl_dataset_rele(ds, FTAG); 722 ds = prev; 723 prev = NULL; 724 } 725 726 if (prev == NULL) { 727 prev = dp->dp_origin_snap; 728 729 /* 730 * The $ORIGIN can't have any data, or the accounting 731 * will be wrong. 732 */ 733 ASSERT0(prev->ds_phys->ds_bp.blk_birth); 734 735 /* The origin doesn't get attached to itself */ 736 if (ds->ds_object == prev->ds_object) { 737 dsl_dataset_rele(ds, FTAG); 738 return (0); 739 } 740 741 dmu_buf_will_dirty(ds->ds_dbuf, tx); 742 ds->ds_phys->ds_prev_snap_obj = prev->ds_object; 743 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg; 744 745 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx); 746 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object; 747 748 dmu_buf_will_dirty(prev->ds_dbuf, tx); 749 prev->ds_phys->ds_num_children++; 750 751 if (ds->ds_phys->ds_next_snap_obj == 0) { 752 ASSERT(ds->ds_prev == NULL); 753 VERIFY0(dsl_dataset_hold_obj(dp, 754 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev)); 755 } 756 } 757 758 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object); 759 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object); 760 761 if (prev->ds_phys->ds_next_clones_obj == 0) { 762 dmu_buf_will_dirty(prev->ds_dbuf, tx); 763 prev->ds_phys->ds_next_clones_obj = 764 zap_create(dp->dp_meta_objset, 765 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx); 766 } 767 VERIFY0(zap_add_int(dp->dp_meta_objset, 768 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx)); 769 770 dsl_dataset_rele(ds, FTAG); 771 if (prev != dp->dp_origin_snap) 772 dsl_dataset_rele(prev, FTAG); 773 return (0); 774} 775 776void 777dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx) 778{ 779 ASSERT(dmu_tx_is_syncing(tx)); 780 ASSERT(dp->dp_origin_snap != NULL); 781 782 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb, 783 tx, DS_FIND_CHILDREN)); 784} 785 786/* ARGSUSED */ 787static int 788upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg) 789{ 790 dmu_tx_t *tx = arg; 791 objset_t *mos = dp->dp_meta_objset; 792 793 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) { 794 dsl_dataset_t *origin; 795 796 VERIFY0(dsl_dataset_hold_obj(dp, 797 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin)); 798 799 if (origin->ds_dir->dd_phys->dd_clones == 0) { 800 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx); 801 origin->ds_dir->dd_phys->dd_clones = zap_create(mos, 802 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx); 803 } 804 805 VERIFY0(zap_add_int(dp->dp_meta_objset, 806 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx)); 807 808 dsl_dataset_rele(origin, FTAG); 809 } 810 return (0); 811} 812 813void 814dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx) 815{ 816 ASSERT(dmu_tx_is_syncing(tx)); 817 uint64_t obj; 818 819 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx); 820 VERIFY0(dsl_pool_open_special_dir(dp, 821 FREE_DIR_NAME, &dp->dp_free_dir)); 822 823 /* 824 * We can't use bpobj_alloc(), because spa_version() still 825 * returns the old version, and we need a new-version bpobj with 826 * subobj support. So call dmu_object_alloc() directly. 827 */ 828 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ, 829 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx); 830 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 831 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx)); 832 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj)); 833 834 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 835 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN)); 836} 837 838void 839dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx) 840{ 841 uint64_t dsobj; 842 dsl_dataset_t *ds; 843 844 ASSERT(dmu_tx_is_syncing(tx)); 845 ASSERT(dp->dp_origin_snap == NULL); 846 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER)); 847 848 /* create the origin dir, ds, & snap-ds */ 849 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME, 850 NULL, 0, kcred, tx); 851 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 852 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx); 853 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj, 854 dp, &dp->dp_origin_snap)); 855 dsl_dataset_rele(ds, FTAG); 856} 857 858taskq_t * 859dsl_pool_vnrele_taskq(dsl_pool_t *dp) 860{ 861 return (dp->dp_vnrele_taskq); 862} 863 864/* 865 * Walk through the pool-wide zap object of temporary snapshot user holds 866 * and release them. 867 */ 868void 869dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp) 870{ 871 zap_attribute_t za; 872 zap_cursor_t zc; 873 objset_t *mos = dp->dp_meta_objset; 874 uint64_t zapobj = dp->dp_tmp_userrefs_obj; 875 nvlist_t *holds; 876 877 if (zapobj == 0) 878 return; 879 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS); 880 881 holds = fnvlist_alloc(); 882 883 for (zap_cursor_init(&zc, mos, zapobj); 884 zap_cursor_retrieve(&zc, &za) == 0; 885 zap_cursor_advance(&zc)) { 886 char *htag; 887 nvlist_t *tags; 888 889 htag = strchr(za.za_name, '-'); 890 *htag = '\0'; 891 ++htag; 892 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) { 893 tags = fnvlist_alloc(); 894 fnvlist_add_boolean(tags, htag); 895 fnvlist_add_nvlist(holds, za.za_name, tags); 896 fnvlist_free(tags); 897 } else { 898 fnvlist_add_boolean(tags, htag); 899 } 900 } 901 dsl_dataset_user_release_tmp(dp, holds); 902 fnvlist_free(holds); 903 zap_cursor_fini(&zc); 904} 905 906/* 907 * Create the pool-wide zap object for storing temporary snapshot holds. 908 */ 909void 910dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx) 911{ 912 objset_t *mos = dp->dp_meta_objset; 913 914 ASSERT(dp->dp_tmp_userrefs_obj == 0); 915 ASSERT(dmu_tx_is_syncing(tx)); 916 917 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS, 918 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx); 919} 920 921static int 922dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj, 923 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding) 924{ 925 objset_t *mos = dp->dp_meta_objset; 926 uint64_t zapobj = dp->dp_tmp_userrefs_obj; 927 char *name; 928 int error; 929 930 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS); 931 ASSERT(dmu_tx_is_syncing(tx)); 932 933 /* 934 * If the pool was created prior to SPA_VERSION_USERREFS, the 935 * zap object for temporary holds might not exist yet. 936 */ 937 if (zapobj == 0) { 938 if (holding) { 939 dsl_pool_user_hold_create_obj(dp, tx); 940 zapobj = dp->dp_tmp_userrefs_obj; 941 } else { 942 return (SET_ERROR(ENOENT)); 943 } 944 } 945 946 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag); 947 if (holding) 948 error = zap_add(mos, zapobj, name, 8, 1, &now, tx); 949 else 950 error = zap_remove(mos, zapobj, name, tx); 951 strfree(name); 952 953 return (error); 954} 955 956/* 957 * Add a temporary hold for the given dataset object and tag. 958 */ 959int 960dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag, 961 uint64_t now, dmu_tx_t *tx) 962{ 963 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE)); 964} 965 966/* 967 * Release a temporary hold for the given dataset object and tag. 968 */ 969int 970dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag, 971 dmu_tx_t *tx) 972{ 973 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0, 974 tx, B_FALSE)); 975} 976 977/* 978 * DSL Pool Configuration Lock 979 * 980 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset 981 * creation / destruction / rename / property setting). It must be held for 982 * read to hold a dataset or dsl_dir. I.e. you must call 983 * dsl_pool_config_enter() or dsl_pool_hold() before calling 984 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock 985 * must be held continuously until all datasets and dsl_dirs are released. 986 * 987 * The only exception to this rule is that if a "long hold" is placed on 988 * a dataset, then the dp_config_rwlock may be dropped while the dataset 989 * is still held. The long hold will prevent the dataset from being 990 * destroyed -- the destroy will fail with EBUSY. A long hold can be 991 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset 992 * (by calling dsl_{dataset,objset}_{try}own{_obj}). 993 * 994 * Legitimate long-holders (including owners) should be long-running, cancelable 995 * tasks that should cause "zfs destroy" to fail. This includes DMU 996 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open), 997 * "zfs send", and "zfs diff". There are several other long-holders whose 998 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()). 999 * 1000 * The usual formula for long-holding would be: 1001 * dsl_pool_hold() 1002 * dsl_dataset_hold() 1003 * ... perform checks ... 1004 * dsl_dataset_long_hold() 1005 * dsl_pool_rele() 1006 * ... perform long-running task ... 1007 * dsl_dataset_long_rele() 1008 * dsl_dataset_rele() 1009 * 1010 * Note that when the long hold is released, the dataset is still held but 1011 * the pool is not held. The dataset may change arbitrarily during this time 1012 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the 1013 * dataset except release it. 1014 * 1015 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only 1016 * or modifying operations. 1017 * 1018 * Modifying operations should generally use dsl_sync_task(). The synctask 1019 * infrastructure enforces proper locking strategy with respect to the 1020 * dp_config_rwlock. See the comment above dsl_sync_task() for details. 1021 * 1022 * Read-only operations will manually hold the pool, then the dataset, obtain 1023 * information from the dataset, then release the pool and dataset. 1024 * dmu_objset_{hold,rele}() are convenience routines that also do the pool 1025 * hold/rele. 1026 */ 1027 1028int 1029dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp) 1030{ 1031 spa_t *spa; 1032 int error; 1033 1034 error = spa_open(name, &spa, tag); 1035 if (error == 0) { 1036 *dp = spa_get_dsl(spa); 1037 dsl_pool_config_enter(*dp, tag); 1038 } 1039 return (error); 1040} 1041 1042void 1043dsl_pool_rele(dsl_pool_t *dp, void *tag) 1044{ 1045 dsl_pool_config_exit(dp, tag); 1046 spa_close(dp->dp_spa, tag); 1047} 1048 1049void 1050dsl_pool_config_enter(dsl_pool_t *dp, void *tag) 1051{ 1052 /* 1053 * We use a "reentrant" reader-writer lock, but not reentrantly. 1054 * 1055 * The rrwlock can (with the track_all flag) track all reading threads, 1056 * which is very useful for debugging which code path failed to release 1057 * the lock, and for verifying that the *current* thread does hold 1058 * the lock. 1059 * 1060 * (Unlike a rwlock, which knows that N threads hold it for 1061 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE 1062 * if any thread holds it for read, even if this thread doesn't). 1063 */ 1064 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER)); 1065 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag); 1066} 1067 1068void 1069dsl_pool_config_exit(dsl_pool_t *dp, void *tag) 1070{ 1071 rrw_exit(&dp->dp_config_rwlock, tag); 1072} 1073 1074boolean_t 1075dsl_pool_config_held(dsl_pool_t *dp) 1076{ 1077 return (RRW_LOCK_HELD(&dp->dp_config_rwlock)); 1078} 1079