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/zvol.h>
  41 #include <sys/dsl_scan.h>
  42 #include <sys/dsl_deadlist.h>
  43
  44 /*
  45  * Enable/disable prefetching of dedup-ed blocks which are going to be freed.
  46  */
  47 int zfs_dedup_prefetch = 1;
  48
  49 static char *dsl_reaper = "the grim reaper";
  50
  51 static dsl_checkfunc_t dsl_dataset_destroy_begin_check;
  52 static dsl_syncfunc_t dsl_dataset_destroy_begin_sync;
  53 static dsl_syncfunc_t dsl_dataset_set_reservation_sync;
  54
  55 #define SWITCH64(x, y) \
  56         { \
  57                 uint64_t __tmp = (x); \
  58                 (x) = (y); \
  59                 (y) = __tmp; \
  60         }
  61
  62 #define DS_REF_MAX      (1ULL << 62)
  63
  64 #define DSL_DEADLIST_BLOCKSIZE  SPA_MAXBLOCKSIZE
  65
  66 #define DSL_DATASET_IS_DESTROYED(ds)    ((ds)->ds_owner == dsl_reaper)
  67
  68
  69 /*
  70  * Figure out how much of this delta should be propogated to the dsl_dir
  71  * layer.  If there's a refreservation, that space has already been
  72  * partially accounted for in our ancestors.
  73  */
  74 static int64_t
  75 parent_delta(dsl_dataset_t *ds, int64_t delta)
  76 {
  77         uint64_t old_bytes, new_bytes;
  78
  79         if (ds->ds_reserved == 0)
  80                 return (delta);
  81
  82         old_bytes = MAX(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
  83         new_bytes = MAX(ds->ds_phys->ds_unique_bytes + delta, ds->ds_reserved);
  84
  85         ASSERT3U(ABS((int64_t)(new_bytes - old_bytes)), <=, ABS(delta));
  86         return (new_bytes - old_bytes);
  87 }
  88
  89 void
  90 dsl_dataset_block_born(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
  91 {
  92         int used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
  93         int compressed = BP_GET_PSIZE(bp);
  94         int uncompressed = BP_GET_UCSIZE(bp);
  95         int64_t delta;
  96
  97         dprintf_bp(bp, "ds=%p", ds);
  98
  99         ASSERT(dmu_tx_is_syncing(tx));
 100         /* It could have been compressed away to nothing */
 101         if (BP_IS_HOLE(bp))
 102                 return;
 103         ASSERT(BP_GET_TYPE(bp) != DMU_OT_NONE);
 104         ASSERT3U(BP_GET_TYPE(bp), <, DMU_OT_NUMTYPES);
 105         if (ds == NULL) {
 106                 /*
 107                  * Account for the meta-objset space in its placeholder
 108                  * dsl_dir.
 109                  */
 110                 ASSERT3U(compressed, ==, uncompressed); /* it's all metadata */
 111                 dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
 112                     used, compressed, uncompressed, tx);
 113                 dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
 114                 return;
 115         }
 116         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 117
 118         mutex_enter(&ds->ds_dir->dd_lock);
 119         mutex_enter(&ds->ds_lock);
 120         delta = parent_delta(ds, used);
 121         ds->ds_phys->ds_used_bytes += used;
 122         ds->ds_phys->ds_compressed_bytes += compressed;
 123         ds->ds_phys->ds_uncompressed_bytes += uncompressed;
 124         ds->ds_phys->ds_unique_bytes += used;
 125         mutex_exit(&ds->ds_lock);
 126         dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD, delta,
 127             compressed, uncompressed, tx);
 128         dsl_dir_transfer_space(ds->ds_dir, used - delta,
 129             DD_USED_REFRSRV, DD_USED_HEAD, tx);
 130         mutex_exit(&ds->ds_dir->dd_lock);
 131 }
 132
 133 int
 134 dsl_dataset_block_kill(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx,
 135     boolean_t async)
 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         int used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
 144         int compressed = BP_GET_PSIZE(bp);
 145         int 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         if (zfs_dedup_prefetch && bp && BP_GET_DEDUP(bp))
 257                 ddt_prefetch(dsl_dataset_get_spa(ds), bp);
 258
 259         return (B_TRUE);
 260 }
 261
 262 /* ARGSUSED */
 263 static void
 264 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
 265 {
 266         dsl_dataset_t *ds = dsv;
 267
 268         ASSERT(ds->ds_owner == NULL || DSL_DATASET_IS_DESTROYED(ds));
 269
 270         unique_remove(ds->ds_fsid_guid);
 271
 272         if (ds->ds_objset != NULL)
 273                 dmu_objset_evict(ds->ds_objset);
 274
 275         if (ds->ds_prev) {
 276                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 277                 ds->ds_prev = NULL;
 278         }
 279
 280         bplist_destroy(&ds->ds_pending_deadlist);
 281         if (db != NULL) {
 282                 dsl_deadlist_close(&ds->ds_deadlist);
 283         } else {
 284                 ASSERT(ds->ds_deadlist.dl_dbuf == NULL);
 285                 ASSERT(!ds->ds_deadlist.dl_oldfmt);
 286         }
 287         if (ds->ds_dir)
 288                 dsl_dir_close(ds->ds_dir, ds);
 289
 290         ASSERT(!list_link_active(&ds->ds_synced_link));
 291
 292         mutex_destroy(&ds->ds_lock);
 293         mutex_destroy(&ds->ds_recvlock);
 294         mutex_destroy(&ds->ds_opening_lock);
 295         rw_destroy(&ds->ds_rwlock);
 296         cv_destroy(&ds->ds_exclusive_cv);
 297
 298         kmem_free(ds, sizeof (dsl_dataset_t));
 299 }
 300
 301 static int
 302 dsl_dataset_get_snapname(dsl_dataset_t *ds)
 303 {
 304         dsl_dataset_phys_t *headphys;
 305         int err;
 306         dmu_buf_t *headdbuf;
 307         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 308         objset_t *mos = dp->dp_meta_objset;
 309
 310         if (ds->ds_snapname[0])
 311                 return (0);
 312         if (ds->ds_phys->ds_next_snap_obj == 0)
 313                 return (0);
 314
 315         err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
 316             FTAG, &headdbuf);
 317         if (err)
 318                 return (err);
 319         headphys = headdbuf->db_data;
 320         err = zap_value_search(dp->dp_meta_objset,
 321             headphys->ds_snapnames_zapobj, ds->ds_object, 0, ds->ds_snapname);
 322         dmu_buf_rele(headdbuf, FTAG);
 323         return (err);
 324 }
 325
 326 static int
 327 dsl_dataset_snap_lookup(dsl_dataset_t *ds, const char *name, uint64_t *value)
 328 {
 329         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 330         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 331         matchtype_t mt;
 332         int err;
 333
 334         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 335                 mt = MT_FIRST;
 336         else
 337                 mt = MT_EXACT;
 338
 339         err = zap_lookup_norm(mos, snapobj, name, 8, 1,
 340             value, mt, NULL, 0, NULL);
 341         if (err == ENOTSUP && mt == MT_FIRST)
 342                 err = zap_lookup(mos, snapobj, name, 8, 1, value);
 343         return (err);
 344 }
 345
 346 static int
 347 dsl_dataset_snap_remove(dsl_dataset_t *ds, char *name, dmu_tx_t *tx)
 348 {
 349         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 350         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 351         matchtype_t mt;
 352         int err;
 353
 354         dsl_dir_snap_cmtime_update(ds->ds_dir);
 355
 356         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 357                 mt = MT_FIRST;
 358         else
 359                 mt = MT_EXACT;
 360
 361         err = zap_remove_norm(mos, snapobj, name, mt, tx);
 362         if (err == ENOTSUP && mt == MT_FIRST)
 363                 err = zap_remove(mos, snapobj, name, tx);
 364         return (err);
 365 }
 366
 367 static int
 368 dsl_dataset_get_ref(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 369     dsl_dataset_t **dsp)
 370 {
 371         objset_t *mos = dp->dp_meta_objset;
 372         dmu_buf_t *dbuf;
 373         dsl_dataset_t *ds;
 374         int err;
 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         ds = dmu_buf_get_user(dbuf);
 383         if (ds == NULL) {
 384                 dsl_dataset_t *winner;
 385
 386                 ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
 387                 ds->ds_dbuf = dbuf;
 388                 ds->ds_object = dsobj;
 389                 ds->ds_phys = dbuf->db_data;
 390
 391                 mutex_init(&ds->ds_lock, NULL, MUTEX_DEFAULT, NULL);
 392                 mutex_init(&ds->ds_recvlock, NULL, MUTEX_DEFAULT, NULL);
 393                 mutex_init(&ds->ds_opening_lock, NULL, MUTEX_DEFAULT, NULL);
 394                 rw_init(&ds->ds_rwlock, 0, 0, 0);
 395                 cv_init(&ds->ds_exclusive_cv, NULL, CV_DEFAULT, NULL);
 396
 397                 bplist_create(&ds->ds_pending_deadlist);
 398                 dsl_deadlist_open(&ds->ds_deadlist,
 399                     mos, ds->ds_phys->ds_deadlist_obj);
 400
 401                 if (err == 0) {
 402                         err = dsl_dir_open_obj(dp,
 403                             ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
 404                 }
 405                 if (err) {
 406                         mutex_destroy(&ds->ds_lock);
 407                         mutex_destroy(&ds->ds_recvlock);
 408                         mutex_destroy(&ds->ds_opening_lock);
 409                         rw_destroy(&ds->ds_rwlock);
 410                         cv_destroy(&ds->ds_exclusive_cv);
 411                         bplist_destroy(&ds->ds_pending_deadlist);
 412                         dsl_deadlist_close(&ds->ds_deadlist);
 413                         kmem_free(ds, sizeof (dsl_dataset_t));
 414                         dmu_buf_rele(dbuf, tag);
 415                         return (err);
 416                 }
 417
 418                 if (!dsl_dataset_is_snapshot(ds)) {
 419                         ds->ds_snapname[0] = '\0';
 420                         if (ds->ds_phys->ds_prev_snap_obj) {
 421                                 err = dsl_dataset_get_ref(dp,
 422                                     ds->ds_phys->ds_prev_snap_obj,
 423                                     ds, &ds->ds_prev);
 424                         }
 425                 } else {
 426                         if (zfs_flags & ZFS_DEBUG_SNAPNAMES)
 427                                 err = dsl_dataset_get_snapname(ds);
 428                         if (err == 0 && ds->ds_phys->ds_userrefs_obj != 0) {
 429                                 err = zap_count(
 430                                     ds->ds_dir->dd_pool->dp_meta_objset,
 431                                     ds->ds_phys->ds_userrefs_obj,
 432                                     &ds->ds_userrefs);
 433                         }
 434                 }
 435
 436                 if (err == 0 && !dsl_dataset_is_snapshot(ds)) {
 437                         /*
 438                          * In sync context, we're called with either no lock
 439                          * or with the write lock.  If we're not syncing,
 440                          * we're always called with the read lock held.
 441                          */
 442                         boolean_t need_lock =
 443                             !RW_WRITE_HELD(&dp->dp_config_rwlock) &&
 444                             dsl_pool_sync_context(dp);
 445
 446                         if (need_lock)
 447                                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 448
 449                         err = dsl_prop_get_ds(ds,
 450                             "refreservation", sizeof (uint64_t), 1,
 451                             &ds->ds_reserved, NULL);
 452                         if (err == 0) {
 453                                 err = dsl_prop_get_ds(ds,
 454                                     "refquota", sizeof (uint64_t), 1,
 455                                     &ds->ds_quota, NULL);
 456                         }
 457
 458                         if (need_lock)
 459                                 rw_exit(&dp->dp_config_rwlock);
 460                 } else {
 461                         ds->ds_reserved = ds->ds_quota = 0;
 462                 }
 463
 464                 if (err == 0) {
 465                         winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
 466                             dsl_dataset_evict);
 467                 }
 468                 if (err || winner) {
 469                         bplist_destroy(&ds->ds_pending_deadlist);
 470                         dsl_deadlist_close(&ds->ds_deadlist);
 471                         if (ds->ds_prev)
 472                                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 473                         dsl_dir_close(ds->ds_dir, ds);
 474                         mutex_destroy(&ds->ds_lock);
 475                         mutex_destroy(&ds->ds_recvlock);
 476                         mutex_destroy(&ds->ds_opening_lock);
 477                         rw_destroy(&ds->ds_rwlock);
 478                         cv_destroy(&ds->ds_exclusive_cv);
 479                         kmem_free(ds, sizeof (dsl_dataset_t));
 480                         if (err) {
 481                                 dmu_buf_rele(dbuf, tag);
 482                                 return (err);
 483                         }
 484                         ds = winner;
 485                 } else {
 486                         ds->ds_fsid_guid =
 487                             unique_insert(ds->ds_phys->ds_fsid_guid);
 488                 }
 489         }
 490         ASSERT3P(ds->ds_dbuf, ==, dbuf);
 491         ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
 492         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0 ||
 493             spa_version(dp->dp_spa) < SPA_VERSION_ORIGIN ||
 494             dp->dp_origin_snap == NULL || ds == dp->dp_origin_snap);
 495         mutex_enter(&ds->ds_lock);
 496         if (!dsl_pool_sync_context(dp) && DSL_DATASET_IS_DESTROYED(ds)) {
 497                 mutex_exit(&ds->ds_lock);
 498                 dmu_buf_rele(ds->ds_dbuf, tag);
 499                 return (ENOENT);
 500         }
 501         mutex_exit(&ds->ds_lock);
 502         *dsp = ds;
 503         return (0);
 504 }
 505
 506 static int
 507 dsl_dataset_hold_ref(dsl_dataset_t *ds, void *tag)
 508 {
 509         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 510
 511         /*
 512          * In syncing context we don't want the rwlock lock: there
 513          * may be an existing writer waiting for sync phase to
 514          * finish.  We don't need to worry about such writers, since
 515          * sync phase is single-threaded, so the writer can't be
 516          * doing anything while we are active.
 517          */
 518         if (dsl_pool_sync_context(dp)) {
 519                 ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
 520                 return (0);
 521         }
 522
 523         /*
 524          * Normal users will hold the ds_rwlock as a READER until they
 525          * are finished (i.e., call dsl_dataset_rele()).  "Owners" will
 526          * drop their READER lock after they set the ds_owner field.
 527          *
 528          * If the dataset is being destroyed, the destroy thread will
 529          * obtain a WRITER lock for exclusive access after it's done its
 530          * open-context work and then change the ds_owner to
 531          * dsl_reaper once destruction is assured.  So threads
 532          * may block here temporarily, until the "destructability" of
 533          * the dataset is determined.
 534          */
 535         ASSERT(!RW_WRITE_HELD(&dp->dp_config_rwlock));
 536         mutex_enter(&ds->ds_lock);
 537         while (!rw_tryenter(&ds->ds_rwlock, RW_READER)) {
 538                 rw_exit(&dp->dp_config_rwlock);
 539                 cv_wait(&ds->ds_exclusive_cv, &ds->ds_lock);
 540                 if (DSL_DATASET_IS_DESTROYED(ds)) {
 541                         mutex_exit(&ds->ds_lock);
 542                         dsl_dataset_drop_ref(ds, tag);
 543                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 544                         return (ENOENT);
 545                 }
 546                 /*
 547                  * The dp_config_rwlock lives above the ds_lock. And
 548                  * we need to check DSL_DATASET_IS_DESTROYED() while
 549                  * holding the ds_lock, so we have to drop and reacquire
 550                  * the ds_lock here.
 551                  */
 552                 mutex_exit(&ds->ds_lock);
 553                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 554                 mutex_enter(&ds->ds_lock);
 555         }
 556         mutex_exit(&ds->ds_lock);
 557         return (0);
 558 }
 559
 560 int
 561 dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 562     dsl_dataset_t **dsp)
 563 {
 564         int err = dsl_dataset_get_ref(dp, dsobj, tag, dsp);
 565
 566         if (err)
 567                 return (err);
 568         return (dsl_dataset_hold_ref(*dsp, tag));
 569 }
 570
 571 int
 572 dsl_dataset_own_obj(dsl_pool_t *dp, uint64_t dsobj, boolean_t inconsistentok,
 573     void *tag, dsl_dataset_t **dsp)
 574 {
 575         int err = dsl_dataset_hold_obj(dp, dsobj, tag, dsp);
 576         if (err)
 577                 return (err);
 578         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 579                 dsl_dataset_rele(*dsp, tag);
 580                 *dsp = NULL;
 581                 return (EBUSY);
 582         }
 583         return (0);
 584 }
 585
 586 int
 587 dsl_dataset_hold(const char *name, void *tag, dsl_dataset_t **dsp)
 588 {
 589         dsl_dir_t *dd;
 590         dsl_pool_t *dp;
 591         const char *snapname;
 592         uint64_t obj;
 593         int err = 0;
 594
 595         err = dsl_dir_open_spa(NULL, name, FTAG, &dd, &snapname);
 596         if (err)
 597                 return (err);
 598
 599         dp = dd->dd_pool;
 600         obj = dd->dd_phys->dd_head_dataset_obj;
 601         rw_enter(&dp->dp_config_rwlock, RW_READER);
 602         if (obj)
 603                 err = dsl_dataset_get_ref(dp, obj, tag, dsp);
 604         else
 605                 err = ENOENT;
 606         if (err)
 607                 goto out;
 608
 609         err = dsl_dataset_hold_ref(*dsp, tag);
 610
 611         /* we may be looking for a snapshot */
 612         if (err == 0 && snapname != NULL) {
 613                 dsl_dataset_t *ds = NULL;
 614
 615                 if (*snapname++ != '@') {
 616                         dsl_dataset_rele(*dsp, tag);
 617                         err = ENOENT;
 618                         goto out;
 619                 }
 620
 621                 dprintf("looking for snapshot '%s'\n", snapname);
 622                 err = dsl_dataset_snap_lookup(*dsp, snapname, &obj);
 623                 if (err == 0)
 624                         err = dsl_dataset_get_ref(dp, obj, tag, &ds);
 625                 dsl_dataset_rele(*dsp, tag);
 626
 627                 ASSERT3U((err == 0), ==, (ds != NULL));
 628
 629                 if (ds) {
 630                         mutex_enter(&ds->ds_lock);
 631                         if (ds->ds_snapname[0] == 0)
 632                                 (void) strlcpy(ds->ds_snapname, snapname,
 633                                     sizeof (ds->ds_snapname));
 634                         mutex_exit(&ds->ds_lock);
 635                         err = dsl_dataset_hold_ref(ds, tag);
 636                         *dsp = err ? NULL : ds;
 637                 }
 638         }
 639 out:
 640         rw_exit(&dp->dp_config_rwlock);
 641         dsl_dir_close(dd, FTAG);
 642         return (err);
 643 }
 644
 645 int
 646 dsl_dataset_own(const char *name, boolean_t inconsistentok,
 647     void *tag, dsl_dataset_t **dsp)
 648 {
 649         int err = dsl_dataset_hold(name, tag, dsp);
 650         if (err)
 651                 return (err);
 652         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 653                 dsl_dataset_rele(*dsp, tag);
 654                 return (EBUSY);
 655         }
 656         return (0);
 657 }
 658
 659 void
 660 dsl_dataset_name(dsl_dataset_t *ds, char *name)
 661 {
 662         if (ds == NULL) {
 663                 (void) strcpy(name, "mos");
 664         } else {
 665                 dsl_dir_name(ds->ds_dir, name);
 666                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 667                 if (ds->ds_snapname[0]) {
 668                         (void) strcat(name, "@");
 669                         /*
 670                          * We use a "recursive" mutex so that we
 671                          * can call dprintf_ds() with ds_lock held.
 672                          */
 673                         if (!MUTEX_HELD(&ds->ds_lock)) {
 674                                 mutex_enter(&ds->ds_lock);
 675                                 (void) strcat(name, ds->ds_snapname);
 676                                 mutex_exit(&ds->ds_lock);
 677                         } else {
 678                                 (void) strcat(name, ds->ds_snapname);
 679                         }
 680                 }
 681         }
 682 }
 683
 684 static int
 685 dsl_dataset_namelen(dsl_dataset_t *ds)
 686 {
 687         int result;
 688
 689         if (ds == NULL) {
 690                 result = 3;     /* "mos" */
 691         } else {
 692                 result = dsl_dir_namelen(ds->ds_dir);
 693                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 694                 if (ds->ds_snapname[0]) {
 695                         ++result;       /* adding one for the @-sign */
 696                         if (!MUTEX_HELD(&ds->ds_lock)) {
 697                                 mutex_enter(&ds->ds_lock);
 698                                 result += strlen(ds->ds_snapname);
 699                                 mutex_exit(&ds->ds_lock);
 700                         } else {
 701                                 result += strlen(ds->ds_snapname);
 702                         }
 703                 }
 704         }
 705
 706         return (result);
 707 }
 708
 709 void
 710 dsl_dataset_drop_ref(dsl_dataset_t *ds, void *tag)
 711 {
 712         dmu_buf_rele(ds->ds_dbuf, tag);
 713 }
 714
 715 void
 716 dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
 717 {
 718         if (!dsl_pool_sync_context(ds->ds_dir->dd_pool)) {
 719                 rw_exit(&ds->ds_rwlock);
 720         }
 721         dsl_dataset_drop_ref(ds, tag);
 722 }
 723
 724 void
 725 dsl_dataset_disown(dsl_dataset_t *ds, void *tag)
 726 {
 727         ASSERT((ds->ds_owner == tag && ds->ds_dbuf) ||
 728             (DSL_DATASET_IS_DESTROYED(ds) && ds->ds_dbuf == NULL));
 729
 730         mutex_enter(&ds->ds_lock);
 731         ds->ds_owner = NULL;
 732         if (RW_WRITE_HELD(&ds->ds_rwlock)) {
 733                 rw_exit(&ds->ds_rwlock);
 734                 cv_broadcast(&ds->ds_exclusive_cv);
 735         }
 736         mutex_exit(&ds->ds_lock);
 737         if (ds->ds_dbuf)
 738                 dsl_dataset_drop_ref(ds, tag);
 739         else
 740                 dsl_dataset_evict(NULL, ds);
 741 }
 742
 743 boolean_t
 744 dsl_dataset_tryown(dsl_dataset_t *ds, boolean_t inconsistentok, void *tag)
 745 {
 746         boolean_t gotit = FALSE;
 747
 748         mutex_enter(&ds->ds_lock);
 749         if (ds->ds_owner == NULL &&
 750             (!DS_IS_INCONSISTENT(ds) || inconsistentok)) {
 751                 ds->ds_owner = tag;
 752                 if (!dsl_pool_sync_context(ds->ds_dir->dd_pool))
 753                         rw_exit(&ds->ds_rwlock);
 754                 gotit = TRUE;
 755         }
 756         mutex_exit(&ds->ds_lock);
 757         return (gotit);
 758 }
 759
 760 void
 761 dsl_dataset_make_exclusive(dsl_dataset_t *ds, void *owner)
 762 {
 763         ASSERT3P(owner, ==, ds->ds_owner);
 764         if (!RW_WRITE_HELD(&ds->ds_rwlock))
 765                 rw_enter(&ds->ds_rwlock, RW_WRITER);
 766 }
 767
 768 uint64_t
 769 dsl_dataset_create_sync_dd(dsl_dir_t *dd, dsl_dataset_t *origin,
 770     uint64_t flags, dmu_tx_t *tx)
 771 {
 772         dsl_pool_t *dp = dd->dd_pool;
 773         dmu_buf_t *dbuf;
 774         dsl_dataset_phys_t *dsphys;
 775         uint64_t dsobj;
 776         objset_t *mos = dp->dp_meta_objset;
 777
 778         if (origin == NULL)
 779                 origin = dp->dp_origin_snap;
 780
 781         ASSERT(origin == NULL || origin->ds_dir->dd_pool == dp);
 782         ASSERT(origin == NULL || origin->ds_phys->ds_num_children > 0);
 783         ASSERT(dmu_tx_is_syncing(tx));
 784         ASSERT(dd->dd_phys->dd_head_dataset_obj == 0);
 785
 786         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
 787             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
 788         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
 789         dmu_buf_will_dirty(dbuf, tx);
 790         dsphys = dbuf->db_data;
 791         bzero(dsphys, sizeof (dsl_dataset_phys_t));
 792         dsphys->ds_dir_obj = dd->dd_object;
 793         dsphys->ds_flags = flags;
 794         dsphys->ds_fsid_guid = unique_create();
 795         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
 796             sizeof (dsphys->ds_guid));
 797         dsphys->ds_snapnames_zapobj =
 798             zap_create_norm(mos, U8_TEXTPREP_TOUPPER, DMU_OT_DSL_DS_SNAP_MAP,
 799             DMU_OT_NONE, 0, tx);
 800         dsphys->ds_creation_time = gethrestime_sec();
 801         dsphys->ds_creation_txg = tx->tx_txg == TXG_INITIAL ? 1 : tx->tx_txg;
 802
 803         if (origin == NULL) {
 804                 dsphys->ds_deadlist_obj = dsl_deadlist_alloc(mos, tx);
 805         } else {
 806                 dsl_dataset_t *ohds;
 807
 808                 dsphys->ds_prev_snap_obj = origin->ds_object;
 809                 dsphys->ds_prev_snap_txg =
 810                     origin->ds_phys->ds_creation_txg;
 811                 dsphys->ds_used_bytes =
 812                     origin->ds_phys->ds_used_bytes;
 813                 dsphys->ds_compressed_bytes =
 814                     origin->ds_phys->ds_compressed_bytes;
 815                 dsphys->ds_uncompressed_bytes =
 816                     origin->ds_phys->ds_uncompressed_bytes;
 817                 dsphys->ds_bp = origin->ds_phys->ds_bp;
 818                 dsphys->ds_flags |= origin->ds_phys->ds_flags;
 819
 820                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
 821                 origin->ds_phys->ds_num_children++;
 822
 823                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
 824                     origin->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ohds));
 825                 dsphys->ds_deadlist_obj = dsl_deadlist_clone(&ohds->ds_deadlist,
 826                     dsphys->ds_prev_snap_txg, dsphys->ds_prev_snap_obj, tx);
 827                 dsl_dataset_rele(ohds, FTAG);
 828
 829                 if (spa_version(dp->dp_spa) >= SPA_VERSION_NEXT_CLONES) {
 830                         if (origin->ds_phys->ds_next_clones_obj == 0) {
 831                                 origin->ds_phys->ds_next_clones_obj =
 832                                     zap_create(mos,
 833                                     DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
 834                         }
 835                         VERIFY(0 == zap_add_int(mos,
 836                             origin->ds_phys->ds_next_clones_obj,
 837                             dsobj, tx));
 838                 }
 839
 840                 dmu_buf_will_dirty(dd->dd_dbuf, tx);
 841                 dd->dd_phys->dd_origin_obj = origin->ds_object;
 842                 if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
 843                         if (origin->ds_dir->dd_phys->dd_clones == 0) {
 844                                 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
 845                                 origin->ds_dir->dd_phys->dd_clones =
 846                                     zap_create(mos,
 847                                     DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
 848                         }
 849                         VERIFY3U(0, ==, zap_add_int(mos,
 850                             origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
 851                 }
 852         }
 853
 854         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
 855                 dsphys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
 856
 857         dmu_buf_rele(dbuf, FTAG);
 858
 859         dmu_buf_will_dirty(dd->dd_dbuf, tx);
 860         dd->dd_phys->dd_head_dataset_obj = dsobj;
 861
 862         return (dsobj);
 863 }
 864
 865 uint64_t
 866 dsl_dataset_create_sync(dsl_dir_t *pdd, const char *lastname,
 867     dsl_dataset_t *origin, uint64_t flags, cred_t *cr, dmu_tx_t *tx)
 868 {
 869         dsl_pool_t *dp = pdd->dd_pool;
 870         uint64_t dsobj, ddobj;
 871         dsl_dir_t *dd;
 872
 873         ASSERT(lastname[0] != '@');
 874
 875         ddobj = dsl_dir_create_sync(dp, pdd, lastname, tx);
 876         VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
 877
 878         dsobj = dsl_dataset_create_sync_dd(dd, origin, flags, tx);
 879
 880         dsl_deleg_set_create_perms(dd, tx, cr);
 881
 882         dsl_dir_close(dd, FTAG);
 883
 884         return (dsobj);
 885 }
 886
 887 struct destroyarg {
 888         dsl_sync_task_group_t *dstg;
 889         char *snapname;
 890         char *failed;
 891         boolean_t defer;
 892 };
 893
 894 static int
 895 dsl_snapshot_destroy_one(const char *name, void *arg)
 896 {
 897         struct destroyarg *da = arg;
 898         dsl_dataset_t *ds;
 899         int err;
 900         char *dsname;
 901
 902         dsname = kmem_asprintf("%s@%s", name, da->snapname);
 903         err = dsl_dataset_own(dsname, B_TRUE, da->dstg, &ds);
 904         strfree(dsname);
 905         if (err == 0) {
 906                 struct dsl_ds_destroyarg *dsda;
 907
 908                 dsl_dataset_make_exclusive(ds, da->dstg);
 909                 dsda = kmem_zalloc(sizeof (struct dsl_ds_destroyarg), KM_SLEEP);
 910                 dsda->ds = ds;
 911                 dsda->defer = da->defer;
 912                 dsl_sync_task_create(da->dstg, dsl_dataset_destroy_check,
 913                     dsl_dataset_destroy_sync, dsda, da->dstg, 0);
 914         } else if (err == ENOENT) {
 915                 err = 0;
 916         } else {
 917                 (void) strcpy(da->failed, name);
 918         }
 919         return (err);
 920 }
 921
 922 /*
 923  * Destroy 'snapname' in all descendants of 'fsname'.
 924  */
 925 #pragma weak dmu_snapshots_destroy = dsl_snapshots_destroy
 926 int
 927 dsl_snapshots_destroy(char *fsname, char *snapname, boolean_t defer)
 928 {
 929         int err;
 930         struct destroyarg da;
 931         dsl_sync_task_t *dst;
 932         spa_t *spa;
 933
 934         err = spa_open(fsname, &spa, FTAG);
 935         if (err)
 936                 return (err);
 937         da.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
 938         da.snapname = snapname;
 939         da.failed = fsname;
 940         da.defer = defer;
 941
 942         err = dmu_objset_find(fsname,
 943             dsl_snapshot_destroy_one, &da, DS_FIND_CHILDREN);
 944
 945         if (err == 0)
 946                 err = dsl_sync_task_group_wait(da.dstg);
 947
 948         for (dst = list_head(&da.dstg->dstg_tasks); dst;
 949             dst = list_next(&da.dstg->dstg_tasks, dst)) {
 950                 struct dsl_ds_destroyarg *dsda = dst->dst_arg1;
 951                 dsl_dataset_t *ds = dsda->ds;
 952
 953                 /*
 954                  * Return the file system name that triggered the error
 955                  */
 956                 if (dst->dst_err) {
 957                         dsl_dataset_name(ds, fsname);
 958                         *strchr(fsname, '@') = '\0';
 959                 }
 960                 ASSERT3P(dsda->rm_origin, ==, NULL);
 961                 dsl_dataset_disown(ds, da.dstg);
 962                 kmem_free(dsda, sizeof (struct dsl_ds_destroyarg));
 963         }
 964
 965         dsl_sync_task_group_destroy(da.dstg);
 966         spa_close(spa, FTAG);
 967         return (err);
 968 }
 969
 970 static boolean_t
 971 dsl_dataset_might_destroy_origin(dsl_dataset_t *ds)
 972 {
 973         boolean_t might_destroy = B_FALSE;
 974
 975         mutex_enter(&ds->ds_lock);
 976         if (ds->ds_phys->ds_num_children == 2 && ds->ds_userrefs == 0 &&
 977             DS_IS_DEFER_DESTROY(ds))
 978                 might_destroy = B_TRUE;
 979         mutex_exit(&ds->ds_lock);
 980
 981         return (might_destroy);
 982 }
 983
 984 /*
 985  * If we're removing a clone, and these three conditions are true:
 986  *      1) the clone's origin has no other children
 987  *      2) the clone's origin has no user references
 988  *      3) the clone's origin has been marked for deferred destruction
 989  * Then, prepare to remove the origin as part of this sync task group.
 990  */
 991 static int
 992 dsl_dataset_origin_rm_prep(struct dsl_ds_destroyarg *dsda, void *tag)
 993 {
 994         dsl_dataset_t *ds = dsda->ds;
 995         dsl_dataset_t *origin = ds->ds_prev;
 996
 997         if (dsl_dataset_might_destroy_origin(origin)) {
 998                 char *name;
 999                 int namelen;
1000                 int error;
1001
1002                 namelen = dsl_dataset_namelen(origin) + 1;
1003                 name = kmem_alloc(namelen, KM_SLEEP);
1004                 dsl_dataset_name(origin, name);
1005 #ifdef _KERNEL
1006                 error = zfs_unmount_snap(name, NULL);
1007                 if (error) {
1008                         kmem_free(name, namelen);
1009                         return (error);
1010                 }
1011 #endif
1012                 error = dsl_dataset_own(name, B_TRUE, tag, &origin);
1013                 kmem_free(name, namelen);
1014                 if (error)
1015                         return (error);
1016                 dsda->rm_origin = origin;
1017                 dsl_dataset_make_exclusive(origin, tag);
1018         }
1019
1020         return (0);
1021 }
1022
1023 /*
1024  * ds must be opened as OWNER.  On return (whether successful or not),
1025  * ds will be closed and caller can no longer dereference it.
1026  */
1027 int
1028 dsl_dataset_destroy(dsl_dataset_t *ds, void *tag, boolean_t defer)
1029 {
1030         int err;
1031         dsl_sync_task_group_t *dstg;
1032         objset_t *os;
1033         dsl_dir_t *dd;
1034         uint64_t obj;
1035         struct dsl_ds_destroyarg dsda = { 0 };
1036         dsl_dataset_t dummy_ds = { 0 };
1037
1038         dsda.ds = ds;
1039
1040         if (dsl_dataset_is_snapshot(ds)) {
1041                 /* Destroying a snapshot is simpler */
1042                 dsl_dataset_make_exclusive(ds, tag);
1043
1044                 dsda.defer = defer;
1045                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1046                     dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
1047                     &dsda, tag, 0);
1048                 ASSERT3P(dsda.rm_origin, ==, NULL);
1049                 goto out;
1050         } else if (defer) {
1051                 err = EINVAL;
1052                 goto out;
1053         }
1054
1055         dd = ds->ds_dir;
1056         dummy_ds.ds_dir = dd;
1057         dummy_ds.ds_object = ds->ds_object;
1058
1059         /*
1060          * Check for errors and mark this ds as inconsistent, in
1061          * case we crash while freeing the objects.
1062          */
1063         err = dsl_sync_task_do(dd->dd_pool, dsl_dataset_destroy_begin_check,
1064             dsl_dataset_destroy_begin_sync, ds, NULL, 0);
1065         if (err)
1066                 goto out;
1067
1068         err = dmu_objset_from_ds(ds, &os);
1069         if (err)
1070                 goto out;
1071
1072         /*
1073          * remove the objects in open context, so that we won't
1074          * have too much to do in syncing context.
1075          */
1076         for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE,
1077             ds->ds_phys->ds_prev_snap_txg)) {
1078                 /*
1079                  * Ignore errors, if there is not enough disk space
1080                  * we will deal with it in dsl_dataset_destroy_sync().
1081                  */
1082                 (void) dmu_free_object(os, obj);
1083         }
1084
1085         /*
1086          * We need to sync out all in-flight IO before we try to evict
1087          * (the dataset evict func is trying to clear the cached entries
1088          * for this dataset in the ARC).
1089          */
1090         txg_wait_synced(dd->dd_pool, 0);
1091
1092         /*
1093          * If we managed to free all the objects in open
1094          * context, the user space accounting should be zero.
1095          */
1096         if (ds->ds_phys->ds_bp.blk_fill == 0 &&
1097             dmu_objset_userused_enabled(os)) {
1098                 uint64_t count;
1099
1100                 ASSERT(zap_count(os, DMU_USERUSED_OBJECT, &count) != 0 ||
1101                     count == 0);
1102                 ASSERT(zap_count(os, DMU_GROUPUSED_OBJECT, &count) != 0 ||
1103                     count == 0);
1104         }
1105
1106         if (err != ESRCH)
1107                 goto out;
1108
1109         rw_enter(&dd->dd_pool->dp_config_rwlock, RW_READER);
1110         err = dsl_dir_open_obj(dd->dd_pool, dd->dd_object, NULL, FTAG, &dd);
1111         rw_exit(&dd->dd_pool->dp_config_rwlock);
1112
1113         if (err)
1114                 goto out;
1115
1116         /*
1117          * Blow away the dsl_dir + head dataset.
1118          */
1119         dsl_dataset_make_exclusive(ds, tag);
1120         /*
1121          * If we're removing a clone, we might also need to remove its
1122          * origin.
1123          */
1124         do {
1125                 dsda.need_prep = B_FALSE;
1126                 if (dsl_dir_is_clone(dd)) {
1127                         err = dsl_dataset_origin_rm_prep(&dsda, tag);
1128                         if (err) {
1129                                 dsl_dir_close(dd, FTAG);
1130                                 goto out;
1131                         }
1132                 }
1133
1134                 dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
1135                 dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
1136                     dsl_dataset_destroy_sync, &dsda, tag, 0);
1137                 dsl_sync_task_create(dstg, dsl_dir_destroy_check,
1138                     dsl_dir_destroy_sync, &dummy_ds, FTAG, 0);
1139                 err = dsl_sync_task_group_wait(dstg);
1140                 dsl_sync_task_group_destroy(dstg);
1141
1142                 /*
1143                  * We could be racing against 'zfs release' or 'zfs destroy -d'
1144                  * on the origin snap, in which case we can get EBUSY if we
1145                  * needed to destroy the origin snap but were not ready to
1146                  * do so.
1147                  */
1148                 if (dsda.need_prep) {
1149                         ASSERT(err == EBUSY);
1150                         ASSERT(dsl_dir_is_clone(dd));
1151                         ASSERT(dsda.rm_origin == NULL);
1152                 }
1153         } while (dsda.need_prep);
1154
1155         if (dsda.rm_origin != NULL)
1156                 dsl_dataset_disown(dsda.rm_origin, tag);
1157
1158         /* if it is successful, dsl_dir_destroy_sync will close the dd */
1159         if (err)
1160                 dsl_dir_close(dd, FTAG);
1161 out:
1162         dsl_dataset_disown(ds, tag);
1163         return (err);
1164 }
1165
1166 blkptr_t *
1167 dsl_dataset_get_blkptr(dsl_dataset_t *ds)
1168 {
1169         return (&ds->ds_phys->ds_bp);
1170 }
1171
1172 void
1173 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
1174 {
1175         ASSERT(dmu_tx_is_syncing(tx));
1176         /* If it's the meta-objset, set dp_meta_rootbp */
1177         if (ds == NULL) {
1178                 tx->tx_pool->dp_meta_rootbp = *bp;
1179         } else {
1180                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1181                 ds->ds_phys->ds_bp = *bp;
1182         }
1183 }
1184
1185 spa_t *
1186 dsl_dataset_get_spa(dsl_dataset_t *ds)
1187 {
1188         return (ds->ds_dir->dd_pool->dp_spa);
1189 }
1190
1191 void
1192 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
1193 {
1194         dsl_pool_t *dp;
1195
1196         if (ds == NULL) /* this is the meta-objset */
1197                 return;
1198
1199         ASSERT(ds->ds_objset != NULL);
1200
1201         if (ds->ds_phys->ds_next_snap_obj != 0)
1202                 panic("dirtying snapshot!");
1203
1204         dp = ds->ds_dir->dd_pool;
1205
1206         if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
1207                 /* up the hold count until we can be written out */
1208                 dmu_buf_add_ref(ds->ds_dbuf, ds);
1209         }
1210 }
1211
1212 /*
1213  * The unique space in the head dataset can be calculated by subtracting
1214  * the space used in the most recent snapshot, that is still being used
1215  * in this file system, from the space currently in use.  To figure out
1216  * the space in the most recent snapshot still in use, we need to take
1217  * the total space used in the snapshot and subtract out the space that
1218  * has been freed up since the snapshot was taken.
1219  */
1220 static void
1221 dsl_dataset_recalc_head_uniq(dsl_dataset_t *ds)
1222 {
1223         uint64_t mrs_used;
1224         uint64_t dlused, dlcomp, dluncomp;
1225
1226         ASSERT(!dsl_dataset_is_snapshot(ds));
1227
1228         if (ds->ds_phys->ds_prev_snap_obj != 0)
1229                 mrs_used = ds->ds_prev->ds_phys->ds_used_bytes;
1230         else
1231                 mrs_used = 0;
1232
1233         dsl_deadlist_space(&ds->ds_deadlist, &dlused, &dlcomp, &dluncomp);
1234
1235         ASSERT3U(dlused, <=, mrs_used);
1236         ds->ds_phys->ds_unique_bytes =
1237             ds->ds_phys->ds_used_bytes - (mrs_used - dlused);
1238
1239         if (spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1240             SPA_VERSION_UNIQUE_ACCURATE)
1241                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1242 }
1243
1244 struct killarg {
1245         dsl_dataset_t *ds;
1246         dmu_tx_t *tx;
1247 };
1248
1249 /* ARGSUSED */
1250 static int
1251 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
1252     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
1253 {
1254         struct killarg *ka = arg;
1255         dmu_tx_t *tx = ka->tx;
1256
1257         if (bp == NULL)
1258                 return (0);
1259
1260         if (zb->zb_level == ZB_ZIL_LEVEL) {
1261                 ASSERT(zilog != NULL);
1262                 /*
1263                  * It's a block in the intent log.  It has no
1264                  * accounting, so just free it.
1265                  */
1266                 dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
1267         } else {
1268                 ASSERT(zilog == NULL);
1269                 ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
1270                 (void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
1271         }
1272
1273         return (0);
1274 }
1275
1276 /* ARGSUSED */
1277 static int
1278 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
1279 {
1280         dsl_dataset_t *ds = arg1;
1281         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1282         uint64_t count;
1283         int err;
1284
1285         /*
1286          * Can't delete a head dataset if there are snapshots of it.
1287          * (Except if the only snapshots are from the branch we cloned
1288          * from.)
1289          */
1290         if (ds->ds_prev != NULL &&
1291             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1292                 return (EBUSY);
1293
1294         /*
1295          * This is really a dsl_dir thing, but check it here so that
1296          * we'll be less likely to leave this dataset inconsistent &
1297          * nearly destroyed.
1298          */
1299         err = zap_count(mos, ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
1300         if (err)
1301                 return (err);
1302         if (count != 0)
1303                 return (EEXIST);
1304
1305         return (0);
1306 }
1307
1308 /* ARGSUSED */
1309 static void
1310 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1311 {
1312         dsl_dataset_t *ds = arg1;
1313         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1314
1315         /* Mark it as inconsistent on-disk, in case we crash */
1316         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1317         ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1318
1319         spa_history_log_internal(LOG_DS_DESTROY_BEGIN, dp->dp_spa, tx,
1320             "dataset = %llu", ds->ds_object);
1321 }
1322
1323 static int
1324 dsl_dataset_origin_check(struct dsl_ds_destroyarg *dsda, void *tag,
1325     dmu_tx_t *tx)
1326 {
1327         dsl_dataset_t *ds = dsda->ds;
1328         dsl_dataset_t *ds_prev = ds->ds_prev;
1329
1330         if (dsl_dataset_might_destroy_origin(ds_prev)) {
1331                 struct dsl_ds_destroyarg ndsda = {0};
1332
1333                 /*
1334                  * If we're not prepared to remove the origin, don't remove
1335                  * the clone either.
1336                  */
1337                 if (dsda->rm_origin == NULL) {
1338                         dsda->need_prep = B_TRUE;
1339                         return (EBUSY);
1340                 }
1341
1342                 ndsda.ds = ds_prev;
1343                 ndsda.is_origin_rm = B_TRUE;
1344                 return (dsl_dataset_destroy_check(&ndsda, tag, tx));
1345         }
1346
1347         /*
1348          * If we're not going to remove the origin after all,
1349          * undo the open context setup.
1350          */
1351         if (dsda->rm_origin != NULL) {
1352                 dsl_dataset_disown(dsda->rm_origin, tag);
1353                 dsda->rm_origin = NULL;
1354         }
1355
1356         return (0);
1357 }
1358
1359 /* ARGSUSED */
1360 int
1361 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1362 {
1363         struct dsl_ds_destroyarg *dsda = arg1;
1364         dsl_dataset_t *ds = dsda->ds;
1365
1366         /* we have an owner hold, so noone else can destroy us */
1367         ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1368
1369         /*
1370          * Only allow deferred destroy on pools that support it.
1371          * NOTE: deferred destroy is only supported on snapshots.
1372          */
1373         if (dsda->defer) {
1374                 if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1375                     SPA_VERSION_USERREFS)
1376                         return (ENOTSUP);
1377                 ASSERT(dsl_dataset_is_snapshot(ds));
1378                 return (0);
1379         }
1380
1381         /*
1382          * Can't delete a head dataset if there are snapshots of it.
1383          * (Except if the only snapshots are from the branch we cloned
1384          * from.)
1385          */
1386         if (ds->ds_prev != NULL &&
1387             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1388                 return (EBUSY);
1389
1390         /*
1391          * If we made changes this txg, traverse_dsl_dataset won't find
1392          * them.  Try again.
1393          */
1394         if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1395                 return (EAGAIN);
1396
1397         if (dsl_dataset_is_snapshot(ds)) {
1398                 /*
1399                  * If this snapshot has an elevated user reference count,
1400                  * we can't destroy it yet.
1401                  */
1402                 if (ds->ds_userrefs > 0 && !dsda->releasing)
1403                         return (EBUSY);
1404
1405                 mutex_enter(&ds->ds_lock);
1406                 /*
1407                  * Can't delete a branch point. However, if we're destroying
1408                  * a clone and removing its origin due to it having a user
1409                  * hold count of 0 and having been marked for deferred destroy,
1410                  * it's OK for the origin to have a single clone.
1411                  */
1412                 if (ds->ds_phys->ds_num_children >
1413                     (dsda->is_origin_rm ? 2 : 1)) {
1414                         mutex_exit(&ds->ds_lock);
1415                         return (EEXIST);
1416                 }
1417                 mutex_exit(&ds->ds_lock);
1418         } else if (dsl_dir_is_clone(ds->ds_dir)) {
1419                 return (dsl_dataset_origin_check(dsda, arg2, tx));
1420         }
1421
1422         /* XXX we should do some i/o error checking... */
1423         return (0);
1424 }
1425
1426 struct refsarg {
1427         kmutex_t lock;
1428         boolean_t gone;
1429         kcondvar_t cv;
1430 };
1431
1432 /* ARGSUSED */
1433 static void
1434 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1435 {
1436         struct refsarg *arg = argv;
1437
1438         mutex_enter(&arg->lock);
1439         arg->gone = TRUE;
1440         cv_signal(&arg->cv);
1441         mutex_exit(&arg->lock);
1442 }
1443
1444 static void
1445 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1446 {
1447         struct refsarg arg;
1448
1449         mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1450         cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1451         arg.gone = FALSE;
1452         (void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1453             dsl_dataset_refs_gone);
1454         dmu_buf_rele(ds->ds_dbuf, tag);
1455         mutex_enter(&arg.lock);
1456         while (!arg.gone)
1457                 cv_wait(&arg.cv, &arg.lock);
1458         ASSERT(arg.gone);
1459         mutex_exit(&arg.lock);
1460         ds->ds_dbuf = NULL;
1461         ds->ds_phys = NULL;
1462         mutex_destroy(&arg.lock);
1463         cv_destroy(&arg.cv);
1464 }
1465
1466 static void
1467 remove_from_next_clones(dsl_dataset_t *ds, uint64_t obj, dmu_tx_t *tx)
1468 {
1469         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1470         uint64_t count;
1471         int err;
1472
1473         ASSERT(ds->ds_phys->ds_num_children >= 2);
1474         err = zap_remove_int(mos, ds->ds_phys->ds_next_clones_obj, obj, tx);
1475         /*
1476          * The err should not be ENOENT, but a bug in a previous version
1477          * of the code could cause upgrade_clones_cb() to not set
1478          * ds_next_snap_obj when it should, leading to a missing entry.
1479          * If we knew that the pool was created after
1480          * SPA_VERSION_NEXT_CLONES, we could assert that it isn't
1481          * ENOENT.  However, at least we can check that we don't have
1482          * too many entries in the next_clones_obj even after failing to
1483          * remove this one.
1484          */
1485         if (err != ENOENT) {
1486                 VERIFY3U(err, ==, 0);
1487         }
1488         ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
1489             &count));
1490         ASSERT3U(count, <=, ds->ds_phys->ds_num_children - 2);
1491 }
1492
1493 static void
1494 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
1495 {
1496         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1497         zap_cursor_t zc;
1498         zap_attribute_t za;
1499
1500         /*
1501          * If it is the old version, dd_clones doesn't exist so we can't
1502          * find the clones, but deadlist_remove_key() is a no-op so it
1503          * doesn't matter.
1504          */
1505         if (ds->ds_dir->dd_phys->dd_clones == 0)
1506                 return;
1507
1508         for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
1509             zap_cursor_retrieve(&zc, &za) == 0;
1510             zap_cursor_advance(&zc)) {
1511                 dsl_dataset_t *clone;
1512
1513                 VERIFY3U(0, ==, dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
1514                     za.za_first_integer, FTAG, &clone));
1515                 if (clone->ds_dir->dd_origin_txg > mintxg) {
1516                         dsl_deadlist_remove_key(&clone->ds_deadlist,
1517                             mintxg, tx);
1518                         dsl_dataset_remove_clones_key(clone, mintxg, tx);
1519                 }
1520                 dsl_dataset_rele(clone, FTAG);
1521         }
1522         zap_cursor_fini(&zc);
1523 }
1524
1525 struct process_old_arg {
1526         dsl_dataset_t *ds;
1527         dsl_dataset_t *ds_prev;
1528         boolean_t after_branch_point;
1529         zio_t *pio;
1530         uint64_t used, comp, uncomp;
1531 };
1532
1533 static int
1534 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1535 {
1536         struct process_old_arg *poa = arg;
1537         dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
1538
1539         if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
1540                 dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
1541                 if (poa->ds_prev && !poa->after_branch_point &&
1542                     bp->blk_birth >
1543                     poa->ds_prev->ds_phys->ds_prev_snap_txg) {
1544                         poa->ds_prev->ds_phys->ds_unique_bytes +=
1545                             bp_get_dsize_sync(dp->dp_spa, bp);
1546                 }
1547         } else {
1548                 poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
1549                 poa->comp += BP_GET_PSIZE(bp);
1550                 poa->uncomp += BP_GET_UCSIZE(bp);
1551                 dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
1552         }
1553         return (0);
1554 }
1555
1556 static void
1557 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
1558     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
1559 {
1560         struct process_old_arg poa = { 0 };
1561         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1562         objset_t *mos = dp->dp_meta_objset;
1563
1564         ASSERT(ds->ds_deadlist.dl_oldfmt);
1565         ASSERT(ds_next->ds_deadlist.dl_oldfmt);
1566
1567         poa.ds = ds;
1568         poa.ds_prev = ds_prev;
1569         poa.after_branch_point = after_branch_point;
1570         poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1571         VERIFY3U(0, ==, bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
1572             process_old_cb, &poa, tx));
1573         VERIFY3U(zio_wait(poa.pio), ==, 0);
1574         ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
1575
1576         /* change snapused */
1577         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1578             -poa.used, -poa.comp, -poa.uncomp, tx);
1579
1580         /* swap next's deadlist to our deadlist */
1581         dsl_deadlist_close(&ds->ds_deadlist);
1582         dsl_deadlist_close(&ds_next->ds_deadlist);
1583         SWITCH64(ds_next->ds_phys->ds_deadlist_obj,
1584             ds->ds_phys->ds_deadlist_obj);
1585         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
1586         dsl_deadlist_open(&ds_next->ds_deadlist, mos,
1587             ds_next->ds_phys->ds_deadlist_obj);
1588 }
1589
1590 void
1591 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1592 {
1593         struct dsl_ds_destroyarg *dsda = arg1;
1594         dsl_dataset_t *ds = dsda->ds;
1595         int err;
1596         int after_branch_point = FALSE;
1597         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1598         objset_t *mos = dp->dp_meta_objset;
1599         dsl_dataset_t *ds_prev = NULL;
1600         uint64_t obj;
1601
1602         ASSERT(ds->ds_owner);
1603         ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1604         ASSERT(ds->ds_prev == NULL ||
1605             ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1606         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1607
1608         if (dsda->defer) {
1609                 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1610                 if (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1) {
1611                         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1612                         ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1613                         return;
1614                 }
1615         }
1616
1617         /* signal any waiters that this dataset is going away */
1618         mutex_enter(&ds->ds_lock);
1619         ds->ds_owner = dsl_reaper;
1620         cv_broadcast(&ds->ds_exclusive_cv);
1621         mutex_exit(&ds->ds_lock);
1622
1623         if (ds->ds_objset) {
1624                 dmu_objset_evict(ds->ds_objset);
1625                 ds->ds_objset = NULL;
1626         }
1627
1628         /* Remove our reservation */
1629         if (ds->ds_reserved != 0) {
1630                 dsl_prop_setarg_t psa;
1631                 uint64_t value = 0;
1632
1633                 dsl_prop_setarg_init_uint64(&psa, "refreservation",
1634                     (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
1635                     &value);
1636                 psa.psa_effective_value = 0;    /* predict default value */
1637
1638                 dsl_dataset_set_reservation_sync(ds, &psa, tx);
1639                 ASSERT3U(ds->ds_reserved, ==, 0);
1640         }
1641
1642         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1643
1644         dsl_scan_ds_destroyed(ds, tx);
1645
1646         obj = ds->ds_object;
1647
1648         if (ds->ds_phys->ds_prev_snap_obj != 0) {
1649                 if (ds->ds_prev) {
1650                         ds_prev = ds->ds_prev;
1651                 } else {
1652                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1653                             ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1654                 }
1655                 after_branch_point =
1656                     (ds_prev->ds_phys->ds_next_snap_obj != obj);
1657
1658                 dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1659                 if (after_branch_point &&
1660                     ds_prev->ds_phys->ds_next_clones_obj != 0) {
1661                         remove_from_next_clones(ds_prev, obj, tx);
1662                         if (ds->ds_phys->ds_next_snap_obj != 0) {
1663                                 VERIFY(0 == zap_add_int(mos,
1664                                     ds_prev->ds_phys->ds_next_clones_obj,
1665                                     ds->ds_phys->ds_next_snap_obj, tx));
1666                         }
1667                 }
1668                 if (after_branch_point &&
1669                     ds->ds_phys->ds_next_snap_obj == 0) {
1670                         /* This clone is toast. */
1671                         ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1672                         ds_prev->ds_phys->ds_num_children--;
1673
1674                         /*
1675                          * If the clone's origin has no other clones, no
1676                          * user holds, and has been marked for deferred
1677                          * deletion, then we should have done the necessary
1678                          * destroy setup for it.
1679                          */
1680                         if (ds_prev->ds_phys->ds_num_children == 1 &&
1681                             ds_prev->ds_userrefs == 0 &&
1682                             DS_IS_DEFER_DESTROY(ds_prev)) {
1683                                 ASSERT3P(dsda->rm_origin, !=, NULL);
1684                         } else {
1685                                 ASSERT3P(dsda->rm_origin, ==, NULL);
1686                         }
1687                 } else if (!after_branch_point) {
1688                         ds_prev->ds_phys->ds_next_snap_obj =
1689                             ds->ds_phys->ds_next_snap_obj;
1690                 }
1691         }
1692
1693         if (dsl_dataset_is_snapshot(ds)) {
1694                 dsl_dataset_t *ds_next;
1695                 uint64_t old_unique;
1696                 uint64_t used = 0, comp = 0, uncomp = 0;
1697
1698                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1699                     ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1700                 ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1701
1702                 old_unique = ds_next->ds_phys->ds_unique_bytes;
1703
1704                 dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1705                 ds_next->ds_phys->ds_prev_snap_obj =
1706                     ds->ds_phys->ds_prev_snap_obj;
1707                 ds_next->ds_phys->ds_prev_snap_txg =
1708                     ds->ds_phys->ds_prev_snap_txg;
1709                 ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1710                     ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1711
1712
1713                 if (ds_next->ds_deadlist.dl_oldfmt) {
1714                         process_old_deadlist(ds, ds_prev, ds_next,
1715                             after_branch_point, tx);
1716                 } else {
1717                         /* Adjust prev's unique space. */
1718                         if (ds_prev && !after_branch_point) {
1719                                 dsl_deadlist_space_range(&ds_next->ds_deadlist,
1720                                     ds_prev->ds_phys->ds_prev_snap_txg,
1721                                     ds->ds_phys->ds_prev_snap_txg,
1722                                     &used, &comp, &uncomp);
1723                                 ds_prev->ds_phys->ds_unique_bytes += used;
1724                         }
1725
1726                         /* Adjust snapused. */
1727                         dsl_deadlist_space_range(&ds_next->ds_deadlist,
1728                             ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
1729                             &used, &comp, &uncomp);
1730                         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1731                             -used, -comp, -uncomp, tx);
1732
1733                         /* Move blocks to be freed to pool's free list. */
1734                         dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
1735                             &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
1736                             tx);
1737                         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
1738                             DD_USED_HEAD, used, comp, uncomp, tx);
1739                         dsl_dir_dirty(tx->tx_pool->dp_free_dir, tx);
1740
1741                         /* Merge our deadlist into next's and free it. */
1742                         dsl_deadlist_merge(&ds_next->ds_deadlist,
1743                             ds->ds_phys->ds_deadlist_obj, tx);
1744                 }
1745                 dsl_deadlist_close(&ds->ds_deadlist);
1746                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1747
1748                 /* Collapse range in clone heads */
1749                 dsl_dataset_remove_clones_key(ds,
1750                     ds->ds_phys->ds_creation_txg, tx);
1751
1752                 if (dsl_dataset_is_snapshot(ds_next)) {
1753                         dsl_dataset_t *ds_nextnext;
1754
1755                         /*
1756                          * Update next's unique to include blocks which
1757                          * were previously shared by only this snapshot
1758                          * and it.  Those blocks will be born after the
1759                          * prev snap and before this snap, and will have
1760                          * died after the next snap and before the one
1761                          * after that (ie. be on the snap after next's
1762                          * deadlist).
1763                          */
1764                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1765                             ds_next->ds_phys->ds_next_snap_obj,
1766                             FTAG, &ds_nextnext));
1767                         dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
1768                             ds->ds_phys->ds_prev_snap_txg,
1769                             ds->ds_phys->ds_creation_txg,
1770                             &used, &comp, &uncomp);
1771                         ds_next->ds_phys->ds_unique_bytes += used;
1772                         dsl_dataset_rele(ds_nextnext, FTAG);
1773                         ASSERT3P(ds_next->ds_prev, ==, NULL);
1774
1775                         /* Collapse range in this head. */
1776                         dsl_dataset_t *hds;
1777                         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
1778                             ds->ds_dir->dd_phys->dd_head_dataset_obj,
1779                             FTAG, &hds));
1780                         dsl_deadlist_remove_key(&hds->ds_deadlist,
1781                             ds->ds_phys->ds_creation_txg, tx);
1782                         dsl_dataset_rele(hds, FTAG);
1783
1784                 } else {
1785                         ASSERT3P(ds_next->ds_prev, ==, ds);
1786                         dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1787                         ds_next->ds_prev = NULL;
1788                         if (ds_prev) {
1789                                 VERIFY(0 == dsl_dataset_get_ref(dp,
1790                                     ds->ds_phys->ds_prev_snap_obj,
1791                                     ds_next, &ds_next->ds_prev));
1792                         }
1793
1794                         dsl_dataset_recalc_head_uniq(ds_next);
1795
1796                         /*
1797                          * Reduce the amount of our unconsmed refreservation
1798                          * being charged to our parent by the amount of
1799                          * new unique data we have gained.
1800                          */
1801                         if (old_unique < ds_next->ds_reserved) {
1802                                 int64_t mrsdelta;
1803                                 uint64_t new_unique =
1804                                     ds_next->ds_phys->ds_unique_bytes;
1805
1806                                 ASSERT(old_unique <= new_unique);
1807                                 mrsdelta = MIN(new_unique - old_unique,
1808                                     ds_next->ds_reserved - old_unique);
1809                                 dsl_dir_diduse_space(ds->ds_dir,
1810                                     DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1811                         }
1812                 }
1813                 dsl_dataset_rele(ds_next, FTAG);
1814         } else {
1815                 /*
1816                  * There's no next snapshot, so this is a head dataset.
1817                  * Destroy the deadlist.  Unless it's a clone, the
1818                  * deadlist should be empty.  (If it's a clone, it's
1819                  * safe to ignore the deadlist contents.)
1820                  */
1821                 struct killarg ka;
1822
1823                 dsl_deadlist_close(&ds->ds_deadlist);
1824                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1825                 ds->ds_phys->ds_deadlist_obj = 0;
1826
1827                 /*
1828                  * Free everything that we point to (that's born after
1829                  * the previous snapshot, if we are a clone)
1830                  *
1831                  * NB: this should be very quick, because we already
1832                  * freed all the objects in open context.
1833                  */
1834                 ka.ds = ds;
1835                 ka.tx = tx;
1836                 err = traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1837                     TRAVERSE_POST, kill_blkptr, &ka);
1838                 ASSERT3U(err, ==, 0);
1839                 ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1840                     ds->ds_phys->ds_unique_bytes == 0);
1841
1842                 if (ds->ds_prev != NULL) {
1843                         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
1844                                 VERIFY3U(0, ==, zap_remove_int(mos,
1845                                     ds->ds_prev->ds_dir->dd_phys->dd_clones,
1846                                     ds->ds_object, tx));
1847                         }
1848                         dsl_dataset_rele(ds->ds_prev, ds);
1849                         ds->ds_prev = ds_prev = NULL;
1850                 }
1851         }
1852
1853         if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1854                 /* Erase the link in the dir */
1855                 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1856                 ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1857                 ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1858                 err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1859                 ASSERT(err == 0);
1860         } else {
1861                 /* remove from snapshot namespace */
1862                 dsl_dataset_t *ds_head;
1863                 ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1864                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1865                     ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1866                 VERIFY(0 == dsl_dataset_get_snapname(ds));
1867 #ifdef ZFS_DEBUG
1868                 {
1869                         uint64_t val;
1870
1871                         err = dsl_dataset_snap_lookup(ds_head,
1872                             ds->ds_snapname, &val);
1873                         ASSERT3U(err, ==, 0);
1874                         ASSERT3U(val, ==, obj);
1875                 }
1876 #endif
1877                 err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1878                 ASSERT(err == 0);
1879                 dsl_dataset_rele(ds_head, FTAG);
1880         }
1881
1882         if (ds_prev && ds->ds_prev != ds_prev)
1883                 dsl_dataset_rele(ds_prev, FTAG);
1884
1885         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1886         spa_history_log_internal(LOG_DS_DESTROY, dp->dp_spa, tx,
1887             "dataset = %llu", ds->ds_object);
1888
1889         if (ds->ds_phys->ds_next_clones_obj != 0) {
1890                 uint64_t count;
1891                 ASSERT(0 == zap_count(mos,
1892                     ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1893                 VERIFY(0 == dmu_object_free(mos,
1894                     ds->ds_phys->ds_next_clones_obj, tx));
1895         }
1896         if (ds->ds_phys->ds_props_obj != 0)
1897                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1898         if (ds->ds_phys->ds_userrefs_obj != 0)
1899                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1900         dsl_dir_close(ds->ds_dir, ds);
1901         ds->ds_dir = NULL;
1902         dsl_dataset_drain_refs(ds, tag);
1903         VERIFY(0 == dmu_object_free(mos, obj, tx));
1904
1905         if (dsda->rm_origin) {
1906                 /*
1907                  * Remove the origin of the clone we just destroyed.
1908                  */
1909                 struct dsl_ds_destroyarg ndsda = {0};
1910
1911                 ndsda.ds = dsda->rm_origin;
1912                 dsl_dataset_destroy_sync(&ndsda, tag, tx);
1913         }
1914 }
1915
1916 static int
1917 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
1918 {
1919         uint64_t asize;
1920
1921         if (!dmu_tx_is_syncing(tx))
1922                 return (0);
1923
1924         /*
1925          * If there's an fs-only reservation, any blocks that might become
1926          * owned by the snapshot dataset must be accommodated by space
1927          * outside of the reservation.
1928          */
1929         ASSERT(ds->ds_reserved == 0 || DS_UNIQUE_IS_ACCURATE(ds));
1930         asize = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
1931         if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, FALSE))
1932                 return (ENOSPC);
1933
1934         /*
1935          * Propogate any reserved space for this snapshot to other
1936          * snapshot checks in this sync group.
1937          */
1938         if (asize > 0)
1939                 dsl_dir_willuse_space(ds->ds_dir, asize, tx);
1940
1941         return (0);
1942 }
1943
1944 int
1945 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1946 {
1947         dsl_dataset_t *ds = arg1;
1948         const char *snapname = arg2;
1949         int err;
1950         uint64_t value;
1951
1952         /*
1953          * We don't allow multiple snapshots of the same txg.  If there
1954          * is already one, try again.
1955          */
1956         if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
1957                 return (EAGAIN);
1958
1959         /*
1960          * Check for conflicting name snapshot name.
1961          */
1962         err = dsl_dataset_snap_lookup(ds, snapname, &value);
1963         if (err == 0)
1964                 return (EEXIST);
1965         if (err != ENOENT)
1966                 return (err);
1967
1968         /*
1969          * Check that the dataset's name is not too long.  Name consists
1970          * of the dataset's length + 1 for the @-sign + snapshot name's length
1971          */
1972         if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
1973                 return (ENAMETOOLONG);
1974
1975         err = dsl_dataset_snapshot_reserve_space(ds, tx);
1976         if (err)
1977                 return (err);
1978
1979         ds->ds_trysnap_txg = tx->tx_txg;
1980         return (0);
1981 }
1982
1983 void
1984 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1985 {
1986         dsl_dataset_t *ds = arg1;
1987         const char *snapname = arg2;
1988         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1989         dmu_buf_t *dbuf;
1990         dsl_dataset_phys_t *dsphys;
1991         uint64_t dsobj, crtxg;
1992         objset_t *mos = dp->dp_meta_objset;
1993         int err;
1994
1995         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1996
1997         /*
1998          * The origin's ds_creation_txg has to be < TXG_INITIAL
1999          */
2000         if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2001                 crtxg = 1;
2002         else
2003                 crtxg = tx->tx_txg;
2004
2005         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2006             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2007         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2008         dmu_buf_will_dirty(dbuf, tx);
2009         dsphys = dbuf->db_data;
2010         bzero(dsphys, sizeof (dsl_dataset_phys_t));
2011         dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2012         dsphys->ds_fsid_guid = unique_create();
2013         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2014             sizeof (dsphys->ds_guid));
2015         dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2016         dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2017         dsphys->ds_next_snap_obj = ds->ds_object;
2018         dsphys->ds_num_children = 1;
2019         dsphys->ds_creation_time = gethrestime_sec();
2020         dsphys->ds_creation_txg = crtxg;
2021         dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2022         dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
2023         dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2024         dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2025         dsphys->ds_flags = ds->ds_phys->ds_flags;
2026         dsphys->ds_bp = ds->ds_phys->ds_bp;
2027         dmu_buf_rele(dbuf, FTAG);
2028
2029         ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2030         if (ds->ds_prev) {
2031                 uint64_t next_clones_obj =
2032                     ds->ds_prev->ds_phys->ds_next_clones_obj;
2033                 ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2034                     ds->ds_object ||
2035                     ds->ds_prev->ds_phys->ds_num_children > 1);
2036                 if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2037                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2038                         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2039                             ds->ds_prev->ds_phys->ds_creation_txg);
2040                         ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2041                 } else if (next_clones_obj != 0) {
2042                         remove_from_next_clones(ds->ds_prev,
2043                             dsphys->ds_next_snap_obj, tx);
2044                         VERIFY3U(0, ==, zap_add_int(mos,
2045                             next_clones_obj, dsobj, tx));
2046                 }
2047         }
2048
2049         /*
2050          * If we have a reference-reservation on this dataset, we will
2051          * need to increase the amount of refreservation being charged
2052          * since our unique space is going to zero.
2053          */
2054         if (ds->ds_reserved) {
2055                 int64_t delta;
2056                 ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
2057                 delta = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2058                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2059                     delta, 0, 0, tx);
2060         }
2061
2062         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2063         zfs_dbgmsg("taking snapshot %s@%s/%llu; newkey=%llu",
2064             ds->ds_dir->dd_myname, snapname, dsobj,
2065             ds->ds_phys->ds_prev_snap_txg);
2066         ds->ds_phys->ds_deadlist_obj = dsl_deadlist_clone(&ds->ds_deadlist,
2067             UINT64_MAX, ds->ds_phys->ds_prev_snap_obj, tx);
2068         dsl_deadlist_close(&ds->ds_deadlist);
2069         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
2070         dsl_deadlist_add_key(&ds->ds_deadlist,
2071             ds->ds_phys->ds_prev_snap_txg, tx);
2072
2073         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2074         ds->ds_phys->ds_prev_snap_obj = dsobj;
2075         ds->ds_phys->ds_prev_snap_txg = crtxg;
2076         ds->ds_phys->ds_unique_bytes = 0;
2077         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2078                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2079
2080         err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2081             snapname, 8, 1, &dsobj, tx);
2082         ASSERT(err == 0);
2083
2084         if (ds->ds_prev)
2085                 dsl_dataset_drop_ref(ds->ds_prev, ds);
2086         VERIFY(0 == dsl_dataset_get_ref(dp,
2087             ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2088
2089         dsl_scan_ds_snapshotted(ds, tx);
2090
2091         dsl_dir_snap_cmtime_update(ds->ds_dir);
2092
2093         spa_history_log_internal(LOG_DS_SNAPSHOT, dp->dp_spa, tx,
2094             "dataset = %llu", dsobj);
2095 }
2096
2097 void
2098 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2099 {
2100         ASSERT(dmu_tx_is_syncing(tx));
2101         ASSERT(ds->ds_objset != NULL);
2102         ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2103
2104         /*
2105          * in case we had to change ds_fsid_guid when we opened it,
2106          * sync it out now.
2107          */
2108         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2109         ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2110
2111         dsl_dir_dirty(ds->ds_dir, tx);
2112         dmu_objset_sync(ds->ds_objset, zio, tx);
2113 }
2114
2115 void
2116 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2117 {
2118         uint64_t refd, avail, uobjs, aobjs;
2119
2120         dsl_dir_stats(ds->ds_dir, nv);
2121
2122         dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2123         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2124         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2125
2126         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2127             ds->ds_phys->ds_creation_time);
2128         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2129             ds->ds_phys->ds_creation_txg);
2130         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2131             ds->ds_quota);
2132         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2133             ds->ds_reserved);
2134         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2135             ds->ds_phys->ds_guid);
2136         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_UNIQUE,
2137             ds->ds_phys->ds_unique_bytes);
2138         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_OBJSETID,
2139             ds->ds_object);
2140         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS,
2141             ds->ds_userrefs);
2142         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2143             DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2144
2145         if (ds->ds_phys->ds_next_snap_obj) {
2146                 /*
2147                  * This is a snapshot; override the dd's space used with
2148                  * our unique space and compression ratio.
2149                  */
2150                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2151                     ds->ds_phys->ds_unique_bytes);
2152                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO,
2153                     ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2154                     (ds->ds_phys->ds_uncompressed_bytes * 100 /
2155                     ds->ds_phys->ds_compressed_bytes));
2156         }
2157 }
2158
2159 void
2160 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2161 {
2162         stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2163         stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2164         stat->dds_guid = ds->ds_phys->ds_guid;
2165         if (ds->ds_phys->ds_next_snap_obj) {
2166                 stat->dds_is_snapshot = B_TRUE;
2167                 stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2168         } else {
2169                 stat->dds_is_snapshot = B_FALSE;
2170                 stat->dds_num_clones = 0;
2171         }
2172
2173         /* clone origin is really a dsl_dir thing... */
2174         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2175         if (dsl_dir_is_clone(ds->ds_dir)) {
2176                 dsl_dataset_t *ods;
2177
2178                 VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2179                     ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2180                 dsl_dataset_name(ods, stat->dds_origin);
2181                 dsl_dataset_drop_ref(ods, FTAG);
2182         } else {
2183                 stat->dds_origin[0] = '\0';
2184         }
2185         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2186 }
2187
2188 uint64_t
2189 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2190 {
2191         return (ds->ds_fsid_guid);
2192 }
2193
2194 void
2195 dsl_dataset_space(dsl_dataset_t *ds,
2196     uint64_t *refdbytesp, uint64_t *availbytesp,
2197     uint64_t *usedobjsp, uint64_t *availobjsp)
2198 {
2199         *refdbytesp = ds->ds_phys->ds_used_bytes;
2200         *availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2201         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2202                 *availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2203         if (ds->ds_quota != 0) {
2204                 /*
2205                  * Adjust available bytes according to refquota
2206                  */
2207                 if (*refdbytesp < ds->ds_quota)
2208                         *availbytesp = MIN(*availbytesp,
2209                             ds->ds_quota - *refdbytesp);
2210                 else
2211                         *availbytesp = 0;
2212         }
2213         *usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2214         *availobjsp = DN_MAX_OBJECT - *usedobjsp;
2215 }
2216
2217 boolean_t
2218 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2219 {
2220         dsl_pool_t *dp = ds->ds_dir->dd_pool;
2221
2222         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2223             dsl_pool_sync_context(dp));
2224         if (ds->ds_prev == NULL)
2225                 return (B_FALSE);
2226         if (ds->ds_phys->ds_bp.blk_birth >
2227             ds->ds_prev->ds_phys->ds_creation_txg)
2228                 return (B_TRUE);
2229         return (B_FALSE);
2230 }
2231
2232 /* ARGSUSED */
2233 static int
2234 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2235 {
2236         dsl_dataset_t *ds = arg1;
2237         char *newsnapname = arg2;
2238         dsl_dir_t *dd = ds->ds_dir;
2239         dsl_dataset_t *hds;
2240         uint64_t val;
2241         int err;
2242
2243         err = dsl_dataset_hold_obj(dd->dd_pool,
2244             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2245         if (err)
2246                 return (err);
2247
2248         /* new name better not be in use */
2249         err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2250         dsl_dataset_rele(hds, FTAG);
2251
2252         if (err == 0)
2253                 err = EEXIST;
2254         else if (err == ENOENT)
2255                 err = 0;
2256
2257         /* dataset name + 1 for the "@" + the new snapshot name must fit */
2258         if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2259                 err = ENAMETOOLONG;
2260
2261         return (err);
2262 }
2263
2264 static void
2265 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2266 {
2267         dsl_dataset_t *ds = arg1;
2268         const char *newsnapname = arg2;
2269         dsl_dir_t *dd = ds->ds_dir;
2270         objset_t *mos = dd->dd_pool->dp_meta_objset;
2271         dsl_dataset_t *hds;
2272         int err;
2273
2274         ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2275
2276         VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2277             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2278
2279         VERIFY(0 == dsl_dataset_get_snapname(ds));
2280         err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2281         ASSERT3U(err, ==, 0);
2282         mutex_enter(&ds->ds_lock);
2283         (void) strcpy(ds->ds_snapname, newsnapname);
2284         mutex_exit(&ds->ds_lock);
2285         err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2286             ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2287         ASSERT3U(err, ==, 0);
2288
2289         spa_history_log_internal(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2290             "dataset = %llu", ds->ds_object);
2291         dsl_dataset_rele(hds, FTAG);
2292 }
2293
2294 struct renamesnaparg {
2295         dsl_sync_task_group_t *dstg;
2296         char failed[MAXPATHLEN];
2297         char *oldsnap;
2298         char *newsnap;
2299 };
2300
2301 static int
2302 dsl_snapshot_rename_one(const char *name, void *arg)
2303 {
2304         struct renamesnaparg *ra = arg;
2305         dsl_dataset_t *ds = NULL;
2306         char *snapname;
2307         int err;
2308
2309         snapname = kmem_asprintf("%s@%s", name, ra->oldsnap);
2310         (void) strlcpy(ra->failed, snapname, sizeof (ra->failed));
2311
2312         /*
2313          * For recursive snapshot renames the parent won't be changing
2314          * so we just pass name for both the to/from argument.
2315          */
2316         err = zfs_secpolicy_rename_perms(snapname, snapname, CRED());
2317         if (err != 0) {
2318                 strfree(snapname);
2319                 return (err == ENOENT ? 0 : err);
2320         }
2321
2322 #ifdef _KERNEL
2323         /*
2324          * For all filesystems undergoing rename, we'll need to unmount it.
2325          */
2326         (void) zfs_unmount_snap(snapname, NULL);
2327 #endif
2328         err = dsl_dataset_hold(snapname, ra->dstg, &ds);
2329         strfree(snapname);
2330         if (err != 0)
2331                 return (err == ENOENT ? 0 : err);
2332
2333         dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2334             dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2335
2336         return (0);
2337 }
2338
2339 static int
2340 dsl_recursive_rename(char *oldname, const char *newname)
2341 {
2342         int err;
2343         struct renamesnaparg *ra;
2344         dsl_sync_task_t *dst;
2345         spa_t *spa;
2346         char *cp, *fsname = spa_strdup(oldname);
2347         int len = strlen(oldname) + 1;
2348
2349         /* truncate the snapshot name to get the fsname */
2350         cp = strchr(fsname, '@');
2351         *cp = '\0';
2352
2353         err = spa_open(fsname, &spa, FTAG);
2354         if (err) {
2355                 kmem_free(fsname, len);
2356                 return (err);
2357         }
2358         ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2359         ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2360
2361         ra->oldsnap = strchr(oldname, '@') + 1;
2362         ra->newsnap = strchr(newname, '@') + 1;
2363         *ra->failed = '\0';
2364
2365         err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2366             DS_FIND_CHILDREN);
2367         kmem_free(fsname, len);
2368
2369         if (err == 0) {
2370                 err = dsl_sync_task_group_wait(ra->dstg);
2371         }
2372
2373         for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2374             dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2375                 dsl_dataset_t *ds = dst->dst_arg1;
2376                 if (dst->dst_err) {
2377                         dsl_dir_name(ds->ds_dir, ra->failed);
2378                         (void) strlcat(ra->failed, "@", sizeof (ra->failed));
2379                         (void) strlcat(ra->failed, ra->newsnap,
2380                             sizeof (ra->failed));
2381                 }
2382                 dsl_dataset_rele(ds, ra->dstg);
2383         }
2384
2385         if (err)
2386                 (void) strlcpy(oldname, ra->failed, sizeof (ra->failed));
2387
2388         dsl_sync_task_group_destroy(ra->dstg);
2389         kmem_free(ra, sizeof (struct renamesnaparg));
2390         spa_close(spa, FTAG);
2391         return (err);
2392 }
2393
2394 static int
2395 dsl_valid_rename(const char *oldname, void *arg)
2396 {
2397         int delta = *(int *)arg;
2398
2399         if (strlen(oldname) + delta >= MAXNAMELEN)
2400                 return (ENAMETOOLONG);
2401
2402         return (0);
2403 }
2404
2405 #pragma weak dmu_objset_rename = dsl_dataset_rename
2406 int
2407 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2408 {
2409         dsl_dir_t *dd;
2410         dsl_dataset_t *ds;
2411         const char *tail;
2412         int err;
2413
2414         err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2415         if (err)
2416                 return (err);
2417
2418         if (tail == NULL) {
2419                 int delta = strlen(newname) - strlen(oldname);
2420
2421                 /* if we're growing, validate child name lengths */
2422                 if (delta > 0)
2423                         err = dmu_objset_find(oldname, dsl_valid_rename,
2424                             &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2425
2426                 if (err == 0)
2427                         err = dsl_dir_rename(dd, newname);
2428                 dsl_dir_close(dd, FTAG);
2429                 return (err);
2430         }
2431
2432         if (tail[0] != '@') {
2433                 /* the name ended in a nonexistent component */
2434                 dsl_dir_close(dd, FTAG);
2435                 return (ENOENT);
2436         }
2437
2438         dsl_dir_close(dd, FTAG);
2439
2440         /* new name must be snapshot in same filesystem */
2441         tail = strchr(newname, '@');
2442         if (tail == NULL)
2443                 return (EINVAL);
2444         tail++;
2445         if (strncmp(oldname, newname, tail - newname) != 0)
2446                 return (EXDEV);
2447
2448         if (recursive) {
2449                 err = dsl_recursive_rename(oldname, newname);
2450         } else {
2451                 err = dsl_dataset_hold(oldname, FTAG, &ds);
2452                 if (err)
2453                         return (err);
2454
2455                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2456                     dsl_dataset_snapshot_rename_check,
2457                     dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2458
2459                 dsl_dataset_rele(ds, FTAG);
2460         }
2461
2462         return (err);
2463 }
2464
2465 struct promotenode {
2466         list_node_t link;
2467         dsl_dataset_t *ds;
2468 };
2469
2470 struct promotearg {
2471         list_t shared_snaps, origin_snaps, clone_snaps;
2472         dsl_dataset_t *origin_origin;
2473         uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2474         char *err_ds;
2475 };
2476
2477 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2478 static boolean_t snaplist_unstable(list_t *l);
2479
2480 static int
2481 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2482 {
2483         dsl_dataset_t *hds = arg1;
2484         struct promotearg *pa = arg2;
2485         struct promotenode *snap = list_head(&pa->shared_snaps);
2486         dsl_dataset_t *origin_ds = snap->ds;
2487         int err;
2488         uint64_t unused;
2489
2490         /* Check that it is a real clone */
2491         if (!dsl_dir_is_clone(hds->ds_dir))
2492                 return (EINVAL);
2493
2494         /* Since this is so expensive, don't do the preliminary check */
2495         if (!dmu_tx_is_syncing(tx))
2496                 return (0);
2497
2498         if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2499                 return (EXDEV);
2500
2501         /* compute origin's new unique space */
2502         snap = list_tail(&pa->clone_snaps);
2503         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2504         dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2505             origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
2506             &pa->unique, &unused, &unused);
2507
2508         /*
2509          * Walk the snapshots that we are moving
2510          *
2511          * Compute space to transfer.  Consider the incremental changes
2512          * to used for each snapshot:
2513          * (my used) = (prev's used) + (blocks born) - (blocks killed)
2514          * So each snapshot gave birth to:
2515          * (blocks born) = (my used) - (prev's used) + (blocks killed)
2516          * So a sequence would look like:
2517          * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2518          * Which simplifies to:
2519          * uN + kN + kN-1 + ... + k1 + k0
2520          * Note however, if we stop before we reach the ORIGIN we get:
2521          * uN + kN + kN-1 + ... + kM - uM-1
2522          */
2523         pa->used = origin_ds->ds_phys->ds_used_bytes;
2524         pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2525         pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2526         for (snap = list_head(&pa->shared_snaps); snap;
2527             snap = list_next(&pa->shared_snaps, snap)) {
2528                 uint64_t val, dlused, dlcomp, dluncomp;
2529                 dsl_dataset_t *ds = snap->ds;
2530
2531                 /* Check that the snapshot name does not conflict */
2532                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2533                 err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2534                 if (err == 0) {
2535                         err = EEXIST;
2536                         goto out;
2537                 }
2538                 if (err != ENOENT)
2539                         goto out;
2540
2541                 /* The very first snapshot does not have a deadlist */
2542                 if (ds->ds_phys->ds_prev_snap_obj == 0)
2543                         continue;
2544
2545                 dsl_deadlist_space(&ds->ds_deadlist,
2546                     &dlused, &dlcomp, &dluncomp);
2547                 pa->used += dlused;
2548                 pa->comp += dlcomp;
2549                 pa->uncomp += dluncomp;
2550         }
2551
2552         /*
2553          * If we are a clone of a clone then we never reached ORIGIN,
2554          * so we need to subtract out the clone origin's used space.
2555          */
2556         if (pa->origin_origin) {
2557                 pa->used -= pa->origin_origin->ds_phys->ds_used_bytes;
2558                 pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2559                 pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2560         }
2561
2562         /* Check that there is enough space here */
2563         err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2564             pa->used);
2565         if (err)
2566                 return (err);
2567
2568         /*
2569          * Compute the amounts of space that will be used by snapshots
2570          * after the promotion (for both origin and clone).  For each,
2571          * it is the amount of space that will be on all of their
2572          * deadlists (that was not born before their new origin).
2573          */
2574         if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2575                 uint64_t space;
2576
2577                 /*
2578                  * Note, typically this will not be a clone of a clone,
2579                  * so dd_origin_txg will be < TXG_INITIAL, so
2580                  * these snaplist_space() -> dsl_deadlist_space_range()
2581                  * calls will be fast because they do not have to
2582                  * iterate over all bps.
2583                  */
2584                 snap = list_head(&pa->origin_snaps);
2585                 err = snaplist_space(&pa->shared_snaps,
2586                     snap->ds->ds_dir->dd_origin_txg, &pa->cloneusedsnap);
2587                 if (err)
2588                         return (err);
2589
2590                 err = snaplist_space(&pa->clone_snaps,
2591                     snap->ds->ds_dir->dd_origin_txg, &space);
2592                 if (err)
2593                         return (err);
2594                 pa->cloneusedsnap += space;
2595         }
2596         if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2597                 err = snaplist_space(&pa->origin_snaps,
2598                     origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2599                 if (err)
2600                         return (err);
2601         }
2602
2603         return (0);
2604 out:
2605         pa->err_ds =  snap->ds->ds_snapname;
2606         return (err);
2607 }
2608
2609 static void
2610 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2611 {
2612         dsl_dataset_t *hds = arg1;
2613         struct promotearg *pa = arg2;
2614         struct promotenode *snap = list_head(&pa->shared_snaps);
2615         dsl_dataset_t *origin_ds = snap->ds;
2616         dsl_dataset_t *origin_head;
2617         dsl_dir_t *dd = hds->ds_dir;
2618         dsl_pool_t *dp = hds->ds_dir->dd_pool;
2619         dsl_dir_t *odd = NULL;
2620         uint64_t oldnext_obj;
2621         int64_t delta;
2622
2623         ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2624
2625         snap = list_head(&pa->origin_snaps);
2626         origin_head = snap->ds;
2627
2628         /*
2629          * We need to explicitly open odd, since origin_ds's dd will be
2630          * changing.
2631          */
2632         VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2633             NULL, FTAG, &odd));
2634
2635         /* change origin's next snap */
2636         dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2637         oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2638         snap = list_tail(&pa->clone_snaps);
2639         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2640         origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2641
2642         /* change the origin's next clone */
2643         if (origin_ds->ds_phys->ds_next_clones_obj) {
2644                 remove_from_next_clones(origin_ds, snap->ds->ds_object, tx);
2645                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2646                     origin_ds->ds_phys->ds_next_clones_obj,
2647                     oldnext_obj, tx));
2648         }
2649
2650         /* change origin */
2651         dmu_buf_will_dirty(dd->dd_dbuf, tx);
2652         ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2653         dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2654         dd->dd_origin_txg = origin_head->ds_dir->dd_origin_txg;
2655         dmu_buf_will_dirty(odd->dd_dbuf, tx);
2656         odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2657         origin_head->ds_dir->dd_origin_txg =
2658             origin_ds->ds_phys->ds_creation_txg;
2659
2660         /* change dd_clone entries */
2661         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2662                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2663                     odd->dd_phys->dd_clones, hds->ds_object, tx));
2664                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2665                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2666                     hds->ds_object, tx));
2667
2668                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2669                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2670                     origin_head->ds_object, tx));
2671                 if (dd->dd_phys->dd_clones == 0) {
2672                         dd->dd_phys->dd_clones = zap_create(dp->dp_meta_objset,
2673                             DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
2674                 }
2675                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2676                     dd->dd_phys->dd_clones, origin_head->ds_object, tx));
2677
2678         }
2679
2680         /* move snapshots to this dir */
2681         for (snap = list_head(&pa->shared_snaps); snap;
2682             snap = list_next(&pa->shared_snaps, snap)) {
2683                 dsl_dataset_t *ds = snap->ds;
2684
2685                 /* unregister props as dsl_dir is changing */
2686                 if (ds->ds_objset) {
2687                         dmu_objset_evict(ds->ds_objset);
2688                         ds->ds_objset = NULL;
2689                 }
2690                 /* move snap name entry */
2691                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2692                 VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2693                     ds->ds_snapname, tx));
2694                 VERIFY(0 == zap_add(dp->dp_meta_objset,
2695                     hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2696                     8, 1, &ds->ds_object, tx));
2697
2698                 /* change containing dsl_dir */
2699                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2700                 ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2701                 ds->ds_phys->ds_dir_obj = dd->dd_object;
2702                 ASSERT3P(ds->ds_dir, ==, odd);
2703                 dsl_dir_close(ds->ds_dir, ds);
2704                 VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2705                     NULL, ds, &ds->ds_dir));
2706
2707                 /* move any clone references */
2708                 if (ds->ds_phys->ds_next_clones_obj &&
2709                     spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2710                         zap_cursor_t zc;
2711                         zap_attribute_t za;
2712
2713                         for (zap_cursor_init(&zc, dp->dp_meta_objset,
2714                             ds->ds_phys->ds_next_clones_obj);
2715                             zap_cursor_retrieve(&zc, &za) == 0;
2716                             zap_cursor_advance(&zc)) {
2717                                 dsl_dataset_t *cnds;
2718                                 uint64_t o;
2719
2720                                 if (za.za_first_integer == oldnext_obj) {
2721                                         /*
2722                                          * We've already moved the
2723                                          * origin's reference.
2724                                          */
2725                                         continue;
2726                                 }
2727
2728                                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
2729                                     za.za_first_integer, FTAG, &cnds));
2730                                 o = cnds->ds_dir->dd_phys->dd_head_dataset_obj;
2731
2732                                 VERIFY3U(zap_remove_int(dp->dp_meta_objset,
2733                                     odd->dd_phys->dd_clones, o, tx), ==, 0);
2734                                 VERIFY3U(zap_add_int(dp->dp_meta_objset,
2735                                     dd->dd_phys->dd_clones, o, tx), ==, 0);
2736                                 dsl_dataset_rele(cnds, FTAG);
2737                         }
2738                         zap_cursor_fini(&zc);
2739                 }
2740
2741                 ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2742         }
2743
2744         /*
2745          * Change space accounting.
2746          * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2747          * both be valid, or both be 0 (resulting in delta == 0).  This
2748          * is true for each of {clone,origin} independently.
2749          */
2750
2751         delta = pa->cloneusedsnap -
2752             dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2753         ASSERT3S(delta, >=, 0);
2754         ASSERT3U(pa->used, >=, delta);
2755         dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2756         dsl_dir_diduse_space(dd, DD_USED_HEAD,
2757             pa->used - delta, pa->comp, pa->uncomp, tx);
2758
2759         delta = pa->originusedsnap -
2760             odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2761         ASSERT3S(delta, <=, 0);
2762         ASSERT3U(pa->used, >=, -delta);
2763         dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2764         dsl_dir_diduse_space(odd, DD_USED_HEAD,
2765             -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2766
2767         origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2768
2769         /* log history record */
2770         spa_history_log_internal(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2771             "dataset = %llu", hds->ds_object);
2772
2773         dsl_dir_close(odd, FTAG);
2774 }
2775
2776 static char *snaplist_tag = "snaplist";
2777 /*
2778  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2779  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2780  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2781  * snapshots back to this dataset's origin.
2782  */
2783 static int
2784 snaplist_make(dsl_pool_t *dp, boolean_t own,
2785     uint64_t first_obj, uint64_t last_obj, list_t *l)
2786 {
2787         uint64_t obj = last_obj;
2788
2789         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2790
2791         list_create(l, sizeof (struct promotenode),
2792             offsetof(struct promotenode, link));
2793
2794         while (obj != first_obj) {
2795                 dsl_dataset_t *ds;
2796                 struct promotenode *snap;
2797                 int err;
2798
2799                 if (own) {
2800                         err = dsl_dataset_own_obj(dp, obj,
2801                             0, snaplist_tag, &ds);
2802                         if (err == 0)
2803                                 dsl_dataset_make_exclusive(ds, snaplist_tag);
2804                 } else {
2805                         err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2806                 }
2807                 if (err == ENOENT) {
2808                         /* lost race with snapshot destroy */
2809                         struct promotenode *last = list_tail(l);
2810                         ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2811                         obj = last->ds->ds_phys->ds_prev_snap_obj;
2812                         continue;
2813                 } else if (err) {
2814                         return (err);
2815                 }
2816
2817                 if (first_obj == 0)
2818                         first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2819
2820                 snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
2821                 snap->ds = ds;
2822                 list_insert_tail(l, snap);
2823                 obj = ds->ds_phys->ds_prev_snap_obj;
2824         }
2825
2826         return (0);
2827 }
2828
2829 static int
2830 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
2831 {
2832         struct promotenode *snap;
2833
2834         *spacep = 0;
2835         for (snap = list_head(l); snap; snap = list_next(l, snap)) {
2836                 uint64_t used, comp, uncomp;
2837                 dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2838                     mintxg, UINT64_MAX, &used, &comp, &uncomp);
2839                 *spacep += used;
2840         }
2841         return (0);
2842 }
2843
2844 static void
2845 snaplist_destroy(list_t *l, boolean_t own)
2846 {
2847         struct promotenode *snap;
2848
2849         if (!l || !list_link_active(&l->list_head))
2850                 return;
2851
2852         while ((snap = list_tail(l)) != NULL) {
2853                 list_remove(l, snap);
2854                 if (own)
2855                         dsl_dataset_disown(snap->ds, snaplist_tag);
2856                 else
2857                         dsl_dataset_rele(snap->ds, snaplist_tag);
2858                 kmem_free(snap, sizeof (struct promotenode));
2859         }
2860         list_destroy(l);
2861 }
2862
2863 /*
2864  * Promote a clone.  Nomenclature note:
2865  * "clone" or "cds": the original clone which is being promoted
2866  * "origin" or "ods": the snapshot which is originally clone's origin
2867  * "origin head" or "ohds": the dataset which is the head
2868  * (filesystem/volume) for the origin
2869  * "origin origin": the origin of the origin's filesystem (typically
2870  * NULL, indicating that the clone is not a clone of a clone).
2871  */
2872 int
2873 dsl_dataset_promote(const char *name, char *conflsnap)
2874 {
2875         dsl_dataset_t *ds;
2876         dsl_dir_t *dd;
2877         dsl_pool_t *dp;
2878         dmu_object_info_t doi;
2879         struct promotearg pa = { 0 };
2880         struct promotenode *snap;
2881         int err;
2882
2883         err = dsl_dataset_hold(name, FTAG, &ds);
2884         if (err)
2885                 return (err);
2886         dd = ds->ds_dir;
2887         dp = dd->dd_pool;
2888
2889         err = dmu_object_info(dp->dp_meta_objset,
2890             ds->ds_phys->ds_snapnames_zapobj, &doi);
2891         if (err) {
2892                 dsl_dataset_rele(ds, FTAG);
2893                 return (err);
2894         }
2895
2896         if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
2897                 dsl_dataset_rele(ds, FTAG);
2898                 return (EINVAL);
2899         }
2900
2901         /*
2902          * We are going to inherit all the snapshots taken before our
2903          * origin (i.e., our new origin will be our parent's origin).
2904          * Take ownership of them so that we can rename them into our
2905          * namespace.
2906          */
2907         rw_enter(&dp->dp_config_rwlock, RW_READER);
2908
2909         err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
2910             &pa.shared_snaps);
2911         if (err != 0)
2912                 goto out;
2913
2914         err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
2915         if (err != 0)
2916                 goto out;
2917
2918         snap = list_head(&pa.shared_snaps);
2919         ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
2920         err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
2921             snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
2922         if (err != 0)
2923                 goto out;
2924
2925         if (snap->ds->ds_dir->dd_phys->dd_origin_obj != 0) {
2926                 err = dsl_dataset_hold_obj(dp,
2927                     snap->ds->ds_dir->dd_phys->dd_origin_obj,
2928                     FTAG, &pa.origin_origin);
2929                 if (err != 0)
2930                         goto out;
2931         }
2932
2933 out:
2934         rw_exit(&dp->dp_config_rwlock);
2935
2936         /*
2937          * Add in 128x the snapnames zapobj size, since we will be moving
2938          * a bunch of snapnames to the promoted ds, and dirtying their
2939          * bonus buffers.
2940          */
2941         if (err == 0) {
2942                 err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
2943                     dsl_dataset_promote_sync, ds, &pa,
2944                     2 + 2 * doi.doi_physical_blocks_512);
2945                 if (err && pa.err_ds && conflsnap)
2946                         (void) strncpy(conflsnap, pa.err_ds, MAXNAMELEN);
2947         }
2948
2949         snaplist_destroy(&pa.shared_snaps, B_TRUE);
2950         snaplist_destroy(&pa.clone_snaps, B_FALSE);
2951         snaplist_destroy(&pa.origin_snaps, B_FALSE);
2952         if (pa.origin_origin)
2953                 dsl_dataset_rele(pa.origin_origin, FTAG);
2954         dsl_dataset_rele(ds, FTAG);
2955         return (err);
2956 }
2957
2958 struct cloneswaparg {
2959         dsl_dataset_t *cds; /* clone dataset */
2960         dsl_dataset_t *ohds; /* origin's head dataset */
2961         boolean_t force;
2962         int64_t unused_refres_delta; /* change in unconsumed refreservation */
2963 };
2964
2965 /* ARGSUSED */
2966 static int
2967 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
2968 {
2969         struct cloneswaparg *csa = arg1;
2970
2971         /* they should both be heads */
2972         if (dsl_dataset_is_snapshot(csa->cds) ||
2973             dsl_dataset_is_snapshot(csa->ohds))
2974                 return (EINVAL);
2975
2976         /* the branch point should be just before them */
2977         if (csa->cds->ds_prev != csa->ohds->ds_prev)
2978                 return (EINVAL);
2979
2980         /* cds should be the clone (unless they are unrelated) */
2981         if (csa->cds->ds_prev != NULL &&
2982             csa->cds->ds_prev != csa->cds->ds_dir->dd_pool->dp_origin_snap &&
2983             csa->ohds->ds_object !=
2984             csa->cds->ds_prev->ds_phys->ds_next_snap_obj)
2985                 return (EINVAL);
2986
2987         /* the clone should be a child of the origin */
2988         if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
2989                 return (EINVAL);
2990
2991         /* ohds shouldn't be modified unless 'force' */
2992         if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
2993                 return (ETXTBSY);
2994
2995         /* adjust amount of any unconsumed refreservation */
2996         csa->unused_refres_delta =
2997             (int64_t)MIN(csa->ohds->ds_reserved,
2998             csa->ohds->ds_phys->ds_unique_bytes) -
2999             (int64_t)MIN(csa->ohds->ds_reserved,
3000             csa->cds->ds_phys->ds_unique_bytes);
3001
3002         if (csa->unused_refres_delta > 0 &&
3003             csa->unused_refres_delta >
3004             dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3005                 return (ENOSPC);
3006
3007         if (csa->ohds->ds_quota != 0 &&
3008             csa->cds->ds_phys->ds_unique_bytes > csa->ohds->ds_quota)
3009                 return (EDQUOT);
3010
3011         return (0);
3012 }
3013
3014 /* ARGSUSED */
3015 static void
3016 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3017 {
3018         struct cloneswaparg *csa = arg1;
3019         dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3020
3021         ASSERT(csa->cds->ds_reserved == 0);
3022         ASSERT(csa->ohds->ds_quota == 0 ||
3023             csa->cds->ds_phys->ds_unique_bytes <= csa->ohds->ds_quota);
3024
3025         dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3026         dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3027
3028         if (csa->cds->ds_objset != NULL) {
3029                 dmu_objset_evict(csa->cds->ds_objset);
3030                 csa->cds->ds_objset = NULL;
3031         }
3032
3033         if (csa->ohds->ds_objset != NULL) {
3034                 dmu_objset_evict(csa->ohds->ds_objset);
3035                 csa->ohds->ds_objset = NULL;
3036         }
3037
3038         /*
3039          * Reset origin's unique bytes, if it exists.
3040          */
3041         if (csa->cds->ds_prev) {
3042                 dsl_dataset_t *origin = csa->cds->ds_prev;
3043                 uint64_t comp, uncomp;
3044
3045                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
3046                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3047                     origin->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3048                     &origin->ds_phys->ds_unique_bytes, &comp, &uncomp);
3049         }
3050
3051         /* swap blkptrs */
3052         {
3053                 blkptr_t tmp;
3054                 tmp = csa->ohds->ds_phys->ds_bp;
3055                 csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3056                 csa->cds->ds_phys->ds_bp = tmp;
3057         }
3058
3059         /* set dd_*_bytes */
3060         {
3061                 int64_t dused, dcomp, duncomp;
3062                 uint64_t cdl_used, cdl_comp, cdl_uncomp;
3063                 uint64_t odl_used, odl_comp, odl_uncomp;
3064
3065                 ASSERT3U(csa->cds->ds_dir->dd_phys->
3066                     dd_used_breakdown[DD_USED_SNAP], ==, 0);
3067
3068                 dsl_deadlist_space(&csa->cds->ds_deadlist,
3069                     &cdl_used, &cdl_comp, &cdl_uncomp);
3070                 dsl_deadlist_space(&csa->ohds->ds_deadlist,
3071                     &odl_used, &odl_comp, &odl_uncomp);
3072
3073                 dused = csa->cds->ds_phys->ds_used_bytes + cdl_used -
3074                     (csa->ohds->ds_phys->ds_used_bytes + odl_used);
3075                 dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3076                     (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3077                 duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3078                     cdl_uncomp -
3079                     (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3080
3081                 dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3082                     dused, dcomp, duncomp, tx);
3083                 dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3084                     -dused, -dcomp, -duncomp, tx);
3085
3086                 /*
3087                  * The difference in the space used by snapshots is the
3088                  * difference in snapshot space due to the head's
3089                  * deadlist (since that's the only thing that's
3090                  * changing that affects the snapused).
3091                  */
3092                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3093                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3094                     &cdl_used, &cdl_comp, &cdl_uncomp);
3095                 dsl_deadlist_space_range(&csa->ohds->ds_deadlist,
3096                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3097                     &odl_used, &odl_comp, &odl_uncomp);
3098                 dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3099                     DD_USED_HEAD, DD_USED_SNAP, tx);
3100         }
3101
3102         /* swap ds_*_bytes */
3103         SWITCH64(csa->ohds->ds_phys->ds_used_bytes,
3104             csa->cds->ds_phys->ds_used_bytes);
3105         SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3106             csa->cds->ds_phys->ds_compressed_bytes);
3107         SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3108             csa->cds->ds_phys->ds_uncompressed_bytes);
3109         SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3110             csa->cds->ds_phys->ds_unique_bytes);
3111
3112         /* apply any parent delta for change in unconsumed refreservation */
3113         dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3114             csa->unused_refres_delta, 0, 0, tx);
3115
3116         /*
3117          * Swap deadlists.
3118          */
3119         dsl_deadlist_close(&csa->cds->ds_deadlist);
3120         dsl_deadlist_close(&csa->ohds->ds_deadlist);
3121         SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3122             csa->cds->ds_phys->ds_deadlist_obj);
3123         dsl_deadlist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3124             csa->cds->ds_phys->ds_deadlist_obj);
3125         dsl_deadlist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3126             csa->ohds->ds_phys->ds_deadlist_obj);
3127
3128         dsl_scan_ds_clone_swapped(csa->ohds, csa->cds, tx);
3129 }
3130
3131 /*
3132  * Swap 'clone' with its origin head datasets.  Used at the end of "zfs
3133  * recv" into an existing fs to swizzle the file system to the new
3134  * version, and by "zfs rollback".  Can also be used to swap two
3135  * independent head datasets if neither has any snapshots.
3136  */
3137 int
3138 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3139     boolean_t force)
3140 {
3141         struct cloneswaparg csa;
3142         int error;
3143
3144         ASSERT(clone->ds_owner);
3145         ASSERT(origin_head->ds_owner);
3146 retry:
3147         /* Need exclusive access for the swap */
3148         rw_enter(&clone->ds_rwlock, RW_WRITER);
3149         if (!rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3150                 rw_exit(&clone->ds_rwlock);
3151                 rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3152                 if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3153                         rw_exit(&origin_head->ds_rwlock);
3154                         goto retry;
3155                 }
3156         }
3157         csa.cds = clone;
3158         csa.ohds = origin_head;
3159         csa.force = force;
3160         error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3161             dsl_dataset_clone_swap_check,
3162             dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3163         return (error);
3164 }
3165
3166 /*
3167  * Given a pool name and a dataset object number in that pool,
3168  * return the name of that dataset.
3169  */
3170 int
3171 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3172 {
3173         spa_t *spa;
3174         dsl_pool_t *dp;
3175         dsl_dataset_t *ds;
3176         int error;
3177
3178         if ((error = spa_open(pname, &spa, FTAG)) != 0)
3179                 return (error);
3180         dp = spa_get_dsl(spa);
3181         rw_enter(&dp->dp_config_rwlock, RW_READER);
3182         if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3183                 dsl_dataset_name(ds, buf);
3184                 dsl_dataset_rele(ds, FTAG);
3185         }
3186         rw_exit(&dp->dp_config_rwlock);
3187         spa_close(spa, FTAG);
3188
3189         return (error);
3190 }
3191
3192 int
3193 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3194     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3195 {
3196         int error = 0;
3197
3198         ASSERT3S(asize, >, 0);
3199
3200         /*
3201          * *ref_rsrv is the portion of asize that will come from any
3202          * unconsumed refreservation space.
3203          */
3204         *ref_rsrv = 0;
3205
3206         mutex_enter(&ds->ds_lock);
3207         /*
3208          * Make a space adjustment for reserved bytes.
3209          */
3210         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3211                 ASSERT3U(*used, >=,
3212                     ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3213                 *used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3214                 *ref_rsrv =
3215                     asize - MIN(asize, parent_delta(ds, asize + inflight));
3216         }
3217
3218         if (!check_quota || ds->ds_quota == 0) {
3219                 mutex_exit(&ds->ds_lock);
3220                 return (0);
3221         }
3222         /*
3223          * If they are requesting more space, and our current estimate
3224          * is over quota, they get to try again unless the actual
3225          * on-disk is over quota and there are no pending changes (which
3226          * may free up space for us).
3227          */
3228         if (ds->ds_phys->ds_used_bytes + inflight >= ds->ds_quota) {
3229                 if (inflight > 0 || ds->ds_phys->ds_used_bytes < ds->ds_quota)
3230                         error = ERESTART;
3231                 else
3232                         error = EDQUOT;
3233         }
3234         mutex_exit(&ds->ds_lock);
3235
3236         return (error);
3237 }
3238
3239 /* ARGSUSED */
3240 static int
3241 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3242 {
3243         dsl_dataset_t *ds = arg1;
3244         dsl_prop_setarg_t *psa = arg2;
3245         int err;
3246
3247         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3248                 return (ENOTSUP);
3249
3250         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3251                 return (err);
3252
3253         if (psa->psa_effective_value == 0)
3254                 return (0);
3255
3256         if (psa->psa_effective_value < ds->ds_phys->ds_used_bytes ||
3257             psa->psa_effective_value < ds->ds_reserved)
3258                 return (ENOSPC);
3259
3260         return (0);
3261 }
3262
3263 extern void dsl_prop_set_sync(void *, void *, dmu_tx_t *);
3264
3265 void
3266 dsl_dataset_set_quota_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3267 {
3268         dsl_dataset_t *ds = arg1;
3269         dsl_prop_setarg_t *psa = arg2;
3270         uint64_t effective_value = psa->psa_effective_value;
3271
3272         dsl_prop_set_sync(ds, psa, tx);
3273         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3274
3275         if (ds->ds_quota != effective_value) {
3276                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3277                 ds->ds_quota = effective_value;
3278
3279                 spa_history_log_internal(LOG_DS_REFQUOTA,
3280                     ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu ",
3281                     (longlong_t)ds->ds_quota, ds->ds_object);
3282         }
3283 }
3284
3285 int
3286 dsl_dataset_set_quota(const char *dsname, zprop_source_t source, uint64_t quota)
3287 {
3288         dsl_dataset_t *ds;
3289         dsl_prop_setarg_t psa;
3290         int err;
3291
3292         dsl_prop_setarg_init_uint64(&psa, "refquota", source, &quota);
3293
3294         err = dsl_dataset_hold(dsname, FTAG, &ds);
3295         if (err)
3296                 return (err);
3297
3298         /*
3299          * If someone removes a file, then tries to set the quota, we
3300          * want to make sure the file freeing takes effect.
3301          */
3302         txg_wait_open(ds->ds_dir->dd_pool, 0);
3303
3304         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3305             dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3306             ds, &psa, 0);
3307
3308         dsl_dataset_rele(ds, FTAG);
3309         return (err);
3310 }
3311
3312 static int
3313 dsl_dataset_set_reservation_check(void *arg1, void *arg2, dmu_tx_t *tx)
3314 {
3315         dsl_dataset_t *ds = arg1;
3316         dsl_prop_setarg_t *psa = arg2;
3317         uint64_t effective_value;
3318         uint64_t unique;
3319         int err;
3320
3321         if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3322             SPA_VERSION_REFRESERVATION)
3323                 return (ENOTSUP);
3324
3325         if (dsl_dataset_is_snapshot(ds))
3326                 return (EINVAL);
3327
3328         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3329                 return (err);
3330
3331         effective_value = psa->psa_effective_value;
3332
3333         /*
3334          * If we are doing the preliminary check in open context, the
3335          * space estimates may be inaccurate.
3336          */
3337         if (!dmu_tx_is_syncing(tx))
3338                 return (0);
3339
3340         mutex_enter(&ds->ds_lock);
3341         if (!DS_UNIQUE_IS_ACCURATE(ds))
3342                 dsl_dataset_recalc_head_uniq(ds);
3343         unique = ds->ds_phys->ds_unique_bytes;
3344         mutex_exit(&ds->ds_lock);
3345
3346         if (MAX(unique, effective_value) > MAX(unique, ds->ds_reserved)) {
3347                 uint64_t delta = MAX(unique, effective_value) -
3348                     MAX(unique, ds->ds_reserved);
3349
3350                 if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3351                         return (ENOSPC);
3352                 if (ds->ds_quota > 0 &&
3353                     effective_value > ds->ds_quota)
3354                         return (ENOSPC);
3355         }
3356
3357         return (0);
3358 }
3359
3360 static void
3361 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3362 {
3363         dsl_dataset_t *ds = arg1;
3364         dsl_prop_setarg_t *psa = arg2;
3365         uint64_t effective_value = psa->psa_effective_value;
3366         uint64_t unique;
3367         int64_t delta;
3368
3369         dsl_prop_set_sync(ds, psa, tx);
3370         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3371
3372         dmu_buf_will_dirty(ds->ds_dbuf, tx);
3373
3374         mutex_enter(&ds->ds_dir->dd_lock);
3375         mutex_enter(&ds->ds_lock);
3376         ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
3377         unique = ds->ds_phys->ds_unique_bytes;
3378         delta = MAX(0, (int64_t)(effective_value - unique)) -
3379             MAX(0, (int64_t)(ds->ds_reserved - unique));
3380         ds->ds_reserved = effective_value;
3381         mutex_exit(&ds->ds_lock);
3382
3383         dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3384         mutex_exit(&ds->ds_dir->dd_lock);
3385
3386         spa_history_log_internal(LOG_DS_REFRESERV,
3387             ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu",
3388             (longlong_t)effective_value, ds->ds_object);
3389 }
3390
3391 int
3392 dsl_dataset_set_reservation(const char *dsname, zprop_source_t source,
3393     uint64_t reservation)
3394 {
3395         dsl_dataset_t *ds;
3396         dsl_prop_setarg_t psa;
3397         int err;
3398
3399         dsl_prop_setarg_init_uint64(&psa, "refreservation", source,
3400             &reservation);
3401
3402         err = dsl_dataset_hold(dsname, FTAG, &ds);
3403         if (err)
3404                 return (err);
3405
3406         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3407             dsl_dataset_set_reservation_check,
3408             dsl_dataset_set_reservation_sync, ds, &psa, 0);
3409
3410         dsl_dataset_rele(ds, FTAG);
3411         return (err);
3412 }
3413
3414 struct dsl_ds_holdarg {
3415         dsl_sync_task_group_t *dstg;
3416         char *htag;
3417         char *snapname;
3418         boolean_t recursive;
3419         boolean_t gotone;
3420         boolean_t temphold;
3421         char failed[MAXPATHLEN];
3422 };
3423
3424 /*
3425  * The max length of a temporary tag prefix is the number of hex digits
3426  * required to express UINT64_MAX plus one for the hyphen.
3427  */
3428 #define MAX_TAG_PREFIX_LEN      17
3429
3430 static int
3431 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3432 {
3433         dsl_dataset_t *ds = arg1;
3434         struct dsl_ds_holdarg *ha = arg2;
3435         char *htag = ha->htag;
3436         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3437         int error = 0;
3438
3439         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3440                 return (ENOTSUP);
3441
3442         if (!dsl_dataset_is_snapshot(ds))
3443                 return (EINVAL);
3444
3445         /* tags must be unique */
3446         mutex_enter(&ds->ds_lock);
3447         if (ds->ds_phys->ds_userrefs_obj) {
3448                 error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3449                     8, 1, tx);
3450                 if (error == 0)
3451                         error = EEXIST;
3452                 else if (error == ENOENT)
3453                         error = 0;
3454         }
3455         mutex_exit(&ds->ds_lock);
3456
3457         if (error == 0 && ha->temphold &&
3458             strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
3459                 error = E2BIG;
3460
3461         return (error);
3462 }
3463
3464 static void
3465 dsl_dataset_user_hold_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3466 {
3467         dsl_dataset_t *ds = arg1;
3468         struct dsl_ds_holdarg *ha = arg2;
3469         char *htag = ha->htag;
3470         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3471         objset_t *mos = dp->dp_meta_objset;
3472         uint64_t now = gethrestime_sec();
3473         uint64_t zapobj;
3474
3475         mutex_enter(&ds->ds_lock);
3476         if (ds->ds_phys->ds_userrefs_obj == 0) {
3477                 /*
3478                  * This is the first user hold for this dataset.  Create
3479                  * the userrefs zap object.
3480                  */
3481                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3482                 zapobj = ds->ds_phys->ds_userrefs_obj =
3483                     zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3484         } else {
3485                 zapobj = ds->ds_phys->ds_userrefs_obj;
3486         }
3487         ds->ds_userrefs++;
3488         mutex_exit(&ds->ds_lock);
3489
3490         VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3491
3492         if (ha->temphold) {
3493                 VERIFY(0 == dsl_pool_user_hold(dp, ds->ds_object,
3494                     htag, &now, tx));
3495         }
3496
3497         spa_history_log_internal(LOG_DS_USER_HOLD,
3498             dp->dp_spa, tx, "<%s> temp = %d dataset = %llu", htag,
3499             (int)ha->temphold, ds->ds_object);
3500 }
3501
3502 static int
3503 dsl_dataset_user_hold_one(const char *dsname, void *arg)
3504 {
3505         struct dsl_ds_holdarg *ha = arg;
3506         dsl_dataset_t *ds;
3507         int error;
3508         char *name;
3509
3510         /* alloc a buffer to hold dsname@snapname plus terminating NULL */
3511         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3512         error = dsl_dataset_hold(name, ha->dstg, &ds);
3513         strfree(name);
3514         if (error == 0) {
3515                 ha->gotone = B_TRUE;
3516                 dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3517                     dsl_dataset_user_hold_sync, ds, ha, 0);
3518         } else if (error == ENOENT && ha->recursive) {
3519                 error = 0;
3520         } else {
3521                 (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3522         }
3523         return (error);
3524 }
3525
3526 int
3527 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3528     boolean_t recursive, boolean_t temphold)
3529 {
3530         struct dsl_ds_holdarg *ha;
3531         dsl_sync_task_t *dst;
3532         spa_t *spa;
3533         int error;
3534
3535         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3536
3537         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3538
3539         error = spa_open(dsname, &spa, FTAG);
3540         if (error) {
3541                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3542                 return (error);
3543         }
3544
3545         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3546         ha->htag = htag;
3547         ha->snapname = snapname;
3548         ha->recursive = recursive;
3549         ha->temphold = temphold;
3550         if (recursive) {
3551                 error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3552                     ha, DS_FIND_CHILDREN);
3553         } else {
3554                 error = dsl_dataset_user_hold_one(dsname, ha);
3555         }
3556         if (error == 0)
3557                 error = dsl_sync_task_group_wait(ha->dstg);
3558
3559         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3560             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3561                 dsl_dataset_t *ds = dst->dst_arg1;
3562
3563                 if (dst->dst_err) {
3564                         dsl_dataset_name(ds, ha->failed);
3565                         *strchr(ha->failed, '@') = '\0';
3566                 }
3567                 dsl_dataset_rele(ds, ha->dstg);
3568         }
3569
3570         if (error == 0 && recursive && !ha->gotone)
3571                 error = ENOENT;
3572
3573         if (error)
3574                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3575
3576         dsl_sync_task_group_destroy(ha->dstg);
3577         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3578         spa_close(spa, FTAG);
3579         return (error);
3580 }
3581
3582 struct dsl_ds_releasearg {
3583         dsl_dataset_t *ds;
3584         const char *htag;
3585         boolean_t own;          /* do we own or just hold ds? */
3586 };
3587
3588 static int
3589 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3590     boolean_t *might_destroy)
3591 {
3592         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3593         uint64_t zapobj;
3594         uint64_t tmp;
3595         int error;
3596
3597         *might_destroy = B_FALSE;
3598
3599         mutex_enter(&ds->ds_lock);
3600         zapobj = ds->ds_phys->ds_userrefs_obj;
3601         if (zapobj == 0) {
3602                 /* The tag can't possibly exist */
3603                 mutex_exit(&ds->ds_lock);
3604                 return (ESRCH);
3605         }
3606
3607         /* Make sure the tag exists */
3608         error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3609         if (error) {
3610                 mutex_exit(&ds->ds_lock);
3611                 if (error == ENOENT)
3612                         error = ESRCH;
3613                 return (error);
3614         }
3615
3616         if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3617             DS_IS_DEFER_DESTROY(ds))
3618                 *might_destroy = B_TRUE;
3619
3620         mutex_exit(&ds->ds_lock);
3621         return (0);
3622 }
3623
3624 static int
3625 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3626 {
3627         struct dsl_ds_releasearg *ra = arg1;
3628         dsl_dataset_t *ds = ra->ds;
3629         boolean_t might_destroy;
3630         int error;
3631
3632         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3633                 return (ENOTSUP);
3634
3635         error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3636         if (error)
3637                 return (error);
3638
3639         if (might_destroy) {
3640                 struct dsl_ds_destroyarg dsda = {0};
3641
3642                 if (dmu_tx_is_syncing(tx)) {
3643                         /*
3644                          * If we're not prepared to remove the snapshot,
3645                          * we can't allow the release to happen right now.
3646                          */
3647                         if (!ra->own)
3648                                 return (EBUSY);
3649                 }
3650                 dsda.ds = ds;
3651                 dsda.releasing = B_TRUE;
3652                 return (dsl_dataset_destroy_check(&dsda, tag, tx));
3653         }
3654
3655         return (0);
3656 }
3657
3658 static void
3659 dsl_dataset_user_release_sync(void *arg1, void *tag, dmu_tx_t *tx)
3660 {
3661         struct dsl_ds_releasearg *ra = arg1;
3662         dsl_dataset_t *ds = ra->ds;
3663         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3664         objset_t *mos = dp->dp_meta_objset;
3665         uint64_t zapobj;
3666         uint64_t dsobj = ds->ds_object;
3667         uint64_t refs;
3668         int error;
3669
3670         if (ds->ds_objset) {
3671                 dmu_objset_evict(ds->ds_objset);
3672                 ds->ds_objset = NULL;
3673         }
3674
3675         mutex_enter(&ds->ds_lock);
3676         ds->ds_userrefs--;
3677         refs = ds->ds_userrefs;
3678         mutex_exit(&ds->ds_lock);
3679         error = dsl_pool_user_release(dp, ds->ds_object, ra->htag, tx);
3680         VERIFY(error == 0 || error == ENOENT);
3681         zapobj = ds->ds_phys->ds_userrefs_obj;
3682         VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3683         if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3684             DS_IS_DEFER_DESTROY(ds)) {
3685                 struct dsl_ds_destroyarg dsda = {0};
3686
3687                 ASSERT(ra->own);
3688                 dsda.ds = ds;
3689                 dsda.releasing = B_TRUE;
3690                 /* We already did the destroy_check */
3691                 dsl_dataset_destroy_sync(&dsda, tag, tx);
3692         }
3693
3694         spa_history_log_internal(LOG_DS_USER_RELEASE,
3695             dp->dp_spa, tx, "<%s> %lld dataset = %llu",
3696             ra->htag, (longlong_t)refs, dsobj);
3697 }
3698
3699 static int
3700 dsl_dataset_user_release_one(const char *dsname, void *arg)
3701 {
3702         struct dsl_ds_holdarg *ha = arg;
3703         struct dsl_ds_releasearg *ra;
3704         dsl_dataset_t *ds;
3705         int error;
3706         void *dtag = ha->dstg;
3707         char *name;
3708         boolean_t own = B_FALSE;
3709         boolean_t might_destroy;
3710
3711         /* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3712         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3713         error = dsl_dataset_hold(name, dtag, &ds);
3714         strfree(name);
3715         if (error == ENOENT && ha->recursive)
3716                 return (0);
3717         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3718         if (error)
3719                 return (error);
3720
3721         ha->gotone = B_TRUE;
3722
3723         ASSERT(dsl_dataset_is_snapshot(ds));
3724
3725         error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3726         if (error) {
3727                 dsl_dataset_rele(ds, dtag);
3728                 return (error);
3729         }
3730
3731         if (might_destroy) {
3732 #ifdef _KERNEL
3733                 name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3734                 error = zfs_unmount_snap(name, NULL);
3735                 strfree(name);
3736                 if (error) {
3737                         dsl_dataset_rele(ds, dtag);
3738                         return (error);
3739                 }
3740 #endif
3741                 if (!dsl_dataset_tryown(ds, B_TRUE, dtag)) {
3742                         dsl_dataset_rele(ds, dtag);
3743                         return (EBUSY);
3744                 } else {
3745                         own = B_TRUE;
3746                         dsl_dataset_make_exclusive(ds, dtag);
3747                 }
3748         }
3749
3750         ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3751         ra->ds = ds;
3752         ra->htag = ha->htag;
3753         ra->own = own;
3754         dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3755             dsl_dataset_user_release_sync, ra, dtag, 0);
3756
3757         return (0);
3758 }
3759
3760 int
3761 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3762     boolean_t recursive)
3763 {
3764         struct dsl_ds_holdarg *ha;
3765         dsl_sync_task_t *dst;
3766         spa_t *spa;
3767         int error;
3768
3769 top:
3770         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3771
3772         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3773
3774         error = spa_open(dsname, &spa, FTAG);
3775         if (error) {
3776                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3777                 return (error);
3778         }
3779
3780         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3781         ha->htag = htag;
3782         ha->snapname = snapname;
3783         ha->recursive = recursive;
3784         if (recursive) {
3785                 error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
3786                     ha, DS_FIND_CHILDREN);
3787         } else {
3788                 error = dsl_dataset_user_release_one(dsname, ha);
3789         }
3790         if (error == 0)
3791                 error = dsl_sync_task_group_wait(ha->dstg);
3792
3793         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3794             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3795                 struct dsl_ds_releasearg *ra = dst->dst_arg1;
3796                 dsl_dataset_t *ds = ra->ds;
3797
3798                 if (dst->dst_err)
3799                         dsl_dataset_name(ds, ha->failed);
3800
3801                 if (ra->own)
3802                         dsl_dataset_disown(ds, ha->dstg);
3803                 else
3804                         dsl_dataset_rele(ds, ha->dstg);
3805
3806                 kmem_free(ra, sizeof (struct dsl_ds_releasearg));
3807         }
3808
3809         if (error == 0 && recursive && !ha->gotone)
3810                 error = ENOENT;
3811
3812         if (error && error != EBUSY)
3813                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3814
3815         dsl_sync_task_group_destroy(ha->dstg);
3816         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3817         spa_close(spa, FTAG);
3818
3819         /*
3820          * We can get EBUSY if we were racing with deferred destroy and
3821          * dsl_dataset_user_release_check() hadn't done the necessary
3822          * open context setup.  We can also get EBUSY if we're racing
3823          * with destroy and that thread is the ds_owner.  Either way
3824          * the busy condition should be transient, and we should retry
3825          * the release operation.
3826          */
3827         if (error == EBUSY)
3828                 goto top;
3829
3830         return (error);
3831 }
3832
3833 /*
3834  * Called at spa_load time to release a stale temporary user hold.
3835  */
3836 int
3837 dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, char *htag)
3838 {
3839         dsl_dataset_t *ds;
3840         char *snap;
3841         char *name;
3842         int namelen;
3843         int error;
3844
3845         rw_enter(&dp->dp_config_rwlock, RW_READER);
3846         error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
3847         rw_exit(&dp->dp_config_rwlock);
3848         if (error)
3849                 return (error);
3850         namelen = dsl_dataset_namelen(ds)+1;
3851         name = kmem_alloc(namelen, KM_SLEEP);
3852         dsl_dataset_name(ds, name);
3853         dsl_dataset_rele(ds, FTAG);
3854
3855         snap = strchr(name, '@');
3856         *snap = '\0';
3857         ++snap;
3858         return (dsl_dataset_user_release(name, snap, htag, B_FALSE));
3859 }
3860
3861 int
3862 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
3863 {
3864         dsl_dataset_t *ds;
3865         int err;
3866
3867         err = dsl_dataset_hold(dsname, FTAG, &ds);
3868         if (err)
3869                 return (err);
3870
3871         VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
3872         if (ds->ds_phys->ds_userrefs_obj != 0) {
3873                 zap_attribute_t *za;
3874                 zap_cursor_t zc;
3875
3876                 za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
3877                 for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
3878                     ds->ds_phys->ds_userrefs_obj);
3879                     zap_cursor_retrieve(&zc, za) == 0;
3880                     zap_cursor_advance(&zc)) {
3881                         VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
3882                             za->za_first_integer));
3883                 }
3884                 zap_cursor_fini(&zc);
3885                 kmem_free(za, sizeof (zap_attribute_t));
3886         }
3887         dsl_dataset_rele(ds, FTAG);
3888         return (0);
3889 }
3890
3891 /*
3892  * Note, this fuction is used as the callback for dmu_objset_find().  We
3893  * always return 0 so that we will continue to find and process
3894  * inconsistent datasets, even if we encounter an error trying to
3895  * process one of them.
3896  */
3897 /* ARGSUSED */
3898 int
3899 dsl_destroy_inconsistent(const char *dsname, void *arg)
3900 {
3901         dsl_dataset_t *ds;
3902
3903         if (dsl_dataset_own(dsname, B_TRUE, FTAG, &ds) == 0) {
3904                 if (DS_IS_INCONSISTENT(ds))
3905                         (void) dsl_dataset_destroy(ds, FTAG, B_FALSE);
3906                 else
3907                         dsl_dataset_disown(ds, FTAG);
3908         }
3909         return (0);
3910 }