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  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  26  */
  27 
  28 #include <sys/dsl_pool.h>
  29 #include <sys/dsl_dataset.h>
  30 #include <sys/dsl_prop.h>
  31 #include <sys/dsl_dir.h>
  32 #include <sys/dsl_synctask.h>
  33 #include <sys/dsl_scan.h>
  34 #include <sys/dnode.h>
  35 #include <sys/dmu_tx.h>
  36 #include <sys/dmu_objset.h>
  37 #include <sys/arc.h>
  38 #include <sys/zap.h>
  39 #include <sys/zio.h>
  40 #include <sys/zfs_context.h>
  41 #include <sys/fs/zfs.h>
  42 #include <sys/zfs_znode.h>
  43 #include <sys/spa_impl.h>
  44 #include <sys/dsl_deadlist.h>
  45 #include <sys/bptree.h>
  46 #include <sys/zfeature.h>
  47 #include <sys/zil_impl.h>
  48 #include <sys/dsl_userhold.h>
  49 
  50 /*
  51  * ZFS Write Throttle
  52  * ------------------
  53  *
  54  * ZFS must limit the rate of incoming writes to the rate at which it is able
  55  * to sync data modifications to the backend storage. Throttling by too much
  56  * creates an artificial limit; throttling by too little can only be sustained
  57  * for short periods and would lead to highly lumpy performance. On a per-pool
  58  * basis, ZFS tracks the amount of modified (dirty) data. As operations change
  59  * data, the amount of dirty data increases; as ZFS syncs out data, the amount
  60  * of dirty data decreases. When the amount of dirty data exceeds a
  61  * predetermined threshold further modifications are blocked until the amount
  62  * of dirty data decreases (as data is synced out).
  63  *
  64  * The limit on dirty data is tunable, and should be adjusted according to
  65  * both the IO capacity and available memory of the system. The larger the
  66  * window, the more ZFS is able to aggregate and amortize metadata (and data)
  67  * changes. However, memory is a limited resource, and allowing for more dirty
  68  * data comes at the cost of keeping other useful data in memory (for example
  69  * ZFS data cached by the ARC).
  70  *
  71  * Implementation
  72  *
  73  * As buffers are modified dsl_pool_willuse_space() increments both the per-
  74  * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
  75  * dirty space used; dsl_pool_dirty_space() decrements those values as data
  76  * is synced out from dsl_pool_sync(). While only the poolwide value is
  77  * relevant, the per-txg value is useful for debugging. The tunable
  78  * zfs_dirty_data_max determines the dirty space limit. Once that value is
  79  * exceeded, new writes are halted until space frees up.
  80  *
  81  * The zfs_dirty_data_sync tunable dictates the threshold at which we
  82  * ensure that there is a txg syncing (see the comment in txg.c for a full
  83  * description of transaction group stages).
  84  *
  85  * The IO scheduler uses both the dirty space limit and current amount of
  86  * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
  87  * issues. See the comment in vdev_queue.c for details of the IO scheduler.
  88  *
  89  * The delay is also calculated based on the amount of dirty data.  See the
  90  * comment above dmu_tx_delay() for details.
  91  */
  92 
  93 /*
  94  * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
  95  * capped at zfs_dirty_data_max_max.  It can also be overridden in /etc/system.
  96  */
  97 uint64_t zfs_dirty_data_max;
  98 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
  99 int zfs_dirty_data_max_percent = 10;
 100 
 101 /*
 102  * If there is at least this much dirty data, push out a txg.
 103  */
 104 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
 105 
 106 /*
 107  * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
 108  * and delay each transaction.
 109  * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
 110  */
 111 int zfs_delay_min_dirty_percent = 60;
 112 
 113 /*
 114  * This controls how quickly the delay approaches infinity.
 115  * Larger values cause it to delay more for a given amount of dirty data.
 116  * Therefore larger values will cause there to be less dirty data for a
 117  * given throughput.
 118  *
 119  * For the smoothest delay, this value should be about 1 billion divided
 120  * by the maximum number of operations per second.  This will smoothly
 121  * handle between 10x and 1/10th this number.
 122  *
 123  * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
 124  * multiply in dmu_tx_delay().
 125  */
 126 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
 127 
 128 
 129 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
 130 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
 131 
 132 int
 133 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
 134 {
 135         uint64_t obj;
 136         int err;
 137 
 138         err = zap_lookup(dp->dp_meta_objset,
 139             dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
 140             name, sizeof (obj), 1, &obj);
 141         if (err)
 142                 return (err);
 143 
 144         return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
 145 }
 146 
 147 static dsl_pool_t *
 148 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
 149 {
 150         dsl_pool_t *dp;
 151         blkptr_t *bp = spa_get_rootblkptr(spa);
 152 
 153         dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
 154         dp->dp_spa = spa;
 155         dp->dp_meta_rootbp = *bp;
 156         rrw_init(&dp->dp_config_rwlock, B_TRUE);
 157         txg_init(dp, txg);
 158 
 159         txg_list_create(&dp->dp_dirty_datasets,
 160             offsetof(dsl_dataset_t, ds_dirty_link));
 161         txg_list_create(&dp->dp_dirty_zilogs,
 162             offsetof(zilog_t, zl_dirty_link));
 163         txg_list_create(&dp->dp_dirty_dirs,
 164             offsetof(dsl_dir_t, dd_dirty_link));
 165         txg_list_create(&dp->dp_sync_tasks,
 166             offsetof(dsl_sync_task_t, dst_node));
 167 
 168         mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
 169         cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
 170 
 171         dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
 172             1, 4, 0);
 173 
 174         return (dp);
 175 }
 176 
 177 int
 178 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
 179 {
 180         int err;
 181         dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
 182 
 183         err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
 184             &dp->dp_meta_objset);
 185         if (err != 0)
 186                 dsl_pool_close(dp);
 187         else
 188                 *dpp = dp;
 189 
 190         return (err);
 191 }
 192 
 193 int
 194 dsl_pool_open(dsl_pool_t *dp)
 195 {
 196         int err;
 197         dsl_dir_t *dd;
 198         dsl_dataset_t *ds;
 199         uint64_t obj;
 200 
 201         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
 202         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 203             DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
 204             &dp->dp_root_dir_obj);
 205         if (err)
 206                 goto out;
 207 
 208         err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
 209             NULL, dp, &dp->dp_root_dir);
 210         if (err)
 211                 goto out;
 212 
 213         err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
 214         if (err)
 215                 goto out;
 216 
 217         if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
 218                 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
 219                 if (err)
 220                         goto out;
 221                 err = dsl_dataset_hold_obj(dp,
 222                     dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
 223                 if (err == 0) {
 224                         err = dsl_dataset_hold_obj(dp,
 225                             dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
 226                             &dp->dp_origin_snap);
 227                         dsl_dataset_rele(ds, FTAG);
 228                 }
 229                 dsl_dir_rele(dd, dp);
 230                 if (err)
 231                         goto out;
 232         }
 233 
 234         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
 235                 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
 236                     &dp->dp_free_dir);
 237                 if (err)
 238                         goto out;
 239 
 240                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 241                     DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
 242                 if (err)
 243                         goto out;
 244                 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
 245                     dp->dp_meta_objset, obj));
 246         }
 247 
 248         /*
 249          * Note: errors ignored, because the leak dir will not exist if we
 250          * have not encountered a leak yet.
 251          */
 252         (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
 253             &dp->dp_leak_dir);
 254 
 255         if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
 256                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 257                     DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
 258                     &dp->dp_bptree_obj);
 259                 if (err != 0)
 260                         goto out;
 261         }
 262 
 263         if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
 264                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 265                     DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
 266                     &dp->dp_empty_bpobj);
 267                 if (err != 0)
 268                         goto out;
 269         }
 270 
 271         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 272             DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
 273             &dp->dp_tmp_userrefs_obj);
 274         if (err == ENOENT)
 275                 err = 0;
 276         if (err)
 277                 goto out;
 278 
 279         err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
 280 
 281 out:
 282         rrw_exit(&dp->dp_config_rwlock, FTAG);
 283         return (err);
 284 }
 285 
 286 void
 287 dsl_pool_close(dsl_pool_t *dp)
 288 {
 289         /*
 290          * Drop our references from dsl_pool_open().
 291          *
 292          * Since we held the origin_snap from "syncing" context (which
 293          * includes pool-opening context), it actually only got a "ref"
 294          * and not a hold, so just drop that here.
 295          */
 296         if (dp->dp_origin_snap)
 297                 dsl_dataset_rele(dp->dp_origin_snap, dp);
 298         if (dp->dp_mos_dir)
 299                 dsl_dir_rele(dp->dp_mos_dir, dp);
 300         if (dp->dp_free_dir)
 301                 dsl_dir_rele(dp->dp_free_dir, dp);
 302         if (dp->dp_leak_dir)
 303                 dsl_dir_rele(dp->dp_leak_dir, dp);
 304         if (dp->dp_root_dir)
 305                 dsl_dir_rele(dp->dp_root_dir, dp);
 306 
 307         bpobj_close(&dp->dp_free_bpobj);
 308 
 309         /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
 310         if (dp->dp_meta_objset)
 311                 dmu_objset_evict(dp->dp_meta_objset);
 312 
 313         txg_list_destroy(&dp->dp_dirty_datasets);
 314         txg_list_destroy(&dp->dp_dirty_zilogs);
 315         txg_list_destroy(&dp->dp_sync_tasks);
 316         txg_list_destroy(&dp->dp_dirty_dirs);
 317 
 318         /*
 319          * We can't set retry to TRUE since we're explicitly specifying
 320          * a spa to flush. This is good enough; any missed buffers for
 321          * this spa won't cause trouble, and they'll eventually fall
 322          * out of the ARC just like any other unused buffer.
 323          */
 324         arc_flush(dp->dp_spa, FALSE);
 325 
 326         txg_fini(dp);
 327         dsl_scan_fini(dp);
 328         dmu_buf_user_evict_wait();
 329 
 330         rrw_destroy(&dp->dp_config_rwlock);
 331         mutex_destroy(&dp->dp_lock);
 332         taskq_destroy(dp->dp_vnrele_taskq);
 333         if (dp->dp_blkstats)
 334                 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 335         kmem_free(dp, sizeof (dsl_pool_t));
 336 }
 337 
 338 dsl_pool_t *
 339 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
 340 {
 341         int err;
 342         dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
 343         dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
 344         objset_t *os;
 345         dsl_dataset_t *ds;
 346         uint64_t obj;
 347 
 348         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
 349 
 350         /* create and open the MOS (meta-objset) */
 351         dp->dp_meta_objset = dmu_objset_create_impl(spa,
 352             NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
 353 
 354         /* create the pool directory */
 355         err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 356             DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
 357         ASSERT0(err);
 358 
 359         /* Initialize scan structures */
 360         VERIFY0(dsl_scan_init(dp, txg));
 361 
 362         /* create and open the root dir */
 363         dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
 364         VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
 365             NULL, dp, &dp->dp_root_dir));
 366 
 367         /* create and open the meta-objset dir */
 368         (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
 369         VERIFY0(dsl_pool_open_special_dir(dp,
 370             MOS_DIR_NAME, &dp->dp_mos_dir));
 371 
 372         if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
 373                 /* create and open the free dir */
 374                 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
 375                     FREE_DIR_NAME, tx);
 376                 VERIFY0(dsl_pool_open_special_dir(dp,
 377                     FREE_DIR_NAME, &dp->dp_free_dir));
 378 
 379                 /* create and open the free_bplist */
 380                 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
 381                 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 382                     DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
 383                 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
 384                     dp->dp_meta_objset, obj));
 385         }
 386 
 387         if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
 388                 dsl_pool_create_origin(dp, tx);
 389 
 390         /* create the root dataset */
 391         obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
 392 
 393         /* create the root objset */
 394         VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
 395         os = dmu_objset_create_impl(dp->dp_spa, ds,
 396             dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
 397 #ifdef _KERNEL
 398         zfs_create_fs(os, kcred, zplprops, tx);
 399 #endif
 400         dsl_dataset_rele(ds, FTAG);
 401 
 402         dmu_tx_commit(tx);
 403 
 404         rrw_exit(&dp->dp_config_rwlock, FTAG);
 405 
 406         return (dp);
 407 }
 408 
 409 /*
 410  * Account for the meta-objset space in its placeholder dsl_dir.
 411  */
 412 void
 413 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
 414     int64_t used, int64_t comp, int64_t uncomp)
 415 {
 416         ASSERT3U(comp, ==, uncomp); /* it's all metadata */
 417         mutex_enter(&dp->dp_lock);
 418         dp->dp_mos_used_delta += used;
 419         dp->dp_mos_compressed_delta += comp;
 420         dp->dp_mos_uncompressed_delta += uncomp;
 421         mutex_exit(&dp->dp_lock);
 422 }
 423 
 424 static int
 425 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
 426 {
 427         dsl_deadlist_t *dl = arg;
 428         dsl_deadlist_insert(dl, bp, tx);
 429         return (0);
 430 }
 431 
 432 static void
 433 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
 434 {
 435         zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 436         dmu_objset_sync(dp->dp_meta_objset, zio, tx);
 437         VERIFY0(zio_wait(zio));
 438         dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
 439         spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
 440 }
 441 
 442 static void
 443 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
 444 {
 445         ASSERT(MUTEX_HELD(&dp->dp_lock));
 446 
 447         if (delta < 0)
 448                 ASSERT3U(-delta, <=, dp->dp_dirty_total);
 449 
 450         dp->dp_dirty_total += delta;
 451 
 452         /*
 453          * Note: we signal even when increasing dp_dirty_total.
 454          * This ensures forward progress -- each thread wakes the next waiter.
 455          */
 456         if (dp->dp_dirty_total <= zfs_dirty_data_max)
 457                 cv_signal(&dp->dp_spaceavail_cv);
 458 }
 459 
 460 void
 461 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
 462 {
 463         zio_t *zio;
 464         dmu_tx_t *tx;
 465         dsl_dir_t *dd;
 466         dsl_dataset_t *ds;
 467         objset_t *mos = dp->dp_meta_objset;
 468         list_t synced_datasets;
 469 
 470         list_create(&synced_datasets, sizeof (dsl_dataset_t),
 471             offsetof(dsl_dataset_t, ds_synced_link));
 472 
 473         tx = dmu_tx_create_assigned(dp, txg);
 474 
 475         /*
 476          * Write out all dirty blocks of dirty datasets.
 477          */
 478         zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 479         while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
 480                 /*
 481                  * We must not sync any non-MOS datasets twice, because
 482                  * we may have taken a snapshot of them.  However, we
 483                  * may sync newly-created datasets on pass 2.
 484                  */
 485                 ASSERT(!list_link_active(&ds->ds_synced_link));
 486                 list_insert_tail(&synced_datasets, ds);
 487                 dsl_dataset_sync(ds, zio, tx);
 488         }
 489         VERIFY0(zio_wait(zio));
 490 
 491         /*
 492          * We have written all of the accounted dirty data, so our
 493          * dp_space_towrite should now be zero.  However, some seldom-used
 494          * code paths do not adhere to this (e.g. dbuf_undirty(), also
 495          * rounding error in dbuf_write_physdone).
 496          * Shore up the accounting of any dirtied space now.
 497          */
 498         dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
 499 
 500         /*
 501          * After the data blocks have been written (ensured by the zio_wait()
 502          * above), update the user/group space accounting.
 503          */
 504         for (ds = list_head(&synced_datasets); ds != NULL;
 505             ds = list_next(&synced_datasets, ds)) {
 506                 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
 507         }
 508 
 509         /*
 510          * Sync the datasets again to push out the changes due to
 511          * userspace updates.  This must be done before we process the
 512          * sync tasks, so that any snapshots will have the correct
 513          * user accounting information (and we won't get confused
 514          * about which blocks are part of the snapshot).
 515          */
 516         zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 517         while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
 518                 ASSERT(list_link_active(&ds->ds_synced_link));
 519                 dmu_buf_rele(ds->ds_dbuf, ds);
 520                 dsl_dataset_sync(ds, zio, tx);
 521         }
 522         VERIFY0(zio_wait(zio));
 523 
 524         /*
 525          * Now that the datasets have been completely synced, we can
 526          * clean up our in-memory structures accumulated while syncing:
 527          *
 528          *  - move dead blocks from the pending deadlist to the on-disk deadlist
 529          *  - release hold from dsl_dataset_dirty()
 530          */
 531         while ((ds = list_remove_head(&synced_datasets)) != NULL) {
 532                 objset_t *os = ds->ds_objset;
 533                 bplist_iterate(&ds->ds_pending_deadlist,
 534                     deadlist_enqueue_cb, &ds->ds_deadlist, tx);
 535                 ASSERT(!dmu_objset_is_dirty(os, txg));
 536                 dmu_buf_rele(ds->ds_dbuf, ds);
 537         }
 538         while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
 539                 dsl_dir_sync(dd, tx);
 540         }
 541 
 542         /*
 543          * The MOS's space is accounted for in the pool/$MOS
 544          * (dp_mos_dir).  We can't modify the mos while we're syncing
 545          * it, so we remember the deltas and apply them here.
 546          */
 547         if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
 548             dp->dp_mos_uncompressed_delta != 0) {
 549                 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
 550                     dp->dp_mos_used_delta,
 551                     dp->dp_mos_compressed_delta,
 552                     dp->dp_mos_uncompressed_delta, tx);
 553                 dp->dp_mos_used_delta = 0;
 554                 dp->dp_mos_compressed_delta = 0;
 555                 dp->dp_mos_uncompressed_delta = 0;
 556         }
 557 
 558         if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
 559             list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
 560                 dsl_pool_sync_mos(dp, tx);
 561         }
 562 
 563         /*
 564          * If we modify a dataset in the same txg that we want to destroy it,
 565          * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
 566          * dsl_dir_destroy_check() will fail if there are unexpected holds.
 567          * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
 568          * and clearing the hold on it) before we process the sync_tasks.
 569          * The MOS data dirtied by the sync_tasks will be synced on the next
 570          * pass.
 571          */
 572         if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
 573                 dsl_sync_task_t *dst;
 574                 /*
 575                  * No more sync tasks should have been added while we
 576                  * were syncing.
 577                  */
 578                 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
 579                 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
 580                         dsl_sync_task_sync(dst, tx);
 581         }
 582 
 583         dmu_tx_commit(tx);
 584 
 585         DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
 586 }
 587 
 588 void
 589 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
 590 {
 591         zilog_t *zilog;
 592 
 593         while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
 594                 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
 595                 zil_clean(zilog, txg);
 596                 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
 597                 dmu_buf_rele(ds->ds_dbuf, zilog);
 598         }
 599         ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
 600 }
 601 
 602 /*
 603  * TRUE if the current thread is the tx_sync_thread or if we
 604  * are being called from SPA context during pool initialization.
 605  */
 606 int
 607 dsl_pool_sync_context(dsl_pool_t *dp)
 608 {
 609         return (curthread == dp->dp_tx.tx_sync_thread ||
 610             spa_is_initializing(dp->dp_spa));
 611 }
 612 
 613 uint64_t
 614 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
 615 {
 616         uint64_t space, resv;
 617 
 618         /*
 619          * If we're trying to assess whether it's OK to do a free,
 620          * cut the reservation in half to allow forward progress
 621          * (e.g. make it possible to rm(1) files from a full pool).
 622          */
 623         space = spa_get_dspace(dp->dp_spa);
 624         resv = spa_get_slop_space(dp->dp_spa);
 625         if (netfree)
 626                 resv >>= 1;
 627 
 628         return (space - resv);
 629 }
 630 
 631 boolean_t
 632 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
 633 {
 634         uint64_t delay_min_bytes =
 635             zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
 636         boolean_t rv;
 637 
 638         mutex_enter(&dp->dp_lock);
 639         if (dp->dp_dirty_total > zfs_dirty_data_sync)
 640                 txg_kick(dp);
 641         rv = (dp->dp_dirty_total > delay_min_bytes);
 642         mutex_exit(&dp->dp_lock);
 643         return (rv);
 644 }
 645 
 646 void
 647 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
 648 {
 649         if (space > 0) {
 650                 mutex_enter(&dp->dp_lock);
 651                 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
 652                 dsl_pool_dirty_delta(dp, space);
 653                 mutex_exit(&dp->dp_lock);
 654         }
 655 }
 656 
 657 void
 658 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
 659 {
 660         ASSERT3S(space, >=, 0);
 661         if (space == 0)
 662                 return;
 663         mutex_enter(&dp->dp_lock);
 664         if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
 665                 /* XXX writing something we didn't dirty? */
 666                 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
 667         }
 668         ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
 669         dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
 670         ASSERT3U(dp->dp_dirty_total, >=, space);
 671         dsl_pool_dirty_delta(dp, -space);
 672         mutex_exit(&dp->dp_lock);
 673 }
 674 
 675 /* ARGSUSED */
 676 static int
 677 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
 678 {
 679         dmu_tx_t *tx = arg;
 680         dsl_dataset_t *ds, *prev = NULL;
 681         int err;
 682 
 683         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
 684         if (err)
 685                 return (err);
 686 
 687         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
 688                 err = dsl_dataset_hold_obj(dp,
 689                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
 690                 if (err) {
 691                         dsl_dataset_rele(ds, FTAG);
 692                         return (err);
 693                 }
 694 
 695                 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
 696                         break;
 697                 dsl_dataset_rele(ds, FTAG);
 698                 ds = prev;
 699                 prev = NULL;
 700         }
 701 
 702         if (prev == NULL) {
 703                 prev = dp->dp_origin_snap;
 704 
 705                 /*
 706                  * The $ORIGIN can't have any data, or the accounting
 707                  * will be wrong.
 708                  */
 709                 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
 710 
 711                 /* The origin doesn't get attached to itself */
 712                 if (ds->ds_object == prev->ds_object) {
 713                         dsl_dataset_rele(ds, FTAG);
 714                         return (0);
 715                 }
 716 
 717                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
 718                 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
 719                 dsl_dataset_phys(ds)->ds_prev_snap_txg =
 720                     dsl_dataset_phys(prev)->ds_creation_txg;
 721 
 722                 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
 723                 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
 724 
 725                 dmu_buf_will_dirty(prev->ds_dbuf, tx);
 726                 dsl_dataset_phys(prev)->ds_num_children++;
 727 
 728                 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
 729                         ASSERT(ds->ds_prev == NULL);
 730                         VERIFY0(dsl_dataset_hold_obj(dp,
 731                             dsl_dataset_phys(ds)->ds_prev_snap_obj,
 732                             ds, &ds->ds_prev));
 733                 }
 734         }
 735 
 736         ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
 737         ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
 738 
 739         if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
 740                 dmu_buf_will_dirty(prev->ds_dbuf, tx);
 741                 dsl_dataset_phys(prev)->ds_next_clones_obj =
 742                     zap_create(dp->dp_meta_objset,
 743                     DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
 744         }
 745         VERIFY0(zap_add_int(dp->dp_meta_objset,
 746             dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
 747 
 748         dsl_dataset_rele(ds, FTAG);
 749         if (prev != dp->dp_origin_snap)
 750                 dsl_dataset_rele(prev, FTAG);
 751         return (0);
 752 }
 753 
 754 void
 755 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
 756 {
 757         ASSERT(dmu_tx_is_syncing(tx));
 758         ASSERT(dp->dp_origin_snap != NULL);
 759 
 760         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
 761             tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
 762 }
 763 
 764 /* ARGSUSED */
 765 static int
 766 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
 767 {
 768         dmu_tx_t *tx = arg;
 769         objset_t *mos = dp->dp_meta_objset;
 770 
 771         if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
 772                 dsl_dataset_t *origin;
 773 
 774                 VERIFY0(dsl_dataset_hold_obj(dp,
 775                     dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
 776 
 777                 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
 778                         dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
 779                         dsl_dir_phys(origin->ds_dir)->dd_clones =
 780                             zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
 781                             0, tx);
 782                 }
 783 
 784                 VERIFY0(zap_add_int(dp->dp_meta_objset,
 785                     dsl_dir_phys(origin->ds_dir)->dd_clones,
 786                     ds->ds_object, tx));
 787 
 788                 dsl_dataset_rele(origin, FTAG);
 789         }
 790         return (0);
 791 }
 792 
 793 void
 794 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
 795 {
 796         ASSERT(dmu_tx_is_syncing(tx));
 797         uint64_t obj;
 798 
 799         (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
 800         VERIFY0(dsl_pool_open_special_dir(dp,
 801             FREE_DIR_NAME, &dp->dp_free_dir));
 802 
 803         /*
 804          * We can't use bpobj_alloc(), because spa_version() still
 805          * returns the old version, and we need a new-version bpobj with
 806          * subobj support.  So call dmu_object_alloc() directly.
 807          */
 808         obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
 809             SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
 810         VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 811             DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
 812         VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
 813 
 814         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
 815             upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
 816 }
 817 
 818 void
 819 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
 820 {
 821         uint64_t dsobj;
 822         dsl_dataset_t *ds;
 823 
 824         ASSERT(dmu_tx_is_syncing(tx));
 825         ASSERT(dp->dp_origin_snap == NULL);
 826         ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
 827 
 828         /* create the origin dir, ds, & snap-ds */
 829         dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
 830             NULL, 0, kcred, tx);
 831         VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 832         dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
 833         VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
 834             dp, &dp->dp_origin_snap));
 835         dsl_dataset_rele(ds, FTAG);
 836 }
 837 
 838 taskq_t *
 839 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
 840 {
 841         return (dp->dp_vnrele_taskq);
 842 }
 843 
 844 /*
 845  * Walk through the pool-wide zap object of temporary snapshot user holds
 846  * and release them.
 847  */
 848 void
 849 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
 850 {
 851         zap_attribute_t za;
 852         zap_cursor_t zc;
 853         objset_t *mos = dp->dp_meta_objset;
 854         uint64_t zapobj = dp->dp_tmp_userrefs_obj;
 855         nvlist_t *holds;
 856 
 857         if (zapobj == 0)
 858                 return;
 859         ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
 860 
 861         holds = fnvlist_alloc();
 862 
 863         for (zap_cursor_init(&zc, mos, zapobj);
 864             zap_cursor_retrieve(&zc, &za) == 0;
 865             zap_cursor_advance(&zc)) {
 866                 char *htag;
 867                 nvlist_t *tags;
 868 
 869                 htag = strchr(za.za_name, '-');
 870                 *htag = '\0';
 871                 ++htag;
 872                 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
 873                         tags = fnvlist_alloc();
 874                         fnvlist_add_boolean(tags, htag);
 875                         fnvlist_add_nvlist(holds, za.za_name, tags);
 876                         fnvlist_free(tags);
 877                 } else {
 878                         fnvlist_add_boolean(tags, htag);
 879                 }
 880         }
 881         dsl_dataset_user_release_tmp(dp, holds);
 882         fnvlist_free(holds);
 883         zap_cursor_fini(&zc);
 884 }
 885 
 886 /*
 887  * Create the pool-wide zap object for storing temporary snapshot holds.
 888  */
 889 void
 890 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
 891 {
 892         objset_t *mos = dp->dp_meta_objset;
 893 
 894         ASSERT(dp->dp_tmp_userrefs_obj == 0);
 895         ASSERT(dmu_tx_is_syncing(tx));
 896 
 897         dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
 898             DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
 899 }
 900 
 901 static int
 902 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
 903     const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
 904 {
 905         objset_t *mos = dp->dp_meta_objset;
 906         uint64_t zapobj = dp->dp_tmp_userrefs_obj;
 907         char *name;
 908         int error;
 909 
 910         ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
 911         ASSERT(dmu_tx_is_syncing(tx));
 912 
 913         /*
 914          * If the pool was created prior to SPA_VERSION_USERREFS, the
 915          * zap object for temporary holds might not exist yet.
 916          */
 917         if (zapobj == 0) {
 918                 if (holding) {
 919                         dsl_pool_user_hold_create_obj(dp, tx);
 920                         zapobj = dp->dp_tmp_userrefs_obj;
 921                 } else {
 922                         return (SET_ERROR(ENOENT));
 923                 }
 924         }
 925 
 926         name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
 927         if (holding)
 928                 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
 929         else
 930                 error = zap_remove(mos, zapobj, name, tx);
 931         strfree(name);
 932 
 933         return (error);
 934 }
 935 
 936 /*
 937  * Add a temporary hold for the given dataset object and tag.
 938  */
 939 int
 940 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
 941     uint64_t now, dmu_tx_t *tx)
 942 {
 943         return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
 944 }
 945 
 946 /*
 947  * Release a temporary hold for the given dataset object and tag.
 948  */
 949 int
 950 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
 951     dmu_tx_t *tx)
 952 {
 953         return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
 954             tx, B_FALSE));
 955 }
 956 
 957 /*
 958  * DSL Pool Configuration Lock
 959  *
 960  * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
 961  * creation / destruction / rename / property setting).  It must be held for
 962  * read to hold a dataset or dsl_dir.  I.e. you must call
 963  * dsl_pool_config_enter() or dsl_pool_hold() before calling
 964  * dsl_{dataset,dir}_hold{_obj}.  In most circumstances, the dp_config_rwlock
 965  * must be held continuously until all datasets and dsl_dirs are released.
 966  *
 967  * The only exception to this rule is that if a "long hold" is placed on
 968  * a dataset, then the dp_config_rwlock may be dropped while the dataset
 969  * is still held.  The long hold will prevent the dataset from being
 970  * destroyed -- the destroy will fail with EBUSY.  A long hold can be
 971  * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
 972  * (by calling dsl_{dataset,objset}_{try}own{_obj}).
 973  *
 974  * Legitimate long-holders (including owners) should be long-running, cancelable
 975  * tasks that should cause "zfs destroy" to fail.  This includes DMU
 976  * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
 977  * "zfs send", and "zfs diff".  There are several other long-holders whose
 978  * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
 979  *
 980  * The usual formula for long-holding would be:
 981  * dsl_pool_hold()
 982  * dsl_dataset_hold()
 983  * ... perform checks ...
 984  * dsl_dataset_long_hold()
 985  * dsl_pool_rele()
 986  * ... perform long-running task ...
 987  * dsl_dataset_long_rele()
 988  * dsl_dataset_rele()
 989  *
 990  * Note that when the long hold is released, the dataset is still held but
 991  * the pool is not held.  The dataset may change arbitrarily during this time
 992  * (e.g. it could be destroyed).  Therefore you shouldn't do anything to the
 993  * dataset except release it.
 994  *
 995  * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
 996  * or modifying operations.
 997  *
 998  * Modifying operations should generally use dsl_sync_task().  The synctask
 999  * infrastructure enforces proper locking strategy with respect to the
1000  * dp_config_rwlock.  See the comment above dsl_sync_task() for details.
1001  *
1002  * Read-only operations will manually hold the pool, then the dataset, obtain
1003  * information from the dataset, then release the pool and dataset.
1004  * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1005  * hold/rele.
1006  */
1007 
1008 int
1009 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1010 {
1011         spa_t *spa;
1012         int error;
1013 
1014         error = spa_open(name, &spa, tag);
1015         if (error == 0) {
1016                 *dp = spa_get_dsl(spa);
1017                 dsl_pool_config_enter(*dp, tag);
1018         }
1019         return (error);
1020 }
1021 
1022 void
1023 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1024 {
1025         dsl_pool_config_exit(dp, tag);
1026         spa_close(dp->dp_spa, tag);
1027 }
1028 
1029 void
1030 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1031 {
1032         /*
1033          * We use a "reentrant" reader-writer lock, but not reentrantly.
1034          *
1035          * The rrwlock can (with the track_all flag) track all reading threads,
1036          * which is very useful for debugging which code path failed to release
1037          * the lock, and for verifying that the *current* thread does hold
1038          * the lock.
1039          *
1040          * (Unlike a rwlock, which knows that N threads hold it for
1041          * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1042          * if any thread holds it for read, even if this thread doesn't).
1043          */
1044         ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1045         rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1046 }
1047 
1048 void
1049 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1050 {
1051         ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1052         rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1053 }
1054 
1055 void
1056 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1057 {
1058         rrw_exit(&dp->dp_config_rwlock, tag);
1059 }
1060 
1061 boolean_t
1062 dsl_pool_config_held(dsl_pool_t *dp)
1063 {
1064         return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1065 }
1066 
1067 boolean_t
1068 dsl_pool_config_held_writer(dsl_pool_t *dp)
1069 {
1070         return (RRW_WRITE_HELD(&dp->dp_config_rwlock));
1071 }