module/zfs/dsl_dataset.c

   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  */
  24
  25 #include <sys/dmu_objset.h>
  26 #include <sys/dsl_dataset.h>
  27 #include <sys/dsl_dir.h>
  28 #include <sys/dsl_prop.h>
  29 #include <sys/dsl_synctask.h>
  30 #include <sys/dmu_traverse.h>
  31 #include <sys/dmu_tx.h>
  32 #include <sys/arc.h>
  33 #include <sys/zio.h>
  34 #include <sys/zap.h>
  35 #include <sys/unique.h>
  36 #include <sys/zfs_context.h>
  37 #include <sys/zfs_ioctl.h>
  38 #include <sys/spa.h>
  39 #include <sys/zfs_znode.h>
  40 #include <sys/zfs_onexit.h>
  41 #include <sys/zvol.h>
  42 #include <sys/dsl_scan.h>
  43 #include <sys/dsl_deadlist.h>
  44
  45 static char *dsl_reaper = "the grim reaper";
  46
  47 static dsl_checkfunc_t dsl_dataset_destroy_begin_check;
  48 static dsl_syncfunc_t dsl_dataset_destroy_begin_sync;
  49 static dsl_syncfunc_t dsl_dataset_set_reservation_sync;
  50
  51 #define SWITCH64(x, y) \
  52         { \
  53                 uint64_t __tmp = (x); \
  54                 (x) = (y); \
  55                 (y) = __tmp; \
  56         }
  57
  58 #define DS_REF_MAX      (1ULL << 62)
  59
  60 #define DSL_DEADLIST_BLOCKSIZE  SPA_MAXBLOCKSIZE
  61
  62 #define DSL_DATASET_IS_DESTROYED(ds)    ((ds)->ds_owner == dsl_reaper)
  63
  64
  65 /*
  66  * Figure out how much of this delta should be propogated to the dsl_dir
  67  * layer.  If there's a refreservation, that space has already been
  68  * partially accounted for in our ancestors.
  69  */
  70 static int64_t
  71 parent_delta(dsl_dataset_t *ds, int64_t delta)
  72 {
  73         uint64_t old_bytes, new_bytes;
  74
  75         if (ds->ds_reserved == 0)
  76                 return (delta);
  77
  78         old_bytes = MAX(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
  79         new_bytes = MAX(ds->ds_phys->ds_unique_bytes + delta, ds->ds_reserved);
  80
  81         ASSERT3U(ABS((int64_t)(new_bytes - old_bytes)), <=, ABS(delta));
  82         return (new_bytes - old_bytes);
  83 }
  84
  85 void
  86 dsl_dataset_block_born(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
  87 {
  88         int used, compressed, uncompressed;
  89         int64_t delta;
  90
  91         used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
  92         compressed = BP_GET_PSIZE(bp);
  93         uncompressed = BP_GET_UCSIZE(bp);
  94
  95         dprintf_bp(bp, "ds=%p", ds);
  96
  97         ASSERT(dmu_tx_is_syncing(tx));
  98         /* It could have been compressed away to nothing */
  99         if (BP_IS_HOLE(bp))
 100                 return;
 101         ASSERT(BP_GET_TYPE(bp) != DMU_OT_NONE);
 102         ASSERT3U(BP_GET_TYPE(bp), <, DMU_OT_NUMTYPES);
 103         if (ds == NULL) {
 104                 /*
 105                  * Account for the meta-objset space in its placeholder
 106                  * dsl_dir.
 107                  */
 108                 ASSERT3U(compressed, ==, uncompressed); /* it's all metadata */
 109                 dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
 110                     used, compressed, uncompressed, tx);
 111                 dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
 112                 return;
 113         }
 114         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 115
 116         mutex_enter(&ds->ds_dir->dd_lock);
 117         mutex_enter(&ds->ds_lock);
 118         delta = parent_delta(ds, used);
 119         ds->ds_phys->ds_used_bytes += used;
 120         ds->ds_phys->ds_compressed_bytes += compressed;
 121         ds->ds_phys->ds_uncompressed_bytes += uncompressed;
 122         ds->ds_phys->ds_unique_bytes += used;
 123         mutex_exit(&ds->ds_lock);
 124         dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD, delta,
 125             compressed, uncompressed, tx);
 126         dsl_dir_transfer_space(ds->ds_dir, used - delta,
 127             DD_USED_REFRSRV, DD_USED_HEAD, tx);
 128         mutex_exit(&ds->ds_dir->dd_lock);
 129 }
 130
 131 int
 132 dsl_dataset_block_kill(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx,
 133     boolean_t async)
 134 {
 135         int used, compressed, uncompressed;
 136
 137         if (BP_IS_HOLE(bp))
 138                 return (0);
 139
 140         ASSERT(dmu_tx_is_syncing(tx));
 141         ASSERT(bp->blk_birth <= tx->tx_txg);
 142
 143         used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
 144         compressed = BP_GET_PSIZE(bp);
 145         uncompressed = BP_GET_UCSIZE(bp);
 146
 147         ASSERT(used > 0);
 148         if (ds == NULL) {
 149                 /*
 150                  * Account for the meta-objset space in its placeholder
 151                  * dataset.
 152                  */
 153                 dsl_free(tx->tx_pool, tx->tx_txg, bp);
 154
 155                 dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
 156                     -used, -compressed, -uncompressed, tx);
 157                 dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
 158                 return (used);
 159         }
 160         ASSERT3P(tx->tx_pool, ==, ds->ds_dir->dd_pool);
 161
 162         ASSERT(!dsl_dataset_is_snapshot(ds));
 163         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 164
 165         if (bp->blk_birth > ds->ds_phys->ds_prev_snap_txg) {
 166                 int64_t delta;
 167
 168                 dprintf_bp(bp, "freeing ds=%llu", ds->ds_object);
 169                 dsl_free(tx->tx_pool, tx->tx_txg, bp);
 170
 171                 mutex_enter(&ds->ds_dir->dd_lock);
 172                 mutex_enter(&ds->ds_lock);
 173                 ASSERT(ds->ds_phys->ds_unique_bytes >= used ||
 174                     !DS_UNIQUE_IS_ACCURATE(ds));
 175                 delta = parent_delta(ds, -used);
 176                 ds->ds_phys->ds_unique_bytes -= used;
 177                 mutex_exit(&ds->ds_lock);
 178                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
 179                     delta, -compressed, -uncompressed, tx);
 180                 dsl_dir_transfer_space(ds->ds_dir, -used - delta,
 181                     DD_USED_REFRSRV, DD_USED_HEAD, tx);
 182                 mutex_exit(&ds->ds_dir->dd_lock);
 183         } else {
 184                 dprintf_bp(bp, "putting on dead list: %s", "");
 185                 if (async) {
 186                         /*
 187                          * We are here as part of zio's write done callback,
 188                          * which means we're a zio interrupt thread.  We can't
 189                          * call dsl_deadlist_insert() now because it may block
 190                          * waiting for I/O.  Instead, put bp on the deferred
 191                          * queue and let dsl_pool_sync() finish the job.
 192                          */
 193                         bplist_append(&ds->ds_pending_deadlist, bp);
 194                 } else {
 195                         dsl_deadlist_insert(&ds->ds_deadlist, bp, tx);
 196                 }
 197                 ASSERT3U(ds->ds_prev->ds_object, ==,
 198                     ds->ds_phys->ds_prev_snap_obj);
 199                 ASSERT(ds->ds_prev->ds_phys->ds_num_children > 0);
 200                 /* if (bp->blk_birth > prev prev snap txg) prev unique += bs */
 201                 if (ds->ds_prev->ds_phys->ds_next_snap_obj ==
 202                     ds->ds_object && bp->blk_birth >
 203                     ds->ds_prev->ds_phys->ds_prev_snap_txg) {
 204                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
 205                         mutex_enter(&ds->ds_prev->ds_lock);
 206                         ds->ds_prev->ds_phys->ds_unique_bytes += used;
 207                         mutex_exit(&ds->ds_prev->ds_lock);
 208                 }
 209                 if (bp->blk_birth > ds->ds_dir->dd_origin_txg) {
 210                         dsl_dir_transfer_space(ds->ds_dir, used,
 211                             DD_USED_HEAD, DD_USED_SNAP, tx);
 212                 }
 213         }
 214         mutex_enter(&ds->ds_lock);
 215         ASSERT3U(ds->ds_phys->ds_used_bytes, >=, used);
 216         ds->ds_phys->ds_used_bytes -= used;
 217         ASSERT3U(ds->ds_phys->ds_compressed_bytes, >=, compressed);
 218         ds->ds_phys->ds_compressed_bytes -= compressed;
 219         ASSERT3U(ds->ds_phys->ds_uncompressed_bytes, >=, uncompressed);
 220         ds->ds_phys->ds_uncompressed_bytes -= uncompressed;
 221         mutex_exit(&ds->ds_lock);
 222
 223         return (used);
 224 }
 225
 226 uint64_t
 227 dsl_dataset_prev_snap_txg(dsl_dataset_t *ds)
 228 {
 229         uint64_t trysnap = 0;
 230
 231         if (ds == NULL)
 232                 return (0);
 233         /*
 234          * The snapshot creation could fail, but that would cause an
 235          * incorrect FALSE return, which would only result in an
 236          * overestimation of the amount of space that an operation would
 237          * consume, which is OK.
 238          *
 239          * There's also a small window where we could miss a pending
 240          * snapshot, because we could set the sync task in the quiescing
 241          * phase.  So this should only be used as a guess.
 242          */
 243         if (ds->ds_trysnap_txg >
 244             spa_last_synced_txg(ds->ds_dir->dd_pool->dp_spa))
 245                 trysnap = ds->ds_trysnap_txg;
 246         return (MAX(ds->ds_phys->ds_prev_snap_txg, trysnap));
 247 }
 248
 249 boolean_t
 250 dsl_dataset_block_freeable(dsl_dataset_t *ds, const blkptr_t *bp,
 251     uint64_t blk_birth)
 252 {
 253         if (blk_birth <= dsl_dataset_prev_snap_txg(ds))
 254                 return (B_FALSE);
 255
 256         ddt_prefetch(dsl_dataset_get_spa(ds), bp);
 257
 258         return (B_TRUE);
 259 }
 260
 261 /* ARGSUSED */
 262 static void
 263 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
 264 {
 265         dsl_dataset_t *ds = dsv;
 266
 267         ASSERT(ds->ds_owner == NULL || DSL_DATASET_IS_DESTROYED(ds));
 268
 269         unique_remove(ds->ds_fsid_guid);
 270
 271         if (ds->ds_objset != NULL)
 272                 dmu_objset_evict(ds->ds_objset);
 273
 274         if (ds->ds_prev) {
 275                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 276                 ds->ds_prev = NULL;
 277         }
 278
 279         bplist_destroy(&ds->ds_pending_deadlist);
 280         if (db != NULL) {
 281                 dsl_deadlist_close(&ds->ds_deadlist);
 282         } else {
 283                 ASSERT(ds->ds_deadlist.dl_dbuf == NULL);
 284                 ASSERT(!ds->ds_deadlist.dl_oldfmt);
 285         }
 286         if (ds->ds_dir)
 287                 dsl_dir_close(ds->ds_dir, ds);
 288
 289         ASSERT(!list_link_active(&ds->ds_synced_link));
 290
 291         mutex_destroy(&ds->ds_lock);
 292         mutex_destroy(&ds->ds_recvlock);
 293         mutex_destroy(&ds->ds_opening_lock);
 294         rw_destroy(&ds->ds_rwlock);
 295         cv_destroy(&ds->ds_exclusive_cv);
 296
 297         kmem_free(ds, sizeof (dsl_dataset_t));
 298 }
 299
 300 static int
 301 dsl_dataset_get_snapname(dsl_dataset_t *ds)
 302 {
 303         dsl_dataset_phys_t *headphys;
 304         int err;
 305         dmu_buf_t *headdbuf;
 306         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 307         objset_t *mos = dp->dp_meta_objset;
 308
 309         if (ds->ds_snapname[0])
 310                 return (0);
 311         if (ds->ds_phys->ds_next_snap_obj == 0)
 312                 return (0);
 313
 314         err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
 315             FTAG, &headdbuf);
 316         if (err)
 317                 return (err);
 318         headphys = headdbuf->db_data;
 319         err = zap_value_search(dp->dp_meta_objset,
 320             headphys->ds_snapnames_zapobj, ds->ds_object, 0, ds->ds_snapname);
 321         dmu_buf_rele(headdbuf, FTAG);
 322         return (err);
 323 }
 324
 325 static int
 326 dsl_dataset_snap_lookup(dsl_dataset_t *ds, const char *name, uint64_t *value)
 327 {
 328         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 329         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 330         matchtype_t mt;
 331         int err;
 332
 333         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 334                 mt = MT_FIRST;
 335         else
 336                 mt = MT_EXACT;
 337
 338         err = zap_lookup_norm(mos, snapobj, name, 8, 1,
 339             value, mt, NULL, 0, NULL);
 340         if (err == ENOTSUP && mt == MT_FIRST)
 341                 err = zap_lookup(mos, snapobj, name, 8, 1, value);
 342         return (err);
 343 }
 344
 345 static int
 346 dsl_dataset_snap_remove(dsl_dataset_t *ds, char *name, dmu_tx_t *tx)
 347 {
 348         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 349         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 350         matchtype_t mt;
 351         int err;
 352
 353         dsl_dir_snap_cmtime_update(ds->ds_dir);
 354
 355         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 356                 mt = MT_FIRST;
 357         else
 358                 mt = MT_EXACT;
 359
 360         err = zap_remove_norm(mos, snapobj, name, mt, tx);
 361         if (err == ENOTSUP && mt == MT_FIRST)
 362                 err = zap_remove(mos, snapobj, name, tx);
 363         return (err);
 364 }
 365
 366 static int
 367 dsl_dataset_get_ref(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 368     dsl_dataset_t **dsp)
 369 {
 370         objset_t *mos = dp->dp_meta_objset;
 371         dmu_buf_t *dbuf;
 372         dsl_dataset_t *ds;
 373         int err;
 374         dmu_object_info_t doi;
 375
 376         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
 377             dsl_pool_sync_context(dp));
 378
 379         err = dmu_bonus_hold(mos, dsobj, tag, &dbuf);
 380         if (err)
 381                 return (err);
 382
 383         /* Make sure dsobj has the correct object type. */
 384         dmu_object_info_from_db(dbuf, &doi);
 385         if (doi.doi_type != DMU_OT_DSL_DATASET)
 386                 return (EINVAL);
 387
 388         ds = dmu_buf_get_user(dbuf);
 389         if (ds == NULL) {
 390                 dsl_dataset_t *winner = NULL;
 391
 392                 ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
 393                 ds->ds_dbuf = dbuf;
 394                 ds->ds_object = dsobj;
 395                 ds->ds_phys = dbuf->db_data;
 396                 list_link_init(&ds->ds_synced_link);
 397
 398                 mutex_init(&ds->ds_lock, NULL, MUTEX_DEFAULT, NULL);
 399                 mutex_init(&ds->ds_recvlock, NULL, MUTEX_DEFAULT, NULL);
 400                 mutex_init(&ds->ds_opening_lock, NULL, MUTEX_DEFAULT, NULL);
 401                 rw_init(&ds->ds_rwlock, NULL, RW_DEFAULT, NULL);
 402                 cv_init(&ds->ds_exclusive_cv, NULL, CV_DEFAULT, NULL);
 403
 404                 bplist_create(&ds->ds_pending_deadlist);
 405                 dsl_deadlist_open(&ds->ds_deadlist,
 406                     mos, ds->ds_phys->ds_deadlist_obj);
 407
 408                 if (err == 0) {
 409                         err = dsl_dir_open_obj(dp,
 410                             ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
 411                 }
 412                 if (err) {
 413                         mutex_destroy(&ds->ds_lock);
 414                         mutex_destroy(&ds->ds_recvlock);
 415                         mutex_destroy(&ds->ds_opening_lock);
 416                         rw_destroy(&ds->ds_rwlock);
 417                         cv_destroy(&ds->ds_exclusive_cv);
 418                         bplist_destroy(&ds->ds_pending_deadlist);
 419                         dsl_deadlist_close(&ds->ds_deadlist);
 420                         kmem_free(ds, sizeof (dsl_dataset_t));
 421                         dmu_buf_rele(dbuf, tag);
 422                         return (err);
 423                 }
 424
 425                 if (!dsl_dataset_is_snapshot(ds)) {
 426                         ds->ds_snapname[0] = '\0';
 427                         if (ds->ds_phys->ds_prev_snap_obj) {
 428                                 err = dsl_dataset_get_ref(dp,
 429                                     ds->ds_phys->ds_prev_snap_obj,
 430                                     ds, &ds->ds_prev);
 431                         }
 432                 } else {
 433                         if (zfs_flags & ZFS_DEBUG_SNAPNAMES)
 434                                 err = dsl_dataset_get_snapname(ds);
 435                         if (err == 0 && ds->ds_phys->ds_userrefs_obj != 0) {
 436                                 err = zap_count(
 437                                     ds->ds_dir->dd_pool->dp_meta_objset,
 438                                     ds->ds_phys->ds_userrefs_obj,
 439                                     &ds->ds_userrefs);
 440                         }
 441                 }
 442
 443                 if (err == 0 && !dsl_dataset_is_snapshot(ds)) {
 444                         /*
 445                          * In sync context, we're called with either no lock
 446                          * or with the write lock.  If we're not syncing,
 447                          * we're always called with the read lock held.
 448                          */
 449                         boolean_t need_lock =
 450                             !RW_WRITE_HELD(&dp->dp_config_rwlock) &&
 451                             dsl_pool_sync_context(dp);
 452
 453                         if (need_lock)
 454                                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 455
 456                         err = dsl_prop_get_ds(ds,
 457                             "refreservation", sizeof (uint64_t), 1,
 458                             &ds->ds_reserved, NULL);
 459                         if (err == 0) {
 460                                 err = dsl_prop_get_ds(ds,
 461                                     "refquota", sizeof (uint64_t), 1,
 462                                     &ds->ds_quota, NULL);
 463                         }
 464
 465                         if (need_lock)
 466                                 rw_exit(&dp->dp_config_rwlock);
 467                 } else {
 468                         ds->ds_reserved = ds->ds_quota = 0;
 469                 }
 470
 471                 if (err == 0) {
 472                         winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
 473                             dsl_dataset_evict);
 474                 }
 475                 if (err || winner) {
 476                         bplist_destroy(&ds->ds_pending_deadlist);
 477                         dsl_deadlist_close(&ds->ds_deadlist);
 478                         if (ds->ds_prev)
 479                                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 480                         dsl_dir_close(ds->ds_dir, ds);
 481                         mutex_destroy(&ds->ds_lock);
 482                         mutex_destroy(&ds->ds_recvlock);
 483                         mutex_destroy(&ds->ds_opening_lock);
 484                         rw_destroy(&ds->ds_rwlock);
 485                         cv_destroy(&ds->ds_exclusive_cv);
 486                         kmem_free(ds, sizeof (dsl_dataset_t));
 487                         if (err) {
 488                                 dmu_buf_rele(dbuf, tag);
 489                                 return (err);
 490                         }
 491                         ds = winner;
 492                 } else {
 493                         ds->ds_fsid_guid =
 494                             unique_insert(ds->ds_phys->ds_fsid_guid);
 495                 }
 496         }
 497         ASSERT3P(ds->ds_dbuf, ==, dbuf);
 498         ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
 499         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0 ||
 500             spa_version(dp->dp_spa) < SPA_VERSION_ORIGIN ||
 501             dp->dp_origin_snap == NULL || ds == dp->dp_origin_snap);
 502         mutex_enter(&ds->ds_lock);
 503         if (!dsl_pool_sync_context(dp) && DSL_DATASET_IS_DESTROYED(ds)) {
 504                 mutex_exit(&ds->ds_lock);
 505                 dmu_buf_rele(ds->ds_dbuf, tag);
 506                 return (ENOENT);
 507         }
 508         mutex_exit(&ds->ds_lock);
 509         *dsp = ds;
 510         return (0);
 511 }
 512
 513 static int
 514 dsl_dataset_hold_ref(dsl_dataset_t *ds, void *tag)
 515 {
 516         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 517
 518         /*
 519          * In syncing context we don't want the rwlock lock: there
 520          * may be an existing writer waiting for sync phase to
 521          * finish.  We don't need to worry about such writers, since
 522          * sync phase is single-threaded, so the writer can't be
 523          * doing anything while we are active.
 524          */
 525         if (dsl_pool_sync_context(dp)) {
 526                 ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
 527                 return (0);
 528         }
 529
 530         /*
 531          * Normal users will hold the ds_rwlock as a READER until they
 532          * are finished (i.e., call dsl_dataset_rele()).  "Owners" will
 533          * drop their READER lock after they set the ds_owner field.
 534          *
 535          * If the dataset is being destroyed, the destroy thread will
 536          * obtain a WRITER lock for exclusive access after it's done its
 537          * open-context work and then change the ds_owner to
 538          * dsl_reaper once destruction is assured.  So threads
 539          * may block here temporarily, until the "destructability" of
 540          * the dataset is determined.
 541          */
 542         ASSERT(!RW_WRITE_HELD(&dp->dp_config_rwlock));
 543         mutex_enter(&ds->ds_lock);
 544         while (!rw_tryenter(&ds->ds_rwlock, RW_READER)) {
 545                 rw_exit(&dp->dp_config_rwlock);
 546                 cv_wait(&ds->ds_exclusive_cv, &ds->ds_lock);
 547                 if (DSL_DATASET_IS_DESTROYED(ds)) {
 548                         mutex_exit(&ds->ds_lock);
 549                         dsl_dataset_drop_ref(ds, tag);
 550                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 551                         return (ENOENT);
 552                 }
 553                 /*
 554                  * The dp_config_rwlock lives above the ds_lock. And
 555                  * we need to check DSL_DATASET_IS_DESTROYED() while
 556                  * holding the ds_lock, so we have to drop and reacquire
 557                  * the ds_lock here.
 558                  */
 559                 mutex_exit(&ds->ds_lock);
 560                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 561                 mutex_enter(&ds->ds_lock);
 562         }
 563         mutex_exit(&ds->ds_lock);
 564         return (0);
 565 }
 566
 567 int
 568 dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 569     dsl_dataset_t **dsp)
 570 {
 571         int err = dsl_dataset_get_ref(dp, dsobj, tag, dsp);
 572
 573         if (err)
 574                 return (err);
 575         return (dsl_dataset_hold_ref(*dsp, tag));
 576 }
 577
 578 int
 579 dsl_dataset_own_obj(dsl_pool_t *dp, uint64_t dsobj, boolean_t inconsistentok,
 580     void *tag, dsl_dataset_t **dsp)
 581 {
 582         int err = dsl_dataset_hold_obj(dp, dsobj, tag, dsp);
 583         if (err)
 584                 return (err);
 585         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 586                 dsl_dataset_rele(*dsp, tag);
 587                 *dsp = NULL;
 588                 return (EBUSY);
 589         }
 590         return (0);
 591 }
 592
 593 int
 594 dsl_dataset_hold(const char *name, void *tag, dsl_dataset_t **dsp)
 595 {
 596         dsl_dir_t *dd;
 597         dsl_pool_t *dp;
 598         const char *snapname;
 599         uint64_t obj;
 600         int err = 0;
 601
 602         err = dsl_dir_open_spa(NULL, name, FTAG, &dd, &snapname);
 603         if (err)
 604                 return (err);
 605
 606         dp = dd->dd_pool;
 607         obj = dd->dd_phys->dd_head_dataset_obj;
 608         rw_enter(&dp->dp_config_rwlock, RW_READER);
 609         if (obj)
 610                 err = dsl_dataset_get_ref(dp, obj, tag, dsp);
 611         else
 612                 err = ENOENT;
 613         if (err)
 614                 goto out;
 615
 616         err = dsl_dataset_hold_ref(*dsp, tag);
 617
 618         /* we may be looking for a snapshot */
 619         if (err == 0 && snapname != NULL) {
 620                 dsl_dataset_t *ds = NULL;
 621
 622                 if (*snapname++ != '@') {
 623                         dsl_dataset_rele(*dsp, tag);
 624                         err = ENOENT;
 625                         goto out;
 626                 }
 627
 628                 dprintf("looking for snapshot '%s'\n", snapname);
 629                 err = dsl_dataset_snap_lookup(*dsp, snapname, &obj);
 630                 if (err == 0)
 631                         err = dsl_dataset_get_ref(dp, obj, tag, &ds);
 632                 dsl_dataset_rele(*dsp, tag);
 633
 634                 ASSERT3U((err == 0), ==, (ds != NULL));
 635
 636                 if (ds) {
 637                         mutex_enter(&ds->ds_lock);
 638                         if (ds->ds_snapname[0] == 0)
 639                                 (void) strlcpy(ds->ds_snapname, snapname,
 640                                     sizeof (ds->ds_snapname));
 641                         mutex_exit(&ds->ds_lock);
 642                         err = dsl_dataset_hold_ref(ds, tag);
 643                         *dsp = err ? NULL : ds;
 644                 }
 645         }
 646 out:
 647         rw_exit(&dp->dp_config_rwlock);
 648         dsl_dir_close(dd, FTAG);
 649         return (err);
 650 }
 651
 652 int
 653 dsl_dataset_own(const char *name, boolean_t inconsistentok,
 654     void *tag, dsl_dataset_t **dsp)
 655 {
 656         int err = dsl_dataset_hold(name, tag, dsp);
 657         if (err)
 658                 return (err);
 659         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 660                 dsl_dataset_rele(*dsp, tag);
 661                 return (EBUSY);
 662         }
 663         return (0);
 664 }
 665
 666 void
 667 dsl_dataset_name(dsl_dataset_t *ds, char *name)
 668 {
 669         if (ds == NULL) {
 670                 (void) strcpy(name, "mos");
 671         } else {
 672                 dsl_dir_name(ds->ds_dir, name);
 673                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 674                 if (ds->ds_snapname[0]) {
 675                         (void) strcat(name, "@");
 676                         /*
 677                          * We use a "recursive" mutex so that we
 678                          * can call dprintf_ds() with ds_lock held.
 679                          */
 680                         if (!MUTEX_HELD(&ds->ds_lock)) {
 681                                 mutex_enter(&ds->ds_lock);
 682                                 (void) strcat(name, ds->ds_snapname);
 683                                 mutex_exit(&ds->ds_lock);
 684                         } else {
 685                                 (void) strcat(name, ds->ds_snapname);
 686                         }
 687                 }
 688         }
 689 }
 690
 691 static int
 692 dsl_dataset_namelen(dsl_dataset_t *ds)
 693 {
 694         int result;
 695
 696         if (ds == NULL) {
 697                 result = 3;     /* "mos" */
 698         } else {
 699                 result = dsl_dir_namelen(ds->ds_dir);
 700                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 701                 if (ds->ds_snapname[0]) {
 702                         ++result;       /* adding one for the @-sign */
 703                         if (!MUTEX_HELD(&ds->ds_lock)) {
 704                                 mutex_enter(&ds->ds_lock);
 705                                 result += strlen(ds->ds_snapname);
 706                                 mutex_exit(&ds->ds_lock);
 707                         } else {
 708                                 result += strlen(ds->ds_snapname);
 709                         }
 710                 }
 711         }
 712
 713         return (result);
 714 }
 715
 716 void
 717 dsl_dataset_drop_ref(dsl_dataset_t *ds, void *tag)
 718 {
 719         dmu_buf_rele(ds->ds_dbuf, tag);
 720 }
 721
 722 void
 723 dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
 724 {
 725         if (!dsl_pool_sync_context(ds->ds_dir->dd_pool)) {
 726                 rw_exit(&ds->ds_rwlock);
 727         }
 728         dsl_dataset_drop_ref(ds, tag);
 729 }
 730
 731 void
 732 dsl_dataset_disown(dsl_dataset_t *ds, void *tag)
 733 {
 734         ASSERT((ds->ds_owner == tag && ds->ds_dbuf) ||
 735             (DSL_DATASET_IS_DESTROYED(ds) && ds->ds_dbuf == NULL));
 736
 737         mutex_enter(&ds->ds_lock);
 738         ds->ds_owner = NULL;
 739         if (RW_WRITE_HELD(&ds->ds_rwlock)) {
 740                 rw_exit(&ds->ds_rwlock);
 741                 cv_broadcast(&ds->ds_exclusive_cv);
 742         }
 743         mutex_exit(&ds->ds_lock);
 744         if (ds->ds_dbuf)
 745                 dsl_dataset_drop_ref(ds, tag);
 746         else
 747                 dsl_dataset_evict(NULL, ds);
 748 }
 749
 750 boolean_t
 751 dsl_dataset_tryown(dsl_dataset_t *ds, boolean_t inconsistentok, void *tag)
 752 {
 753         boolean_t gotit = FALSE;
 754
 755         mutex_enter(&ds->ds_lock);
 756         if (ds->ds_owner == NULL &&
 757             (!DS_IS_INCONSISTENT(ds) || inconsistentok)) {
 758                 ds->ds_owner = tag;
 759                 if (!dsl_pool_sync_context(ds->ds_dir->dd_pool))
 760                         rw_exit(&ds->ds_rwlock);
 761                 gotit = TRUE;
 762         }
 763         mutex_exit(&ds->ds_lock);
 764         return (gotit);
 765 }
 766
 767 void
 768 dsl_dataset_make_exclusive(dsl_dataset_t *ds, void *owner)
 769 {
 770         ASSERT3P(owner, ==, ds->ds_owner);
 771         if (!RW_WRITE_HELD(&ds->ds_rwlock))
 772                 rw_enter(&ds->ds_rwlock, RW_WRITER);
 773 }
 774
 775 uint64_t
 776 dsl_dataset_create_sync_dd(dsl_dir_t *dd, dsl_dataset_t *origin,
 777     uint64_t flags, dmu_tx_t *tx)
 778 {
 779         dsl_pool_t *dp = dd->dd_pool;
 780         dmu_buf_t *dbuf;
 781         dsl_dataset_phys_t *dsphys;
 782         uint64_t dsobj;
 783         objset_t *mos = dp->dp_meta_objset;
 784
 785         if (origin == NULL)
 786                 origin = dp->dp_origin_snap;
 787
 788         ASSERT(origin == NULL || origin->ds_dir->dd_pool == dp);
 789         ASSERT(origin == NULL || origin->ds_phys->ds_num_children > 0);
 790         ASSERT(dmu_tx_is_syncing(tx));
 791         ASSERT(dd->dd_phys->dd_head_dataset_obj == 0);
 792
 793         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
 794             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
 795         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
 796         dmu_buf_will_dirty(dbuf, tx);
 797         dsphys = dbuf->db_data;
 798         bzero(dsphys, sizeof (dsl_dataset_phys_t));
 799         dsphys->ds_dir_obj = dd->dd_object;
 800         dsphys->ds_flags = flags;
 801         dsphys->ds_fsid_guid = unique_create();
 802         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
 803             sizeof (dsphys->ds_guid));
 804         dsphys->ds_snapnames_zapobj =
 805             zap_create_norm(mos, U8_TEXTPREP_TOUPPER, DMU_OT_DSL_DS_SNAP_MAP,
 806             DMU_OT_NONE, 0, tx);
 807         dsphys->ds_creation_time = gethrestime_sec();
 808         dsphys->ds_creation_txg = tx->tx_txg == TXG_INITIAL ? 1 : tx->tx_txg;
 809
 810         if (origin == NULL) {
 811                 dsphys->ds_deadlist_obj = dsl_deadlist_alloc(mos, tx);
 812         } else {
 813                 dsl_dataset_t *ohds;
 814
 815                 dsphys->ds_prev_snap_obj = origin->ds_object;
 816                 dsphys->ds_prev_snap_txg =
 817                     origin->ds_phys->ds_creation_txg;
 818                 dsphys->ds_used_bytes =
 819                     origin->ds_phys->ds_used_bytes;
 820                 dsphys->ds_compressed_bytes =
 821                     origin->ds_phys->ds_compressed_bytes;
 822                 dsphys->ds_uncompressed_bytes =
 823                     origin->ds_phys->ds_uncompressed_bytes;
 824                 dsphys->ds_bp = origin->ds_phys->ds_bp;
 825                 dsphys->ds_flags |= origin->ds_phys->ds_flags;
 826
 827                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
 828                 origin->ds_phys->ds_num_children++;
 829
 830                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
 831                     origin->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ohds));
 832                 dsphys->ds_deadlist_obj = dsl_deadlist_clone(&ohds->ds_deadlist,
 833                     dsphys->ds_prev_snap_txg, dsphys->ds_prev_snap_obj, tx);
 834                 dsl_dataset_rele(ohds, FTAG);
 835
 836                 if (spa_version(dp->dp_spa) >= SPA_VERSION_NEXT_CLONES) {
 837                         if (origin->ds_phys->ds_next_clones_obj == 0) {
 838                                 origin->ds_phys->ds_next_clones_obj =
 839                                     zap_create(mos,
 840                                     DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
 841                         }
 842                         VERIFY(0 == zap_add_int(mos,
 843                             origin->ds_phys->ds_next_clones_obj,
 844                             dsobj, tx));
 845                 }
 846
 847                 dmu_buf_will_dirty(dd->dd_dbuf, tx);
 848                 dd->dd_phys->dd_origin_obj = origin->ds_object;
 849                 if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
 850                         if (origin->ds_dir->dd_phys->dd_clones == 0) {
 851                                 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
 852                                 origin->ds_dir->dd_phys->dd_clones =
 853                                     zap_create(mos,
 854                                     DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
 855                         }
 856                         VERIFY3U(0, ==, zap_add_int(mos,
 857                             origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
 858                 }
 859         }
 860
 861         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
 862                 dsphys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
 863
 864         dmu_buf_rele(dbuf, FTAG);
 865
 866         dmu_buf_will_dirty(dd->dd_dbuf, tx);
 867         dd->dd_phys->dd_head_dataset_obj = dsobj;
 868
 869         return (dsobj);
 870 }
 871
 872 uint64_t
 873 dsl_dataset_create_sync(dsl_dir_t *pdd, const char *lastname,
 874     dsl_dataset_t *origin, uint64_t flags, cred_t *cr, dmu_tx_t *tx)
 875 {
 876         dsl_pool_t *dp = pdd->dd_pool;
 877         uint64_t dsobj, ddobj;
 878         dsl_dir_t *dd;
 879
 880         ASSERT(lastname[0] != '@');
 881
 882         ddobj = dsl_dir_create_sync(dp, pdd, lastname, tx);
 883         VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
 884
 885         dsobj = dsl_dataset_create_sync_dd(dd, origin, flags, tx);
 886
 887         dsl_deleg_set_create_perms(dd, tx, cr);
 888
 889         dsl_dir_close(dd, FTAG);
 890
 891         /*
 892          * If we are creating a clone, make sure we zero out any stale
 893          * data from the origin snapshots zil header.
 894          */
 895         if (origin != NULL) {
 896                 dsl_dataset_t *ds;
 897                 objset_t *os;
 898
 899                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 900                 VERIFY3U(0, ==, dmu_objset_from_ds(ds, &os));
 901                 bzero(&os->os_zil_header, sizeof (os->os_zil_header));
 902                 dsl_dataset_dirty(ds, tx);
 903                 dsl_dataset_rele(ds, FTAG);
 904         }
 905
 906         return (dsobj);
 907 }
 908
 909 struct destroyarg {
 910         dsl_sync_task_group_t *dstg;
 911         char *snapname;
 912         char *failed;
 913         boolean_t defer;
 914 };
 915
 916 static int
 917 dsl_snapshot_destroy_one(const char *name, void *arg)
 918 {
 919         struct destroyarg *da = arg;
 920         dsl_dataset_t *ds;
 921         int err;
 922         char *dsname;
 923
 924         dsname = kmem_asprintf("%s@%s", name, da->snapname);
 925         err = dsl_dataset_own(dsname, B_TRUE, da->dstg, &ds);
 926         strfree(dsname);
 927         if (err == 0) {
 928                 struct dsl_ds_destroyarg *dsda;
 929
 930                 dsl_dataset_make_exclusive(ds, da->dstg);
 931                 dsda = kmem_zalloc(sizeof (struct dsl_ds_destroyarg), KM_SLEEP);
 932                 dsda->ds = ds;
 933                 dsda->defer = da->defer;
 934                 dsl_sync_task_create(da->dstg, dsl_dataset_destroy_check,
 935                     dsl_dataset_destroy_sync, dsda, da->dstg, 0);
 936         } else if (err == ENOENT) {
 937                 err = 0;
 938         } else {
 939                 (void) strcpy(da->failed, name);
 940         }
 941         return (err);
 942 }
 943
 944 /*
 945  * Destroy 'snapname' in all descendants of 'fsname'.
 946  */
 947 #pragma weak dmu_snapshots_destroy = dsl_snapshots_destroy
 948 int
 949 dsl_snapshots_destroy(char *fsname, char *snapname, boolean_t defer)
 950 {
 951         int err;
 952         struct destroyarg da;
 953         dsl_sync_task_t *dst;
 954         spa_t *spa;
 955
 956         err = spa_open(fsname, &spa, FTAG);
 957         if (err)
 958                 return (err);
 959         da.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
 960         da.snapname = snapname;
 961         da.failed = fsname;
 962         da.defer = defer;
 963
 964         err = dmu_objset_find(fsname,
 965             dsl_snapshot_destroy_one, &da, DS_FIND_CHILDREN);
 966
 967         if (err == 0)
 968                 err = dsl_sync_task_group_wait(da.dstg);
 969
 970         for (dst = list_head(&da.dstg->dstg_tasks); dst;
 971             dst = list_next(&da.dstg->dstg_tasks, dst)) {
 972                 struct dsl_ds_destroyarg *dsda = dst->dst_arg1;
 973                 dsl_dataset_t *ds = dsda->ds;
 974
 975                 /*
 976                  * Return the file system name that triggered the error
 977                  */
 978                 if (dst->dst_err) {
 979                         dsl_dataset_name(ds, fsname);
 980                         *strchr(fsname, '@') = '\0';
 981                 }
 982                 ASSERT3P(dsda->rm_origin, ==, NULL);
 983                 dsl_dataset_disown(ds, da.dstg);
 984                 kmem_free(dsda, sizeof (struct dsl_ds_destroyarg));
 985         }
 986
 987         dsl_sync_task_group_destroy(da.dstg);
 988         spa_close(spa, FTAG);
 989         return (err);
 990 }
 991
 992 static boolean_t
 993 dsl_dataset_might_destroy_origin(dsl_dataset_t *ds)
 994 {
 995         boolean_t might_destroy = B_FALSE;
 996
 997         mutex_enter(&ds->ds_lock);
 998         if (ds->ds_phys->ds_num_children == 2 && ds->ds_userrefs == 0 &&
 999             DS_IS_DEFER_DESTROY(ds))
1000                 might_destroy = B_TRUE;
1001         mutex_exit(&ds->ds_lock);
1002
1003         return (might_destroy);
1004 }
1005
1006 /*
1007  * If we're removing a clone, and these three conditions are true:
1008  *      1) the clone's origin has no other children
1009  *      2) the clone's origin has no user references
1010  *      3) the clone's origin has been marked for deferred destruction
1011  * Then, prepare to remove the origin as part of this sync task group.
1012  */
1013 static int
1014 dsl_dataset_origin_rm_prep(struct dsl_ds_destroyarg *dsda, void *tag)
1015 {
1016         dsl_dataset_t *ds = dsda->ds;
1017         dsl_dataset_t *origin = ds->ds_prev;
1018
1019         if (dsl_dataset_might_destroy_origin(origin)) {
1020                 char *name;
1021                 int namelen;
1022                 int error;
1023
1024                 namelen = dsl_dataset_namelen(origin) + 1;
1025                 name = kmem_alloc(namelen, KM_SLEEP);
1026                 dsl_dataset_name(origin, name);
1027 #ifdef _KERNEL
1028                 error = zfs_unmount_snap(name, NULL);
1029                 if (error) {
1030                         kmem_free(name, namelen);
1031                         return (error);
1032                 }
1033 #endif
1034                 error = dsl_dataset_own(name, B_TRUE, tag, &origin);
1035                 kmem_free(name, namelen);
1036                 if (error)
1037                         return (error);
1038                 dsda->rm_origin = origin;
1039                 dsl_dataset_make_exclusive(origin, tag);
1040         }
1041
1042         return (0);
1043 }
1044
1045 /*
1046  * ds must be opened as OWNER.  On return (whether successful or not),
1047  * ds will be closed and caller can no longer dereference it.
1048  */
1049 int
1050 dsl_dataset_destroy(dsl_dataset_t *ds, void *tag, boolean_t defer)
1051 {
1052         int err;
1053         dsl_sync_task_group_t *dstg;
1054         objset_t *os;
1055         dsl_dir_t *dd;
1056         uint64_t obj;
1057         struct dsl_ds_destroyarg dsda = { 0 };
1058         dsl_dataset_t *dummy_ds;
1059
1060         dsda.ds = ds;
1061
1062         if (dsl_dataset_is_snapshot(ds)) {
1063                 /* Destroying a snapshot is simpler */
1064                 dsl_dataset_make_exclusive(ds, tag);
1065
1066                 dsda.defer = defer;
1067                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1068                     dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
1069                     &dsda, tag, 0);
1070                 ASSERT3P(dsda.rm_origin, ==, NULL);
1071                 goto out;
1072         } else if (defer) {
1073                 err = EINVAL;
1074                 goto out;
1075         }
1076
1077         dd = ds->ds_dir;
1078         dummy_ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
1079         dummy_ds->ds_dir = dd;
1080         dummy_ds->ds_object = ds->ds_object;
1081
1082         /*
1083          * Check for errors and mark this ds as inconsistent, in
1084          * case we crash while freeing the objects.
1085          */
1086         err = dsl_sync_task_do(dd->dd_pool, dsl_dataset_destroy_begin_check,
1087             dsl_dataset_destroy_begin_sync, ds, NULL, 0);
1088         if (err)
1089                 goto out_free;
1090
1091         err = dmu_objset_from_ds(ds, &os);
1092         if (err)
1093                 goto out_free;
1094
1095         /*
1096          * remove the objects in open context, so that we won't
1097          * have too much to do in syncing context.
1098          */
1099         for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE,
1100             ds->ds_phys->ds_prev_snap_txg)) {
1101                 /*
1102                  * Ignore errors, if there is not enough disk space
1103                  * we will deal with it in dsl_dataset_destroy_sync().
1104                  */
1105                 (void) dmu_free_object(os, obj);
1106         }
1107         if (err != ESRCH)
1108                 goto out_free;
1109
1110         /*
1111          * Only the ZIL knows how to free log blocks.
1112          */
1113         zil_destroy(dmu_objset_zil(os), B_FALSE);
1114
1115         /*
1116          * Sync out all in-flight IO.
1117          */
1118         txg_wait_synced(dd->dd_pool, 0);
1119
1120         /*
1121          * If we managed to free all the objects in open
1122          * context, the user space accounting should be zero.
1123          */
1124         if (ds->ds_phys->ds_bp.blk_fill == 0 &&
1125             dmu_objset_userused_enabled(os)) {
1126                 ASSERTV(uint64_t count);
1127                 ASSERT(zap_count(os, DMU_USERUSED_OBJECT, &count) != 0 ||
1128                     count == 0);
1129                 ASSERT(zap_count(os, DMU_GROUPUSED_OBJECT, &count) != 0 ||
1130                     count == 0);
1131         }
1132
1133         rw_enter(&dd->dd_pool->dp_config_rwlock, RW_READER);
1134         err = dsl_dir_open_obj(dd->dd_pool, dd->dd_object, NULL, FTAG, &dd);
1135         rw_exit(&dd->dd_pool->dp_config_rwlock);
1136
1137         if (err)
1138                 goto out_free;
1139
1140         /*
1141          * Blow away the dsl_dir + head dataset.
1142          */
1143         dsl_dataset_make_exclusive(ds, tag);
1144         /*
1145          * If we're removing a clone, we might also need to remove its
1146          * origin.
1147          */
1148         do {
1149                 dsda.need_prep = B_FALSE;
1150                 if (dsl_dir_is_clone(dd)) {
1151                         err = dsl_dataset_origin_rm_prep(&dsda, tag);
1152                         if (err) {
1153                                 dsl_dir_close(dd, FTAG);
1154                                 goto out_free;
1155                         }
1156                 }
1157
1158                 dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
1159                 dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
1160                     dsl_dataset_destroy_sync, &dsda, tag, 0);
1161                 dsl_sync_task_create(dstg, dsl_dir_destroy_check,
1162                     dsl_dir_destroy_sync, dummy_ds, FTAG, 0);
1163                 err = dsl_sync_task_group_wait(dstg);
1164                 dsl_sync_task_group_destroy(dstg);
1165
1166                 /*
1167                  * We could be racing against 'zfs release' or 'zfs destroy -d'
1168                  * on the origin snap, in which case we can get EBUSY if we
1169                  * needed to destroy the origin snap but were not ready to
1170                  * do so.
1171                  */
1172                 if (dsda.need_prep) {
1173                         ASSERT(err == EBUSY);
1174                         ASSERT(dsl_dir_is_clone(dd));
1175                         ASSERT(dsda.rm_origin == NULL);
1176                 }
1177         } while (dsda.need_prep);
1178
1179         if (dsda.rm_origin != NULL)
1180                 dsl_dataset_disown(dsda.rm_origin, tag);
1181
1182         /* if it is successful, dsl_dir_destroy_sync will close the dd */
1183         if (err)
1184                 dsl_dir_close(dd, FTAG);
1185
1186 out_free:
1187         kmem_free(dummy_ds, sizeof (dsl_dataset_t));
1188 out:
1189         dsl_dataset_disown(ds, tag);
1190         return (err);
1191 }
1192
1193 blkptr_t *
1194 dsl_dataset_get_blkptr(dsl_dataset_t *ds)
1195 {
1196         return (&ds->ds_phys->ds_bp);
1197 }
1198
1199 void
1200 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
1201 {
1202         ASSERT(dmu_tx_is_syncing(tx));
1203         /* If it's the meta-objset, set dp_meta_rootbp */
1204         if (ds == NULL) {
1205                 tx->tx_pool->dp_meta_rootbp = *bp;
1206         } else {
1207                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1208                 ds->ds_phys->ds_bp = *bp;
1209         }
1210 }
1211
1212 spa_t *
1213 dsl_dataset_get_spa(dsl_dataset_t *ds)
1214 {
1215         return (ds->ds_dir->dd_pool->dp_spa);
1216 }
1217
1218 void
1219 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
1220 {
1221         dsl_pool_t *dp;
1222
1223         if (ds == NULL) /* this is the meta-objset */
1224                 return;
1225
1226         ASSERT(ds->ds_objset != NULL);
1227
1228         if (ds->ds_phys->ds_next_snap_obj != 0)
1229                 panic("dirtying snapshot!");
1230
1231         dp = ds->ds_dir->dd_pool;
1232
1233         if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
1234                 /* up the hold count until we can be written out */
1235                 dmu_buf_add_ref(ds->ds_dbuf, ds);
1236         }
1237 }
1238
1239 /*
1240  * The unique space in the head dataset can be calculated by subtracting
1241  * the space used in the most recent snapshot, that is still being used
1242  * in this file system, from the space currently in use.  To figure out
1243  * the space in the most recent snapshot still in use, we need to take
1244  * the total space used in the snapshot and subtract out the space that
1245  * has been freed up since the snapshot was taken.
1246  */
1247 static void
1248 dsl_dataset_recalc_head_uniq(dsl_dataset_t *ds)
1249 {
1250         uint64_t mrs_used;
1251         uint64_t dlused, dlcomp, dluncomp;
1252
1253         ASSERT(!dsl_dataset_is_snapshot(ds));
1254
1255         if (ds->ds_phys->ds_prev_snap_obj != 0)
1256                 mrs_used = ds->ds_prev->ds_phys->ds_used_bytes;
1257         else
1258                 mrs_used = 0;
1259
1260         dsl_deadlist_space(&ds->ds_deadlist, &dlused, &dlcomp, &dluncomp);
1261
1262         ASSERT3U(dlused, <=, mrs_used);
1263         ds->ds_phys->ds_unique_bytes =
1264             ds->ds_phys->ds_used_bytes - (mrs_used - dlused);
1265
1266         if (spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1267             SPA_VERSION_UNIQUE_ACCURATE)
1268                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1269 }
1270
1271 struct killarg {
1272         dsl_dataset_t *ds;
1273         dmu_tx_t *tx;
1274 };
1275
1276 /* ARGSUSED */
1277 static int
1278 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
1279     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
1280 {
1281         struct killarg *ka = arg;
1282         dmu_tx_t *tx = ka->tx;
1283
1284         if (bp == NULL)
1285                 return (0);
1286
1287         if (zb->zb_level == ZB_ZIL_LEVEL) {
1288                 ASSERT(zilog != NULL);
1289                 /*
1290                  * It's a block in the intent log.  It has no
1291                  * accounting, so just free it.
1292                  */
1293                 dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
1294         } else {
1295                 ASSERT(zilog == NULL);
1296                 ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
1297                 (void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
1298         }
1299
1300         return (0);
1301 }
1302
1303 /* ARGSUSED */
1304 static int
1305 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
1306 {
1307         dsl_dataset_t *ds = arg1;
1308         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1309         uint64_t count;
1310         int err;
1311
1312         /*
1313          * Can't delete a head dataset if there are snapshots of it.
1314          * (Except if the only snapshots are from the branch we cloned
1315          * from.)
1316          */
1317         if (ds->ds_prev != NULL &&
1318             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1319                 return (EBUSY);
1320
1321         /*
1322          * This is really a dsl_dir thing, but check it here so that
1323          * we'll be less likely to leave this dataset inconsistent &
1324          * nearly destroyed.
1325          */
1326         err = zap_count(mos, ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
1327         if (err)
1328                 return (err);
1329         if (count != 0)
1330                 return (EEXIST);
1331
1332         return (0);
1333 }
1334
1335 /* ARGSUSED */
1336 static void
1337 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1338 {
1339         dsl_dataset_t *ds = arg1;
1340         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1341
1342         /* Mark it as inconsistent on-disk, in case we crash */
1343         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1344         ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1345
1346         spa_history_log_internal(LOG_DS_DESTROY_BEGIN, dp->dp_spa, tx,
1347             "dataset = %llu", ds->ds_object);
1348 }
1349
1350 static int
1351 dsl_dataset_origin_check(struct dsl_ds_destroyarg *dsda, void *tag,
1352     dmu_tx_t *tx)
1353 {
1354         dsl_dataset_t *ds = dsda->ds;
1355         dsl_dataset_t *ds_prev = ds->ds_prev;
1356
1357         if (dsl_dataset_might_destroy_origin(ds_prev)) {
1358                 struct dsl_ds_destroyarg ndsda = {0};
1359
1360                 /*
1361                  * If we're not prepared to remove the origin, don't remove
1362                  * the clone either.
1363                  */
1364                 if (dsda->rm_origin == NULL) {
1365                         dsda->need_prep = B_TRUE;
1366                         return (EBUSY);
1367                 }
1368
1369                 ndsda.ds = ds_prev;
1370                 ndsda.is_origin_rm = B_TRUE;
1371                 return (dsl_dataset_destroy_check(&ndsda, tag, tx));
1372         }
1373
1374         /*
1375          * If we're not going to remove the origin after all,
1376          * undo the open context setup.
1377          */
1378         if (dsda->rm_origin != NULL) {
1379                 dsl_dataset_disown(dsda->rm_origin, tag);
1380                 dsda->rm_origin = NULL;
1381         }
1382
1383         return (0);
1384 }
1385
1386 /*
1387  * If you add new checks here, you may need to add
1388  * additional checks to the "temporary" case in
1389  * snapshot_check() in dmu_objset.c.
1390  */
1391 /* ARGSUSED */
1392 int
1393 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1394 {
1395         struct dsl_ds_destroyarg *dsda = arg1;
1396         dsl_dataset_t *ds = dsda->ds;
1397
1398         /* we have an owner hold, so noone else can destroy us */
1399         ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1400
1401         /*
1402          * Only allow deferred destroy on pools that support it.
1403          * NOTE: deferred destroy is only supported on snapshots.
1404          */
1405         if (dsda->defer) {
1406                 if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1407                     SPA_VERSION_USERREFS)
1408                         return (ENOTSUP);
1409                 ASSERT(dsl_dataset_is_snapshot(ds));
1410                 return (0);
1411         }
1412
1413         /*
1414          * Can't delete a head dataset if there are snapshots of it.
1415          * (Except if the only snapshots are from the branch we cloned
1416          * from.)
1417          */
1418         if (ds->ds_prev != NULL &&
1419             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1420                 return (EBUSY);
1421
1422         /*
1423          * If we made changes this txg, traverse_dsl_dataset won't find
1424          * them.  Try again.
1425          */
1426         if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1427                 return (EAGAIN);
1428
1429         if (dsl_dataset_is_snapshot(ds)) {
1430                 /*
1431                  * If this snapshot has an elevated user reference count,
1432                  * we can't destroy it yet.
1433                  */
1434                 if (ds->ds_userrefs > 0 && !dsda->releasing)
1435                         return (EBUSY);
1436
1437                 mutex_enter(&ds->ds_lock);
1438                 /*
1439                  * Can't delete a branch point. However, if we're destroying
1440                  * a clone and removing its origin due to it having a user
1441                  * hold count of 0 and having been marked for deferred destroy,
1442                  * it's OK for the origin to have a single clone.
1443                  */
1444                 if (ds->ds_phys->ds_num_children >
1445                     (dsda->is_origin_rm ? 2 : 1)) {
1446                         mutex_exit(&ds->ds_lock);
1447                         return (EEXIST);
1448                 }
1449                 mutex_exit(&ds->ds_lock);
1450         } else if (dsl_dir_is_clone(ds->ds_dir)) {
1451                 return (dsl_dataset_origin_check(dsda, arg2, tx));
1452         }
1453
1454         /* XXX we should do some i/o error checking... */
1455         return (0);
1456 }
1457
1458 struct refsarg {
1459         kmutex_t lock;
1460         boolean_t gone;
1461         kcondvar_t cv;
1462 };
1463
1464 /* ARGSUSED */
1465 static void
1466 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1467 {
1468         struct refsarg *arg = argv;
1469
1470         mutex_enter(&arg->lock);
1471         arg->gone = TRUE;
1472         cv_signal(&arg->cv);
1473         mutex_exit(&arg->lock);
1474 }
1475
1476 static void
1477 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1478 {
1479         struct refsarg arg;
1480
1481         mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1482         cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1483         arg.gone = FALSE;
1484         (void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1485             dsl_dataset_refs_gone);
1486         dmu_buf_rele(ds->ds_dbuf, tag);
1487         mutex_enter(&arg.lock);
1488         while (!arg.gone)
1489                 cv_wait(&arg.cv, &arg.lock);
1490         ASSERT(arg.gone);
1491         mutex_exit(&arg.lock);
1492         ds->ds_dbuf = NULL;
1493         ds->ds_phys = NULL;
1494         mutex_destroy(&arg.lock);
1495         cv_destroy(&arg.cv);
1496 }
1497
1498 static void
1499 remove_from_next_clones(dsl_dataset_t *ds, uint64_t obj, dmu_tx_t *tx)
1500 {
1501         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1502         int err;
1503         ASSERTV(uint64_t count);
1504
1505         ASSERT(ds->ds_phys->ds_num_children >= 2);
1506         err = zap_remove_int(mos, ds->ds_phys->ds_next_clones_obj, obj, tx);
1507         /*
1508          * The err should not be ENOENT, but a bug in a previous version
1509          * of the code could cause upgrade_clones_cb() to not set
1510          * ds_next_snap_obj when it should, leading to a missing entry.
1511          * If we knew that the pool was created after
1512          * SPA_VERSION_NEXT_CLONES, we could assert that it isn't
1513          * ENOENT.  However, at least we can check that we don't have
1514          * too many entries in the next_clones_obj even after failing to
1515          * remove this one.
1516          */
1517         if (err != ENOENT) {
1518                 VERIFY3U(err, ==, 0);
1519         }
1520         ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
1521             &count));
1522         ASSERT3U(count, <=, ds->ds_phys->ds_num_children - 2);
1523 }
1524
1525 static void
1526 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
1527 {
1528         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1529         zap_cursor_t zc;
1530         zap_attribute_t za;
1531
1532         /*
1533          * If it is the old version, dd_clones doesn't exist so we can't
1534          * find the clones, but deadlist_remove_key() is a no-op so it
1535          * doesn't matter.
1536          */
1537         if (ds->ds_dir->dd_phys->dd_clones == 0)
1538                 return;
1539
1540         for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
1541             zap_cursor_retrieve(&zc, &za) == 0;
1542             zap_cursor_advance(&zc)) {
1543                 dsl_dataset_t *clone;
1544
1545                 VERIFY3U(0, ==, dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
1546                     za.za_first_integer, FTAG, &clone));
1547                 if (clone->ds_dir->dd_origin_txg > mintxg) {
1548                         dsl_deadlist_remove_key(&clone->ds_deadlist,
1549                             mintxg, tx);
1550                         dsl_dataset_remove_clones_key(clone, mintxg, tx);
1551                 }
1552                 dsl_dataset_rele(clone, FTAG);
1553         }
1554         zap_cursor_fini(&zc);
1555 }
1556
1557 struct process_old_arg {
1558         dsl_dataset_t *ds;
1559         dsl_dataset_t *ds_prev;
1560         boolean_t after_branch_point;
1561         zio_t *pio;
1562         uint64_t used, comp, uncomp;
1563 };
1564
1565 static int
1566 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1567 {
1568         struct process_old_arg *poa = arg;
1569         dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
1570
1571         if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
1572                 dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
1573                 if (poa->ds_prev && !poa->after_branch_point &&
1574                     bp->blk_birth >
1575                     poa->ds_prev->ds_phys->ds_prev_snap_txg) {
1576                         poa->ds_prev->ds_phys->ds_unique_bytes +=
1577                             bp_get_dsize_sync(dp->dp_spa, bp);
1578                 }
1579         } else {
1580                 poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
1581                 poa->comp += BP_GET_PSIZE(bp);
1582                 poa->uncomp += BP_GET_UCSIZE(bp);
1583                 dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
1584         }
1585         return (0);
1586 }
1587
1588 static void
1589 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
1590     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
1591 {
1592         struct process_old_arg poa = { 0 };
1593         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1594         objset_t *mos = dp->dp_meta_objset;
1595
1596         ASSERT(ds->ds_deadlist.dl_oldfmt);
1597         ASSERT(ds_next->ds_deadlist.dl_oldfmt);
1598
1599         poa.ds = ds;
1600         poa.ds_prev = ds_prev;
1601         poa.after_branch_point = after_branch_point;
1602         poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1603         VERIFY3U(0, ==, bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
1604             process_old_cb, &poa, tx));
1605         VERIFY3U(zio_wait(poa.pio), ==, 0);
1606         ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
1607
1608         /* change snapused */
1609         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1610             -poa.used, -poa.comp, -poa.uncomp, tx);
1611
1612         /* swap next's deadlist to our deadlist */
1613         dsl_deadlist_close(&ds->ds_deadlist);
1614         dsl_deadlist_close(&ds_next->ds_deadlist);
1615         SWITCH64(ds_next->ds_phys->ds_deadlist_obj,
1616             ds->ds_phys->ds_deadlist_obj);
1617         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
1618         dsl_deadlist_open(&ds_next->ds_deadlist, mos,
1619             ds_next->ds_phys->ds_deadlist_obj);
1620 }
1621
1622 void
1623 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1624 {
1625         struct dsl_ds_destroyarg *dsda = arg1;
1626         dsl_dataset_t *ds = dsda->ds;
1627         int err;
1628         int after_branch_point = FALSE;
1629         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1630         objset_t *mos = dp->dp_meta_objset;
1631         dsl_dataset_t *ds_prev = NULL;
1632         boolean_t wont_destroy;
1633         uint64_t obj;
1634
1635         wont_destroy = (dsda->defer &&
1636             (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1));
1637
1638         ASSERT(ds->ds_owner || wont_destroy);
1639         ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1640         ASSERT(ds->ds_prev == NULL ||
1641             ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1642         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1643
1644         if (wont_destroy) {
1645                 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1646                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1647                 ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1648                 return;
1649         }
1650
1651         /* signal any waiters that this dataset is going away */
1652         mutex_enter(&ds->ds_lock);
1653         ds->ds_owner = dsl_reaper;
1654         cv_broadcast(&ds->ds_exclusive_cv);
1655         mutex_exit(&ds->ds_lock);
1656
1657         /* Remove our reservation */
1658         if (ds->ds_reserved != 0) {
1659                 dsl_prop_setarg_t psa;
1660                 uint64_t value = 0;
1661
1662                 dsl_prop_setarg_init_uint64(&psa, "refreservation",
1663                     (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
1664                     &value);
1665                 psa.psa_effective_value = 0;    /* predict default value */
1666
1667                 dsl_dataset_set_reservation_sync(ds, &psa, tx);
1668                 ASSERT3U(ds->ds_reserved, ==, 0);
1669         }
1670
1671         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1672
1673         dsl_scan_ds_destroyed(ds, tx);
1674
1675         obj = ds->ds_object;
1676
1677         if (ds->ds_phys->ds_prev_snap_obj != 0) {
1678                 if (ds->ds_prev) {
1679                         ds_prev = ds->ds_prev;
1680                 } else {
1681                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1682                             ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1683                 }
1684                 after_branch_point =
1685                     (ds_prev->ds_phys->ds_next_snap_obj != obj);
1686
1687                 dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1688                 if (after_branch_point &&
1689                     ds_prev->ds_phys->ds_next_clones_obj != 0) {
1690                         remove_from_next_clones(ds_prev, obj, tx);
1691                         if (ds->ds_phys->ds_next_snap_obj != 0) {
1692                                 VERIFY(0 == zap_add_int(mos,
1693                                     ds_prev->ds_phys->ds_next_clones_obj,
1694                                     ds->ds_phys->ds_next_snap_obj, tx));
1695                         }
1696                 }
1697                 if (after_branch_point &&
1698                     ds->ds_phys->ds_next_snap_obj == 0) {
1699                         /* This clone is toast. */
1700                         ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1701                         ds_prev->ds_phys->ds_num_children--;
1702
1703                         /*
1704                          * If the clone's origin has no other clones, no
1705                          * user holds, and has been marked for deferred
1706                          * deletion, then we should have done the necessary
1707                          * destroy setup for it.
1708                          */
1709                         if (ds_prev->ds_phys->ds_num_children == 1 &&
1710                             ds_prev->ds_userrefs == 0 &&
1711                             DS_IS_DEFER_DESTROY(ds_prev)) {
1712                                 ASSERT3P(dsda->rm_origin, !=, NULL);
1713                         } else {
1714                                 ASSERT3P(dsda->rm_origin, ==, NULL);
1715                         }
1716                 } else if (!after_branch_point) {
1717                         ds_prev->ds_phys->ds_next_snap_obj =
1718                             ds->ds_phys->ds_next_snap_obj;
1719                 }
1720         }
1721
1722         if (dsl_dataset_is_snapshot(ds)) {
1723                 dsl_dataset_t *ds_next;
1724                 uint64_t old_unique;
1725                 uint64_t used = 0, comp = 0, uncomp = 0;
1726
1727                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1728                     ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1729                 ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1730
1731                 old_unique = ds_next->ds_phys->ds_unique_bytes;
1732
1733                 dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1734                 ds_next->ds_phys->ds_prev_snap_obj =
1735                     ds->ds_phys->ds_prev_snap_obj;
1736                 ds_next->ds_phys->ds_prev_snap_txg =
1737                     ds->ds_phys->ds_prev_snap_txg;
1738                 ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1739                     ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1740
1741
1742                 if (ds_next->ds_deadlist.dl_oldfmt) {
1743                         process_old_deadlist(ds, ds_prev, ds_next,
1744                             after_branch_point, tx);
1745                 } else {
1746                         /* Adjust prev's unique space. */
1747                         if (ds_prev && !after_branch_point) {
1748                                 dsl_deadlist_space_range(&ds_next->ds_deadlist,
1749                                     ds_prev->ds_phys->ds_prev_snap_txg,
1750                                     ds->ds_phys->ds_prev_snap_txg,
1751                                     &used, &comp, &uncomp);
1752                                 ds_prev->ds_phys->ds_unique_bytes += used;
1753                         }
1754
1755                         /* Adjust snapused. */
1756                         dsl_deadlist_space_range(&ds_next->ds_deadlist,
1757                             ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
1758                             &used, &comp, &uncomp);
1759                         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1760                             -used, -comp, -uncomp, tx);
1761
1762                         /* Move blocks to be freed to pool's free list. */
1763                         dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
1764                             &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
1765                             tx);
1766                         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
1767                             DD_USED_HEAD, used, comp, uncomp, tx);
1768                         dsl_dir_dirty(tx->tx_pool->dp_free_dir, tx);
1769
1770                         /* Merge our deadlist into next's and free it. */
1771                         dsl_deadlist_merge(&ds_next->ds_deadlist,
1772                             ds->ds_phys->ds_deadlist_obj, tx);
1773                 }
1774                 dsl_deadlist_close(&ds->ds_deadlist);
1775                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1776
1777                 /* Collapse range in clone heads */
1778                 dsl_dataset_remove_clones_key(ds,
1779                     ds->ds_phys->ds_creation_txg, tx);
1780
1781                 if (dsl_dataset_is_snapshot(ds_next)) {
1782                         dsl_dataset_t *ds_nextnext;
1783                         dsl_dataset_t *hds;
1784
1785                         /*
1786                          * Update next's unique to include blocks which
1787                          * were previously shared by only this snapshot
1788                          * and it.  Those blocks will be born after the
1789                          * prev snap and before this snap, and will have
1790                          * died after the next snap and before the one
1791                          * after that (ie. be on the snap after next's
1792                          * deadlist).
1793                          */
1794                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1795                             ds_next->ds_phys->ds_next_snap_obj,
1796                             FTAG, &ds_nextnext));
1797                         dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
1798                             ds->ds_phys->ds_prev_snap_txg,
1799                             ds->ds_phys->ds_creation_txg,
1800                             &used, &comp, &uncomp);
1801                         ds_next->ds_phys->ds_unique_bytes += used;
1802                         dsl_dataset_rele(ds_nextnext, FTAG);
1803                         ASSERT3P(ds_next->ds_prev, ==, NULL);
1804
1805                         /* Collapse range in this head. */
1806                         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
1807                             ds->ds_dir->dd_phys->dd_head_dataset_obj,
1808                             FTAG, &hds));
1809                         dsl_deadlist_remove_key(&hds->ds_deadlist,
1810                             ds->ds_phys->ds_creation_txg, tx);
1811                         dsl_dataset_rele(hds, FTAG);
1812
1813                 } else {
1814                         ASSERT3P(ds_next->ds_prev, ==, ds);
1815                         dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1816                         ds_next->ds_prev = NULL;
1817                         if (ds_prev) {
1818                                 VERIFY(0 == dsl_dataset_get_ref(dp,
1819                                     ds->ds_phys->ds_prev_snap_obj,
1820                                     ds_next, &ds_next->ds_prev));
1821                         }
1822
1823                         dsl_dataset_recalc_head_uniq(ds_next);
1824
1825                         /*
1826                          * Reduce the amount of our unconsmed refreservation
1827                          * being charged to our parent by the amount of
1828                          * new unique data we have gained.
1829                          */
1830                         if (old_unique < ds_next->ds_reserved) {
1831                                 int64_t mrsdelta;
1832                                 uint64_t new_unique =
1833                                     ds_next->ds_phys->ds_unique_bytes;
1834
1835                                 ASSERT(old_unique <= new_unique);
1836                                 mrsdelta = MIN(new_unique - old_unique,
1837                                     ds_next->ds_reserved - old_unique);
1838                                 dsl_dir_diduse_space(ds->ds_dir,
1839                                     DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1840                         }
1841                 }
1842                 dsl_dataset_rele(ds_next, FTAG);
1843         } else {
1844                 /*
1845                  * There's no next snapshot, so this is a head dataset.
1846                  * Destroy the deadlist.  Unless it's a clone, the
1847                  * deadlist should be empty.  (If it's a clone, it's
1848                  * safe to ignore the deadlist contents.)
1849                  */
1850                 struct killarg ka;
1851
1852                 dsl_deadlist_close(&ds->ds_deadlist);
1853                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1854                 ds->ds_phys->ds_deadlist_obj = 0;
1855
1856                 /*
1857                  * Free everything that we point to (that's born after
1858                  * the previous snapshot, if we are a clone)
1859                  *
1860                  * NB: this should be very quick, because we already
1861                  * freed all the objects in open context.
1862                  */
1863                 ka.ds = ds;
1864                 ka.tx = tx;
1865                 err = traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1866                     TRAVERSE_POST, kill_blkptr, &ka);
1867                 ASSERT3U(err, ==, 0);
1868                 ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1869                     ds->ds_phys->ds_unique_bytes == 0);
1870
1871                 if (ds->ds_prev != NULL) {
1872                         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
1873                                 VERIFY3U(0, ==, zap_remove_int(mos,
1874                                     ds->ds_prev->ds_dir->dd_phys->dd_clones,
1875                                     ds->ds_object, tx));
1876                         }
1877                         dsl_dataset_rele(ds->ds_prev, ds);
1878                         ds->ds_prev = ds_prev = NULL;
1879                 }
1880         }
1881
1882         /*
1883          * This must be done after the dsl_traverse(), because it will
1884          * re-open the objset.
1885          */
1886         if (ds->ds_objset) {
1887                 dmu_objset_evict(ds->ds_objset);
1888                 ds->ds_objset = NULL;
1889         }
1890
1891         if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1892                 /* Erase the link in the dir */
1893                 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1894                 ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1895                 ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1896                 err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1897                 ASSERT(err == 0);
1898         } else {
1899                 /* remove from snapshot namespace */
1900                 dsl_dataset_t *ds_head;
1901                 ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1902                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1903                     ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1904                 VERIFY(0 == dsl_dataset_get_snapname(ds));
1905 #ifdef ZFS_DEBUG
1906                 {
1907                         uint64_t val;
1908
1909                         err = dsl_dataset_snap_lookup(ds_head,
1910                             ds->ds_snapname, &val);
1911                         ASSERT3U(err, ==, 0);
1912                         ASSERT3U(val, ==, obj);
1913                 }
1914 #endif
1915                 err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1916                 ASSERT(err == 0);
1917                 dsl_dataset_rele(ds_head, FTAG);
1918         }
1919
1920         if (ds_prev && ds->ds_prev != ds_prev)
1921                 dsl_dataset_rele(ds_prev, FTAG);
1922
1923         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1924         spa_history_log_internal(LOG_DS_DESTROY, dp->dp_spa, tx,
1925             "dataset = %llu", ds->ds_object);
1926
1927         if (ds->ds_phys->ds_next_clones_obj != 0) {
1928                 ASSERTV(uint64_t count);
1929                 ASSERT(0 == zap_count(mos,
1930                     ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1931                 VERIFY(0 == dmu_object_free(mos,
1932                     ds->ds_phys->ds_next_clones_obj, tx));
1933         }
1934         if (ds->ds_phys->ds_props_obj != 0)
1935                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1936         if (ds->ds_phys->ds_userrefs_obj != 0)
1937                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1938         dsl_dir_close(ds->ds_dir, ds);
1939         ds->ds_dir = NULL;
1940         dsl_dataset_drain_refs(ds, tag);
1941         VERIFY(0 == dmu_object_free(mos, obj, tx));
1942
1943         if (dsda->rm_origin) {
1944                 /*
1945                  * Remove the origin of the clone we just destroyed.
1946                  */
1947                 struct dsl_ds_destroyarg ndsda = {0};
1948
1949                 ndsda.ds = dsda->rm_origin;
1950                 dsl_dataset_destroy_sync(&ndsda, tag, tx);
1951         }
1952 }
1953
1954 static int
1955 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
1956 {
1957         uint64_t asize;
1958
1959         if (!dmu_tx_is_syncing(tx))
1960                 return (0);
1961
1962         /*
1963          * If there's an fs-only reservation, any blocks that might become
1964          * owned by the snapshot dataset must be accommodated by space
1965          * outside of the reservation.
1966          */
1967         ASSERT(ds->ds_reserved == 0 || DS_UNIQUE_IS_ACCURATE(ds));
1968         asize = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
1969         if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
1970                 return (ENOSPC);
1971
1972         /*
1973          * Propogate any reserved space for this snapshot to other
1974          * snapshot checks in this sync group.
1975          */
1976         if (asize > 0)
1977                 dsl_dir_willuse_space(ds->ds_dir, asize, tx);
1978
1979         return (0);
1980 }
1981
1982 int
1983 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1984 {
1985         dsl_dataset_t *ds = arg1;
1986         const char *snapname = arg2;
1987         int err;
1988         uint64_t value;
1989
1990         /*
1991          * We don't allow multiple snapshots of the same txg.  If there
1992          * is already one, try again.
1993          */
1994         if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
1995                 return (EAGAIN);
1996
1997         /*
1998          * Check for conflicting name snapshot name.
1999          */
2000         err = dsl_dataset_snap_lookup(ds, snapname, &value);
2001         if (err == 0)
2002                 return (EEXIST);
2003         if (err != ENOENT)
2004                 return (err);
2005
2006         /*
2007          * Check that the dataset's name is not too long.  Name consists
2008          * of the dataset's length + 1 for the @-sign + snapshot name's length
2009          */
2010         if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
2011                 return (ENAMETOOLONG);
2012
2013         err = dsl_dataset_snapshot_reserve_space(ds, tx);
2014         if (err)
2015                 return (err);
2016
2017         ds->ds_trysnap_txg = tx->tx_txg;
2018         return (0);
2019 }
2020
2021 void
2022 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2023 {
2024         dsl_dataset_t *ds = arg1;
2025         const char *snapname = arg2;
2026         dsl_pool_t *dp = ds->ds_dir->dd_pool;
2027         dmu_buf_t *dbuf;
2028         dsl_dataset_phys_t *dsphys;
2029         uint64_t dsobj, crtxg;
2030         objset_t *mos = dp->dp_meta_objset;
2031         int err;
2032
2033         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
2034
2035         /*
2036          * The origin's ds_creation_txg has to be < TXG_INITIAL
2037          */
2038         if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2039                 crtxg = 1;
2040         else
2041                 crtxg = tx->tx_txg;
2042
2043         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2044             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2045         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2046         dmu_buf_will_dirty(dbuf, tx);
2047         dsphys = dbuf->db_data;
2048         bzero(dsphys, sizeof (dsl_dataset_phys_t));
2049         dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2050         dsphys->ds_fsid_guid = unique_create();
2051         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2052             sizeof (dsphys->ds_guid));
2053         dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2054         dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2055         dsphys->ds_next_snap_obj = ds->ds_object;
2056         dsphys->ds_num_children = 1;
2057         dsphys->ds_creation_time = gethrestime_sec();
2058         dsphys->ds_creation_txg = crtxg;
2059         dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2060         dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
2061         dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2062         dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2063         dsphys->ds_flags = ds->ds_phys->ds_flags;
2064         dsphys->ds_bp = ds->ds_phys->ds_bp;
2065         dmu_buf_rele(dbuf, FTAG);
2066
2067         ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2068         if (ds->ds_prev) {
2069                 uint64_t next_clones_obj =
2070                     ds->ds_prev->ds_phys->ds_next_clones_obj;
2071                 ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2072                     ds->ds_object ||
2073                     ds->ds_prev->ds_phys->ds_num_children > 1);
2074                 if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2075                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2076                         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2077                             ds->ds_prev->ds_phys->ds_creation_txg);
2078                         ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2079                 } else if (next_clones_obj != 0) {
2080                         remove_from_next_clones(ds->ds_prev,
2081                             dsphys->ds_next_snap_obj, tx);
2082                         VERIFY3U(0, ==, zap_add_int(mos,
2083                             next_clones_obj, dsobj, tx));
2084                 }
2085         }
2086
2087         /*
2088          * If we have a reference-reservation on this dataset, we will
2089          * need to increase the amount of refreservation being charged
2090          * since our unique space is going to zero.
2091          */
2092         if (ds->ds_reserved) {
2093                 int64_t delta;
2094                 ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
2095                 delta = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2096                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2097                     delta, 0, 0, tx);
2098         }
2099
2100         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2101         zfs_dbgmsg("taking snapshot %s@%s/%llu; newkey=%llu",
2102             ds->ds_dir->dd_myname, snapname, dsobj,
2103             ds->ds_phys->ds_prev_snap_txg);
2104         ds->ds_phys->ds_deadlist_obj = dsl_deadlist_clone(&ds->ds_deadlist,
2105             UINT64_MAX, ds->ds_phys->ds_prev_snap_obj, tx);
2106         dsl_deadlist_close(&ds->ds_deadlist);
2107         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
2108         dsl_deadlist_add_key(&ds->ds_deadlist,
2109             ds->ds_phys->ds_prev_snap_txg, tx);
2110
2111         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2112         ds->ds_phys->ds_prev_snap_obj = dsobj;
2113         ds->ds_phys->ds_prev_snap_txg = crtxg;
2114         ds->ds_phys->ds_unique_bytes = 0;
2115         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2116                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2117
2118         err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2119             snapname, 8, 1, &dsobj, tx);
2120         ASSERT(err == 0);
2121
2122         if (ds->ds_prev)
2123                 dsl_dataset_drop_ref(ds->ds_prev, ds);
2124         VERIFY(0 == dsl_dataset_get_ref(dp,
2125             ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2126
2127         dsl_scan_ds_snapshotted(ds, tx);
2128
2129         dsl_dir_snap_cmtime_update(ds->ds_dir);
2130
2131         spa_history_log_internal(LOG_DS_SNAPSHOT, dp->dp_spa, tx,
2132             "dataset = %llu", dsobj);
2133 }
2134
2135 void
2136 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2137 {
2138         ASSERT(dmu_tx_is_syncing(tx));
2139         ASSERT(ds->ds_objset != NULL);
2140         ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2141
2142         /*
2143          * in case we had to change ds_fsid_guid when we opened it,
2144          * sync it out now.
2145          */
2146         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2147         ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2148
2149         dsl_dir_dirty(ds->ds_dir, tx);
2150         dmu_objset_sync(ds->ds_objset, zio, tx);
2151 }
2152
2153 void
2154 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2155 {
2156         uint64_t refd, avail, uobjs, aobjs;
2157
2158         dsl_dir_stats(ds->ds_dir, nv);
2159
2160         dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2161         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2162         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2163
2164         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2165             ds->ds_phys->ds_creation_time);
2166         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2167             ds->ds_phys->ds_creation_txg);
2168         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2169             ds->ds_quota);
2170         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2171             ds->ds_reserved);
2172         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2173             ds->ds_phys->ds_guid);
2174         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_UNIQUE,
2175             ds->ds_phys->ds_unique_bytes);
2176         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_OBJSETID,
2177             ds->ds_object);
2178         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS,
2179             ds->ds_userrefs);
2180         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2181             DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2182
2183         if (ds->ds_phys->ds_next_snap_obj) {
2184                 /*
2185                  * This is a snapshot; override the dd's space used with
2186                  * our unique space and compression ratio.
2187                  */
2188                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2189                     ds->ds_phys->ds_unique_bytes);
2190                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO,
2191                     ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2192                     (ds->ds_phys->ds_uncompressed_bytes * 100 /
2193                     ds->ds_phys->ds_compressed_bytes));
2194         }
2195 }
2196
2197 void
2198 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2199 {
2200         stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2201         stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2202         stat->dds_guid = ds->ds_phys->ds_guid;
2203         if (ds->ds_phys->ds_next_snap_obj) {
2204                 stat->dds_is_snapshot = B_TRUE;
2205                 stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2206         } else {
2207                 stat->dds_is_snapshot = B_FALSE;
2208                 stat->dds_num_clones = 0;
2209         }
2210
2211         /* clone origin is really a dsl_dir thing... */
2212         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2213         if (dsl_dir_is_clone(ds->ds_dir)) {
2214                 dsl_dataset_t *ods;
2215
2216                 VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2217                     ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2218                 dsl_dataset_name(ods, stat->dds_origin);
2219                 dsl_dataset_drop_ref(ods, FTAG);
2220         } else {
2221                 stat->dds_origin[0] = '\0';
2222         }
2223         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2224 }
2225
2226 uint64_t
2227 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2228 {
2229         return (ds->ds_fsid_guid);
2230 }
2231
2232 void
2233 dsl_dataset_space(dsl_dataset_t *ds,
2234     uint64_t *refdbytesp, uint64_t *availbytesp,
2235     uint64_t *usedobjsp, uint64_t *availobjsp)
2236 {
2237         *refdbytesp = ds->ds_phys->ds_used_bytes;
2238         *availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2239         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2240                 *availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2241         if (ds->ds_quota != 0) {
2242                 /*
2243                  * Adjust available bytes according to refquota
2244                  */
2245                 if (*refdbytesp < ds->ds_quota)
2246                         *availbytesp = MIN(*availbytesp,
2247                             ds->ds_quota - *refdbytesp);
2248                 else
2249                         *availbytesp = 0;
2250         }
2251         *usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2252         *availobjsp = DN_MAX_OBJECT - *usedobjsp;
2253 }
2254
2255 boolean_t
2256 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2257 {
2258         ASSERTV(dsl_pool_t *dp = ds->ds_dir->dd_pool);
2259
2260         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2261             dsl_pool_sync_context(dp));
2262         if (ds->ds_prev == NULL)
2263                 return (B_FALSE);
2264         if (ds->ds_phys->ds_bp.blk_birth >
2265             ds->ds_prev->ds_phys->ds_creation_txg) {
2266                 objset_t *os, *os_prev;
2267                 /*
2268                  * It may be that only the ZIL differs, because it was
2269                  * reset in the head.  Don't count that as being
2270                  * modified.
2271                  */
2272                 if (dmu_objset_from_ds(ds, &os) != 0)
2273                         return (B_TRUE);
2274                 if (dmu_objset_from_ds(ds->ds_prev, &os_prev) != 0)
2275                         return (B_TRUE);
2276                 return (bcmp(&os->os_phys->os_meta_dnode,
2277                     &os_prev->os_phys->os_meta_dnode,
2278                     sizeof (os->os_phys->os_meta_dnode)) != 0);
2279         }
2280         return (B_FALSE);
2281 }
2282
2283 /* ARGSUSED */
2284 static int
2285 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2286 {
2287         dsl_dataset_t *ds = arg1;
2288         char *newsnapname = arg2;
2289         dsl_dir_t *dd = ds->ds_dir;
2290         dsl_dataset_t *hds;
2291         uint64_t val;
2292         int err;
2293
2294         err = dsl_dataset_hold_obj(dd->dd_pool,
2295             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2296         if (err)
2297                 return (err);
2298
2299         /* new name better not be in use */
2300         err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2301         dsl_dataset_rele(hds, FTAG);
2302
2303         if (err == 0)
2304                 err = EEXIST;
2305         else if (err == ENOENT)
2306                 err = 0;
2307
2308         /* dataset name + 1 for the "@" + the new snapshot name must fit */
2309         if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2310                 err = ENAMETOOLONG;
2311
2312         return (err);
2313 }
2314
2315 static void
2316 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2317 {
2318         dsl_dataset_t *ds = arg1;
2319         const char *newsnapname = arg2;
2320         dsl_dir_t *dd = ds->ds_dir;
2321         objset_t *mos = dd->dd_pool->dp_meta_objset;
2322         dsl_dataset_t *hds;
2323         int err;
2324
2325         ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2326
2327         VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2328             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2329
2330         VERIFY(0 == dsl_dataset_get_snapname(ds));
2331         err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2332         ASSERT3U(err, ==, 0);
2333         mutex_enter(&ds->ds_lock);
2334         (void) strcpy(ds->ds_snapname, newsnapname);
2335         mutex_exit(&ds->ds_lock);
2336         err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2337             ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2338         ASSERT3U(err, ==, 0);
2339
2340         spa_history_log_internal(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2341             "dataset = %llu", ds->ds_object);
2342         dsl_dataset_rele(hds, FTAG);
2343 }
2344
2345 struct renamesnaparg {
2346         dsl_sync_task_group_t *dstg;
2347         char failed[MAXPATHLEN];
2348         char *oldsnap;
2349         char *newsnap;
2350 };
2351
2352 static int
2353 dsl_snapshot_rename_one(const char *name, void *arg)
2354 {
2355         struct renamesnaparg *ra = arg;
2356         dsl_dataset_t *ds = NULL;
2357         char *snapname;
2358         int err;
2359
2360         snapname = kmem_asprintf("%s@%s", name, ra->oldsnap);
2361         (void) strlcpy(ra->failed, snapname, sizeof (ra->failed));
2362
2363         /*
2364          * For recursive snapshot renames the parent won't be changing
2365          * so we just pass name for both the to/from argument.
2366          */
2367         err = zfs_secpolicy_rename_perms(snapname, snapname, CRED());
2368         if (err != 0) {
2369                 strfree(snapname);
2370                 return (err == ENOENT ? 0 : err);
2371         }
2372
2373 /* XXX: Ignore for SPL version until mounting the FS is supported */
2374 #if defined(_KERNEL) && !defined(HAVE_SPL)
2375         /*
2376          * For all filesystems undergoing rename, we'll need to unmount it.
2377          */
2378         (void) zfs_unmount_snap(snapname, NULL);
2379 #endif
2380         err = dsl_dataset_hold(snapname, ra->dstg, &ds);
2381         strfree(snapname);
2382         if (err != 0)
2383                 return (err == ENOENT ? 0 : err);
2384
2385         dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2386             dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2387
2388         return (0);
2389 }
2390
2391 static int
2392 dsl_recursive_rename(char *oldname, const char *newname)
2393 {
2394         int err;
2395         struct renamesnaparg *ra;
2396         dsl_sync_task_t *dst;
2397         spa_t *spa;
2398         char *cp, *fsname = spa_strdup(oldname);
2399         int len = strlen(oldname) + 1;
2400
2401         /* truncate the snapshot name to get the fsname */
2402         cp = strchr(fsname, '@');
2403         *cp = '\0';
2404
2405         err = spa_open(fsname, &spa, FTAG);
2406         if (err) {
2407                 kmem_free(fsname, len);
2408                 return (err);
2409         }
2410         ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2411         ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2412
2413         ra->oldsnap = strchr(oldname, '@') + 1;
2414         ra->newsnap = strchr(newname, '@') + 1;
2415         *ra->failed = '\0';
2416
2417         err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2418             DS_FIND_CHILDREN);
2419         kmem_free(fsname, len);
2420
2421         if (err == 0) {
2422                 err = dsl_sync_task_group_wait(ra->dstg);
2423         }
2424
2425         for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2426             dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2427                 dsl_dataset_t *ds = dst->dst_arg1;
2428                 if (dst->dst_err) {
2429                         dsl_dir_name(ds->ds_dir, ra->failed);
2430                         (void) strlcat(ra->failed, "@", sizeof (ra->failed));
2431                         (void) strlcat(ra->failed, ra->newsnap,
2432                             sizeof (ra->failed));
2433                 }
2434                 dsl_dataset_rele(ds, ra->dstg);
2435         }
2436
2437         if (err)
2438                 (void) strlcpy(oldname, ra->failed, sizeof (ra->failed));
2439
2440         dsl_sync_task_group_destroy(ra->dstg);
2441         kmem_free(ra, sizeof (struct renamesnaparg));
2442         spa_close(spa, FTAG);
2443         return (err);
2444 }
2445
2446 static int
2447 dsl_valid_rename(const char *oldname, void *arg)
2448 {
2449         int delta = *(int *)arg;
2450
2451         if (strlen(oldname) + delta >= MAXNAMELEN)
2452                 return (ENAMETOOLONG);
2453
2454         return (0);
2455 }
2456
2457 #pragma weak dmu_objset_rename = dsl_dataset_rename
2458 int
2459 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2460 {
2461         dsl_dir_t *dd;
2462         dsl_dataset_t *ds;
2463         const char *tail;
2464         int err;
2465
2466         err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2467         if (err)
2468                 return (err);
2469
2470         if (tail == NULL) {
2471                 int delta = strlen(newname) - strlen(oldname);
2472
2473                 /* if we're growing, validate child name lengths */
2474                 if (delta > 0)
2475                         err = dmu_objset_find(oldname, dsl_valid_rename,
2476                             &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2477
2478                 if (err == 0)
2479                         err = dsl_dir_rename(dd, newname);
2480                 dsl_dir_close(dd, FTAG);
2481                 return (err);
2482         }
2483
2484         if (tail[0] != '@') {
2485                 /* the name ended in a nonexistent component */
2486                 dsl_dir_close(dd, FTAG);
2487                 return (ENOENT);
2488         }
2489
2490         dsl_dir_close(dd, FTAG);
2491
2492         /* new name must be snapshot in same filesystem */
2493         tail = strchr(newname, '@');
2494         if (tail == NULL)
2495                 return (EINVAL);
2496         tail++;
2497         if (strncmp(oldname, newname, tail - newname) != 0)
2498                 return (EXDEV);
2499
2500         if (recursive) {
2501                 err = dsl_recursive_rename(oldname, newname);
2502         } else {
2503                 err = dsl_dataset_hold(oldname, FTAG, &ds);
2504                 if (err)
2505                         return (err);
2506
2507                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2508                     dsl_dataset_snapshot_rename_check,
2509                     dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2510
2511                 dsl_dataset_rele(ds, FTAG);
2512         }
2513
2514         return (err);
2515 }
2516
2517 struct promotenode {
2518         list_node_t link;
2519         dsl_dataset_t *ds;
2520 };
2521
2522 struct promotearg {
2523         list_t shared_snaps, origin_snaps, clone_snaps;
2524         dsl_dataset_t *origin_origin;
2525         uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2526         char *err_ds;
2527 };
2528
2529 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2530
2531 static int
2532 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2533 {
2534         dsl_dataset_t *hds = arg1;
2535         struct promotearg *pa = arg2;
2536         struct promotenode *snap = list_head(&pa->shared_snaps);
2537         dsl_dataset_t *origin_ds = snap->ds;
2538         int err;
2539         uint64_t unused;
2540
2541         /* Check that it is a real clone */
2542         if (!dsl_dir_is_clone(hds->ds_dir))
2543                 return (EINVAL);
2544
2545         /* Since this is so expensive, don't do the preliminary check */
2546         if (!dmu_tx_is_syncing(tx))
2547                 return (0);
2548
2549         if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2550                 return (EXDEV);
2551
2552         /* compute origin's new unique space */
2553         snap = list_tail(&pa->clone_snaps);
2554         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2555         dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2556             origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
2557             &pa->unique, &unused, &unused);
2558
2559         /*
2560          * Walk the snapshots that we are moving
2561          *
2562          * Compute space to transfer.  Consider the incremental changes
2563          * to used for each snapshot:
2564          * (my used) = (prev's used) + (blocks born) - (blocks killed)
2565          * So each snapshot gave birth to:
2566          * (blocks born) = (my used) - (prev's used) + (blocks killed)
2567          * So a sequence would look like:
2568          * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2569          * Which simplifies to:
2570          * uN + kN + kN-1 + ... + k1 + k0
2571          * Note however, if we stop before we reach the ORIGIN we get:
2572          * uN + kN + kN-1 + ... + kM - uM-1
2573          */
2574         pa->used = origin_ds->ds_phys->ds_used_bytes;
2575         pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2576         pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2577         for (snap = list_head(&pa->shared_snaps); snap;
2578             snap = list_next(&pa->shared_snaps, snap)) {
2579                 uint64_t val, dlused, dlcomp, dluncomp;
2580                 dsl_dataset_t *ds = snap->ds;
2581
2582                 /* Check that the snapshot name does not conflict */
2583                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2584                 err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2585                 if (err == 0) {
2586                         err = EEXIST;
2587                         goto out;
2588                 }
2589                 if (err != ENOENT)
2590                         goto out;
2591
2592                 /* The very first snapshot does not have a deadlist */
2593                 if (ds->ds_phys->ds_prev_snap_obj == 0)
2594                         continue;
2595
2596                 dsl_deadlist_space(&ds->ds_deadlist,
2597                     &dlused, &dlcomp, &dluncomp);
2598                 pa->used += dlused;
2599                 pa->comp += dlcomp;
2600                 pa->uncomp += dluncomp;
2601         }
2602
2603         /*
2604          * If we are a clone of a clone then we never reached ORIGIN,
2605          * so we need to subtract out the clone origin's used space.
2606          */
2607         if (pa->origin_origin) {
2608                 pa->used -= pa->origin_origin->ds_phys->ds_used_bytes;
2609                 pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2610                 pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2611         }
2612
2613         /* Check that there is enough space here */
2614         err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2615             pa->used);
2616         if (err)
2617                 return (err);
2618
2619         /*
2620          * Compute the amounts of space that will be used by snapshots
2621          * after the promotion (for both origin and clone).  For each,
2622          * it is the amount of space that will be on all of their
2623          * deadlists (that was not born before their new origin).
2624          */
2625         if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2626                 uint64_t space;
2627
2628                 /*
2629                  * Note, typically this will not be a clone of a clone,
2630                  * so dd_origin_txg will be < TXG_INITIAL, so
2631                  * these snaplist_space() -> dsl_deadlist_space_range()
2632                  * calls will be fast because they do not have to
2633                  * iterate over all bps.
2634                  */
2635                 snap = list_head(&pa->origin_snaps);
2636                 err = snaplist_space(&pa->shared_snaps,
2637                     snap->ds->ds_dir->dd_origin_txg, &pa->cloneusedsnap);
2638                 if (err)
2639                         return (err);
2640
2641                 err = snaplist_space(&pa->clone_snaps,
2642                     snap->ds->ds_dir->dd_origin_txg, &space);
2643                 if (err)
2644                         return (err);
2645                 pa->cloneusedsnap += space;
2646         }
2647         if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2648                 err = snaplist_space(&pa->origin_snaps,
2649                     origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2650                 if (err)
2651                         return (err);
2652         }
2653
2654         return (0);
2655 out:
2656         pa->err_ds =  snap->ds->ds_snapname;
2657         return (err);
2658 }
2659
2660 static void
2661 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2662 {
2663         dsl_dataset_t *hds = arg1;
2664         struct promotearg *pa = arg2;
2665         struct promotenode *snap = list_head(&pa->shared_snaps);
2666         dsl_dataset_t *origin_ds = snap->ds;
2667         dsl_dataset_t *origin_head;
2668         dsl_dir_t *dd = hds->ds_dir;
2669         dsl_pool_t *dp = hds->ds_dir->dd_pool;
2670         dsl_dir_t *odd = NULL;
2671         uint64_t oldnext_obj;
2672         int64_t delta;
2673
2674         ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2675
2676         snap = list_head(&pa->origin_snaps);
2677         origin_head = snap->ds;
2678
2679         /*
2680          * We need to explicitly open odd, since origin_ds's dd will be
2681          * changing.
2682          */
2683         VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2684             NULL, FTAG, &odd));
2685
2686         /* change origin's next snap */
2687         dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2688         oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2689         snap = list_tail(&pa->clone_snaps);
2690         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2691         origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2692
2693         /* change the origin's next clone */
2694         if (origin_ds->ds_phys->ds_next_clones_obj) {
2695                 remove_from_next_clones(origin_ds, snap->ds->ds_object, tx);
2696                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2697                     origin_ds->ds_phys->ds_next_clones_obj,
2698                     oldnext_obj, tx));
2699         }
2700
2701         /* change origin */
2702         dmu_buf_will_dirty(dd->dd_dbuf, tx);
2703         ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2704         dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2705         dd->dd_origin_txg = origin_head->ds_dir->dd_origin_txg;
2706         dmu_buf_will_dirty(odd->dd_dbuf, tx);
2707         odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2708         origin_head->ds_dir->dd_origin_txg =
2709             origin_ds->ds_phys->ds_creation_txg;
2710
2711         /* change dd_clone entries */
2712         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2713                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2714                     odd->dd_phys->dd_clones, hds->ds_object, tx));
2715                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2716                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2717                     hds->ds_object, tx));
2718
2719                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2720                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2721                     origin_head->ds_object, tx));
2722                 if (dd->dd_phys->dd_clones == 0) {
2723                         dd->dd_phys->dd_clones = zap_create(dp->dp_meta_objset,
2724                             DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
2725                 }
2726                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2727                     dd->dd_phys->dd_clones, origin_head->ds_object, tx));
2728
2729         }
2730
2731         /* move snapshots to this dir */
2732         for (snap = list_head(&pa->shared_snaps); snap;
2733             snap = list_next(&pa->shared_snaps, snap)) {
2734                 dsl_dataset_t *ds = snap->ds;
2735
2736                 /* unregister props as dsl_dir is changing */
2737                 if (ds->ds_objset) {
2738                         dmu_objset_evict(ds->ds_objset);
2739                         ds->ds_objset = NULL;
2740                 }
2741                 /* move snap name entry */
2742                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2743                 VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2744                     ds->ds_snapname, tx));
2745                 VERIFY(0 == zap_add(dp->dp_meta_objset,
2746                     hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2747                     8, 1, &ds->ds_object, tx));
2748
2749                 /* change containing dsl_dir */
2750                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2751                 ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2752                 ds->ds_phys->ds_dir_obj = dd->dd_object;
2753                 ASSERT3P(ds->ds_dir, ==, odd);
2754                 dsl_dir_close(ds->ds_dir, ds);
2755                 VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2756                     NULL, ds, &ds->ds_dir));
2757
2758                 /* move any clone references */
2759                 if (ds->ds_phys->ds_next_clones_obj &&
2760                     spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2761                         zap_cursor_t zc;
2762                         zap_attribute_t za;
2763
2764                         for (zap_cursor_init(&zc, dp->dp_meta_objset,
2765                             ds->ds_phys->ds_next_clones_obj);
2766                             zap_cursor_retrieve(&zc, &za) == 0;
2767                             zap_cursor_advance(&zc)) {
2768                                 dsl_dataset_t *cnds;
2769                                 uint64_t o;
2770
2771                                 if (za.za_first_integer == oldnext_obj) {
2772                                         /*
2773                                          * We've already moved the
2774                                          * origin's reference.
2775                                          */
2776                                         continue;
2777                                 }
2778
2779                                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
2780                                     za.za_first_integer, FTAG, &cnds));
2781                                 o = cnds->ds_dir->dd_phys->dd_head_dataset_obj;
2782
2783                                 VERIFY3U(zap_remove_int(dp->dp_meta_objset,
2784                                     odd->dd_phys->dd_clones, o, tx), ==, 0);
2785                                 VERIFY3U(zap_add_int(dp->dp_meta_objset,
2786                                     dd->dd_phys->dd_clones, o, tx), ==, 0);
2787                                 dsl_dataset_rele(cnds, FTAG);
2788                         }
2789                         zap_cursor_fini(&zc);
2790                 }
2791
2792                 ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2793         }
2794
2795         /*
2796          * Change space accounting.
2797          * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2798          * both be valid, or both be 0 (resulting in delta == 0).  This
2799          * is true for each of {clone,origin} independently.
2800          */
2801
2802         delta = pa->cloneusedsnap -
2803             dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2804         ASSERT3S(delta, >=, 0);
2805         ASSERT3U(pa->used, >=, delta);
2806         dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2807         dsl_dir_diduse_space(dd, DD_USED_HEAD,
2808             pa->used - delta, pa->comp, pa->uncomp, tx);
2809
2810         delta = pa->originusedsnap -
2811             odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2812         ASSERT3S(delta, <=, 0);
2813         ASSERT3U(pa->used, >=, -delta);
2814         dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2815         dsl_dir_diduse_space(odd, DD_USED_HEAD,
2816             -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2817
2818         origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2819
2820         /* log history record */
2821         spa_history_log_internal(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2822             "dataset = %llu", hds->ds_object);
2823
2824         dsl_dir_close(odd, FTAG);
2825 }
2826
2827 static char *snaplist_tag = "snaplist";
2828 /*
2829  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2830  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2831  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2832  * snapshots back to this dataset's origin.
2833  */
2834 static int
2835 snaplist_make(dsl_pool_t *dp, boolean_t own,
2836     uint64_t first_obj, uint64_t last_obj, list_t *l)
2837 {
2838         uint64_t obj = last_obj;
2839
2840         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2841
2842         list_create(l, sizeof (struct promotenode),
2843             offsetof(struct promotenode, link));
2844
2845         while (obj != first_obj) {
2846                 dsl_dataset_t *ds;
2847                 struct promotenode *snap;
2848                 int err;
2849
2850                 if (own) {
2851                         err = dsl_dataset_own_obj(dp, obj,
2852                             0, snaplist_tag, &ds);
2853                         if (err == 0)
2854                                 dsl_dataset_make_exclusive(ds, snaplist_tag);
2855                 } else {
2856                         err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2857                 }
2858                 if (err == ENOENT) {
2859                         /* lost race with snapshot destroy */
2860                         struct promotenode *last = list_tail(l);
2861                         ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2862                         obj = last->ds->ds_phys->ds_prev_snap_obj;
2863                         continue;
2864                 } else if (err) {
2865                         return (err);
2866                 }
2867
2868                 if (first_obj == 0)
2869                         first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2870
2871                 snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
2872                 snap->ds = ds;
2873                 list_insert_tail(l, snap);
2874                 obj = ds->ds_phys->ds_prev_snap_obj;
2875         }
2876
2877         return (0);
2878 }
2879
2880 static int
2881 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
2882 {
2883         struct promotenode *snap;
2884
2885         *spacep = 0;
2886         for (snap = list_head(l); snap; snap = list_next(l, snap)) {
2887                 uint64_t used, comp, uncomp;
2888                 dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2889                     mintxg, UINT64_MAX, &used, &comp, &uncomp);
2890                 *spacep += used;
2891         }
2892         return (0);
2893 }
2894
2895 static void
2896 snaplist_destroy(list_t *l, boolean_t own)
2897 {
2898         struct promotenode *snap;
2899
2900         if (!l || !list_link_active(&l->list_head))
2901                 return;
2902
2903         while ((snap = list_tail(l)) != NULL) {
2904                 list_remove(l, snap);
2905                 if (own)
2906                         dsl_dataset_disown(snap->ds, snaplist_tag);
2907                 else
2908                         dsl_dataset_rele(snap->ds, snaplist_tag);
2909                 kmem_free(snap, sizeof (struct promotenode));
2910         }
2911         list_destroy(l);
2912 }
2913
2914 /*
2915  * Promote a clone.  Nomenclature note:
2916  * "clone" or "cds": the original clone which is being promoted
2917  * "origin" or "ods": the snapshot which is originally clone's origin
2918  * "origin head" or "ohds": the dataset which is the head
2919  * (filesystem/volume) for the origin
2920  * "origin origin": the origin of the origin's filesystem (typically
2921  * NULL, indicating that the clone is not a clone of a clone).
2922  */
2923 int
2924 dsl_dataset_promote(const char *name, char *conflsnap)
2925 {
2926         dsl_dataset_t *ds;
2927         dsl_dir_t *dd;
2928         dsl_pool_t *dp;
2929         dmu_object_info_t doi;
2930         struct promotearg pa;
2931         struct promotenode *snap;
2932         int err;
2933
2934         bzero(&pa, sizeof(struct promotearg));
2935         err = dsl_dataset_hold(name, FTAG, &ds);
2936         if (err)
2937                 return (err);
2938         dd = ds->ds_dir;
2939         dp = dd->dd_pool;
2940
2941         err = dmu_object_info(dp->dp_meta_objset,
2942             ds->ds_phys->ds_snapnames_zapobj, &doi);
2943         if (err) {
2944                 dsl_dataset_rele(ds, FTAG);
2945                 return (err);
2946         }
2947
2948         if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
2949                 dsl_dataset_rele(ds, FTAG);
2950                 return (EINVAL);
2951         }
2952
2953         /*
2954          * We are going to inherit all the snapshots taken before our
2955          * origin (i.e., our new origin will be our parent's origin).
2956          * Take ownership of them so that we can rename them into our
2957          * namespace.
2958          */
2959         rw_enter(&dp->dp_config_rwlock, RW_READER);
2960
2961         err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
2962             &pa.shared_snaps);
2963         if (err != 0)
2964                 goto out;
2965
2966         err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
2967         if (err != 0)
2968                 goto out;
2969
2970         snap = list_head(&pa.shared_snaps);
2971         ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
2972         err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
2973             snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
2974         if (err != 0)
2975                 goto out;
2976
2977         if (snap->ds->ds_dir->dd_phys->dd_origin_obj != 0) {
2978                 err = dsl_dataset_hold_obj(dp,
2979                     snap->ds->ds_dir->dd_phys->dd_origin_obj,
2980                     FTAG, &pa.origin_origin);
2981                 if (err != 0)
2982                         goto out;
2983         }
2984
2985 out:
2986         rw_exit(&dp->dp_config_rwlock);
2987
2988         /*
2989          * Add in 128x the snapnames zapobj size, since we will be moving
2990          * a bunch of snapnames to the promoted ds, and dirtying their
2991          * bonus buffers.
2992          */
2993         if (err == 0) {
2994                 err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
2995                     dsl_dataset_promote_sync, ds, &pa,
2996                     2 + 2 * doi.doi_physical_blocks_512);
2997                 if (err && pa.err_ds && conflsnap)
2998                         (void) strncpy(conflsnap, pa.err_ds, MAXNAMELEN);
2999         }
3000
3001         snaplist_destroy(&pa.shared_snaps, B_TRUE);
3002         snaplist_destroy(&pa.clone_snaps, B_FALSE);
3003         snaplist_destroy(&pa.origin_snaps, B_FALSE);
3004         if (pa.origin_origin)
3005                 dsl_dataset_rele(pa.origin_origin, FTAG);
3006         dsl_dataset_rele(ds, FTAG);
3007         return (err);
3008 }
3009
3010 struct cloneswaparg {
3011         dsl_dataset_t *cds; /* clone dataset */
3012         dsl_dataset_t *ohds; /* origin's head dataset */
3013         boolean_t force;
3014         int64_t unused_refres_delta; /* change in unconsumed refreservation */
3015 };
3016
3017 /* ARGSUSED */
3018 static int
3019 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
3020 {
3021         struct cloneswaparg *csa = arg1;
3022
3023         /* they should both be heads */
3024         if (dsl_dataset_is_snapshot(csa->cds) ||
3025             dsl_dataset_is_snapshot(csa->ohds))
3026                 return (EINVAL);
3027
3028         /* the branch point should be just before them */
3029         if (csa->cds->ds_prev != csa->ohds->ds_prev)
3030                 return (EINVAL);
3031
3032         /* cds should be the clone (unless they are unrelated) */
3033         if (csa->cds->ds_prev != NULL &&
3034             csa->cds->ds_prev != csa->cds->ds_dir->dd_pool->dp_origin_snap &&
3035             csa->ohds->ds_object !=
3036             csa->cds->ds_prev->ds_phys->ds_next_snap_obj)
3037                 return (EINVAL);
3038
3039         /* the clone should be a child of the origin */
3040         if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
3041                 return (EINVAL);
3042
3043         /* ohds shouldn't be modified unless 'force' */
3044         if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
3045                 return (ETXTBSY);
3046
3047         /* adjust amount of any unconsumed refreservation */
3048         csa->unused_refres_delta =
3049             (int64_t)MIN(csa->ohds->ds_reserved,
3050             csa->ohds->ds_phys->ds_unique_bytes) -
3051             (int64_t)MIN(csa->ohds->ds_reserved,
3052             csa->cds->ds_phys->ds_unique_bytes);
3053
3054         if (csa->unused_refres_delta > 0 &&
3055             csa->unused_refres_delta >
3056             dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3057                 return (ENOSPC);
3058
3059         if (csa->ohds->ds_quota != 0 &&
3060             csa->cds->ds_phys->ds_unique_bytes > csa->ohds->ds_quota)
3061                 return (EDQUOT);
3062
3063         return (0);
3064 }
3065
3066 /* ARGSUSED */
3067 static void
3068 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3069 {
3070         struct cloneswaparg *csa = arg1;
3071         dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3072
3073         ASSERT(csa->cds->ds_reserved == 0);
3074         ASSERT(csa->ohds->ds_quota == 0 ||
3075             csa->cds->ds_phys->ds_unique_bytes <= csa->ohds->ds_quota);
3076
3077         dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3078         dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3079
3080         if (csa->cds->ds_objset != NULL) {
3081                 dmu_objset_evict(csa->cds->ds_objset);
3082                 csa->cds->ds_objset = NULL;
3083         }
3084
3085         if (csa->ohds->ds_objset != NULL) {
3086                 dmu_objset_evict(csa->ohds->ds_objset);
3087                 csa->ohds->ds_objset = NULL;
3088         }
3089
3090         /*
3091          * Reset origin's unique bytes, if it exists.
3092          */
3093         if (csa->cds->ds_prev) {
3094                 dsl_dataset_t *origin = csa->cds->ds_prev;
3095                 uint64_t comp, uncomp;
3096
3097                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
3098                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3099                     origin->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3100                     &origin->ds_phys->ds_unique_bytes, &comp, &uncomp);
3101         }
3102
3103         /* swap blkptrs */
3104         {
3105                 blkptr_t tmp;
3106                 tmp = csa->ohds->ds_phys->ds_bp;
3107                 csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3108                 csa->cds->ds_phys->ds_bp = tmp;
3109         }
3110
3111         /* set dd_*_bytes */
3112         {
3113                 int64_t dused, dcomp, duncomp;
3114                 uint64_t cdl_used, cdl_comp, cdl_uncomp;
3115                 uint64_t odl_used, odl_comp, odl_uncomp;
3116
3117                 ASSERT3U(csa->cds->ds_dir->dd_phys->
3118                     dd_used_breakdown[DD_USED_SNAP], ==, 0);
3119
3120                 dsl_deadlist_space(&csa->cds->ds_deadlist,
3121                     &cdl_used, &cdl_comp, &cdl_uncomp);
3122                 dsl_deadlist_space(&csa->ohds->ds_deadlist,
3123                     &odl_used, &odl_comp, &odl_uncomp);
3124
3125                 dused = csa->cds->ds_phys->ds_used_bytes + cdl_used -
3126                     (csa->ohds->ds_phys->ds_used_bytes + odl_used);
3127                 dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3128                     (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3129                 duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3130                     cdl_uncomp -
3131                     (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3132
3133                 dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3134                     dused, dcomp, duncomp, tx);
3135                 dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3136                     -dused, -dcomp, -duncomp, tx);
3137
3138                 /*
3139                  * The difference in the space used by snapshots is the
3140                  * difference in snapshot space due to the head's
3141                  * deadlist (since that's the only thing that's
3142                  * changing that affects the snapused).
3143                  */
3144                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3145                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3146                     &cdl_used, &cdl_comp, &cdl_uncomp);
3147                 dsl_deadlist_space_range(&csa->ohds->ds_deadlist,
3148                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3149                     &odl_used, &odl_comp, &odl_uncomp);
3150                 dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3151                     DD_USED_HEAD, DD_USED_SNAP, tx);
3152         }
3153
3154         /* swap ds_*_bytes */
3155         SWITCH64(csa->ohds->ds_phys->ds_used_bytes,
3156             csa->cds->ds_phys->ds_used_bytes);
3157         SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3158             csa->cds->ds_phys->ds_compressed_bytes);
3159         SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3160             csa->cds->ds_phys->ds_uncompressed_bytes);
3161         SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3162             csa->cds->ds_phys->ds_unique_bytes);
3163
3164         /* apply any parent delta for change in unconsumed refreservation */
3165         dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3166             csa->unused_refres_delta, 0, 0, tx);
3167
3168         /*
3169          * Swap deadlists.
3170          */
3171         dsl_deadlist_close(&csa->cds->ds_deadlist);
3172         dsl_deadlist_close(&csa->ohds->ds_deadlist);
3173         SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3174             csa->cds->ds_phys->ds_deadlist_obj);
3175         dsl_deadlist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3176             csa->cds->ds_phys->ds_deadlist_obj);
3177         dsl_deadlist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3178             csa->ohds->ds_phys->ds_deadlist_obj);
3179
3180         dsl_scan_ds_clone_swapped(csa->ohds, csa->cds, tx);
3181 }
3182
3183 /*
3184  * Swap 'clone' with its origin head datasets.  Used at the end of "zfs
3185  * recv" into an existing fs to swizzle the file system to the new
3186  * version, and by "zfs rollback".  Can also be used to swap two
3187  * independent head datasets if neither has any snapshots.
3188  */
3189 int
3190 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3191     boolean_t force)
3192 {
3193         struct cloneswaparg csa;
3194         int error;
3195
3196         ASSERT(clone->ds_owner);
3197         ASSERT(origin_head->ds_owner);
3198 retry:
3199         /*
3200          * Need exclusive access for the swap. If we're swapping these
3201          * datasets back after an error, we already hold the locks.
3202          */
3203         if (!RW_WRITE_HELD(&clone->ds_rwlock))
3204                 rw_enter(&clone->ds_rwlock, RW_WRITER);
3205         if (!RW_WRITE_HELD(&origin_head->ds_rwlock) &&
3206             !rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3207                 rw_exit(&clone->ds_rwlock);
3208                 rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3209                 if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3210                         rw_exit(&origin_head->ds_rwlock);
3211                         goto retry;
3212                 }
3213         }
3214         csa.cds = clone;
3215         csa.ohds = origin_head;
3216         csa.force = force;
3217         error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3218             dsl_dataset_clone_swap_check,
3219             dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3220         return (error);
3221 }
3222
3223 /*
3224  * Given a pool name and a dataset object number in that pool,
3225  * return the name of that dataset.
3226  */
3227 int
3228 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3229 {
3230         spa_t *spa;
3231         dsl_pool_t *dp;
3232         dsl_dataset_t *ds;
3233         int error;
3234
3235         if ((error = spa_open(pname, &spa, FTAG)) != 0)
3236                 return (error);
3237         dp = spa_get_dsl(spa);
3238         rw_enter(&dp->dp_config_rwlock, RW_READER);
3239         if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3240                 dsl_dataset_name(ds, buf);
3241                 dsl_dataset_rele(ds, FTAG);
3242         }
3243         rw_exit(&dp->dp_config_rwlock);
3244         spa_close(spa, FTAG);
3245
3246         return (error);
3247 }
3248
3249 int
3250 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3251     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3252 {
3253         int error = 0;
3254
3255         ASSERT3S(asize, >, 0);
3256
3257         /*
3258          * *ref_rsrv is the portion of asize that will come from any
3259          * unconsumed refreservation space.
3260          */
3261         *ref_rsrv = 0;
3262
3263         mutex_enter(&ds->ds_lock);
3264         /*
3265          * Make a space adjustment for reserved bytes.
3266          */
3267         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3268                 ASSERT3U(*used, >=,
3269                     ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3270                 *used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3271                 *ref_rsrv =
3272                     asize - MIN(asize, parent_delta(ds, asize + inflight));
3273         }
3274
3275         if (!check_quota || ds->ds_quota == 0) {
3276                 mutex_exit(&ds->ds_lock);
3277                 return (0);
3278         }
3279         /*
3280          * If they are requesting more space, and our current estimate
3281          * is over quota, they get to try again unless the actual
3282          * on-disk is over quota and there are no pending changes (which
3283          * may free up space for us).
3284          */
3285         if (ds->ds_phys->ds_used_bytes + inflight >= ds->ds_quota) {
3286                 if (inflight > 0 || ds->ds_phys->ds_used_bytes < ds->ds_quota)
3287                         error = ERESTART;
3288                 else
3289                         error = EDQUOT;
3290         }
3291         mutex_exit(&ds->ds_lock);
3292
3293         return (error);
3294 }
3295
3296 /* ARGSUSED */
3297 static int
3298 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3299 {
3300         dsl_dataset_t *ds = arg1;
3301         dsl_prop_setarg_t *psa = arg2;
3302         int err;
3303
3304         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3305                 return (ENOTSUP);
3306
3307         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3308                 return (err);
3309
3310         if (psa->psa_effective_value == 0)
3311                 return (0);
3312
3313         if (psa->psa_effective_value < ds->ds_phys->ds_used_bytes ||
3314             psa->psa_effective_value < ds->ds_reserved)
3315                 return (ENOSPC);
3316
3317         return (0);
3318 }
3319
3320 extern void dsl_prop_set_sync(void *, void *, dmu_tx_t *);
3321
3322 void
3323 dsl_dataset_set_quota_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3324 {
3325         dsl_dataset_t *ds = arg1;
3326         dsl_prop_setarg_t *psa = arg2;
3327         uint64_t effective_value = psa->psa_effective_value;
3328
3329         dsl_prop_set_sync(ds, psa, tx);
3330         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3331
3332         if (ds->ds_quota != effective_value) {
3333                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3334                 ds->ds_quota = effective_value;
3335
3336                 spa_history_log_internal(LOG_DS_REFQUOTA,
3337                     ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu ",
3338                     (longlong_t)ds->ds_quota, ds->ds_object);
3339         }
3340 }
3341
3342 int
3343 dsl_dataset_set_quota(const char *dsname, zprop_source_t source, uint64_t quota)
3344 {
3345         dsl_dataset_t *ds;
3346         dsl_prop_setarg_t psa;
3347         int err;
3348
3349         dsl_prop_setarg_init_uint64(&psa, "refquota", source, &quota);
3350
3351         err = dsl_dataset_hold(dsname, FTAG, &ds);
3352         if (err)
3353                 return (err);
3354
3355         /*
3356          * If someone removes a file, then tries to set the quota, we
3357          * want to make sure the file freeing takes effect.
3358          */
3359         txg_wait_open(ds->ds_dir->dd_pool, 0);
3360
3361         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3362             dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3363             ds, &psa, 0);
3364
3365         dsl_dataset_rele(ds, FTAG);
3366         return (err);
3367 }
3368
3369 static int
3370 dsl_dataset_set_reservation_check(void *arg1, void *arg2, dmu_tx_t *tx)
3371 {
3372         dsl_dataset_t *ds = arg1;
3373         dsl_prop_setarg_t *psa = arg2;
3374         uint64_t effective_value;
3375         uint64_t unique;
3376         int err;
3377
3378         if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3379             SPA_VERSION_REFRESERVATION)
3380                 return (ENOTSUP);
3381
3382         if (dsl_dataset_is_snapshot(ds))
3383                 return (EINVAL);
3384
3385         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3386                 return (err);
3387
3388         effective_value = psa->psa_effective_value;
3389
3390         /*
3391          * If we are doing the preliminary check in open context, the
3392          * space estimates may be inaccurate.
3393          */
3394         if (!dmu_tx_is_syncing(tx))
3395                 return (0);
3396
3397         mutex_enter(&ds->ds_lock);
3398         if (!DS_UNIQUE_IS_ACCURATE(ds))
3399                 dsl_dataset_recalc_head_uniq(ds);
3400         unique = ds->ds_phys->ds_unique_bytes;
3401         mutex_exit(&ds->ds_lock);
3402
3403         if (MAX(unique, effective_value) > MAX(unique, ds->ds_reserved)) {
3404                 uint64_t delta = MAX(unique, effective_value) -
3405                     MAX(unique, ds->ds_reserved);
3406
3407                 if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3408                         return (ENOSPC);
3409                 if (ds->ds_quota > 0 &&
3410                     effective_value > ds->ds_quota)
3411                         return (ENOSPC);
3412         }
3413
3414         return (0);
3415 }
3416
3417 static void
3418 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3419 {
3420         dsl_dataset_t *ds = arg1;
3421         dsl_prop_setarg_t *psa = arg2;
3422         uint64_t effective_value = psa->psa_effective_value;
3423         uint64_t unique;
3424         int64_t delta;
3425
3426         dsl_prop_set_sync(ds, psa, tx);
3427         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3428
3429         dmu_buf_will_dirty(ds->ds_dbuf, tx);
3430
3431         mutex_enter(&ds->ds_dir->dd_lock);
3432         mutex_enter(&ds->ds_lock);
3433         ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
3434         unique = ds->ds_phys->ds_unique_bytes;
3435         delta = MAX(0, (int64_t)(effective_value - unique)) -
3436             MAX(0, (int64_t)(ds->ds_reserved - unique));
3437         ds->ds_reserved = effective_value;
3438         mutex_exit(&ds->ds_lock);
3439
3440         dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3441         mutex_exit(&ds->ds_dir->dd_lock);
3442
3443         spa_history_log_internal(LOG_DS_REFRESERV,
3444             ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu",
3445             (longlong_t)effective_value, ds->ds_object);
3446 }
3447
3448 int
3449 dsl_dataset_set_reservation(const char *dsname, zprop_source_t source,
3450     uint64_t reservation)
3451 {
3452         dsl_dataset_t *ds;
3453         dsl_prop_setarg_t psa;
3454         int err;
3455
3456         dsl_prop_setarg_init_uint64(&psa, "refreservation", source,
3457             &reservation);
3458
3459         err = dsl_dataset_hold(dsname, FTAG, &ds);
3460         if (err)
3461                 return (err);
3462
3463         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3464             dsl_dataset_set_reservation_check,
3465             dsl_dataset_set_reservation_sync, ds, &psa, 0);
3466
3467         dsl_dataset_rele(ds, FTAG);
3468         return (err);
3469 }
3470
3471 typedef struct zfs_hold_cleanup_arg {
3472         dsl_pool_t *dp;
3473         uint64_t dsobj;
3474         char htag[MAXNAMELEN];
3475 } zfs_hold_cleanup_arg_t;
3476
3477 static void
3478 dsl_dataset_user_release_onexit(void *arg)
3479 {
3480         zfs_hold_cleanup_arg_t *ca = arg;
3481
3482         (void) dsl_dataset_user_release_tmp(ca->dp, ca->dsobj, ca->htag,
3483             B_TRUE);
3484         kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t));
3485 }
3486
3487 void
3488 dsl_register_onexit_hold_cleanup(dsl_dataset_t *ds, const char *htag,
3489     minor_t minor)
3490 {
3491         zfs_hold_cleanup_arg_t *ca;
3492
3493         ca = kmem_alloc(sizeof (zfs_hold_cleanup_arg_t), KM_SLEEP);
3494         ca->dp = ds->ds_dir->dd_pool;
3495         ca->dsobj = ds->ds_object;
3496         (void) strlcpy(ca->htag, htag, sizeof (ca->htag));
3497         VERIFY3U(0, ==, zfs_onexit_add_cb(minor,
3498             dsl_dataset_user_release_onexit, ca, NULL));
3499 }
3500
3501 /*
3502  * If you add new checks here, you may need to add
3503  * additional checks to the "temporary" case in
3504  * snapshot_check() in dmu_objset.c.
3505  */
3506 static int
3507 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3508 {
3509         dsl_dataset_t *ds = arg1;
3510         struct dsl_ds_holdarg *ha = arg2;
3511         char *htag = ha->htag;
3512         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3513         int error = 0;
3514
3515         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3516                 return (ENOTSUP);
3517
3518         if (!dsl_dataset_is_snapshot(ds))
3519                 return (EINVAL);
3520
3521         /* tags must be unique */
3522         mutex_enter(&ds->ds_lock);
3523         if (ds->ds_phys->ds_userrefs_obj) {
3524                 error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3525                     8, 1, tx);
3526                 if (error == 0)
3527                         error = EEXIST;
3528                 else if (error == ENOENT)
3529                         error = 0;
3530         }
3531         mutex_exit(&ds->ds_lock);
3532
3533         if (error == 0 && ha->temphold &&
3534             strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
3535                 error = E2BIG;
3536
3537         return (error);
3538 }
3539
3540 void
3541 dsl_dataset_user_hold_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3542 {
3543         dsl_dataset_t *ds = arg1;
3544         struct dsl_ds_holdarg *ha = arg2;
3545         char *htag = ha->htag;
3546         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3547         objset_t *mos = dp->dp_meta_objset;
3548         uint64_t now = gethrestime_sec();
3549         uint64_t zapobj;
3550
3551         mutex_enter(&ds->ds_lock);
3552         if (ds->ds_phys->ds_userrefs_obj == 0) {
3553                 /*
3554                  * This is the first user hold for this dataset.  Create
3555                  * the userrefs zap object.
3556                  */
3557                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3558                 zapobj = ds->ds_phys->ds_userrefs_obj =
3559                     zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3560         } else {
3561                 zapobj = ds->ds_phys->ds_userrefs_obj;
3562         }
3563         ds->ds_userrefs++;
3564         mutex_exit(&ds->ds_lock);
3565
3566         VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3567
3568         if (ha->temphold) {
3569                 VERIFY(0 == dsl_pool_user_hold(dp, ds->ds_object,
3570                     htag, &now, tx));
3571         }
3572
3573         spa_history_log_internal(LOG_DS_USER_HOLD,
3574             dp->dp_spa, tx, "<%s> temp = %d dataset = %llu", htag,
3575             (int)ha->temphold, ds->ds_object);
3576 }
3577
3578 static int
3579 dsl_dataset_user_hold_one(const char *dsname, void *arg)
3580 {
3581         struct dsl_ds_holdarg *ha = arg;
3582         dsl_dataset_t *ds;
3583         int error;
3584         char *name;
3585
3586         /* alloc a buffer to hold dsname@snapname plus terminating NULL */
3587         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3588         error = dsl_dataset_hold(name, ha->dstg, &ds);
3589         strfree(name);
3590         if (error == 0) {
3591                 ha->gotone = B_TRUE;
3592                 dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3593                     dsl_dataset_user_hold_sync, ds, ha, 0);
3594         } else if (error == ENOENT && ha->recursive) {
3595                 error = 0;
3596         } else {
3597                 (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3598         }
3599         return (error);
3600 }
3601
3602 int
3603 dsl_dataset_user_hold_for_send(dsl_dataset_t *ds, char *htag,
3604     boolean_t temphold)
3605 {
3606         struct dsl_ds_holdarg *ha;
3607         int error;
3608
3609         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3610         ha->htag = htag;
3611         ha->temphold = temphold;
3612         error = dsl_sync_task_do(ds->ds_dir->dd_pool,
3613             dsl_dataset_user_hold_check, dsl_dataset_user_hold_sync,
3614             ds, ha, 0);
3615         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3616
3617         return (error);
3618 }
3619
3620 int
3621 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3622     boolean_t recursive, boolean_t temphold, int cleanup_fd)
3623 {
3624         struct dsl_ds_holdarg *ha;
3625         dsl_sync_task_t *dst;
3626         spa_t *spa;
3627         int error;
3628         minor_t minor = 0;
3629
3630         if (cleanup_fd != -1) {
3631                 /* Currently we only support cleanup-on-exit of tempholds. */
3632                 if (!temphold)
3633                         return (EINVAL);
3634                 error = zfs_onexit_fd_hold(cleanup_fd, &minor);
3635                 if (error)
3636                         return (error);
3637         }
3638
3639         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3640
3641         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3642
3643         error = spa_open(dsname, &spa, FTAG);
3644         if (error) {
3645                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3646                 if (cleanup_fd != -1)
3647                         zfs_onexit_fd_rele(cleanup_fd);
3648                 return (error);
3649         }
3650
3651         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3652         ha->htag = htag;
3653         ha->snapname = snapname;
3654         ha->recursive = recursive;
3655         ha->temphold = temphold;
3656
3657         if (recursive) {
3658                 error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3659                     ha, DS_FIND_CHILDREN);
3660         } else {
3661                 error = dsl_dataset_user_hold_one(dsname, ha);
3662         }
3663         if (error == 0)
3664                 error = dsl_sync_task_group_wait(ha->dstg);
3665
3666         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3667             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3668                 dsl_dataset_t *ds = dst->dst_arg1;
3669
3670                 if (dst->dst_err) {
3671                         dsl_dataset_name(ds, ha->failed);
3672                         *strchr(ha->failed, '@') = '\0';
3673                 } else if (error == 0 && minor != 0 && temphold) {
3674                         /*
3675                          * If this hold is to be released upon process exit,
3676                          * register that action now.
3677                          */
3678                         dsl_register_onexit_hold_cleanup(ds, htag, minor);
3679                 }
3680                 dsl_dataset_rele(ds, ha->dstg);
3681         }
3682
3683         if (error == 0 && recursive && !ha->gotone)
3684                 error = ENOENT;
3685
3686         if (error)
3687                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3688
3689         dsl_sync_task_group_destroy(ha->dstg);
3690
3691         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3692         spa_close(spa, FTAG);
3693         if (cleanup_fd != -1)
3694                 zfs_onexit_fd_rele(cleanup_fd);
3695         return (error);
3696 }
3697
3698 struct dsl_ds_releasearg {
3699         dsl_dataset_t *ds;
3700         const char *htag;
3701         boolean_t own;          /* do we own or just hold ds? */
3702 };
3703
3704 static int
3705 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3706     boolean_t *might_destroy)
3707 {
3708         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3709         uint64_t zapobj;
3710         uint64_t tmp;
3711         int error;
3712
3713         *might_destroy = B_FALSE;
3714
3715         mutex_enter(&ds->ds_lock);
3716         zapobj = ds->ds_phys->ds_userrefs_obj;
3717         if (zapobj == 0) {
3718                 /* The tag can't possibly exist */
3719                 mutex_exit(&ds->ds_lock);
3720                 return (ESRCH);
3721         }
3722
3723         /* Make sure the tag exists */
3724         error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3725         if (error) {
3726                 mutex_exit(&ds->ds_lock);
3727                 if (error == ENOENT)
3728                         error = ESRCH;
3729                 return (error);
3730         }
3731
3732         if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3733             DS_IS_DEFER_DESTROY(ds))
3734                 *might_destroy = B_TRUE;
3735
3736         mutex_exit(&ds->ds_lock);
3737         return (0);
3738 }
3739
3740 static int
3741 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3742 {
3743         struct dsl_ds_releasearg *ra = arg1;
3744         dsl_dataset_t *ds = ra->ds;
3745         boolean_t might_destroy;
3746         int error;
3747
3748         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3749                 return (ENOTSUP);
3750
3751         error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3752         if (error)
3753                 return (error);
3754
3755         if (might_destroy) {
3756                 struct dsl_ds_destroyarg dsda = {0};
3757
3758                 if (dmu_tx_is_syncing(tx)) {
3759                         /*
3760                          * If we're not prepared to remove the snapshot,
3761                          * we can't allow the release to happen right now.
3762                          */
3763                         if (!ra->own)
3764                                 return (EBUSY);
3765                 }
3766                 dsda.ds = ds;
3767                 dsda.releasing = B_TRUE;
3768                 return (dsl_dataset_destroy_check(&dsda, tag, tx));
3769         }
3770
3771         return (0);
3772 }
3773
3774 static void
3775 dsl_dataset_user_release_sync(void *arg1, void *tag, dmu_tx_t *tx)
3776 {
3777         struct dsl_ds_releasearg *ra = arg1;
3778         dsl_dataset_t *ds = ra->ds;
3779         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3780         objset_t *mos = dp->dp_meta_objset;
3781         uint64_t zapobj;
3782         uint64_t dsobj = ds->ds_object;
3783         uint64_t refs;
3784         int error;
3785
3786         mutex_enter(&ds->ds_lock);
3787         ds->ds_userrefs--;
3788         refs = ds->ds_userrefs;
3789         mutex_exit(&ds->ds_lock);
3790         error = dsl_pool_user_release(dp, ds->ds_object, ra->htag, tx);
3791         VERIFY(error == 0 || error == ENOENT);
3792         zapobj = ds->ds_phys->ds_userrefs_obj;
3793         VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3794         if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3795             DS_IS_DEFER_DESTROY(ds)) {
3796                 struct dsl_ds_destroyarg dsda = {0};
3797
3798                 ASSERT(ra->own);
3799                 dsda.ds = ds;
3800                 dsda.releasing = B_TRUE;
3801                 /* We already did the destroy_check */
3802                 dsl_dataset_destroy_sync(&dsda, tag, tx);
3803         }
3804
3805         spa_history_log_internal(LOG_DS_USER_RELEASE,
3806             dp->dp_spa, tx, "<%s> %lld dataset = %llu",
3807             ra->htag, (longlong_t)refs, dsobj);
3808 }
3809
3810 static int
3811 dsl_dataset_user_release_one(const char *dsname, void *arg)
3812 {
3813         struct dsl_ds_holdarg *ha = arg;
3814         struct dsl_ds_releasearg *ra;
3815         dsl_dataset_t *ds;
3816         int error;
3817         void *dtag = ha->dstg;
3818         char *name;
3819         boolean_t own = B_FALSE;
3820         boolean_t might_destroy;
3821
3822         /* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3823         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3824         error = dsl_dataset_hold(name, dtag, &ds);
3825         strfree(name);
3826         if (error == ENOENT && ha->recursive)
3827                 return (0);
3828         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3829         if (error)
3830                 return (error);
3831
3832         ha->gotone = B_TRUE;
3833
3834         ASSERT(dsl_dataset_is_snapshot(ds));
3835
3836         error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3837         if (error) {
3838                 dsl_dataset_rele(ds, dtag);
3839                 return (error);
3840         }
3841
3842         if (might_destroy) {
3843 #ifdef _KERNEL
3844                 name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3845                 error = zfs_unmount_snap(name, NULL);
3846                 strfree(name);
3847                 if (error) {
3848                         dsl_dataset_rele(ds, dtag);
3849                         return (error);
3850                 }
3851 #endif
3852                 if (!dsl_dataset_tryown(ds, B_TRUE, dtag)) {
3853                         dsl_dataset_rele(ds, dtag);
3854                         return (EBUSY);
3855                 } else {
3856                         own = B_TRUE;
3857                         dsl_dataset_make_exclusive(ds, dtag);
3858                 }
3859         }
3860
3861         ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3862         ra->ds = ds;
3863         ra->htag = ha->htag;
3864         ra->own = own;
3865         dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3866             dsl_dataset_user_release_sync, ra, dtag, 0);
3867
3868         return (0);
3869 }
3870
3871 int
3872 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3873     boolean_t recursive)
3874 {
3875         struct dsl_ds_holdarg *ha;
3876         dsl_sync_task_t *dst;
3877         spa_t *spa;
3878         int error;
3879
3880 top:
3881         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3882
3883         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3884
3885         error = spa_open(dsname, &spa, FTAG);
3886         if (error) {
3887                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3888                 return (error);
3889         }
3890
3891         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3892         ha->htag = htag;
3893         ha->snapname = snapname;
3894         ha->recursive = recursive;
3895         if (recursive) {
3896                 error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
3897                     ha, DS_FIND_CHILDREN);
3898         } else {
3899                 error = dsl_dataset_user_release_one(dsname, ha);
3900         }
3901         if (error == 0)
3902                 error = dsl_sync_task_group_wait(ha->dstg);
3903
3904         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3905             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3906                 struct dsl_ds_releasearg *ra = dst->dst_arg1;
3907                 dsl_dataset_t *ds = ra->ds;
3908
3909                 if (dst->dst_err)
3910                         dsl_dataset_name(ds, ha->failed);
3911
3912                 if (ra->own)
3913                         dsl_dataset_disown(ds, ha->dstg);
3914                 else
3915                         dsl_dataset_rele(ds, ha->dstg);
3916
3917                 kmem_free(ra, sizeof (struct dsl_ds_releasearg));
3918         }
3919
3920         if (error == 0 && recursive && !ha->gotone)
3921                 error = ENOENT;
3922
3923         if (error && error != EBUSY)
3924                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3925
3926         dsl_sync_task_group_destroy(ha->dstg);
3927         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3928         spa_close(spa, FTAG);
3929
3930         /*
3931          * We can get EBUSY if we were racing with deferred destroy and
3932          * dsl_dataset_user_release_check() hadn't done the necessary
3933          * open context setup.  We can also get EBUSY if we're racing
3934          * with destroy and that thread is the ds_owner.  Either way
3935          * the busy condition should be transient, and we should retry
3936          * the release operation.
3937          */
3938         if (error == EBUSY)
3939                 goto top;
3940
3941         return (error);
3942 }
3943
3944 /*
3945  * Called at spa_load time (with retry == B_FALSE) to release a stale
3946  * temporary user hold. Also called by the onexit code (with retry == B_TRUE).
3947  */
3948 int
3949 dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, char *htag,
3950     boolean_t retry)
3951 {
3952         dsl_dataset_t *ds;
3953         char *snap;
3954         char *name;
3955         int namelen;
3956         int error;
3957
3958         do {
3959                 rw_enter(&dp->dp_config_rwlock, RW_READER);
3960                 error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
3961                 rw_exit(&dp->dp_config_rwlock);
3962                 if (error)
3963                         return (error);
3964                 namelen = dsl_dataset_namelen(ds)+1;
3965                 name = kmem_alloc(namelen, KM_SLEEP);
3966                 dsl_dataset_name(ds, name);
3967                 dsl_dataset_rele(ds, FTAG);
3968
3969                 snap = strchr(name, '@');
3970                 *snap = '\0';
3971                 ++snap;
3972                 error = dsl_dataset_user_release(name, snap, htag, B_FALSE);
3973                 kmem_free(name, namelen);
3974
3975                 /*
3976                  * The object can't have been destroyed because we have a hold,
3977                  * but it might have been renamed, resulting in ENOENT.  Retry
3978                  * if we've been requested to do so.
3979                  *
3980                  * It would be nice if we could use the dsobj all the way
3981                  * through and avoid ENOENT entirely.  But we might need to
3982                  * unmount the snapshot, and there's currently no way to lookup
3983                  * a vfsp using a ZFS object id.
3984                  */
3985         } while ((error == ENOENT) && retry);
3986
3987         return (error);
3988 }
3989
3990 int
3991 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
3992 {
3993         dsl_dataset_t *ds;
3994         int err;
3995
3996         err = dsl_dataset_hold(dsname, FTAG, &ds);
3997         if (err)
3998                 return (err);
3999
4000         VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
4001         if (ds->ds_phys->ds_userrefs_obj != 0) {
4002                 zap_attribute_t *za;
4003                 zap_cursor_t zc;
4004
4005                 za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
4006                 for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
4007                     ds->ds_phys->ds_userrefs_obj);
4008                     zap_cursor_retrieve(&zc, za) == 0;
4009                     zap_cursor_advance(&zc)) {
4010                         VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
4011                             za->za_first_integer));
4012                 }
4013                 zap_cursor_fini(&zc);
4014                 kmem_free(za, sizeof (zap_attribute_t));
4015         }
4016         dsl_dataset_rele(ds, FTAG);
4017         return (0);
4018 }
4019
4020 /*
4021  * Note, this fuction is used as the callback for dmu_objset_find().  We
4022  * always return 0 so that we will continue to find and process
4023  * inconsistent datasets, even if we encounter an error trying to
4024  * process one of them.
4025  */
4026 /* ARGSUSED */
4027 int
4028 dsl_destroy_inconsistent(const char *dsname, void *arg)
4029 {
4030         dsl_dataset_t *ds;
4031
4032         if (dsl_dataset_own(dsname, B_TRUE, FTAG, &ds) == 0) {
4033                 if (DS_IS_INCONSISTENT(ds))
4034                         (void) dsl_dataset_destroy(ds, FTAG, B_FALSE);
4035                 else
4036                         dsl_dataset_disown(ds, FTAG);
4037         }
4038         return (0);
4039 }
4040
4041 #if defined(_KERNEL) && defined(HAVE_SPL)
4042 EXPORT_SYMBOL(dsl_dataset_hold);
4043 EXPORT_SYMBOL(dsl_dataset_hold_obj);
4044 EXPORT_SYMBOL(dsl_dataset_own);
4045 EXPORT_SYMBOL(dsl_dataset_own_obj);
4046 EXPORT_SYMBOL(dsl_dataset_name);
4047 EXPORT_SYMBOL(dsl_dataset_rele);
4048 EXPORT_SYMBOL(dsl_dataset_disown);
4049 EXPORT_SYMBOL(dsl_dataset_drop_ref);
4050 EXPORT_SYMBOL(dsl_dataset_tryown);
4051 EXPORT_SYMBOL(dsl_dataset_make_exclusive);
4052 EXPORT_SYMBOL(dsl_dataset_create_sync);
4053 EXPORT_SYMBOL(dsl_dataset_create_sync_dd);
4054 EXPORT_SYMBOL(dsl_dataset_destroy);
4055 EXPORT_SYMBOL(dsl_snapshots_destroy);
4056 EXPORT_SYMBOL(dsl_dataset_destroy_check);
4057 EXPORT_SYMBOL(dsl_dataset_destroy_sync);
4058 EXPORT_SYMBOL(dsl_dataset_snapshot_check);
4059 EXPORT_SYMBOL(dsl_dataset_snapshot_sync);
4060 EXPORT_SYMBOL(dsl_dataset_rename);
4061 EXPORT_SYMBOL(dsl_dataset_promote);
4062 EXPORT_SYMBOL(dsl_dataset_clone_swap);
4063 EXPORT_SYMBOL(dsl_dataset_user_hold);
4064 EXPORT_SYMBOL(dsl_dataset_user_release);
4065 EXPORT_SYMBOL(dsl_dataset_user_release_tmp);
4066 EXPORT_SYMBOL(dsl_dataset_get_holds);
4067 EXPORT_SYMBOL(dsl_dataset_get_blkptr);
4068 EXPORT_SYMBOL(dsl_dataset_set_blkptr);
4069 EXPORT_SYMBOL(dsl_dataset_get_spa);
4070 EXPORT_SYMBOL(dsl_dataset_modified_since_lastsnap);
4071 EXPORT_SYMBOL(dsl_dataset_sync);
4072 EXPORT_SYMBOL(dsl_dataset_block_born);
4073 EXPORT_SYMBOL(dsl_dataset_block_kill);
4074 EXPORT_SYMBOL(dsl_dataset_block_freeable);
4075 EXPORT_SYMBOL(dsl_dataset_prev_snap_txg);
4076 EXPORT_SYMBOL(dsl_dataset_dirty);
4077 EXPORT_SYMBOL(dsl_dataset_stats);
4078 EXPORT_SYMBOL(dsl_dataset_fast_stat);
4079 EXPORT_SYMBOL(dsl_dataset_space);
4080 EXPORT_SYMBOL(dsl_dataset_fsid_guid);
4081 EXPORT_SYMBOL(dsl_dsobj_to_dsname);
4082 EXPORT_SYMBOL(dsl_dataset_check_quota);
4083 EXPORT_SYMBOL(dsl_dataset_set_quota);
4084 EXPORT_SYMBOL(dsl_dataset_set_quota_sync);
4085 EXPORT_SYMBOL(dsl_dataset_set_reservation);
4086 EXPORT_SYMBOL(dsl_destroy_inconsistent);
4087 #endif