sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_dataset.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  */
  24
  25 #include <sys/dmu_objset.h>
  26 #include <sys/dsl_dataset.h>
  27 #include <sys/dsl_dir.h>
  28 #include <sys/dsl_prop.h>
  29 #include <sys/dsl_synctask.h>
  30 #include <sys/dmu_traverse.h>
  31 #include <sys/dmu_tx.h>
  32 #include <sys/arc.h>
  33 #include <sys/zio.h>
  34 #include <sys/zap.h>
  35 #include <sys/unique.h>
  36 #include <sys/zfs_context.h>
  37 #include <sys/zfs_ioctl.h>
  38 #include <sys/spa.h>
  39 #include <sys/zfs_znode.h>
  40 #include <sys/zfs_onexit.h>
  41 #include <sys/zvol.h>
  42 #include <sys/dsl_scan.h>
  43 #include <sys/dsl_deadlist.h>
  44
  45 static char *dsl_reaper = "the grim reaper";
  46
  47 static dsl_checkfunc_t dsl_dataset_destroy_begin_check;
  48 static dsl_syncfunc_t dsl_dataset_destroy_begin_sync;
  49 static dsl_syncfunc_t dsl_dataset_set_reservation_sync;
  50
  51 #define SWITCH64(x, y) \
  52         { \
  53                 uint64_t __tmp = (x); \
  54                 (x) = (y); \
  55                 (y) = __tmp; \
  56         }
  57
  58 #define DS_REF_MAX      (1ULL << 62)
  59
  60 #define DSL_DEADLIST_BLOCKSIZE  SPA_MAXBLOCKSIZE
  61
  62 #define DSL_DATASET_IS_DESTROYED(ds)    ((ds)->ds_owner == dsl_reaper)
  63
  64
  65 /*
  66  * Figure out how much of this delta should be propogated to the dsl_dir
  67  * layer.  If there's a refreservation, that space has already been
  68  * partially accounted for in our ancestors.
  69  */
  70 static int64_t
  71 parent_delta(dsl_dataset_t *ds, int64_t delta)
  72 {
  73         uint64_t old_bytes, new_bytes;
  74
  75         if (ds->ds_reserved == 0)
  76                 return (delta);
  77
  78         old_bytes = MAX(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
  79         new_bytes = MAX(ds->ds_phys->ds_unique_bytes + delta, ds->ds_reserved);
  80
  81         ASSERT3U(ABS((int64_t)(new_bytes - old_bytes)), <=, ABS(delta));
  82         return (new_bytes - old_bytes);
  83 }
  84
  85 void
  86 dsl_dataset_block_born(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
  87 {
  88         int used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
  89         int compressed = BP_GET_PSIZE(bp);
  90         int uncompressed = BP_GET_UCSIZE(bp);
  91         int64_t delta;
  92
  93         dprintf_bp(bp, "ds=%p", ds);
  94
  95         ASSERT(dmu_tx_is_syncing(tx));
  96         /* It could have been compressed away to nothing */
  97         if (BP_IS_HOLE(bp))
  98                 return;
  99         ASSERT(BP_GET_TYPE(bp) != DMU_OT_NONE);
 100         ASSERT3U(BP_GET_TYPE(bp), <, DMU_OT_NUMTYPES);
 101         if (ds == NULL) {
 102                 /*
 103                  * Account for the meta-objset space in its placeholder
 104                  * dsl_dir.
 105                  */
 106                 ASSERT3U(compressed, ==, uncompressed); /* it's all metadata */
 107                 dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
 108                     used, compressed, uncompressed, tx);
 109                 dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
 110                 return;
 111         }
 112         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 113
 114         mutex_enter(&ds->ds_dir->dd_lock);
 115         mutex_enter(&ds->ds_lock);
 116         delta = parent_delta(ds, used);
 117         ds->ds_phys->ds_used_bytes += used;
 118         ds->ds_phys->ds_compressed_bytes += compressed;
 119         ds->ds_phys->ds_uncompressed_bytes += uncompressed;
 120         ds->ds_phys->ds_unique_bytes += used;
 121         mutex_exit(&ds->ds_lock);
 122         dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD, delta,
 123             compressed, uncompressed, tx);
 124         dsl_dir_transfer_space(ds->ds_dir, used - delta,
 125             DD_USED_REFRSRV, DD_USED_HEAD, tx);
 126         mutex_exit(&ds->ds_dir->dd_lock);
 127 }
 128
 129 int
 130 dsl_dataset_block_kill(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx,
 131     boolean_t async)
 132 {
 133         if (BP_IS_HOLE(bp))
 134                 return (0);
 135
 136         ASSERT(dmu_tx_is_syncing(tx));
 137         ASSERT(bp->blk_birth <= tx->tx_txg);
 138
 139         int used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
 140         int compressed = BP_GET_PSIZE(bp);
 141         int uncompressed = BP_GET_UCSIZE(bp);
 142
 143         ASSERT(used > 0);
 144         if (ds == NULL) {
 145                 /*
 146                  * Account for the meta-objset space in its placeholder
 147                  * dataset.
 148                  */
 149                 dsl_free(tx->tx_pool, tx->tx_txg, bp);
 150
 151                 dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
 152                     -used, -compressed, -uncompressed, tx);
 153                 dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
 154                 return (used);
 155         }
 156         ASSERT3P(tx->tx_pool, ==, ds->ds_dir->dd_pool);
 157
 158         ASSERT(!dsl_dataset_is_snapshot(ds));
 159         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 160
 161         if (bp->blk_birth > ds->ds_phys->ds_prev_snap_txg) {
 162                 int64_t delta;
 163
 164                 dprintf_bp(bp, "freeing ds=%llu", ds->ds_object);
 165                 dsl_free(tx->tx_pool, tx->tx_txg, bp);
 166
 167                 mutex_enter(&ds->ds_dir->dd_lock);
 168                 mutex_enter(&ds->ds_lock);
 169                 ASSERT(ds->ds_phys->ds_unique_bytes >= used ||
 170                     !DS_UNIQUE_IS_ACCURATE(ds));
 171                 delta = parent_delta(ds, -used);
 172                 ds->ds_phys->ds_unique_bytes -= used;
 173                 mutex_exit(&ds->ds_lock);
 174                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
 175                     delta, -compressed, -uncompressed, tx);
 176                 dsl_dir_transfer_space(ds->ds_dir, -used - delta,
 177                     DD_USED_REFRSRV, DD_USED_HEAD, tx);
 178                 mutex_exit(&ds->ds_dir->dd_lock);
 179         } else {
 180                 dprintf_bp(bp, "putting on dead list: %s", "");
 181                 if (async) {
 182                         /*
 183                          * We are here as part of zio's write done callback,
 184                          * which means we're a zio interrupt thread.  We can't
 185                          * call dsl_deadlist_insert() now because it may block
 186                          * waiting for I/O.  Instead, put bp on the deferred
 187                          * queue and let dsl_pool_sync() finish the job.
 188                          */
 189                         bplist_append(&ds->ds_pending_deadlist, bp);
 190                 } else {
 191                         dsl_deadlist_insert(&ds->ds_deadlist, bp, tx);
 192                 }
 193                 ASSERT3U(ds->ds_prev->ds_object, ==,
 194                     ds->ds_phys->ds_prev_snap_obj);
 195                 ASSERT(ds->ds_prev->ds_phys->ds_num_children > 0);
 196                 /* if (bp->blk_birth > prev prev snap txg) prev unique += bs */
 197                 if (ds->ds_prev->ds_phys->ds_next_snap_obj ==
 198                     ds->ds_object && bp->blk_birth >
 199                     ds->ds_prev->ds_phys->ds_prev_snap_txg) {
 200                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
 201                         mutex_enter(&ds->ds_prev->ds_lock);
 202                         ds->ds_prev->ds_phys->ds_unique_bytes += used;
 203                         mutex_exit(&ds->ds_prev->ds_lock);
 204                 }
 205                 if (bp->blk_birth > ds->ds_dir->dd_origin_txg) {
 206                         dsl_dir_transfer_space(ds->ds_dir, used,
 207                             DD_USED_HEAD, DD_USED_SNAP, tx);
 208                 }
 209         }
 210         mutex_enter(&ds->ds_lock);
 211         ASSERT3U(ds->ds_phys->ds_used_bytes, >=, used);
 212         ds->ds_phys->ds_used_bytes -= used;
 213         ASSERT3U(ds->ds_phys->ds_compressed_bytes, >=, compressed);
 214         ds->ds_phys->ds_compressed_bytes -= compressed;
 215         ASSERT3U(ds->ds_phys->ds_uncompressed_bytes, >=, uncompressed);
 216         ds->ds_phys->ds_uncompressed_bytes -= uncompressed;
 217         mutex_exit(&ds->ds_lock);
 218
 219         return (used);
 220 }
 221
 222 uint64_t
 223 dsl_dataset_prev_snap_txg(dsl_dataset_t *ds)
 224 {
 225         uint64_t trysnap = 0;
 226
 227         if (ds == NULL)
 228                 return (0);
 229         /*
 230          * The snapshot creation could fail, but that would cause an
 231          * incorrect FALSE return, which would only result in an
 232          * overestimation of the amount of space that an operation would
 233          * consume, which is OK.
 234          *
 235          * There's also a small window where we could miss a pending
 236          * snapshot, because we could set the sync task in the quiescing
 237          * phase.  So this should only be used as a guess.
 238          */
 239         if (ds->ds_trysnap_txg >
 240             spa_last_synced_txg(ds->ds_dir->dd_pool->dp_spa))
 241                 trysnap = ds->ds_trysnap_txg;
 242         return (MAX(ds->ds_phys->ds_prev_snap_txg, trysnap));
 243 }
 244
 245 boolean_t
 246 dsl_dataset_block_freeable(dsl_dataset_t *ds, const blkptr_t *bp,
 247     uint64_t blk_birth)
 248 {
 249         if (blk_birth <= dsl_dataset_prev_snap_txg(ds))
 250                 return (B_FALSE);
 251
 252         ddt_prefetch(dsl_dataset_get_spa(ds), bp);
 253
 254         return (B_TRUE);
 255 }
 256
 257 /* ARGSUSED */
 258 static void
 259 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
 260 {
 261         dsl_dataset_t *ds = dsv;
 262
 263         ASSERT(ds->ds_owner == NULL || DSL_DATASET_IS_DESTROYED(ds));
 264
 265         unique_remove(ds->ds_fsid_guid);
 266
 267         if (ds->ds_objset != NULL)
 268                 dmu_objset_evict(ds->ds_objset);
 269
 270         if (ds->ds_prev) {
 271                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 272                 ds->ds_prev = NULL;
 273         }
 274
 275         bplist_destroy(&ds->ds_pending_deadlist);
 276         if (db != NULL) {
 277                 dsl_deadlist_close(&ds->ds_deadlist);
 278         } else {
 279                 ASSERT(ds->ds_deadlist.dl_dbuf == NULL);
 280                 ASSERT(!ds->ds_deadlist.dl_oldfmt);
 281         }
 282         if (ds->ds_dir)
 283                 dsl_dir_close(ds->ds_dir, ds);
 284
 285         ASSERT(!list_link_active(&ds->ds_synced_link));
 286
 287         if (mutex_owned(&ds->ds_lock))
 288                 mutex_exit(&ds->ds_lock);
 289         mutex_destroy(&ds->ds_lock);
 290         mutex_destroy(&ds->ds_recvlock);
 291         if (mutex_owned(&ds->ds_opening_lock))
 292                 mutex_exit(&ds->ds_opening_lock);
 293         mutex_destroy(&ds->ds_opening_lock);
 294         rw_destroy(&ds->ds_rwlock);
 295         cv_destroy(&ds->ds_exclusive_cv);
 296
 297         kmem_free(ds, sizeof (dsl_dataset_t));
 298 }
 299
 300 static int
 301 dsl_dataset_get_snapname(dsl_dataset_t *ds)
 302 {
 303         dsl_dataset_phys_t *headphys;
 304         int err;
 305         dmu_buf_t *headdbuf;
 306         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 307         objset_t *mos = dp->dp_meta_objset;
 308
 309         if (ds->ds_snapname[0])
 310                 return (0);
 311         if (ds->ds_phys->ds_next_snap_obj == 0)
 312                 return (0);
 313
 314         err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
 315             FTAG, &headdbuf);
 316         if (err)
 317                 return (err);
 318         headphys = headdbuf->db_data;
 319         err = zap_value_search(dp->dp_meta_objset,
 320             headphys->ds_snapnames_zapobj, ds->ds_object, 0, ds->ds_snapname);
 321         dmu_buf_rele(headdbuf, FTAG);
 322         return (err);
 323 }
 324
 325 static int
 326 dsl_dataset_snap_lookup(dsl_dataset_t *ds, const char *name, uint64_t *value)
 327 {
 328         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 329         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 330         matchtype_t mt;
 331         int err;
 332
 333         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 334                 mt = MT_FIRST;
 335         else
 336                 mt = MT_EXACT;
 337
 338         err = zap_lookup_norm(mos, snapobj, name, 8, 1,
 339             value, mt, NULL, 0, NULL);
 340         if (err == ENOTSUP && mt == MT_FIRST)
 341                 err = zap_lookup(mos, snapobj, name, 8, 1, value);
 342         return (err);
 343 }
 344
 345 static int
 346 dsl_dataset_snap_remove(dsl_dataset_t *ds, char *name, dmu_tx_t *tx)
 347 {
 348         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 349         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 350         matchtype_t mt;
 351         int err;
 352
 353         dsl_dir_snap_cmtime_update(ds->ds_dir);
 354
 355         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 356                 mt = MT_FIRST;
 357         else
 358                 mt = MT_EXACT;
 359
 360         err = zap_remove_norm(mos, snapobj, name, mt, tx);
 361         if (err == ENOTSUP && mt == MT_FIRST)
 362                 err = zap_remove(mos, snapobj, name, tx);
 363         return (err);
 364 }
 365
 366 static int
 367 dsl_dataset_get_ref(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 368     dsl_dataset_t **dsp)
 369 {
 370         objset_t *mos = dp->dp_meta_objset;
 371         dmu_buf_t *dbuf;
 372         dsl_dataset_t *ds;
 373         int err;
 374         dmu_object_info_t doi;
 375
 376         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
 377             dsl_pool_sync_context(dp));
 378
 379         err = dmu_bonus_hold(mos, dsobj, tag, &dbuf);
 380         if (err)
 381                 return (err);
 382
 383         /* Make sure dsobj has the correct object type. */
 384         dmu_object_info_from_db(dbuf, &doi);
 385         if (doi.doi_type != DMU_OT_DSL_DATASET)
 386                 return (EINVAL);
 387
 388         ds = dmu_buf_get_user(dbuf);
 389         if (ds == NULL) {
 390                 dsl_dataset_t *winner;
 391
 392                 ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
 393                 ds->ds_dbuf = dbuf;
 394                 ds->ds_object = dsobj;
 395                 ds->ds_phys = dbuf->db_data;
 396
 397                 mutex_init(&ds->ds_lock, NULL, MUTEX_DEFAULT, NULL);
 398                 mutex_init(&ds->ds_recvlock, NULL, MUTEX_DEFAULT, NULL);
 399                 mutex_init(&ds->ds_opening_lock, NULL, MUTEX_DEFAULT, NULL);
 400                 rw_init(&ds->ds_rwlock, 0, 0, 0);
 401                 cv_init(&ds->ds_exclusive_cv, NULL, CV_DEFAULT, NULL);
 402
 403                 bplist_create(&ds->ds_pending_deadlist);
 404                 dsl_deadlist_open(&ds->ds_deadlist,
 405                     mos, ds->ds_phys->ds_deadlist_obj);
 406
 407                 if (err == 0) {
 408                         err = dsl_dir_open_obj(dp,
 409                             ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
 410                 }
 411                 if (err) {
 412                         mutex_destroy(&ds->ds_lock);
 413                         mutex_destroy(&ds->ds_recvlock);
 414                         mutex_destroy(&ds->ds_opening_lock);
 415                         rw_destroy(&ds->ds_rwlock);
 416                         cv_destroy(&ds->ds_exclusive_cv);
 417                         bplist_destroy(&ds->ds_pending_deadlist);
 418                         dsl_deadlist_close(&ds->ds_deadlist);
 419                         kmem_free(ds, sizeof (dsl_dataset_t));
 420                         dmu_buf_rele(dbuf, tag);
 421                         return (err);
 422                 }
 423
 424                 if (!dsl_dataset_is_snapshot(ds)) {
 425                         ds->ds_snapname[0] = '\0';
 426                         if (ds->ds_phys->ds_prev_snap_obj) {
 427                                 err = dsl_dataset_get_ref(dp,
 428                                     ds->ds_phys->ds_prev_snap_obj,
 429                                     ds, &ds->ds_prev);
 430                         }
 431                 } else {
 432                         if (zfs_flags & ZFS_DEBUG_SNAPNAMES)
 433                                 err = dsl_dataset_get_snapname(ds);
 434                         if (err == 0 && ds->ds_phys->ds_userrefs_obj != 0) {
 435                                 err = zap_count(
 436                                     ds->ds_dir->dd_pool->dp_meta_objset,
 437                                     ds->ds_phys->ds_userrefs_obj,
 438                                     &ds->ds_userrefs);
 439                         }
 440                 }
 441
 442                 if (err == 0 && !dsl_dataset_is_snapshot(ds)) {
 443                         /*
 444                          * In sync context, we're called with either no lock
 445                          * or with the write lock.  If we're not syncing,
 446                          * we're always called with the read lock held.
 447                          */
 448                         boolean_t need_lock =
 449                             !RW_WRITE_HELD(&dp->dp_config_rwlock) &&
 450                             dsl_pool_sync_context(dp);
 451
 452                         if (need_lock)
 453                                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 454
 455                         err = dsl_prop_get_ds(ds,
 456                             "refreservation", sizeof (uint64_t), 1,
 457                             &ds->ds_reserved, NULL);
 458                         if (err == 0) {
 459                                 err = dsl_prop_get_ds(ds,
 460                                     "refquota", sizeof (uint64_t), 1,
 461                                     &ds->ds_quota, NULL);
 462                         }
 463
 464                         if (need_lock)
 465                                 rw_exit(&dp->dp_config_rwlock);
 466                 } else {
 467                         ds->ds_reserved = ds->ds_quota = 0;
 468                 }
 469
 470                 if (err == 0) {
 471                         winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
 472                             dsl_dataset_evict);
 473                 }
 474                 if (err || winner) {
 475                         bplist_destroy(&ds->ds_pending_deadlist);
 476                         dsl_deadlist_close(&ds->ds_deadlist);
 477                         if (ds->ds_prev)
 478                                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 479                         dsl_dir_close(ds->ds_dir, ds);
 480                         mutex_destroy(&ds->ds_lock);
 481                         mutex_destroy(&ds->ds_recvlock);
 482                         mutex_destroy(&ds->ds_opening_lock);
 483                         rw_destroy(&ds->ds_rwlock);
 484                         cv_destroy(&ds->ds_exclusive_cv);
 485                         kmem_free(ds, sizeof (dsl_dataset_t));
 486                         if (err) {
 487                                 dmu_buf_rele(dbuf, tag);
 488                                 return (err);
 489                         }
 490                         ds = winner;
 491                 } else {
 492                         ds->ds_fsid_guid =
 493                             unique_insert(ds->ds_phys->ds_fsid_guid);
 494                 }
 495         }
 496         ASSERT3P(ds->ds_dbuf, ==, dbuf);
 497         ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
 498         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0 ||
 499             spa_version(dp->dp_spa) < SPA_VERSION_ORIGIN ||
 500             dp->dp_origin_snap == NULL || ds == dp->dp_origin_snap);
 501         mutex_enter(&ds->ds_lock);
 502         if (!dsl_pool_sync_context(dp) && DSL_DATASET_IS_DESTROYED(ds)) {
 503                 mutex_exit(&ds->ds_lock);
 504                 dmu_buf_rele(ds->ds_dbuf, tag);
 505                 return (ENOENT);
 506         }
 507         mutex_exit(&ds->ds_lock);
 508         *dsp = ds;
 509         return (0);
 510 }
 511
 512 static int
 513 dsl_dataset_hold_ref(dsl_dataset_t *ds, void *tag)
 514 {
 515         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 516
 517         /*
 518          * In syncing context we don't want the rwlock lock: there
 519          * may be an existing writer waiting for sync phase to
 520          * finish.  We don't need to worry about such writers, since
 521          * sync phase is single-threaded, so the writer can't be
 522          * doing anything while we are active.
 523          */
 524         if (dsl_pool_sync_context(dp)) {
 525                 ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
 526                 return (0);
 527         }
 528
 529         /*
 530          * Normal users will hold the ds_rwlock as a READER until they
 531          * are finished (i.e., call dsl_dataset_rele()).  "Owners" will
 532          * drop their READER lock after they set the ds_owner field.
 533          *
 534          * If the dataset is being destroyed, the destroy thread will
 535          * obtain a WRITER lock for exclusive access after it's done its
 536          * open-context work and then change the ds_owner to
 537          * dsl_reaper once destruction is assured.  So threads
 538          * may block here temporarily, until the "destructability" of
 539          * the dataset is determined.
 540          */
 541         ASSERT(!RW_WRITE_HELD(&dp->dp_config_rwlock));
 542         mutex_enter(&ds->ds_lock);
 543         while (!rw_tryenter(&ds->ds_rwlock, RW_READER)) {
 544                 rw_exit(&dp->dp_config_rwlock);
 545                 cv_wait(&ds->ds_exclusive_cv, &ds->ds_lock);
 546                 if (DSL_DATASET_IS_DESTROYED(ds)) {
 547                         mutex_exit(&ds->ds_lock);
 548                         dsl_dataset_drop_ref(ds, tag);
 549                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 550                         return (ENOENT);
 551                 }
 552                 /*
 553                  * The dp_config_rwlock lives above the ds_lock. And
 554                  * we need to check DSL_DATASET_IS_DESTROYED() while
 555                  * holding the ds_lock, so we have to drop and reacquire
 556                  * the ds_lock here.
 557                  */
 558                 mutex_exit(&ds->ds_lock);
 559                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 560                 mutex_enter(&ds->ds_lock);
 561         }
 562         mutex_exit(&ds->ds_lock);
 563         return (0);
 564 }
 565
 566 int
 567 dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 568     dsl_dataset_t **dsp)
 569 {
 570         int err = dsl_dataset_get_ref(dp, dsobj, tag, dsp);
 571
 572         if (err)
 573                 return (err);
 574         return (dsl_dataset_hold_ref(*dsp, tag));
 575 }
 576
 577 int
 578 dsl_dataset_own_obj(dsl_pool_t *dp, uint64_t dsobj, boolean_t inconsistentok,
 579     void *tag, dsl_dataset_t **dsp)
 580 {
 581         int err = dsl_dataset_hold_obj(dp, dsobj, tag, dsp);
 582         if (err)
 583                 return (err);
 584         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 585                 dsl_dataset_rele(*dsp, tag);
 586                 *dsp = NULL;
 587                 return (EBUSY);
 588         }
 589         return (0);
 590 }
 591
 592 int
 593 dsl_dataset_hold(const char *name, void *tag, dsl_dataset_t **dsp)
 594 {
 595         dsl_dir_t *dd;
 596         dsl_pool_t *dp;
 597         const char *snapname;
 598         uint64_t obj;
 599         int err = 0;
 600
 601         err = dsl_dir_open_spa(NULL, name, FTAG, &dd, &snapname);
 602         if (err)
 603                 return (err);
 604
 605         dp = dd->dd_pool;
 606         obj = dd->dd_phys->dd_head_dataset_obj;
 607         rw_enter(&dp->dp_config_rwlock, RW_READER);
 608         if (obj)
 609                 err = dsl_dataset_get_ref(dp, obj, tag, dsp);
 610         else
 611                 err = ENOENT;
 612         if (err)
 613                 goto out;
 614
 615         err = dsl_dataset_hold_ref(*dsp, tag);
 616
 617         /* we may be looking for a snapshot */
 618         if (err == 0 && snapname != NULL) {
 619                 dsl_dataset_t *ds = NULL;
 620
 621                 if (*snapname++ != '@') {
 622                         dsl_dataset_rele(*dsp, tag);
 623                         err = ENOENT;
 624                         goto out;
 625                 }
 626
 627                 dprintf("looking for snapshot '%s'\n", snapname);
 628                 err = dsl_dataset_snap_lookup(*dsp, snapname, &obj);
 629                 if (err == 0)
 630                         err = dsl_dataset_get_ref(dp, obj, tag, &ds);
 631                 dsl_dataset_rele(*dsp, tag);
 632
 633                 ASSERT3U((err == 0), ==, (ds != NULL));
 634
 635                 if (ds) {
 636                         mutex_enter(&ds->ds_lock);
 637                         if (ds->ds_snapname[0] == 0)
 638                                 (void) strlcpy(ds->ds_snapname, snapname,
 639                                     sizeof (ds->ds_snapname));
 640                         mutex_exit(&ds->ds_lock);
 641                         err = dsl_dataset_hold_ref(ds, tag);
 642                         *dsp = err ? NULL : ds;
 643                 }
 644         }
 645 out:
 646         rw_exit(&dp->dp_config_rwlock);
 647         dsl_dir_close(dd, FTAG);
 648         return (err);
 649 }
 650
 651 int
 652 dsl_dataset_own(const char *name, boolean_t inconsistentok,
 653     void *tag, dsl_dataset_t **dsp)
 654 {
 655         int err = dsl_dataset_hold(name, tag, dsp);
 656         if (err)
 657                 return (err);
 658         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 659                 dsl_dataset_rele(*dsp, tag);
 660                 return (EBUSY);
 661         }
 662         return (0);
 663 }
 664
 665 void
 666 dsl_dataset_name(dsl_dataset_t *ds, char *name)
 667 {
 668         if (ds == NULL) {
 669                 (void) strcpy(name, "mos");
 670         } else {
 671                 dsl_dir_name(ds->ds_dir, name);
 672                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 673                 if (ds->ds_snapname[0]) {
 674                         (void) strcat(name, "@");
 675                         /*
 676                          * We use a "recursive" mutex so that we
 677                          * can call dprintf_ds() with ds_lock held.
 678                          */
 679                         if (!MUTEX_HELD(&ds->ds_lock)) {
 680                                 mutex_enter(&ds->ds_lock);
 681                                 (void) strcat(name, ds->ds_snapname);
 682                                 mutex_exit(&ds->ds_lock);
 683                         } else {
 684                                 (void) strcat(name, ds->ds_snapname);
 685                         }
 686                 }
 687         }
 688 }
 689
 690 static int
 691 dsl_dataset_namelen(dsl_dataset_t *ds)
 692 {
 693         int result;
 694
 695         if (ds == NULL) {
 696                 result = 3;     /* "mos" */
 697         } else {
 698                 result = dsl_dir_namelen(ds->ds_dir);
 699                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 700                 if (ds->ds_snapname[0]) {
 701                         ++result;       /* adding one for the @-sign */
 702                         if (!MUTEX_HELD(&ds->ds_lock)) {
 703                                 mutex_enter(&ds->ds_lock);
 704                                 result += strlen(ds->ds_snapname);
 705                                 mutex_exit(&ds->ds_lock);
 706                         } else {
 707                                 result += strlen(ds->ds_snapname);
 708                         }
 709                 }
 710         }
 711
 712         return (result);
 713 }
 714
 715 void
 716 dsl_dataset_drop_ref(dsl_dataset_t *ds, void *tag)
 717 {
 718         dmu_buf_rele(ds->ds_dbuf, tag);
 719 }
 720
 721 void
 722 dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
 723 {
 724         if (!dsl_pool_sync_context(ds->ds_dir->dd_pool)) {
 725                 rw_exit(&ds->ds_rwlock);
 726         }
 727         dsl_dataset_drop_ref(ds, tag);
 728 }
 729
 730 void
 731 dsl_dataset_disown(dsl_dataset_t *ds, void *tag)
 732 {
 733         ASSERT((ds->ds_owner == tag && ds->ds_dbuf) ||
 734             (DSL_DATASET_IS_DESTROYED(ds) && ds->ds_dbuf == NULL));
 735
 736         mutex_enter(&ds->ds_lock);
 737         ds->ds_owner = NULL;
 738         if (RW_WRITE_HELD(&ds->ds_rwlock)) {
 739                 rw_exit(&ds->ds_rwlock);
 740                 cv_broadcast(&ds->ds_exclusive_cv);
 741         }
 742         mutex_exit(&ds->ds_lock);
 743         if (ds->ds_dbuf)
 744                 dsl_dataset_drop_ref(ds, tag);
 745         else
 746                 dsl_dataset_evict(NULL, ds);
 747 }
 748
 749 boolean_t
 750 dsl_dataset_tryown(dsl_dataset_t *ds, boolean_t inconsistentok, void *tag)
 751 {
 752         boolean_t gotit = FALSE;
 753
 754         mutex_enter(&ds->ds_lock);
 755         if (ds->ds_owner == NULL &&
 756             (!DS_IS_INCONSISTENT(ds) || inconsistentok)) {
 757                 ds->ds_owner = tag;
 758                 if (!dsl_pool_sync_context(ds->ds_dir->dd_pool))
 759                         rw_exit(&ds->ds_rwlock);
 760                 gotit = TRUE;
 761         }
 762         mutex_exit(&ds->ds_lock);
 763         return (gotit);
 764 }
 765
 766 void
 767 dsl_dataset_make_exclusive(dsl_dataset_t *ds, void *owner)
 768 {
 769         ASSERT3P(owner, ==, ds->ds_owner);
 770         if (!RW_WRITE_HELD(&ds->ds_rwlock))
 771                 rw_enter(&ds->ds_rwlock, RW_WRITER);
 772 }
 773
 774 uint64_t
 775 dsl_dataset_create_sync_dd(dsl_dir_t *dd, dsl_dataset_t *origin,
 776     uint64_t flags, dmu_tx_t *tx)
 777 {
 778         dsl_pool_t *dp = dd->dd_pool;
 779         dmu_buf_t *dbuf;
 780         dsl_dataset_phys_t *dsphys;
 781         uint64_t dsobj;
 782         objset_t *mos = dp->dp_meta_objset;
 783
 784         if (origin == NULL)
 785                 origin = dp->dp_origin_snap;
 786
 787         ASSERT(origin == NULL || origin->ds_dir->dd_pool == dp);
 788         ASSERT(origin == NULL || origin->ds_phys->ds_num_children > 0);
 789         ASSERT(dmu_tx_is_syncing(tx));
 790         ASSERT(dd->dd_phys->dd_head_dataset_obj == 0);
 791
 792         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
 793             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
 794         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
 795         dmu_buf_will_dirty(dbuf, tx);
 796         dsphys = dbuf->db_data;
 797         bzero(dsphys, sizeof (dsl_dataset_phys_t));
 798         dsphys->ds_dir_obj = dd->dd_object;
 799         dsphys->ds_flags = flags;
 800         dsphys->ds_fsid_guid = unique_create();
 801         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
 802             sizeof (dsphys->ds_guid));
 803         dsphys->ds_snapnames_zapobj =
 804             zap_create_norm(mos, U8_TEXTPREP_TOUPPER, DMU_OT_DSL_DS_SNAP_MAP,
 805             DMU_OT_NONE, 0, tx);
 806         dsphys->ds_creation_time = gethrestime_sec();
 807         dsphys->ds_creation_txg = tx->tx_txg == TXG_INITIAL ? 1 : tx->tx_txg;
 808
 809         if (origin == NULL) {
 810                 dsphys->ds_deadlist_obj = dsl_deadlist_alloc(mos, tx);
 811         } else {
 812                 dsl_dataset_t *ohds;
 813
 814                 dsphys->ds_prev_snap_obj = origin->ds_object;
 815                 dsphys->ds_prev_snap_txg =
 816                     origin->ds_phys->ds_creation_txg;
 817                 dsphys->ds_used_bytes =
 818                     origin->ds_phys->ds_used_bytes;
 819                 dsphys->ds_compressed_bytes =
 820                     origin->ds_phys->ds_compressed_bytes;
 821                 dsphys->ds_uncompressed_bytes =
 822                     origin->ds_phys->ds_uncompressed_bytes;
 823                 dsphys->ds_bp = origin->ds_phys->ds_bp;
 824                 dsphys->ds_flags |= origin->ds_phys->ds_flags;
 825
 826                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
 827                 origin->ds_phys->ds_num_children++;
 828
 829                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
 830                     origin->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ohds));
 831                 dsphys->ds_deadlist_obj = dsl_deadlist_clone(&ohds->ds_deadlist,
 832                     dsphys->ds_prev_snap_txg, dsphys->ds_prev_snap_obj, tx);
 833                 dsl_dataset_rele(ohds, FTAG);
 834
 835                 if (spa_version(dp->dp_spa) >= SPA_VERSION_NEXT_CLONES) {
 836                         if (origin->ds_phys->ds_next_clones_obj == 0) {
 837                                 origin->ds_phys->ds_next_clones_obj =
 838                                     zap_create(mos,
 839                                     DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
 840                         }
 841                         VERIFY(0 == zap_add_int(mos,
 842                             origin->ds_phys->ds_next_clones_obj,
 843                             dsobj, tx));
 844                 }
 845
 846                 dmu_buf_will_dirty(dd->dd_dbuf, tx);
 847                 dd->dd_phys->dd_origin_obj = origin->ds_object;
 848                 if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
 849                         if (origin->ds_dir->dd_phys->dd_clones == 0) {
 850                                 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
 851                                 origin->ds_dir->dd_phys->dd_clones =
 852                                     zap_create(mos,
 853                                     DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
 854                         }
 855                         VERIFY3U(0, ==, zap_add_int(mos,
 856                             origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
 857                 }
 858         }
 859
 860         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
 861                 dsphys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
 862
 863         dmu_buf_rele(dbuf, FTAG);
 864
 865         dmu_buf_will_dirty(dd->dd_dbuf, tx);
 866         dd->dd_phys->dd_head_dataset_obj = dsobj;
 867
 868         return (dsobj);
 869 }
 870
 871 uint64_t
 872 dsl_dataset_create_sync(dsl_dir_t *pdd, const char *lastname,
 873     dsl_dataset_t *origin, uint64_t flags, cred_t *cr, dmu_tx_t *tx)
 874 {
 875         dsl_pool_t *dp = pdd->dd_pool;
 876         uint64_t dsobj, ddobj;
 877         dsl_dir_t *dd;
 878
 879         ASSERT(lastname[0] != '@');
 880
 881         ddobj = dsl_dir_create_sync(dp, pdd, lastname, tx);
 882         VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
 883
 884         dsobj = dsl_dataset_create_sync_dd(dd, origin, flags, tx);
 885
 886         dsl_deleg_set_create_perms(dd, tx, cr);
 887
 888         dsl_dir_close(dd, FTAG);
 889
 890         /*
 891          * If we are creating a clone, make sure we zero out any stale
 892          * data from the origin snapshots zil header.
 893          */
 894         if (origin != NULL) {
 895                 dsl_dataset_t *ds;
 896                 objset_t *os;
 897
 898                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 899                 VERIFY3U(0, ==, dmu_objset_from_ds(ds, &os));
 900                 bzero(&os->os_zil_header, sizeof (os->os_zil_header));
 901                 dsl_dataset_dirty(ds, tx);
 902                 dsl_dataset_rele(ds, FTAG);
 903         }
 904
 905         return (dsobj);
 906 }
 907
 908 struct destroyarg {
 909         dsl_sync_task_group_t *dstg;
 910         char *snapname;
 911         char *failed;
 912         boolean_t defer;
 913 };
 914
 915 static int
 916 dsl_snapshot_destroy_one(const char *name, void *arg)
 917 {
 918         struct destroyarg *da = arg;
 919         dsl_dataset_t *ds;
 920         int err;
 921         char *dsname;
 922
 923         dsname = kmem_asprintf("%s@%s", name, da->snapname);
 924         err = dsl_dataset_own(dsname, B_TRUE, da->dstg, &ds);
 925         strfree(dsname);
 926         if (err == 0) {
 927                 struct dsl_ds_destroyarg *dsda;
 928
 929                 dsl_dataset_make_exclusive(ds, da->dstg);
 930                 dsda = kmem_zalloc(sizeof (struct dsl_ds_destroyarg), KM_SLEEP);
 931                 dsda->ds = ds;
 932                 dsda->defer = da->defer;
 933                 dsl_sync_task_create(da->dstg, dsl_dataset_destroy_check,
 934                     dsl_dataset_destroy_sync, dsda, da->dstg, 0);
 935         } else if (err == ENOENT) {
 936                 err = 0;
 937         } else {
 938                 (void) strcpy(da->failed, name);
 939         }
 940         return (err);
 941 }
 942
 943 /*
 944  * Destroy 'snapname' in all descendants of 'fsname'.
 945  */
 946 #pragma weak dmu_snapshots_destroy = dsl_snapshots_destroy
 947 int
 948 dsl_snapshots_destroy(char *fsname, char *snapname, boolean_t defer)
 949 {
 950         int err;
 951         struct destroyarg da;
 952         dsl_sync_task_t *dst;
 953         spa_t *spa;
 954
 955         err = spa_open(fsname, &spa, FTAG);
 956         if (err)
 957                 return (err);
 958         da.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
 959         da.snapname = snapname;
 960         da.failed = fsname;
 961         da.defer = defer;
 962
 963         err = dmu_objset_find(fsname,
 964             dsl_snapshot_destroy_one, &da, DS_FIND_CHILDREN);
 965
 966         if (err == 0)
 967                 err = dsl_sync_task_group_wait(da.dstg);
 968
 969         for (dst = list_head(&da.dstg->dstg_tasks); dst;
 970             dst = list_next(&da.dstg->dstg_tasks, dst)) {
 971                 struct dsl_ds_destroyarg *dsda = dst->dst_arg1;
 972                 dsl_dataset_t *ds = dsda->ds;
 973
 974                 /*
 975                  * Return the file system name that triggered the error
 976                  */
 977                 if (dst->dst_err) {
 978                         dsl_dataset_name(ds, fsname);
 979                         *strchr(fsname, '@') = '\0';
 980                 }
 981                 ASSERT3P(dsda->rm_origin, ==, NULL);
 982                 dsl_dataset_disown(ds, da.dstg);
 983                 kmem_free(dsda, sizeof (struct dsl_ds_destroyarg));
 984         }
 985
 986         dsl_sync_task_group_destroy(da.dstg);
 987         spa_close(spa, FTAG);
 988         return (err);
 989 }
 990
 991 static boolean_t
 992 dsl_dataset_might_destroy_origin(dsl_dataset_t *ds)
 993 {
 994         boolean_t might_destroy = B_FALSE;
 995
 996         mutex_enter(&ds->ds_lock);
 997         if (ds->ds_phys->ds_num_children == 2 && ds->ds_userrefs == 0 &&
 998             DS_IS_DEFER_DESTROY(ds))
 999                 might_destroy = B_TRUE;
1000         mutex_exit(&ds->ds_lock);
1001
1002         return (might_destroy);
1003 }
1004
1005 /*
1006  * If we're removing a clone, and these three conditions are true:
1007  *      1) the clone's origin has no other children
1008  *      2) the clone's origin has no user references
1009  *      3) the clone's origin has been marked for deferred destruction
1010  * Then, prepare to remove the origin as part of this sync task group.
1011  */
1012 static int
1013 dsl_dataset_origin_rm_prep(struct dsl_ds_destroyarg *dsda, void *tag)
1014 {
1015         dsl_dataset_t *ds = dsda->ds;
1016         dsl_dataset_t *origin = ds->ds_prev;
1017
1018         if (dsl_dataset_might_destroy_origin(origin)) {
1019                 char *name;
1020                 int namelen;
1021                 int error;
1022
1023                 namelen = dsl_dataset_namelen(origin) + 1;
1024                 name = kmem_alloc(namelen, KM_SLEEP);
1025                 dsl_dataset_name(origin, name);
1026 #ifdef _KERNEL
1027                 error = zfs_unmount_snap(name, NULL);
1028                 if (error) {
1029                         kmem_free(name, namelen);
1030                         return (error);
1031                 }
1032 #endif
1033                 error = dsl_dataset_own(name, B_TRUE, tag, &origin);
1034                 kmem_free(name, namelen);
1035                 if (error)
1036                         return (error);
1037                 dsda->rm_origin = origin;
1038                 dsl_dataset_make_exclusive(origin, tag);
1039         }
1040
1041         return (0);
1042 }
1043
1044 /*
1045  * ds must be opened as OWNER.  On return (whether successful or not),
1046  * ds will be closed and caller can no longer dereference it.
1047  */
1048 int
1049 dsl_dataset_destroy(dsl_dataset_t *ds, void *tag, boolean_t defer)
1050 {
1051         int err;
1052         dsl_sync_task_group_t *dstg;
1053         objset_t *os;
1054         dsl_dir_t *dd;
1055         uint64_t obj;
1056         struct dsl_ds_destroyarg dsda = { 0 };
1057         dsl_dataset_t dummy_ds = { 0 };
1058
1059         dsda.ds = ds;
1060
1061         if (dsl_dataset_is_snapshot(ds)) {
1062                 /* Destroying a snapshot is simpler */
1063                 dsl_dataset_make_exclusive(ds, tag);
1064
1065                 dsda.defer = defer;
1066                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1067                     dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
1068                     &dsda, tag, 0);
1069                 ASSERT3P(dsda.rm_origin, ==, NULL);
1070                 goto out;
1071         } else if (defer) {
1072                 err = EINVAL;
1073                 goto out;
1074         }
1075
1076         dd = ds->ds_dir;
1077         dummy_ds.ds_dir = dd;
1078         dummy_ds.ds_object = ds->ds_object;
1079
1080         /*
1081          * Check for errors and mark this ds as inconsistent, in
1082          * case we crash while freeing the objects.
1083          */
1084         err = dsl_sync_task_do(dd->dd_pool, dsl_dataset_destroy_begin_check,
1085             dsl_dataset_destroy_begin_sync, ds, NULL, 0);
1086         if (err)
1087                 goto out;
1088
1089         err = dmu_objset_from_ds(ds, &os);
1090         if (err)
1091                 goto out;
1092
1093         /*
1094          * remove the objects in open context, so that we won't
1095          * have too much to do in syncing context.
1096          */
1097         for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE,
1098             ds->ds_phys->ds_prev_snap_txg)) {
1099                 /*
1100                  * Ignore errors, if there is not enough disk space
1101                  * we will deal with it in dsl_dataset_destroy_sync().
1102                  */
1103                 (void) dmu_free_object(os, obj);
1104         }
1105         if (err != ESRCH)
1106                 goto out;
1107
1108         /*
1109          * Only the ZIL knows how to free log blocks.
1110          */
1111         zil_destroy(dmu_objset_zil(os), B_FALSE);
1112
1113         /*
1114          * Sync out all in-flight IO.
1115          */
1116         txg_wait_synced(dd->dd_pool, 0);
1117
1118         /*
1119          * If we managed to free all the objects in open
1120          * context, the user space accounting should be zero.
1121          */
1122         if (ds->ds_phys->ds_bp.blk_fill == 0 &&
1123             dmu_objset_userused_enabled(os)) {
1124                 uint64_t count;
1125
1126                 ASSERT(zap_count(os, DMU_USERUSED_OBJECT, &count) != 0 ||
1127                     count == 0);
1128                 ASSERT(zap_count(os, DMU_GROUPUSED_OBJECT, &count) != 0 ||
1129                     count == 0);
1130         }
1131
1132         rw_enter(&dd->dd_pool->dp_config_rwlock, RW_READER);
1133         err = dsl_dir_open_obj(dd->dd_pool, dd->dd_object, NULL, FTAG, &dd);
1134         rw_exit(&dd->dd_pool->dp_config_rwlock);
1135
1136         if (err)
1137                 goto out;
1138
1139         /*
1140          * Blow away the dsl_dir + head dataset.
1141          */
1142         dsl_dataset_make_exclusive(ds, tag);
1143         /*
1144          * If we're removing a clone, we might also need to remove its
1145          * origin.
1146          */
1147         do {
1148                 dsda.need_prep = B_FALSE;
1149                 if (dsl_dir_is_clone(dd)) {
1150                         err = dsl_dataset_origin_rm_prep(&dsda, tag);
1151                         if (err) {
1152                                 dsl_dir_close(dd, FTAG);
1153                                 goto out;
1154                         }
1155                 }
1156
1157                 dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
1158                 dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
1159                     dsl_dataset_destroy_sync, &dsda, tag, 0);
1160                 dsl_sync_task_create(dstg, dsl_dir_destroy_check,
1161                     dsl_dir_destroy_sync, &dummy_ds, FTAG, 0);
1162                 err = dsl_sync_task_group_wait(dstg);
1163                 dsl_sync_task_group_destroy(dstg);
1164
1165                 /*
1166                  * We could be racing against 'zfs release' or 'zfs destroy -d'
1167                  * on the origin snap, in which case we can get EBUSY if we
1168                  * needed to destroy the origin snap but were not ready to
1169                  * do so.
1170                  */
1171                 if (dsda.need_prep) {
1172                         ASSERT(err == EBUSY);
1173                         ASSERT(dsl_dir_is_clone(dd));
1174                         ASSERT(dsda.rm_origin == NULL);
1175                 }
1176         } while (dsda.need_prep);
1177
1178         if (dsda.rm_origin != NULL)
1179                 dsl_dataset_disown(dsda.rm_origin, tag);
1180
1181         /* if it is successful, dsl_dir_destroy_sync will close the dd */
1182         if (err)
1183                 dsl_dir_close(dd, FTAG);
1184 out:
1185         dsl_dataset_disown(ds, tag);
1186         return (err);
1187 }
1188
1189 blkptr_t *
1190 dsl_dataset_get_blkptr(dsl_dataset_t *ds)
1191 {
1192         return (&ds->ds_phys->ds_bp);
1193 }
1194
1195 void
1196 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
1197 {
1198         ASSERT(dmu_tx_is_syncing(tx));
1199         /* If it's the meta-objset, set dp_meta_rootbp */
1200         if (ds == NULL) {
1201                 tx->tx_pool->dp_meta_rootbp = *bp;
1202         } else {
1203                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1204                 ds->ds_phys->ds_bp = *bp;
1205         }
1206 }
1207
1208 spa_t *
1209 dsl_dataset_get_spa(dsl_dataset_t *ds)
1210 {
1211         return (ds->ds_dir->dd_pool->dp_spa);
1212 }
1213
1214 void
1215 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
1216 {
1217         dsl_pool_t *dp;
1218
1219         if (ds == NULL) /* this is the meta-objset */
1220                 return;
1221
1222         ASSERT(ds->ds_objset != NULL);
1223
1224         if (ds->ds_phys->ds_next_snap_obj != 0)
1225                 panic("dirtying snapshot!");
1226
1227         dp = ds->ds_dir->dd_pool;
1228
1229         if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
1230                 /* up the hold count until we can be written out */
1231                 dmu_buf_add_ref(ds->ds_dbuf, ds);
1232         }
1233 }
1234
1235 /*
1236  * The unique space in the head dataset can be calculated by subtracting
1237  * the space used in the most recent snapshot, that is still being used
1238  * in this file system, from the space currently in use.  To figure out
1239  * the space in the most recent snapshot still in use, we need to take
1240  * the total space used in the snapshot and subtract out the space that
1241  * has been freed up since the snapshot was taken.
1242  */
1243 static void
1244 dsl_dataset_recalc_head_uniq(dsl_dataset_t *ds)
1245 {
1246         uint64_t mrs_used;
1247         uint64_t dlused, dlcomp, dluncomp;
1248
1249         ASSERT(!dsl_dataset_is_snapshot(ds));
1250
1251         if (ds->ds_phys->ds_prev_snap_obj != 0)
1252                 mrs_used = ds->ds_prev->ds_phys->ds_used_bytes;
1253         else
1254                 mrs_used = 0;
1255
1256         dsl_deadlist_space(&ds->ds_deadlist, &dlused, &dlcomp, &dluncomp);
1257
1258         ASSERT3U(dlused, <=, mrs_used);
1259         ds->ds_phys->ds_unique_bytes =
1260             ds->ds_phys->ds_used_bytes - (mrs_used - dlused);
1261
1262         if (spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1263             SPA_VERSION_UNIQUE_ACCURATE)
1264                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1265 }
1266
1267 struct killarg {
1268         dsl_dataset_t *ds;
1269         dmu_tx_t *tx;
1270 };
1271
1272 /* ARGSUSED */
1273 static int
1274 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
1275     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
1276 {
1277         struct killarg *ka = arg;
1278         dmu_tx_t *tx = ka->tx;
1279
1280         if (bp == NULL)
1281                 return (0);
1282
1283         if (zb->zb_level == ZB_ZIL_LEVEL) {
1284                 ASSERT(zilog != NULL);
1285                 /*
1286                  * It's a block in the intent log.  It has no
1287                  * accounting, so just free it.
1288                  */
1289                 dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
1290         } else {
1291                 ASSERT(zilog == NULL);
1292                 ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
1293                 (void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
1294         }
1295
1296         return (0);
1297 }
1298
1299 /* ARGSUSED */
1300 static int
1301 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
1302 {
1303         dsl_dataset_t *ds = arg1;
1304         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1305         uint64_t count;
1306         int err;
1307
1308         /*
1309          * Can't delete a head dataset if there are snapshots of it.
1310          * (Except if the only snapshots are from the branch we cloned
1311          * from.)
1312          */
1313         if (ds->ds_prev != NULL &&
1314             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1315                 return (EBUSY);
1316
1317         /*
1318          * This is really a dsl_dir thing, but check it here so that
1319          * we'll be less likely to leave this dataset inconsistent &
1320          * nearly destroyed.
1321          */
1322         err = zap_count(mos, ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
1323         if (err)
1324                 return (err);
1325         if (count != 0)
1326                 return (EEXIST);
1327
1328         return (0);
1329 }
1330
1331 /* ARGSUSED */
1332 static void
1333 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1334 {
1335         dsl_dataset_t *ds = arg1;
1336         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1337
1338         /* Mark it as inconsistent on-disk, in case we crash */
1339         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1340         ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1341
1342         spa_history_log_internal(LOG_DS_DESTROY_BEGIN, dp->dp_spa, tx,
1343             "dataset = %llu", ds->ds_object);
1344 }
1345
1346 static int
1347 dsl_dataset_origin_check(struct dsl_ds_destroyarg *dsda, void *tag,
1348     dmu_tx_t *tx)
1349 {
1350         dsl_dataset_t *ds = dsda->ds;
1351         dsl_dataset_t *ds_prev = ds->ds_prev;
1352
1353         if (dsl_dataset_might_destroy_origin(ds_prev)) {
1354                 struct dsl_ds_destroyarg ndsda = {0};
1355
1356                 /*
1357                  * If we're not prepared to remove the origin, don't remove
1358                  * the clone either.
1359                  */
1360                 if (dsda->rm_origin == NULL) {
1361                         dsda->need_prep = B_TRUE;
1362                         return (EBUSY);
1363                 }
1364
1365                 ndsda.ds = ds_prev;
1366                 ndsda.is_origin_rm = B_TRUE;
1367                 return (dsl_dataset_destroy_check(&ndsda, tag, tx));
1368         }
1369
1370         /*
1371          * If we're not going to remove the origin after all,
1372          * undo the open context setup.
1373          */
1374         if (dsda->rm_origin != NULL) {
1375                 dsl_dataset_disown(dsda->rm_origin, tag);
1376                 dsda->rm_origin = NULL;
1377         }
1378
1379         return (0);
1380 }
1381
1382 /*
1383  * If you add new checks here, you may need to add
1384  * additional checks to the "temporary" case in
1385  * snapshot_check() in dmu_objset.c.
1386  */
1387 /* ARGSUSED */
1388 int
1389 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1390 {
1391         struct dsl_ds_destroyarg *dsda = arg1;
1392         dsl_dataset_t *ds = dsda->ds;
1393
1394         /* we have an owner hold, so noone else can destroy us */
1395         ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1396
1397         /*
1398          * Only allow deferred destroy on pools that support it.
1399          * NOTE: deferred destroy is only supported on snapshots.
1400          */
1401         if (dsda->defer) {
1402                 if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1403                     SPA_VERSION_USERREFS)
1404                         return (ENOTSUP);
1405                 ASSERT(dsl_dataset_is_snapshot(ds));
1406                 return (0);
1407         }
1408
1409         /*
1410          * Can't delete a head dataset if there are snapshots of it.
1411          * (Except if the only snapshots are from the branch we cloned
1412          * from.)
1413          */
1414         if (ds->ds_prev != NULL &&
1415             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1416                 return (EBUSY);
1417
1418         /*
1419          * If we made changes this txg, traverse_dsl_dataset won't find
1420          * them.  Try again.
1421          */
1422         if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1423                 return (EAGAIN);
1424
1425         if (dsl_dataset_is_snapshot(ds)) {
1426                 /*
1427                  * If this snapshot has an elevated user reference count,
1428                  * we can't destroy it yet.
1429                  */
1430                 if (ds->ds_userrefs > 0 && !dsda->releasing)
1431                         return (EBUSY);
1432
1433                 mutex_enter(&ds->ds_lock);
1434                 /*
1435                  * Can't delete a branch point. However, if we're destroying
1436                  * a clone and removing its origin due to it having a user
1437                  * hold count of 0 and having been marked for deferred destroy,
1438                  * it's OK for the origin to have a single clone.
1439                  */
1440                 if (ds->ds_phys->ds_num_children >
1441                     (dsda->is_origin_rm ? 2 : 1)) {
1442                         mutex_exit(&ds->ds_lock);
1443                         return (EEXIST);
1444                 }
1445                 mutex_exit(&ds->ds_lock);
1446         } else if (dsl_dir_is_clone(ds->ds_dir)) {
1447                 return (dsl_dataset_origin_check(dsda, arg2, tx));
1448         }
1449
1450         /* XXX we should do some i/o error checking... */
1451         return (0);
1452 }
1453
1454 struct refsarg {
1455         kmutex_t lock;
1456         boolean_t gone;
1457         kcondvar_t cv;
1458 };
1459
1460 /* ARGSUSED */
1461 static void
1462 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1463 {
1464         struct refsarg *arg = argv;
1465
1466         mutex_enter(&arg->lock);
1467         arg->gone = TRUE;
1468         cv_signal(&arg->cv);
1469         mutex_exit(&arg->lock);
1470 }
1471
1472 static void
1473 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1474 {
1475         struct refsarg arg;
1476
1477         bzero(&arg, sizeof(arg));
1478         mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1479         cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1480         arg.gone = FALSE;
1481         (void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1482             dsl_dataset_refs_gone);
1483         dmu_buf_rele(ds->ds_dbuf, tag);
1484         mutex_enter(&arg.lock);
1485         while (!arg.gone)
1486                 cv_wait(&arg.cv, &arg.lock);
1487         ASSERT(arg.gone);
1488         mutex_exit(&arg.lock);
1489         ds->ds_dbuf = NULL;
1490         ds->ds_phys = NULL;
1491         mutex_destroy(&arg.lock);
1492         cv_destroy(&arg.cv);
1493 }
1494
1495 static void
1496 remove_from_next_clones(dsl_dataset_t *ds, uint64_t obj, dmu_tx_t *tx)
1497 {
1498         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1499         uint64_t count;
1500         int err;
1501
1502         ASSERT(ds->ds_phys->ds_num_children >= 2);
1503         err = zap_remove_int(mos, ds->ds_phys->ds_next_clones_obj, obj, tx);
1504         /*
1505          * The err should not be ENOENT, but a bug in a previous version
1506          * of the code could cause upgrade_clones_cb() to not set
1507          * ds_next_snap_obj when it should, leading to a missing entry.
1508          * If we knew that the pool was created after
1509          * SPA_VERSION_NEXT_CLONES, we could assert that it isn't
1510          * ENOENT.  However, at least we can check that we don't have
1511          * too many entries in the next_clones_obj even after failing to
1512          * remove this one.
1513          */
1514         if (err != ENOENT) {
1515                 VERIFY3U(err, ==, 0);
1516         }
1517         ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
1518             &count));
1519         ASSERT3U(count, <=, ds->ds_phys->ds_num_children - 2);
1520 }
1521
1522 static void
1523 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
1524 {
1525         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1526         zap_cursor_t zc;
1527         zap_attribute_t za;
1528
1529         /*
1530          * If it is the old version, dd_clones doesn't exist so we can't
1531          * find the clones, but deadlist_remove_key() is a no-op so it
1532          * doesn't matter.
1533          */
1534         if (ds->ds_dir->dd_phys->dd_clones == 0)
1535                 return;
1536
1537         for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
1538             zap_cursor_retrieve(&zc, &za) == 0;
1539             zap_cursor_advance(&zc)) {
1540                 dsl_dataset_t *clone;
1541
1542                 VERIFY3U(0, ==, dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
1543                     za.za_first_integer, FTAG, &clone));
1544                 if (clone->ds_dir->dd_origin_txg > mintxg) {
1545                         dsl_deadlist_remove_key(&clone->ds_deadlist,
1546                             mintxg, tx);
1547                         dsl_dataset_remove_clones_key(clone, mintxg, tx);
1548                 }
1549                 dsl_dataset_rele(clone, FTAG);
1550         }
1551         zap_cursor_fini(&zc);
1552 }
1553
1554 struct process_old_arg {
1555         dsl_dataset_t *ds;
1556         dsl_dataset_t *ds_prev;
1557         boolean_t after_branch_point;
1558         zio_t *pio;
1559         uint64_t used, comp, uncomp;
1560 };
1561
1562 static int
1563 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1564 {
1565         struct process_old_arg *poa = arg;
1566         dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
1567
1568         if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
1569                 dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
1570                 if (poa->ds_prev && !poa->after_branch_point &&
1571                     bp->blk_birth >
1572                     poa->ds_prev->ds_phys->ds_prev_snap_txg) {
1573                         poa->ds_prev->ds_phys->ds_unique_bytes +=
1574                             bp_get_dsize_sync(dp->dp_spa, bp);
1575                 }
1576         } else {
1577                 poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
1578                 poa->comp += BP_GET_PSIZE(bp);
1579                 poa->uncomp += BP_GET_UCSIZE(bp);
1580                 dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
1581         }
1582         return (0);
1583 }
1584
1585 static void
1586 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
1587     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
1588 {
1589         struct process_old_arg poa = { 0 };
1590         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1591         objset_t *mos = dp->dp_meta_objset;
1592
1593         ASSERT(ds->ds_deadlist.dl_oldfmt);
1594         ASSERT(ds_next->ds_deadlist.dl_oldfmt);
1595
1596         poa.ds = ds;
1597         poa.ds_prev = ds_prev;
1598         poa.after_branch_point = after_branch_point;
1599         poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1600         VERIFY3U(0, ==, bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
1601             process_old_cb, &poa, tx));
1602         VERIFY3U(zio_wait(poa.pio), ==, 0);
1603         ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
1604
1605         /* change snapused */
1606         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1607             -poa.used, -poa.comp, -poa.uncomp, tx);
1608
1609         /* swap next's deadlist to our deadlist */
1610         dsl_deadlist_close(&ds->ds_deadlist);
1611         dsl_deadlist_close(&ds_next->ds_deadlist);
1612         SWITCH64(ds_next->ds_phys->ds_deadlist_obj,
1613             ds->ds_phys->ds_deadlist_obj);
1614         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
1615         dsl_deadlist_open(&ds_next->ds_deadlist, mos,
1616             ds_next->ds_phys->ds_deadlist_obj);
1617 }
1618
1619 void
1620 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1621 {
1622         struct dsl_ds_destroyarg *dsda = arg1;
1623         dsl_dataset_t *ds = dsda->ds;
1624         int err;
1625         int after_branch_point = FALSE;
1626         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1627         objset_t *mos = dp->dp_meta_objset;
1628         dsl_dataset_t *ds_prev = NULL;
1629         boolean_t wont_destroy;
1630         uint64_t obj;
1631
1632         wont_destroy = (dsda->defer &&
1633             (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1));
1634
1635         ASSERT(ds->ds_owner || wont_destroy);
1636         ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1637         ASSERT(ds->ds_prev == NULL ||
1638             ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1639         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1640
1641         if (wont_destroy) {
1642                 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1643                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1644                 ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1645                 return;
1646         }
1647
1648         /* signal any waiters that this dataset is going away */
1649         mutex_enter(&ds->ds_lock);
1650         ds->ds_owner = dsl_reaper;
1651         cv_broadcast(&ds->ds_exclusive_cv);
1652         mutex_exit(&ds->ds_lock);
1653
1654         /* Remove our reservation */
1655         if (ds->ds_reserved != 0) {
1656                 dsl_prop_setarg_t psa;
1657                 uint64_t value = 0;
1658
1659                 dsl_prop_setarg_init_uint64(&psa, "refreservation",
1660                     (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
1661                     &value);
1662                 psa.psa_effective_value = 0;    /* predict default value */
1663
1664                 dsl_dataset_set_reservation_sync(ds, &psa, tx);
1665                 ASSERT3U(ds->ds_reserved, ==, 0);
1666         }
1667
1668         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1669
1670         dsl_scan_ds_destroyed(ds, tx);
1671
1672         obj = ds->ds_object;
1673
1674         if (ds->ds_phys->ds_prev_snap_obj != 0) {
1675                 if (ds->ds_prev) {
1676                         ds_prev = ds->ds_prev;
1677                 } else {
1678                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1679                             ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1680                 }
1681                 after_branch_point =
1682                     (ds_prev->ds_phys->ds_next_snap_obj != obj);
1683
1684                 dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1685                 if (after_branch_point &&
1686                     ds_prev->ds_phys->ds_next_clones_obj != 0) {
1687                         remove_from_next_clones(ds_prev, obj, tx);
1688                         if (ds->ds_phys->ds_next_snap_obj != 0) {
1689                                 VERIFY(0 == zap_add_int(mos,
1690                                     ds_prev->ds_phys->ds_next_clones_obj,
1691                                     ds->ds_phys->ds_next_snap_obj, tx));
1692                         }
1693                 }
1694                 if (after_branch_point &&
1695                     ds->ds_phys->ds_next_snap_obj == 0) {
1696                         /* This clone is toast. */
1697                         ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1698                         ds_prev->ds_phys->ds_num_children--;
1699
1700                         /*
1701                          * If the clone's origin has no other clones, no
1702                          * user holds, and has been marked for deferred
1703                          * deletion, then we should have done the necessary
1704                          * destroy setup for it.
1705                          */
1706                         if (ds_prev->ds_phys->ds_num_children == 1 &&
1707                             ds_prev->ds_userrefs == 0 &&
1708                             DS_IS_DEFER_DESTROY(ds_prev)) {
1709                                 ASSERT3P(dsda->rm_origin, !=, NULL);
1710                         } else {
1711                                 ASSERT3P(dsda->rm_origin, ==, NULL);
1712                         }
1713                 } else if (!after_branch_point) {
1714                         ds_prev->ds_phys->ds_next_snap_obj =
1715                             ds->ds_phys->ds_next_snap_obj;
1716                 }
1717         }
1718
1719         if (dsl_dataset_is_snapshot(ds)) {
1720                 dsl_dataset_t *ds_next;
1721                 uint64_t old_unique;
1722                 uint64_t used = 0, comp = 0, uncomp = 0;
1723
1724                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1725                     ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1726                 ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1727
1728                 old_unique = ds_next->ds_phys->ds_unique_bytes;
1729
1730                 dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1731                 ds_next->ds_phys->ds_prev_snap_obj =
1732                     ds->ds_phys->ds_prev_snap_obj;
1733                 ds_next->ds_phys->ds_prev_snap_txg =
1734                     ds->ds_phys->ds_prev_snap_txg;
1735                 ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1736                     ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1737
1738
1739                 if (ds_next->ds_deadlist.dl_oldfmt) {
1740                         process_old_deadlist(ds, ds_prev, ds_next,
1741                             after_branch_point, tx);
1742                 } else {
1743                         /* Adjust prev's unique space. */
1744                         if (ds_prev && !after_branch_point) {
1745                                 dsl_deadlist_space_range(&ds_next->ds_deadlist,
1746                                     ds_prev->ds_phys->ds_prev_snap_txg,
1747                                     ds->ds_phys->ds_prev_snap_txg,
1748                                     &used, &comp, &uncomp);
1749                                 ds_prev->ds_phys->ds_unique_bytes += used;
1750                         }
1751
1752                         /* Adjust snapused. */
1753                         dsl_deadlist_space_range(&ds_next->ds_deadlist,
1754                             ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
1755                             &used, &comp, &uncomp);
1756                         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1757                             -used, -comp, -uncomp, tx);
1758
1759                         /* Move blocks to be freed to pool's free list. */
1760                         dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
1761                             &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
1762                             tx);
1763                         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
1764                             DD_USED_HEAD, used, comp, uncomp, tx);
1765                         dsl_dir_dirty(tx->tx_pool->dp_free_dir, tx);
1766
1767                         /* Merge our deadlist into next's and free it. */
1768                         dsl_deadlist_merge(&ds_next->ds_deadlist,
1769                             ds->ds_phys->ds_deadlist_obj, tx);
1770                 }
1771                 dsl_deadlist_close(&ds->ds_deadlist);
1772                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1773
1774                 /* Collapse range in clone heads */
1775                 dsl_dataset_remove_clones_key(ds,
1776                     ds->ds_phys->ds_creation_txg, tx);
1777
1778                 if (dsl_dataset_is_snapshot(ds_next)) {
1779                         dsl_dataset_t *ds_nextnext;
1780
1781                         /*
1782                          * Update next's unique to include blocks which
1783                          * were previously shared by only this snapshot
1784                          * and it.  Those blocks will be born after the
1785                          * prev snap and before this snap, and will have
1786                          * died after the next snap and before the one
1787                          * after that (ie. be on the snap after next's
1788                          * deadlist).
1789                          */
1790                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1791                             ds_next->ds_phys->ds_next_snap_obj,
1792                             FTAG, &ds_nextnext));
1793                         dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
1794                             ds->ds_phys->ds_prev_snap_txg,
1795                             ds->ds_phys->ds_creation_txg,
1796                             &used, &comp, &uncomp);
1797                         ds_next->ds_phys->ds_unique_bytes += used;
1798                         dsl_dataset_rele(ds_nextnext, FTAG);
1799                         ASSERT3P(ds_next->ds_prev, ==, NULL);
1800
1801                         /* Collapse range in this head. */
1802                         dsl_dataset_t *hds;
1803                         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
1804                             ds->ds_dir->dd_phys->dd_head_dataset_obj,
1805                             FTAG, &hds));
1806                         dsl_deadlist_remove_key(&hds->ds_deadlist,
1807                             ds->ds_phys->ds_creation_txg, tx);
1808                         dsl_dataset_rele(hds, FTAG);
1809
1810                 } else {
1811                         ASSERT3P(ds_next->ds_prev, ==, ds);
1812                         dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1813                         ds_next->ds_prev = NULL;
1814                         if (ds_prev) {
1815                                 VERIFY(0 == dsl_dataset_get_ref(dp,
1816                                     ds->ds_phys->ds_prev_snap_obj,
1817                                     ds_next, &ds_next->ds_prev));
1818                         }
1819
1820                         dsl_dataset_recalc_head_uniq(ds_next);
1821
1822                         /*
1823                          * Reduce the amount of our unconsmed refreservation
1824                          * being charged to our parent by the amount of
1825                          * new unique data we have gained.
1826                          */
1827                         if (old_unique < ds_next->ds_reserved) {
1828                                 int64_t mrsdelta;
1829                                 uint64_t new_unique =
1830                                     ds_next->ds_phys->ds_unique_bytes;
1831
1832                                 ASSERT(old_unique <= new_unique);
1833                                 mrsdelta = MIN(new_unique - old_unique,
1834                                     ds_next->ds_reserved - old_unique);
1835                                 dsl_dir_diduse_space(ds->ds_dir,
1836                                     DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1837                         }
1838                 }
1839                 dsl_dataset_rele(ds_next, FTAG);
1840         } else {
1841                 /*
1842                  * There's no next snapshot, so this is a head dataset.
1843                  * Destroy the deadlist.  Unless it's a clone, the
1844                  * deadlist should be empty.  (If it's a clone, it's
1845                  * safe to ignore the deadlist contents.)
1846                  */
1847                 struct killarg ka;
1848
1849                 dsl_deadlist_close(&ds->ds_deadlist);
1850                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1851                 ds->ds_phys->ds_deadlist_obj = 0;
1852
1853                 /*
1854                  * Free everything that we point to (that's born after
1855                  * the previous snapshot, if we are a clone)
1856                  *
1857                  * NB: this should be very quick, because we already
1858                  * freed all the objects in open context.
1859                  */
1860                 ka.ds = ds;
1861                 ka.tx = tx;
1862                 err = traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1863                     TRAVERSE_POST, kill_blkptr, &ka);
1864                 ASSERT3U(err, ==, 0);
1865                 ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1866                     ds->ds_phys->ds_unique_bytes == 0);
1867
1868                 if (ds->ds_prev != NULL) {
1869                         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
1870                                 VERIFY3U(0, ==, zap_remove_int(mos,
1871                                     ds->ds_prev->ds_dir->dd_phys->dd_clones,
1872                                     ds->ds_object, tx));
1873                         }
1874                         dsl_dataset_rele(ds->ds_prev, ds);
1875                         ds->ds_prev = ds_prev = NULL;
1876                 }
1877         }
1878
1879         /*
1880          * This must be done after the dsl_traverse(), because it will
1881          * re-open the objset.
1882          */
1883         if (ds->ds_objset) {
1884                 dmu_objset_evict(ds->ds_objset);
1885                 ds->ds_objset = NULL;
1886         }
1887
1888         if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1889                 /* Erase the link in the dir */
1890                 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1891                 ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1892                 ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1893                 err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1894                 ASSERT(err == 0);
1895         } else {
1896                 /* remove from snapshot namespace */
1897                 dsl_dataset_t *ds_head;
1898                 ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1899                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1900                     ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1901                 VERIFY(0 == dsl_dataset_get_snapname(ds));
1902 #ifdef ZFS_DEBUG
1903                 {
1904                         uint64_t val;
1905
1906                         err = dsl_dataset_snap_lookup(ds_head,
1907                             ds->ds_snapname, &val);
1908                         ASSERT3U(err, ==, 0);
1909                         ASSERT3U(val, ==, obj);
1910                 }
1911 #endif
1912                 err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1913                 ASSERT(err == 0);
1914                 dsl_dataset_rele(ds_head, FTAG);
1915         }
1916
1917         if (ds_prev && ds->ds_prev != ds_prev)
1918                 dsl_dataset_rele(ds_prev, FTAG);
1919
1920         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1921         spa_history_log_internal(LOG_DS_DESTROY, dp->dp_spa, tx,
1922             "dataset = %llu", ds->ds_object);
1923
1924         if (ds->ds_phys->ds_next_clones_obj != 0) {
1925                 uint64_t count;
1926                 ASSERT(0 == zap_count(mos,
1927                     ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1928                 VERIFY(0 == dmu_object_free(mos,
1929                     ds->ds_phys->ds_next_clones_obj, tx));
1930         }
1931         if (ds->ds_phys->ds_props_obj != 0)
1932                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1933         if (ds->ds_phys->ds_userrefs_obj != 0)
1934                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1935         dsl_dir_close(ds->ds_dir, ds);
1936         ds->ds_dir = NULL;
1937         dsl_dataset_drain_refs(ds, tag);
1938         VERIFY(0 == dmu_object_free(mos, obj, tx));
1939
1940         if (dsda->rm_origin) {
1941                 /*
1942                  * Remove the origin of the clone we just destroyed.
1943                  */
1944                 struct dsl_ds_destroyarg ndsda = {0};
1945
1946                 ndsda.ds = dsda->rm_origin;
1947                 dsl_dataset_destroy_sync(&ndsda, tag, tx);
1948         }
1949 }
1950
1951 static int
1952 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
1953 {
1954         uint64_t asize;
1955
1956         if (!dmu_tx_is_syncing(tx))
1957                 return (0);
1958
1959         /*
1960          * If there's an fs-only reservation, any blocks that might become
1961          * owned by the snapshot dataset must be accommodated by space
1962          * outside of the reservation.
1963          */
1964         ASSERT(ds->ds_reserved == 0 || DS_UNIQUE_IS_ACCURATE(ds));
1965         asize = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
1966         if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
1967                 return (ENOSPC);
1968
1969         /*
1970          * Propogate any reserved space for this snapshot to other
1971          * snapshot checks in this sync group.
1972          */
1973         if (asize > 0)
1974                 dsl_dir_willuse_space(ds->ds_dir, asize, tx);
1975
1976         return (0);
1977 }
1978
1979 int
1980 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1981 {
1982         dsl_dataset_t *ds = arg1;
1983         const char *snapname = arg2;
1984         int err;
1985         uint64_t value;
1986
1987         /*
1988          * We don't allow multiple snapshots of the same txg.  If there
1989          * is already one, try again.
1990          */
1991         if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
1992                 return (EAGAIN);
1993
1994         /*
1995          * Check for conflicting name snapshot name.
1996          */
1997         err = dsl_dataset_snap_lookup(ds, snapname, &value);
1998         if (err == 0)
1999                 return (EEXIST);
2000         if (err != ENOENT)
2001                 return (err);
2002
2003         /*
2004          * Check that the dataset's name is not too long.  Name consists
2005          * of the dataset's length + 1 for the @-sign + snapshot name's length
2006          */
2007         if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
2008                 return (ENAMETOOLONG);
2009
2010         err = dsl_dataset_snapshot_reserve_space(ds, tx);
2011         if (err)
2012                 return (err);
2013
2014         ds->ds_trysnap_txg = tx->tx_txg;
2015         return (0);
2016 }
2017
2018 void
2019 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2020 {
2021         dsl_dataset_t *ds = arg1;
2022         const char *snapname = arg2;
2023         dsl_pool_t *dp = ds->ds_dir->dd_pool;
2024         dmu_buf_t *dbuf;
2025         dsl_dataset_phys_t *dsphys;
2026         uint64_t dsobj, crtxg;
2027         objset_t *mos = dp->dp_meta_objset;
2028         int err;
2029
2030         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
2031
2032         /*
2033          * The origin's ds_creation_txg has to be < TXG_INITIAL
2034          */
2035         if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2036                 crtxg = 1;
2037         else
2038                 crtxg = tx->tx_txg;
2039
2040         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2041             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2042         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2043         dmu_buf_will_dirty(dbuf, tx);
2044         dsphys = dbuf->db_data;
2045         bzero(dsphys, sizeof (dsl_dataset_phys_t));
2046         dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2047         dsphys->ds_fsid_guid = unique_create();
2048         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2049             sizeof (dsphys->ds_guid));
2050         dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2051         dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2052         dsphys->ds_next_snap_obj = ds->ds_object;
2053         dsphys->ds_num_children = 1;
2054         dsphys->ds_creation_time = gethrestime_sec();
2055         dsphys->ds_creation_txg = crtxg;
2056         dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2057         dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
2058         dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2059         dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2060         dsphys->ds_flags = ds->ds_phys->ds_flags;
2061         dsphys->ds_bp = ds->ds_phys->ds_bp;
2062         dmu_buf_rele(dbuf, FTAG);
2063
2064         ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2065         if (ds->ds_prev) {
2066                 uint64_t next_clones_obj =
2067                     ds->ds_prev->ds_phys->ds_next_clones_obj;
2068                 ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2069                     ds->ds_object ||
2070                     ds->ds_prev->ds_phys->ds_num_children > 1);
2071                 if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2072                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2073                         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2074                             ds->ds_prev->ds_phys->ds_creation_txg);
2075                         ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2076                 } else if (next_clones_obj != 0) {
2077                         remove_from_next_clones(ds->ds_prev,
2078                             dsphys->ds_next_snap_obj, tx);
2079                         VERIFY3U(0, ==, zap_add_int(mos,
2080                             next_clones_obj, dsobj, tx));
2081                 }
2082         }
2083
2084         /*
2085          * If we have a reference-reservation on this dataset, we will
2086          * need to increase the amount of refreservation being charged
2087          * since our unique space is going to zero.
2088          */
2089         if (ds->ds_reserved) {
2090                 int64_t delta;
2091                 ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
2092                 delta = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2093                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2094                     delta, 0, 0, tx);
2095         }
2096
2097         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2098         zfs_dbgmsg("taking snapshot %s@%s/%llu; newkey=%llu",
2099             ds->ds_dir->dd_myname, snapname, dsobj,
2100             ds->ds_phys->ds_prev_snap_txg);
2101         ds->ds_phys->ds_deadlist_obj = dsl_deadlist_clone(&ds->ds_deadlist,
2102             UINT64_MAX, ds->ds_phys->ds_prev_snap_obj, tx);
2103         dsl_deadlist_close(&ds->ds_deadlist);
2104         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
2105         dsl_deadlist_add_key(&ds->ds_deadlist,
2106             ds->ds_phys->ds_prev_snap_txg, tx);
2107
2108         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2109         ds->ds_phys->ds_prev_snap_obj = dsobj;
2110         ds->ds_phys->ds_prev_snap_txg = crtxg;
2111         ds->ds_phys->ds_unique_bytes = 0;
2112         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2113                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2114
2115         err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2116             snapname, 8, 1, &dsobj, tx);
2117         ASSERT(err == 0);
2118
2119         if (ds->ds_prev)
2120                 dsl_dataset_drop_ref(ds->ds_prev, ds);
2121         VERIFY(0 == dsl_dataset_get_ref(dp,
2122             ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2123
2124         dsl_scan_ds_snapshotted(ds, tx);
2125
2126         dsl_dir_snap_cmtime_update(ds->ds_dir);
2127
2128         spa_history_log_internal(LOG_DS_SNAPSHOT, dp->dp_spa, tx,
2129             "dataset = %llu", dsobj);
2130 }
2131
2132 void
2133 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2134 {
2135         ASSERT(dmu_tx_is_syncing(tx));
2136         ASSERT(ds->ds_objset != NULL);
2137         ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2138
2139         /*
2140          * in case we had to change ds_fsid_guid when we opened it,
2141          * sync it out now.
2142          */
2143         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2144         ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2145
2146         dsl_dir_dirty(ds->ds_dir, tx);
2147         dmu_objset_sync(ds->ds_objset, zio, tx);
2148 }
2149
2150 void
2151 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2152 {
2153         uint64_t refd, avail, uobjs, aobjs;
2154
2155         dsl_dir_stats(ds->ds_dir, nv);
2156
2157         dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2158         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2159         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2160
2161         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2162             ds->ds_phys->ds_creation_time);
2163         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2164             ds->ds_phys->ds_creation_txg);
2165         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2166             ds->ds_quota);
2167         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2168             ds->ds_reserved);
2169         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2170             ds->ds_phys->ds_guid);
2171         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_UNIQUE,
2172             ds->ds_phys->ds_unique_bytes);
2173         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_OBJSETID,
2174             ds->ds_object);
2175         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS,
2176             ds->ds_userrefs);
2177         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2178             DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2179
2180         if (ds->ds_phys->ds_next_snap_obj) {
2181                 /*
2182                  * This is a snapshot; override the dd's space used with
2183                  * our unique space and compression ratio.
2184                  */
2185                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2186                     ds->ds_phys->ds_unique_bytes);
2187                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO,
2188                     ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2189                     (ds->ds_phys->ds_uncompressed_bytes * 100 /
2190                     ds->ds_phys->ds_compressed_bytes));
2191         }
2192 }
2193
2194 void
2195 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2196 {
2197         stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2198         stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2199         stat->dds_guid = ds->ds_phys->ds_guid;
2200         if (ds->ds_phys->ds_next_snap_obj) {
2201                 stat->dds_is_snapshot = B_TRUE;
2202                 stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2203         } else {
2204                 stat->dds_is_snapshot = B_FALSE;
2205                 stat->dds_num_clones = 0;
2206         }
2207
2208         /* clone origin is really a dsl_dir thing... */
2209         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2210         if (dsl_dir_is_clone(ds->ds_dir)) {
2211                 dsl_dataset_t *ods;
2212
2213                 VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2214                     ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2215                 dsl_dataset_name(ods, stat->dds_origin);
2216                 dsl_dataset_drop_ref(ods, FTAG);
2217         } else {
2218                 stat->dds_origin[0] = '\0';
2219         }
2220         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2221 }
2222
2223 uint64_t
2224 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2225 {
2226         return (ds->ds_fsid_guid);
2227 }
2228
2229 void
2230 dsl_dataset_space(dsl_dataset_t *ds,
2231     uint64_t *refdbytesp, uint64_t *availbytesp,
2232     uint64_t *usedobjsp, uint64_t *availobjsp)
2233 {
2234         *refdbytesp = ds->ds_phys->ds_used_bytes;
2235         *availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2236         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2237                 *availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2238         if (ds->ds_quota != 0) {
2239                 /*
2240                  * Adjust available bytes according to refquota
2241                  */
2242                 if (*refdbytesp < ds->ds_quota)
2243                         *availbytesp = MIN(*availbytesp,
2244                             ds->ds_quota - *refdbytesp);
2245                 else
2246                         *availbytesp = 0;
2247         }
2248         *usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2249         *availobjsp = DN_MAX_OBJECT - *usedobjsp;
2250 }
2251
2252 boolean_t
2253 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2254 {
2255         dsl_pool_t *dp = ds->ds_dir->dd_pool;
2256
2257         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2258             dsl_pool_sync_context(dp));
2259         if (ds->ds_prev == NULL)
2260                 return (B_FALSE);
2261         if (ds->ds_phys->ds_bp.blk_birth >
2262             ds->ds_prev->ds_phys->ds_creation_txg) {
2263                 objset_t *os, *os_prev;
2264                 /*
2265                  * It may be that only the ZIL differs, because it was
2266                  * reset in the head.  Don't count that as being
2267                  * modified.
2268                  */
2269                 if (dmu_objset_from_ds(ds, &os) != 0)
2270                         return (B_TRUE);
2271                 if (dmu_objset_from_ds(ds->ds_prev, &os_prev) != 0)
2272                         return (B_TRUE);
2273                 return (bcmp(&os->os_phys->os_meta_dnode,
2274                     &os_prev->os_phys->os_meta_dnode,
2275                     sizeof (os->os_phys->os_meta_dnode)) != 0);
2276         }
2277         return (B_FALSE);
2278 }
2279
2280 /* ARGSUSED */
2281 static int
2282 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2283 {
2284         dsl_dataset_t *ds = arg1;
2285         char *newsnapname = arg2;
2286         dsl_dir_t *dd = ds->ds_dir;
2287         dsl_dataset_t *hds;
2288         uint64_t val;
2289         int err;
2290
2291         err = dsl_dataset_hold_obj(dd->dd_pool,
2292             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2293         if (err)
2294                 return (err);
2295
2296         /* new name better not be in use */
2297         err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2298         dsl_dataset_rele(hds, FTAG);
2299
2300         if (err == 0)
2301                 err = EEXIST;
2302         else if (err == ENOENT)
2303                 err = 0;
2304
2305         /* dataset name + 1 for the "@" + the new snapshot name must fit */
2306         if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2307                 err = ENAMETOOLONG;
2308
2309         return (err);
2310 }
2311
2312 static void
2313 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2314 {
2315         char oldname[MAXPATHLEN], newname[MAXPATHLEN];
2316         dsl_dataset_t *ds = arg1;
2317         const char *newsnapname = arg2;
2318         dsl_dir_t *dd = ds->ds_dir;
2319         objset_t *mos = dd->dd_pool->dp_meta_objset;
2320         dsl_dataset_t *hds;
2321         int err;
2322
2323         ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2324
2325         VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2326             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2327
2328         VERIFY(0 == dsl_dataset_get_snapname(ds));
2329         err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2330         ASSERT3U(err, ==, 0);
2331         dsl_dataset_name(ds, oldname);
2332         mutex_enter(&ds->ds_lock);
2333         (void) strcpy(ds->ds_snapname, newsnapname);
2334         mutex_exit(&ds->ds_lock);
2335         err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2336             ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2337         ASSERT3U(err, ==, 0);
2338         dsl_dataset_name(ds, newname);
2339 #ifdef _KERNEL
2340         zvol_rename_minors(oldname, newname);
2341 #endif
2342
2343         spa_history_log_internal(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2344             "dataset = %llu", ds->ds_object);
2345         dsl_dataset_rele(hds, FTAG);
2346 }
2347
2348 struct renamesnaparg {
2349         dsl_sync_task_group_t *dstg;
2350         char failed[MAXPATHLEN];
2351         char *oldsnap;
2352         char *newsnap;
2353 };
2354
2355 static int
2356 dsl_snapshot_rename_one(const char *name, void *arg)
2357 {
2358         struct renamesnaparg *ra = arg;
2359         dsl_dataset_t *ds = NULL;
2360         char *snapname;
2361         int err;
2362
2363         snapname = kmem_asprintf("%s@%s", name, ra->oldsnap);
2364         (void) strlcpy(ra->failed, snapname, sizeof (ra->failed));
2365
2366         /*
2367          * For recursive snapshot renames the parent won't be changing
2368          * so we just pass name for both the to/from argument.
2369          */
2370         err = zfs_secpolicy_rename_perms(snapname, snapname, CRED());
2371         if (err != 0) {
2372                 strfree(snapname);
2373                 return (err == ENOENT ? 0 : err);
2374         }
2375
2376 #ifdef _KERNEL
2377         /*
2378          * For all filesystems undergoing rename, we'll need to unmount it.
2379          */
2380         (void) zfs_unmount_snap(snapname, NULL);
2381 #endif
2382         err = dsl_dataset_hold(snapname, ra->dstg, &ds);
2383         strfree(snapname);
2384         if (err != 0)
2385                 return (err == ENOENT ? 0 : err);
2386
2387         dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2388             dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2389
2390         return (0);
2391 }
2392
2393 static int
2394 dsl_recursive_rename(char *oldname, const char *newname)
2395 {
2396         int err;
2397         struct renamesnaparg *ra;
2398         dsl_sync_task_t *dst;
2399         spa_t *spa;
2400         char *cp, *fsname = spa_strdup(oldname);
2401         int len = strlen(oldname) + 1;
2402
2403         /* truncate the snapshot name to get the fsname */
2404         cp = strchr(fsname, '@');
2405         *cp = '\0';
2406
2407         err = spa_open(fsname, &spa, FTAG);
2408         if (err) {
2409                 kmem_free(fsname, len);
2410                 return (err);
2411         }
2412         ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2413         ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2414
2415         ra->oldsnap = strchr(oldname, '@') + 1;
2416         ra->newsnap = strchr(newname, '@') + 1;
2417         *ra->failed = '\0';
2418
2419         err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2420             DS_FIND_CHILDREN);
2421         kmem_free(fsname, len);
2422
2423         if (err == 0) {
2424                 err = dsl_sync_task_group_wait(ra->dstg);
2425         }
2426
2427         for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2428             dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2429                 dsl_dataset_t *ds = dst->dst_arg1;
2430                 if (dst->dst_err) {
2431                         dsl_dir_name(ds->ds_dir, ra->failed);
2432                         (void) strlcat(ra->failed, "@", sizeof (ra->failed));
2433                         (void) strlcat(ra->failed, ra->newsnap,
2434                             sizeof (ra->failed));
2435                 }
2436                 dsl_dataset_rele(ds, ra->dstg);
2437         }
2438
2439         if (err)
2440                 (void) strlcpy(oldname, ra->failed, sizeof (ra->failed));
2441
2442         dsl_sync_task_group_destroy(ra->dstg);
2443         kmem_free(ra, sizeof (struct renamesnaparg));
2444         spa_close(spa, FTAG);
2445         return (err);
2446 }
2447
2448 static int
2449 dsl_valid_rename(const char *oldname, void *arg)
2450 {
2451         int delta = *(int *)arg;
2452
2453         if (strlen(oldname) + delta >= MAXNAMELEN)
2454                 return (ENAMETOOLONG);
2455
2456         return (0);
2457 }
2458
2459 #pragma weak dmu_objset_rename = dsl_dataset_rename
2460 int
2461 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2462 {
2463         dsl_dir_t *dd;
2464         dsl_dataset_t *ds;
2465         const char *tail;
2466         int err;
2467
2468         err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2469         if (err)
2470                 return (err);
2471
2472         if (tail == NULL) {
2473                 int delta = strlen(newname) - strlen(oldname);
2474
2475                 /* if we're growing, validate child name lengths */
2476                 if (delta > 0)
2477                         err = dmu_objset_find(oldname, dsl_valid_rename,
2478                             &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2479
2480                 if (err == 0)
2481                         err = dsl_dir_rename(dd, newname);
2482                 dsl_dir_close(dd, FTAG);
2483                 return (err);
2484         }
2485
2486         if (tail[0] != '@') {
2487                 /* the name ended in a nonexistent component */
2488                 dsl_dir_close(dd, FTAG);
2489                 return (ENOENT);
2490         }
2491
2492         dsl_dir_close(dd, FTAG);
2493
2494         /* new name must be snapshot in same filesystem */
2495         tail = strchr(newname, '@');
2496         if (tail == NULL)
2497                 return (EINVAL);
2498         tail++;
2499         if (strncmp(oldname, newname, tail - newname) != 0)
2500                 return (EXDEV);
2501
2502         if (recursive) {
2503                 err = dsl_recursive_rename(oldname, newname);
2504         } else {
2505                 err = dsl_dataset_hold(oldname, FTAG, &ds);
2506                 if (err)
2507                         return (err);
2508
2509                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2510                     dsl_dataset_snapshot_rename_check,
2511                     dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2512
2513                 dsl_dataset_rele(ds, FTAG);
2514         }
2515
2516         return (err);
2517 }
2518
2519 struct promotenode {
2520         list_node_t link;
2521         dsl_dataset_t *ds;
2522 };
2523
2524 struct promotearg {
2525         list_t shared_snaps, origin_snaps, clone_snaps;
2526         dsl_dataset_t *origin_origin;
2527         uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2528         char *err_ds;
2529 };
2530
2531 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2532 static boolean_t snaplist_unstable(list_t *l);
2533
2534 static int
2535 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2536 {
2537         dsl_dataset_t *hds = arg1;
2538         struct promotearg *pa = arg2;
2539         struct promotenode *snap = list_head(&pa->shared_snaps);
2540         dsl_dataset_t *origin_ds = snap->ds;
2541         int err;
2542         uint64_t unused;
2543
2544         /* Check that it is a real clone */
2545         if (!dsl_dir_is_clone(hds->ds_dir))
2546                 return (EINVAL);
2547
2548         /* Since this is so expensive, don't do the preliminary check */
2549         if (!dmu_tx_is_syncing(tx))
2550                 return (0);
2551
2552         if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2553                 return (EXDEV);
2554
2555         /* compute origin's new unique space */
2556         snap = list_tail(&pa->clone_snaps);
2557         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2558         dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2559             origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
2560             &pa->unique, &unused, &unused);
2561
2562         /*
2563          * Walk the snapshots that we are moving
2564          *
2565          * Compute space to transfer.  Consider the incremental changes
2566          * to used for each snapshot:
2567          * (my used) = (prev's used) + (blocks born) - (blocks killed)
2568          * So each snapshot gave birth to:
2569          * (blocks born) = (my used) - (prev's used) + (blocks killed)
2570          * So a sequence would look like:
2571          * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2572          * Which simplifies to:
2573          * uN + kN + kN-1 + ... + k1 + k0
2574          * Note however, if we stop before we reach the ORIGIN we get:
2575          * uN + kN + kN-1 + ... + kM - uM-1
2576          */
2577         pa->used = origin_ds->ds_phys->ds_used_bytes;
2578         pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2579         pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2580         for (snap = list_head(&pa->shared_snaps); snap;
2581             snap = list_next(&pa->shared_snaps, snap)) {
2582                 uint64_t val, dlused, dlcomp, dluncomp;
2583                 dsl_dataset_t *ds = snap->ds;
2584
2585                 /* Check that the snapshot name does not conflict */
2586                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2587                 err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2588                 if (err == 0) {
2589                         err = EEXIST;
2590                         goto out;
2591                 }
2592                 if (err != ENOENT)
2593                         goto out;
2594
2595                 /* The very first snapshot does not have a deadlist */
2596                 if (ds->ds_phys->ds_prev_snap_obj == 0)
2597                         continue;
2598
2599                 dsl_deadlist_space(&ds->ds_deadlist,
2600                     &dlused, &dlcomp, &dluncomp);
2601                 pa->used += dlused;
2602                 pa->comp += dlcomp;
2603                 pa->uncomp += dluncomp;
2604         }
2605
2606         /*
2607          * If we are a clone of a clone then we never reached ORIGIN,
2608          * so we need to subtract out the clone origin's used space.
2609          */
2610         if (pa->origin_origin) {
2611                 pa->used -= pa->origin_origin->ds_phys->ds_used_bytes;
2612                 pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2613                 pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2614         }
2615
2616         /* Check that there is enough space here */
2617         err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2618             pa->used);
2619         if (err)
2620                 return (err);
2621
2622         /*
2623          * Compute the amounts of space that will be used by snapshots
2624          * after the promotion (for both origin and clone).  For each,
2625          * it is the amount of space that will be on all of their
2626          * deadlists (that was not born before their new origin).
2627          */
2628         if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2629                 uint64_t space;
2630
2631                 /*
2632                  * Note, typically this will not be a clone of a clone,
2633                  * so dd_origin_txg will be < TXG_INITIAL, so
2634                  * these snaplist_space() -> dsl_deadlist_space_range()
2635                  * calls will be fast because they do not have to
2636                  * iterate over all bps.
2637                  */
2638                 snap = list_head(&pa->origin_snaps);
2639                 err = snaplist_space(&pa->shared_snaps,
2640                     snap->ds->ds_dir->dd_origin_txg, &pa->cloneusedsnap);
2641                 if (err)
2642                         return (err);
2643
2644                 err = snaplist_space(&pa->clone_snaps,
2645                     snap->ds->ds_dir->dd_origin_txg, &space);
2646                 if (err)
2647                         return (err);
2648                 pa->cloneusedsnap += space;
2649         }
2650         if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2651                 err = snaplist_space(&pa->origin_snaps,
2652                     origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2653                 if (err)
2654                         return (err);
2655         }
2656
2657         return (0);
2658 out:
2659         pa->err_ds =  snap->ds->ds_snapname;
2660         return (err);
2661 }
2662
2663 static void
2664 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2665 {
2666         dsl_dataset_t *hds = arg1;
2667         struct promotearg *pa = arg2;
2668         struct promotenode *snap = list_head(&pa->shared_snaps);
2669         dsl_dataset_t *origin_ds = snap->ds;
2670         dsl_dataset_t *origin_head;
2671         dsl_dir_t *dd = hds->ds_dir;
2672         dsl_pool_t *dp = hds->ds_dir->dd_pool;
2673         dsl_dir_t *odd = NULL;
2674         uint64_t oldnext_obj;
2675         int64_t delta;
2676
2677         ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2678
2679         snap = list_head(&pa->origin_snaps);
2680         origin_head = snap->ds;
2681
2682         /*
2683          * We need to explicitly open odd, since origin_ds's dd will be
2684          * changing.
2685          */
2686         VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2687             NULL, FTAG, &odd));
2688
2689         /* change origin's next snap */
2690         dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2691         oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2692         snap = list_tail(&pa->clone_snaps);
2693         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2694         origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2695
2696         /* change the origin's next clone */
2697         if (origin_ds->ds_phys->ds_next_clones_obj) {
2698                 remove_from_next_clones(origin_ds, snap->ds->ds_object, tx);
2699                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2700                     origin_ds->ds_phys->ds_next_clones_obj,
2701                     oldnext_obj, tx));
2702         }
2703
2704         /* change origin */
2705         dmu_buf_will_dirty(dd->dd_dbuf, tx);
2706         ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2707         dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2708         dd->dd_origin_txg = origin_head->ds_dir->dd_origin_txg;
2709         dmu_buf_will_dirty(odd->dd_dbuf, tx);
2710         odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2711         origin_head->ds_dir->dd_origin_txg =
2712             origin_ds->ds_phys->ds_creation_txg;
2713
2714         /* change dd_clone entries */
2715         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2716                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2717                     odd->dd_phys->dd_clones, hds->ds_object, tx));
2718                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2719                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2720                     hds->ds_object, tx));
2721
2722                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2723                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2724                     origin_head->ds_object, tx));
2725                 if (dd->dd_phys->dd_clones == 0) {
2726                         dd->dd_phys->dd_clones = zap_create(dp->dp_meta_objset,
2727                             DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
2728                 }
2729                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2730                     dd->dd_phys->dd_clones, origin_head->ds_object, tx));
2731
2732         }
2733
2734         /* move snapshots to this dir */
2735         for (snap = list_head(&pa->shared_snaps); snap;
2736             snap = list_next(&pa->shared_snaps, snap)) {
2737                 dsl_dataset_t *ds = snap->ds;
2738
2739                 /* unregister props as dsl_dir is changing */
2740                 if (ds->ds_objset) {
2741                         dmu_objset_evict(ds->ds_objset);
2742                         ds->ds_objset = NULL;
2743                 }
2744                 /* move snap name entry */
2745                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2746                 VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2747                     ds->ds_snapname, tx));
2748                 VERIFY(0 == zap_add(dp->dp_meta_objset,
2749                     hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2750                     8, 1, &ds->ds_object, tx));
2751
2752                 /* change containing dsl_dir */
2753                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2754                 ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2755                 ds->ds_phys->ds_dir_obj = dd->dd_object;
2756                 ASSERT3P(ds->ds_dir, ==, odd);
2757                 dsl_dir_close(ds->ds_dir, ds);
2758                 VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2759                     NULL, ds, &ds->ds_dir));
2760
2761                 /* move any clone references */
2762                 if (ds->ds_phys->ds_next_clones_obj &&
2763                     spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2764                         zap_cursor_t zc;
2765                         zap_attribute_t za;
2766
2767                         for (zap_cursor_init(&zc, dp->dp_meta_objset,
2768                             ds->ds_phys->ds_next_clones_obj);
2769                             zap_cursor_retrieve(&zc, &za) == 0;
2770                             zap_cursor_advance(&zc)) {
2771                                 dsl_dataset_t *cnds;
2772                                 uint64_t o;
2773
2774                                 if (za.za_first_integer == oldnext_obj) {
2775                                         /*
2776                                          * We've already moved the
2777                                          * origin's reference.
2778                                          */
2779                                         continue;
2780                                 }
2781
2782                                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
2783                                     za.za_first_integer, FTAG, &cnds));
2784                                 o = cnds->ds_dir->dd_phys->dd_head_dataset_obj;
2785
2786                                 VERIFY3U(zap_remove_int(dp->dp_meta_objset,
2787                                     odd->dd_phys->dd_clones, o, tx), ==, 0);
2788                                 VERIFY3U(zap_add_int(dp->dp_meta_objset,
2789                                     dd->dd_phys->dd_clones, o, tx), ==, 0);
2790                                 dsl_dataset_rele(cnds, FTAG);
2791                         }
2792                         zap_cursor_fini(&zc);
2793                 }
2794
2795                 ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2796         }
2797
2798         /*
2799          * Change space accounting.
2800          * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2801          * both be valid, or both be 0 (resulting in delta == 0).  This
2802          * is true for each of {clone,origin} independently.
2803          */
2804
2805         delta = pa->cloneusedsnap -
2806             dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2807         ASSERT3S(delta, >=, 0);
2808         ASSERT3U(pa->used, >=, delta);
2809         dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2810         dsl_dir_diduse_space(dd, DD_USED_HEAD,
2811             pa->used - delta, pa->comp, pa->uncomp, tx);
2812
2813         delta = pa->originusedsnap -
2814             odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2815         ASSERT3S(delta, <=, 0);
2816         ASSERT3U(pa->used, >=, -delta);
2817         dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2818         dsl_dir_diduse_space(odd, DD_USED_HEAD,
2819             -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2820
2821         origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2822
2823         /* log history record */
2824         spa_history_log_internal(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2825             "dataset = %llu", hds->ds_object);
2826
2827         dsl_dir_close(odd, FTAG);
2828 }
2829
2830 static char *snaplist_tag = "snaplist";
2831 /*
2832  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2833  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2834  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2835  * snapshots back to this dataset's origin.
2836  */
2837 static int
2838 snaplist_make(dsl_pool_t *dp, boolean_t own,
2839     uint64_t first_obj, uint64_t last_obj, list_t *l)
2840 {
2841         uint64_t obj = last_obj;
2842
2843         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2844
2845         list_create(l, sizeof (struct promotenode),
2846             offsetof(struct promotenode, link));
2847
2848         while (obj != first_obj) {
2849                 dsl_dataset_t *ds;
2850                 struct promotenode *snap;
2851                 int err;
2852
2853                 if (own) {
2854                         err = dsl_dataset_own_obj(dp, obj,
2855                             0, snaplist_tag, &ds);
2856                         if (err == 0)
2857                                 dsl_dataset_make_exclusive(ds, snaplist_tag);
2858                 } else {
2859                         err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2860                 }
2861                 if (err == ENOENT) {
2862                         /* lost race with snapshot destroy */
2863                         struct promotenode *last = list_tail(l);
2864                         ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2865                         obj = last->ds->ds_phys->ds_prev_snap_obj;
2866                         continue;
2867                 } else if (err) {
2868                         return (err);
2869                 }
2870
2871                 if (first_obj == 0)
2872                         first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2873
2874                 snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
2875                 snap->ds = ds;
2876                 list_insert_tail(l, snap);
2877                 obj = ds->ds_phys->ds_prev_snap_obj;
2878         }
2879
2880         return (0);
2881 }
2882
2883 static int
2884 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
2885 {
2886         struct promotenode *snap;
2887
2888         *spacep = 0;
2889         for (snap = list_head(l); snap; snap = list_next(l, snap)) {
2890                 uint64_t used, comp, uncomp;
2891                 dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2892                     mintxg, UINT64_MAX, &used, &comp, &uncomp);
2893                 *spacep += used;
2894         }
2895         return (0);
2896 }
2897
2898 static void
2899 snaplist_destroy(list_t *l, boolean_t own)
2900 {
2901         struct promotenode *snap;
2902
2903         if (!l || !list_link_active(&l->list_head))
2904                 return;
2905
2906         while ((snap = list_tail(l)) != NULL) {
2907                 list_remove(l, snap);
2908                 if (own)
2909                         dsl_dataset_disown(snap->ds, snaplist_tag);
2910                 else
2911                         dsl_dataset_rele(snap->ds, snaplist_tag);
2912                 kmem_free(snap, sizeof (struct promotenode));
2913         }
2914         list_destroy(l);
2915 }
2916
2917 /*
2918  * Promote a clone.  Nomenclature note:
2919  * "clone" or "cds": the original clone which is being promoted
2920  * "origin" or "ods": the snapshot which is originally clone's origin
2921  * "origin head" or "ohds": the dataset which is the head
2922  * (filesystem/volume) for the origin
2923  * "origin origin": the origin of the origin's filesystem (typically
2924  * NULL, indicating that the clone is not a clone of a clone).
2925  */
2926 int
2927 dsl_dataset_promote(const char *name, char *conflsnap)
2928 {
2929         dsl_dataset_t *ds;
2930         dsl_dir_t *dd;
2931         dsl_pool_t *dp;
2932         dmu_object_info_t doi;
2933         struct promotearg pa = { 0 };
2934         struct promotenode *snap;
2935         int err;
2936
2937         err = dsl_dataset_hold(name, FTAG, &ds);
2938         if (err)
2939                 return (err);
2940         dd = ds->ds_dir;
2941         dp = dd->dd_pool;
2942
2943         err = dmu_object_info(dp->dp_meta_objset,
2944             ds->ds_phys->ds_snapnames_zapobj, &doi);
2945         if (err) {
2946                 dsl_dataset_rele(ds, FTAG);
2947                 return (err);
2948         }
2949
2950         if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
2951                 dsl_dataset_rele(ds, FTAG);
2952                 return (EINVAL);
2953         }
2954
2955         /*
2956          * We are going to inherit all the snapshots taken before our
2957          * origin (i.e., our new origin will be our parent's origin).
2958          * Take ownership of them so that we can rename them into our
2959          * namespace.
2960          */
2961         rw_enter(&dp->dp_config_rwlock, RW_READER);
2962
2963         err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
2964             &pa.shared_snaps);
2965         if (err != 0)
2966                 goto out;
2967
2968         err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
2969         if (err != 0)
2970                 goto out;
2971
2972         snap = list_head(&pa.shared_snaps);
2973         ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
2974         err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
2975             snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
2976         if (err != 0)
2977                 goto out;
2978
2979         if (snap->ds->ds_dir->dd_phys->dd_origin_obj != 0) {
2980                 err = dsl_dataset_hold_obj(dp,
2981                     snap->ds->ds_dir->dd_phys->dd_origin_obj,
2982                     FTAG, &pa.origin_origin);
2983                 if (err != 0)
2984                         goto out;
2985         }
2986
2987 out:
2988         rw_exit(&dp->dp_config_rwlock);
2989
2990         /*
2991          * Add in 128x the snapnames zapobj size, since we will be moving
2992          * a bunch of snapnames to the promoted ds, and dirtying their
2993          * bonus buffers.
2994          */
2995         if (err == 0) {
2996                 err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
2997                     dsl_dataset_promote_sync, ds, &pa,
2998                     2 + 2 * doi.doi_physical_blocks_512);
2999                 if (err && pa.err_ds && conflsnap)
3000                         (void) strncpy(conflsnap, pa.err_ds, MAXNAMELEN);
3001         }
3002
3003         snaplist_destroy(&pa.shared_snaps, B_TRUE);
3004         snaplist_destroy(&pa.clone_snaps, B_FALSE);
3005         snaplist_destroy(&pa.origin_snaps, B_FALSE);
3006         if (pa.origin_origin)
3007                 dsl_dataset_rele(pa.origin_origin, FTAG);
3008         dsl_dataset_rele(ds, FTAG);
3009         return (err);
3010 }
3011
3012 struct cloneswaparg {
3013         dsl_dataset_t *cds; /* clone dataset */
3014         dsl_dataset_t *ohds; /* origin's head dataset */
3015         boolean_t force;
3016         int64_t unused_refres_delta; /* change in unconsumed refreservation */
3017 };
3018
3019 /* ARGSUSED */
3020 static int
3021 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
3022 {
3023         struct cloneswaparg *csa = arg1;
3024
3025         /* they should both be heads */
3026         if (dsl_dataset_is_snapshot(csa->cds) ||
3027             dsl_dataset_is_snapshot(csa->ohds))
3028                 return (EINVAL);
3029
3030         /* the branch point should be just before them */
3031         if (csa->cds->ds_prev != csa->ohds->ds_prev)
3032                 return (EINVAL);
3033
3034         /* cds should be the clone (unless they are unrelated) */
3035         if (csa->cds->ds_prev != NULL &&
3036             csa->cds->ds_prev != csa->cds->ds_dir->dd_pool->dp_origin_snap &&
3037             csa->ohds->ds_object !=
3038             csa->cds->ds_prev->ds_phys->ds_next_snap_obj)
3039                 return (EINVAL);
3040
3041         /* the clone should be a child of the origin */
3042         if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
3043                 return (EINVAL);
3044
3045         /* ohds shouldn't be modified unless 'force' */
3046         if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
3047                 return (ETXTBSY);
3048
3049         /* adjust amount of any unconsumed refreservation */
3050         csa->unused_refres_delta =
3051             (int64_t)MIN(csa->ohds->ds_reserved,
3052             csa->ohds->ds_phys->ds_unique_bytes) -
3053             (int64_t)MIN(csa->ohds->ds_reserved,
3054             csa->cds->ds_phys->ds_unique_bytes);
3055
3056         if (csa->unused_refres_delta > 0 &&
3057             csa->unused_refres_delta >
3058             dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3059                 return (ENOSPC);
3060
3061         if (csa->ohds->ds_quota != 0 &&
3062             csa->cds->ds_phys->ds_unique_bytes > csa->ohds->ds_quota)
3063                 return (EDQUOT);
3064
3065         return (0);
3066 }
3067
3068 /* ARGSUSED */
3069 static void
3070 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3071 {
3072         struct cloneswaparg *csa = arg1;
3073         dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3074
3075         ASSERT(csa->cds->ds_reserved == 0);
3076         ASSERT(csa->ohds->ds_quota == 0 ||
3077             csa->cds->ds_phys->ds_unique_bytes <= csa->ohds->ds_quota);
3078
3079         dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3080         dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3081
3082         if (csa->cds->ds_objset != NULL) {
3083                 dmu_objset_evict(csa->cds->ds_objset);
3084                 csa->cds->ds_objset = NULL;
3085         }
3086
3087         if (csa->ohds->ds_objset != NULL) {
3088                 dmu_objset_evict(csa->ohds->ds_objset);
3089                 csa->ohds->ds_objset = NULL;
3090         }
3091
3092         /*
3093          * Reset origin's unique bytes, if it exists.
3094          */
3095         if (csa->cds->ds_prev) {
3096                 dsl_dataset_t *origin = csa->cds->ds_prev;
3097                 uint64_t comp, uncomp;
3098
3099                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
3100                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3101                     origin->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3102                     &origin->ds_phys->ds_unique_bytes, &comp, &uncomp);
3103         }
3104
3105         /* swap blkptrs */
3106         {
3107                 blkptr_t tmp;
3108                 tmp = csa->ohds->ds_phys->ds_bp;
3109                 csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3110                 csa->cds->ds_phys->ds_bp = tmp;
3111         }
3112
3113         /* set dd_*_bytes */
3114         {
3115                 int64_t dused, dcomp, duncomp;
3116                 uint64_t cdl_used, cdl_comp, cdl_uncomp;
3117                 uint64_t odl_used, odl_comp, odl_uncomp;
3118
3119                 ASSERT3U(csa->cds->ds_dir->dd_phys->
3120                     dd_used_breakdown[DD_USED_SNAP], ==, 0);
3121
3122                 dsl_deadlist_space(&csa->cds->ds_deadlist,
3123                     &cdl_used, &cdl_comp, &cdl_uncomp);
3124                 dsl_deadlist_space(&csa->ohds->ds_deadlist,
3125                     &odl_used, &odl_comp, &odl_uncomp);
3126
3127                 dused = csa->cds->ds_phys->ds_used_bytes + cdl_used -
3128                     (csa->ohds->ds_phys->ds_used_bytes + odl_used);
3129                 dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3130                     (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3131                 duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3132                     cdl_uncomp -
3133                     (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3134
3135                 dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3136                     dused, dcomp, duncomp, tx);
3137                 dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3138                     -dused, -dcomp, -duncomp, tx);
3139
3140                 /*
3141                  * The difference in the space used by snapshots is the
3142                  * difference in snapshot space due to the head's
3143                  * deadlist (since that's the only thing that's
3144                  * changing that affects the snapused).
3145                  */
3146                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3147                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3148                     &cdl_used, &cdl_comp, &cdl_uncomp);
3149                 dsl_deadlist_space_range(&csa->ohds->ds_deadlist,
3150                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3151                     &odl_used, &odl_comp, &odl_uncomp);
3152                 dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3153                     DD_USED_HEAD, DD_USED_SNAP, tx);
3154         }
3155
3156         /* swap ds_*_bytes */
3157         SWITCH64(csa->ohds->ds_phys->ds_used_bytes,
3158             csa->cds->ds_phys->ds_used_bytes);
3159         SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3160             csa->cds->ds_phys->ds_compressed_bytes);
3161         SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3162             csa->cds->ds_phys->ds_uncompressed_bytes);
3163         SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3164             csa->cds->ds_phys->ds_unique_bytes);
3165
3166         /* apply any parent delta for change in unconsumed refreservation */
3167         dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3168             csa->unused_refres_delta, 0, 0, tx);
3169
3170         /*
3171          * Swap deadlists.
3172          */
3173         dsl_deadlist_close(&csa->cds->ds_deadlist);
3174         dsl_deadlist_close(&csa->ohds->ds_deadlist);
3175         SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3176             csa->cds->ds_phys->ds_deadlist_obj);
3177         dsl_deadlist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3178             csa->cds->ds_phys->ds_deadlist_obj);
3179         dsl_deadlist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3180             csa->ohds->ds_phys->ds_deadlist_obj);
3181
3182         dsl_scan_ds_clone_swapped(csa->ohds, csa->cds, tx);
3183 }
3184
3185 /*
3186  * Swap 'clone' with its origin head datasets.  Used at the end of "zfs
3187  * recv" into an existing fs to swizzle the file system to the new
3188  * version, and by "zfs rollback".  Can also be used to swap two
3189  * independent head datasets if neither has any snapshots.
3190  */
3191 int
3192 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3193     boolean_t force)
3194 {
3195         struct cloneswaparg csa;
3196         int error;
3197
3198         ASSERT(clone->ds_owner);
3199         ASSERT(origin_head->ds_owner);
3200 retry:
3201         /*
3202          * Need exclusive access for the swap. If we're swapping these
3203          * datasets back after an error, we already hold the locks.
3204          */
3205         if (!RW_WRITE_HELD(&clone->ds_rwlock))
3206                 rw_enter(&clone->ds_rwlock, RW_WRITER);
3207         if (!RW_WRITE_HELD(&origin_head->ds_rwlock) &&
3208             !rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3209                 rw_exit(&clone->ds_rwlock);
3210                 rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3211                 if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3212                         rw_exit(&origin_head->ds_rwlock);
3213                         goto retry;
3214                 }
3215         }
3216         csa.cds = clone;
3217         csa.ohds = origin_head;
3218         csa.force = force;
3219         error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3220             dsl_dataset_clone_swap_check,
3221             dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3222         return (error);
3223 }
3224
3225 /*
3226  * Given a pool name and a dataset object number in that pool,
3227  * return the name of that dataset.
3228  */
3229 int
3230 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3231 {
3232         spa_t *spa;
3233         dsl_pool_t *dp;
3234         dsl_dataset_t *ds;
3235         int error;
3236
3237         if ((error = spa_open(pname, &spa, FTAG)) != 0)
3238                 return (error);
3239         dp = spa_get_dsl(spa);
3240         rw_enter(&dp->dp_config_rwlock, RW_READER);
3241         if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3242                 dsl_dataset_name(ds, buf);
3243                 dsl_dataset_rele(ds, FTAG);
3244         }
3245         rw_exit(&dp->dp_config_rwlock);
3246         spa_close(spa, FTAG);
3247
3248         return (error);
3249 }
3250
3251 int
3252 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3253     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3254 {
3255         int error = 0;
3256
3257         ASSERT3S(asize, >, 0);
3258
3259         /*
3260          * *ref_rsrv is the portion of asize that will come from any
3261          * unconsumed refreservation space.
3262          */
3263         *ref_rsrv = 0;
3264
3265         mutex_enter(&ds->ds_lock);
3266         /*
3267          * Make a space adjustment for reserved bytes.
3268          */
3269         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3270                 ASSERT3U(*used, >=,
3271                     ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3272                 *used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3273                 *ref_rsrv =
3274                     asize - MIN(asize, parent_delta(ds, asize + inflight));
3275         }
3276
3277         if (!check_quota || ds->ds_quota == 0) {
3278                 mutex_exit(&ds->ds_lock);
3279                 return (0);
3280         }
3281         /*
3282          * If they are requesting more space, and our current estimate
3283          * is over quota, they get to try again unless the actual
3284          * on-disk is over quota and there are no pending changes (which
3285          * may free up space for us).
3286          */
3287         if (ds->ds_phys->ds_used_bytes + inflight >= ds->ds_quota) {
3288                 if (inflight > 0 || ds->ds_phys->ds_used_bytes < ds->ds_quota)
3289                         error = ERESTART;
3290                 else
3291                         error = EDQUOT;
3292         }
3293         mutex_exit(&ds->ds_lock);
3294
3295         return (error);
3296 }
3297
3298 /* ARGSUSED */
3299 static int
3300 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3301 {
3302         dsl_dataset_t *ds = arg1;
3303         dsl_prop_setarg_t *psa = arg2;
3304         int err;
3305
3306         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3307                 return (ENOTSUP);
3308
3309         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3310                 return (err);
3311
3312         if (psa->psa_effective_value == 0)
3313                 return (0);
3314
3315         if (psa->psa_effective_value < ds->ds_phys->ds_used_bytes ||
3316             psa->psa_effective_value < ds->ds_reserved)
3317                 return (ENOSPC);
3318
3319         return (0);
3320 }
3321
3322 extern void dsl_prop_set_sync(void *, void *, dmu_tx_t *);
3323
3324 void
3325 dsl_dataset_set_quota_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3326 {
3327         dsl_dataset_t *ds = arg1;
3328         dsl_prop_setarg_t *psa = arg2;
3329         uint64_t effective_value = psa->psa_effective_value;
3330
3331         dsl_prop_set_sync(ds, psa, tx);
3332         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3333
3334         if (ds->ds_quota != effective_value) {
3335                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3336                 ds->ds_quota = effective_value;
3337
3338                 spa_history_log_internal(LOG_DS_REFQUOTA,
3339                     ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu ",
3340                     (longlong_t)ds->ds_quota, ds->ds_object);
3341         }
3342 }
3343
3344 int
3345 dsl_dataset_set_quota(const char *dsname, zprop_source_t source, uint64_t quota)
3346 {
3347         dsl_dataset_t *ds;
3348         dsl_prop_setarg_t psa;
3349         int err;
3350
3351         dsl_prop_setarg_init_uint64(&psa, "refquota", source, &quota);
3352
3353         err = dsl_dataset_hold(dsname, FTAG, &ds);
3354         if (err)
3355                 return (err);
3356
3357         /*
3358          * If someone removes a file, then tries to set the quota, we
3359          * want to make sure the file freeing takes effect.
3360          */
3361         txg_wait_open(ds->ds_dir->dd_pool, 0);
3362
3363         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3364             dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3365             ds, &psa, 0);
3366
3367         dsl_dataset_rele(ds, FTAG);
3368         return (err);
3369 }
3370
3371 static int
3372 dsl_dataset_set_reservation_check(void *arg1, void *arg2, dmu_tx_t *tx)
3373 {
3374         dsl_dataset_t *ds = arg1;
3375         dsl_prop_setarg_t *psa = arg2;
3376         uint64_t effective_value;
3377         uint64_t unique;
3378         int err;
3379
3380         if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3381             SPA_VERSION_REFRESERVATION)
3382                 return (ENOTSUP);
3383
3384         if (dsl_dataset_is_snapshot(ds))
3385                 return (EINVAL);
3386
3387         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3388                 return (err);
3389
3390         effective_value = psa->psa_effective_value;
3391
3392         /*
3393          * If we are doing the preliminary check in open context, the
3394          * space estimates may be inaccurate.
3395          */
3396         if (!dmu_tx_is_syncing(tx))
3397                 return (0);
3398
3399         mutex_enter(&ds->ds_lock);
3400         if (!DS_UNIQUE_IS_ACCURATE(ds))
3401                 dsl_dataset_recalc_head_uniq(ds);
3402         unique = ds->ds_phys->ds_unique_bytes;
3403         mutex_exit(&ds->ds_lock);
3404
3405         if (MAX(unique, effective_value) > MAX(unique, ds->ds_reserved)) {
3406                 uint64_t delta = MAX(unique, effective_value) -
3407                     MAX(unique, ds->ds_reserved);
3408
3409                 if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3410                         return (ENOSPC);
3411                 if (ds->ds_quota > 0 &&
3412                     effective_value > ds->ds_quota)
3413                         return (ENOSPC);
3414         }
3415
3416         return (0);
3417 }
3418
3419 static void
3420 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3421 {
3422         dsl_dataset_t *ds = arg1;
3423         dsl_prop_setarg_t *psa = arg2;
3424         uint64_t effective_value = psa->psa_effective_value;
3425         uint64_t unique;
3426         int64_t delta;
3427
3428         dsl_prop_set_sync(ds, psa, tx);
3429         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3430
3431         dmu_buf_will_dirty(ds->ds_dbuf, tx);
3432
3433         mutex_enter(&ds->ds_dir->dd_lock);
3434         mutex_enter(&ds->ds_lock);
3435         ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
3436         unique = ds->ds_phys->ds_unique_bytes;
3437         delta = MAX(0, (int64_t)(effective_value - unique)) -
3438             MAX(0, (int64_t)(ds->ds_reserved - unique));
3439         ds->ds_reserved = effective_value;
3440         mutex_exit(&ds->ds_lock);
3441
3442         dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3443         mutex_exit(&ds->ds_dir->dd_lock);
3444
3445         spa_history_log_internal(LOG_DS_REFRESERV,
3446             ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu",
3447             (longlong_t)effective_value, ds->ds_object);
3448 }
3449
3450 int
3451 dsl_dataset_set_reservation(const char *dsname, zprop_source_t source,
3452     uint64_t reservation)
3453 {
3454         dsl_dataset_t *ds;
3455         dsl_prop_setarg_t psa;
3456         int err;
3457
3458         dsl_prop_setarg_init_uint64(&psa, "refreservation", source,
3459             &reservation);
3460
3461         err = dsl_dataset_hold(dsname, FTAG, &ds);
3462         if (err)
3463                 return (err);
3464
3465         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3466             dsl_dataset_set_reservation_check,
3467             dsl_dataset_set_reservation_sync, ds, &psa, 0);
3468
3469         dsl_dataset_rele(ds, FTAG);
3470         return (err);
3471 }
3472
3473 typedef struct zfs_hold_cleanup_arg {
3474         dsl_pool_t *dp;
3475         uint64_t dsobj;
3476         char htag[MAXNAMELEN];
3477 } zfs_hold_cleanup_arg_t;
3478
3479 static void
3480 dsl_dataset_user_release_onexit(void *arg)
3481 {
3482         zfs_hold_cleanup_arg_t *ca = arg;
3483
3484         (void) dsl_dataset_user_release_tmp(ca->dp, ca->dsobj, ca->htag,
3485             B_TRUE);
3486         kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t));
3487 }
3488
3489 void
3490 dsl_register_onexit_hold_cleanup(dsl_dataset_t *ds, const char *htag,
3491     minor_t minor)
3492 {
3493         zfs_hold_cleanup_arg_t *ca;
3494
3495         ca = kmem_alloc(sizeof (zfs_hold_cleanup_arg_t), KM_SLEEP);
3496         ca->dp = ds->ds_dir->dd_pool;
3497         ca->dsobj = ds->ds_object;
3498         (void) strlcpy(ca->htag, htag, sizeof (ca->htag));
3499         VERIFY3U(0, ==, zfs_onexit_add_cb(minor,
3500             dsl_dataset_user_release_onexit, ca, NULL));
3501 }
3502
3503 /*
3504  * If you add new checks here, you may need to add
3505  * additional checks to the "temporary" case in
3506  * snapshot_check() in dmu_objset.c.
3507  */
3508 static int
3509 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3510 {
3511         dsl_dataset_t *ds = arg1;
3512         struct dsl_ds_holdarg *ha = arg2;
3513         char *htag = ha->htag;
3514         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3515         int error = 0;
3516
3517         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3518                 return (ENOTSUP);
3519
3520         if (!dsl_dataset_is_snapshot(ds))
3521                 return (EINVAL);
3522
3523         /* tags must be unique */
3524         mutex_enter(&ds->ds_lock);
3525         if (ds->ds_phys->ds_userrefs_obj) {
3526                 error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3527                     8, 1, tx);
3528                 if (error == 0)
3529                         error = EEXIST;
3530                 else if (error == ENOENT)
3531                         error = 0;
3532         }
3533         mutex_exit(&ds->ds_lock);
3534
3535         if (error == 0 && ha->temphold &&
3536             strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
3537                 error = E2BIG;
3538
3539         return (error);
3540 }
3541
3542 void
3543 dsl_dataset_user_hold_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3544 {
3545         dsl_dataset_t *ds = arg1;
3546         struct dsl_ds_holdarg *ha = arg2;
3547         char *htag = ha->htag;
3548         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3549         objset_t *mos = dp->dp_meta_objset;
3550         uint64_t now = gethrestime_sec();
3551         uint64_t zapobj;
3552
3553         mutex_enter(&ds->ds_lock);
3554         if (ds->ds_phys->ds_userrefs_obj == 0) {
3555                 /*
3556                  * This is the first user hold for this dataset.  Create
3557                  * the userrefs zap object.
3558                  */
3559                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3560                 zapobj = ds->ds_phys->ds_userrefs_obj =
3561                     zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3562         } else {
3563                 zapobj = ds->ds_phys->ds_userrefs_obj;
3564         }
3565         ds->ds_userrefs++;
3566         mutex_exit(&ds->ds_lock);
3567
3568         VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3569
3570         if (ha->temphold) {
3571                 VERIFY(0 == dsl_pool_user_hold(dp, ds->ds_object,
3572                     htag, &now, tx));
3573         }
3574
3575         spa_history_log_internal(LOG_DS_USER_HOLD,
3576             dp->dp_spa, tx, "<%s> temp = %d dataset = %llu", htag,
3577             (int)ha->temphold, ds->ds_object);
3578 }
3579
3580 static int
3581 dsl_dataset_user_hold_one(const char *dsname, void *arg)
3582 {
3583         struct dsl_ds_holdarg *ha = arg;
3584         dsl_dataset_t *ds;
3585         int error;
3586         char *name;
3587
3588         /* alloc a buffer to hold dsname@snapname plus terminating NULL */
3589         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3590         error = dsl_dataset_hold(name, ha->dstg, &ds);
3591         strfree(name);
3592         if (error == 0) {
3593                 ha->gotone = B_TRUE;
3594                 dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3595                     dsl_dataset_user_hold_sync, ds, ha, 0);
3596         } else if (error == ENOENT && ha->recursive) {
3597                 error = 0;
3598         } else {
3599                 (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3600         }
3601         return (error);
3602 }
3603
3604 int
3605 dsl_dataset_user_hold_for_send(dsl_dataset_t *ds, char *htag,
3606     boolean_t temphold)
3607 {
3608         struct dsl_ds_holdarg *ha;
3609         int error;
3610
3611         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3612         ha->htag = htag;
3613         ha->temphold = temphold;
3614         error = dsl_sync_task_do(ds->ds_dir->dd_pool,
3615             dsl_dataset_user_hold_check, dsl_dataset_user_hold_sync,
3616             ds, ha, 0);
3617         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3618
3619         return (error);
3620 }
3621
3622 int
3623 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3624     boolean_t recursive, boolean_t temphold, int cleanup_fd)
3625 {
3626         struct dsl_ds_holdarg *ha;
3627         dsl_sync_task_t *dst;
3628         spa_t *spa;
3629         int error;
3630         minor_t minor = 0;
3631
3632         if (cleanup_fd != -1) {
3633                 /* Currently we only support cleanup-on-exit of tempholds. */
3634                 if (!temphold)
3635                         return (EINVAL);
3636                 error = zfs_onexit_fd_hold(cleanup_fd, &minor);
3637                 if (error)
3638                         return (error);
3639         }
3640
3641         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3642
3643         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3644
3645         error = spa_open(dsname, &spa, FTAG);
3646         if (error) {
3647                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3648                 if (cleanup_fd != -1)
3649                         zfs_onexit_fd_rele(cleanup_fd);
3650                 return (error);
3651         }
3652
3653         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3654         ha->htag = htag;
3655         ha->snapname = snapname;
3656         ha->recursive = recursive;
3657         ha->temphold = temphold;
3658
3659         if (recursive) {
3660                 error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3661                     ha, DS_FIND_CHILDREN);
3662         } else {
3663                 error = dsl_dataset_user_hold_one(dsname, ha);
3664         }
3665         if (error == 0)
3666                 error = dsl_sync_task_group_wait(ha->dstg);
3667
3668         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3669             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3670                 dsl_dataset_t *ds = dst->dst_arg1;
3671
3672                 if (dst->dst_err) {
3673                         dsl_dataset_name(ds, ha->failed);
3674                         *strchr(ha->failed, '@') = '\0';
3675                 } else if (error == 0 && minor != 0 && temphold) {
3676                         /*
3677                          * If this hold is to be released upon process exit,
3678                          * register that action now.
3679                          */
3680                         dsl_register_onexit_hold_cleanup(ds, htag, minor);
3681                 }
3682                 dsl_dataset_rele(ds, ha->dstg);
3683         }
3684
3685         if (error == 0 && recursive && !ha->gotone)
3686                 error = ENOENT;
3687
3688         if (error)
3689                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3690
3691         dsl_sync_task_group_destroy(ha->dstg);
3692
3693         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3694         spa_close(spa, FTAG);
3695         if (cleanup_fd != -1)
3696                 zfs_onexit_fd_rele(cleanup_fd);
3697         return (error);
3698 }
3699
3700 struct dsl_ds_releasearg {
3701         dsl_dataset_t *ds;
3702         const char *htag;
3703         boolean_t own;          /* do we own or just hold ds? */
3704 };
3705
3706 static int
3707 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3708     boolean_t *might_destroy)
3709 {
3710         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3711         uint64_t zapobj;
3712         uint64_t tmp;
3713         int error;
3714
3715         *might_destroy = B_FALSE;
3716
3717         mutex_enter(&ds->ds_lock);
3718         zapobj = ds->ds_phys->ds_userrefs_obj;
3719         if (zapobj == 0) {
3720                 /* The tag can't possibly exist */
3721                 mutex_exit(&ds->ds_lock);
3722                 return (ESRCH);
3723         }
3724
3725         /* Make sure the tag exists */
3726         error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3727         if (error) {
3728                 mutex_exit(&ds->ds_lock);
3729                 if (error == ENOENT)
3730                         error = ESRCH;
3731                 return (error);
3732         }
3733
3734         if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3735             DS_IS_DEFER_DESTROY(ds))
3736                 *might_destroy = B_TRUE;
3737
3738         mutex_exit(&ds->ds_lock);
3739         return (0);
3740 }
3741
3742 static int
3743 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3744 {
3745         struct dsl_ds_releasearg *ra = arg1;
3746         dsl_dataset_t *ds = ra->ds;
3747         boolean_t might_destroy;
3748         int error;
3749
3750         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3751                 return (ENOTSUP);
3752
3753         error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3754         if (error)
3755                 return (error);
3756
3757         if (might_destroy) {
3758                 struct dsl_ds_destroyarg dsda = {0};
3759
3760                 if (dmu_tx_is_syncing(tx)) {
3761                         /*
3762                          * If we're not prepared to remove the snapshot,
3763                          * we can't allow the release to happen right now.
3764                          */
3765                         if (!ra->own)
3766                                 return (EBUSY);
3767                 }
3768                 dsda.ds = ds;
3769                 dsda.releasing = B_TRUE;
3770                 return (dsl_dataset_destroy_check(&dsda, tag, tx));
3771         }
3772
3773         return (0);
3774 }
3775
3776 static void
3777 dsl_dataset_user_release_sync(void *arg1, void *tag, dmu_tx_t *tx)
3778 {
3779         struct dsl_ds_releasearg *ra = arg1;
3780         dsl_dataset_t *ds = ra->ds;
3781         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3782         objset_t *mos = dp->dp_meta_objset;
3783         uint64_t zapobj;
3784         uint64_t dsobj = ds->ds_object;
3785         uint64_t refs;
3786         int error;
3787
3788         mutex_enter(&ds->ds_lock);
3789         ds->ds_userrefs--;
3790         refs = ds->ds_userrefs;
3791         mutex_exit(&ds->ds_lock);
3792         error = dsl_pool_user_release(dp, ds->ds_object, ra->htag, tx);
3793         VERIFY(error == 0 || error == ENOENT);
3794         zapobj = ds->ds_phys->ds_userrefs_obj;
3795         VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3796         if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3797             DS_IS_DEFER_DESTROY(ds)) {
3798                 struct dsl_ds_destroyarg dsda = {0};
3799
3800                 ASSERT(ra->own);
3801                 dsda.ds = ds;
3802                 dsda.releasing = B_TRUE;
3803                 /* We already did the destroy_check */
3804                 dsl_dataset_destroy_sync(&dsda, tag, tx);
3805         }
3806
3807         spa_history_log_internal(LOG_DS_USER_RELEASE,
3808             dp->dp_spa, tx, "<%s> %lld dataset = %llu",
3809             ra->htag, (longlong_t)refs, dsobj);
3810 }
3811
3812 static int
3813 dsl_dataset_user_release_one(const char *dsname, void *arg)
3814 {
3815         struct dsl_ds_holdarg *ha = arg;
3816         struct dsl_ds_releasearg *ra;
3817         dsl_dataset_t *ds;
3818         int error;
3819         void *dtag = ha->dstg;
3820         char *name;
3821         boolean_t own = B_FALSE;
3822         boolean_t might_destroy;
3823
3824         /* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3825         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3826         error = dsl_dataset_hold(name, dtag, &ds);
3827         strfree(name);
3828         if (error == ENOENT && ha->recursive)
3829                 return (0);
3830         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3831         if (error)
3832                 return (error);
3833
3834         ha->gotone = B_TRUE;
3835
3836         ASSERT(dsl_dataset_is_snapshot(ds));
3837
3838         error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3839         if (error) {
3840                 dsl_dataset_rele(ds, dtag);
3841                 return (error);
3842         }
3843
3844         if (might_destroy) {
3845 #ifdef _KERNEL
3846                 name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3847                 error = zfs_unmount_snap(name, NULL);
3848                 strfree(name);
3849                 if (error) {
3850                         dsl_dataset_rele(ds, dtag);
3851                         return (error);
3852                 }
3853 #endif
3854                 if (!dsl_dataset_tryown(ds, B_TRUE, dtag)) {
3855                         dsl_dataset_rele(ds, dtag);
3856                         return (EBUSY);
3857                 } else {
3858                         own = B_TRUE;
3859                         dsl_dataset_make_exclusive(ds, dtag);
3860                 }
3861         }
3862
3863         ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3864         ra->ds = ds;
3865         ra->htag = ha->htag;
3866         ra->own = own;
3867         dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3868             dsl_dataset_user_release_sync, ra, dtag, 0);
3869
3870         return (0);
3871 }
3872
3873 int
3874 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3875     boolean_t recursive)
3876 {
3877         struct dsl_ds_holdarg *ha;
3878         dsl_sync_task_t *dst;
3879         spa_t *spa;
3880         int error;
3881
3882 top:
3883         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3884
3885         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3886
3887         error = spa_open(dsname, &spa, FTAG);
3888         if (error) {
3889                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3890                 return (error);
3891         }
3892
3893         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3894         ha->htag = htag;
3895         ha->snapname = snapname;
3896         ha->recursive = recursive;
3897         if (recursive) {
3898                 error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
3899                     ha, DS_FIND_CHILDREN);
3900         } else {
3901                 error = dsl_dataset_user_release_one(dsname, ha);
3902         }
3903         if (error == 0)
3904                 error = dsl_sync_task_group_wait(ha->dstg);
3905
3906         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3907             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3908                 struct dsl_ds_releasearg *ra = dst->dst_arg1;
3909                 dsl_dataset_t *ds = ra->ds;
3910
3911                 if (dst->dst_err)
3912                         dsl_dataset_name(ds, ha->failed);
3913
3914                 if (ra->own)
3915                         dsl_dataset_disown(ds, ha->dstg);
3916                 else
3917                         dsl_dataset_rele(ds, ha->dstg);
3918
3919                 kmem_free(ra, sizeof (struct dsl_ds_releasearg));
3920         }
3921
3922         if (error == 0 && recursive && !ha->gotone)
3923                 error = ENOENT;
3924
3925         if (error && error != EBUSY)
3926                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3927
3928         dsl_sync_task_group_destroy(ha->dstg);
3929         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3930         spa_close(spa, FTAG);
3931
3932         /*
3933          * We can get EBUSY if we were racing with deferred destroy and
3934          * dsl_dataset_user_release_check() hadn't done the necessary
3935          * open context setup.  We can also get EBUSY if we're racing
3936          * with destroy and that thread is the ds_owner.  Either way
3937          * the busy condition should be transient, and we should retry
3938          * the release operation.
3939          */
3940         if (error == EBUSY)
3941                 goto top;
3942
3943         return (error);
3944 }
3945
3946 /*
3947  * Called at spa_load time (with retry == B_FALSE) to release a stale
3948  * temporary user hold. Also called by the onexit code (with retry == B_TRUE).
3949  */
3950 int
3951 dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, char *htag,
3952     boolean_t retry)
3953 {
3954         dsl_dataset_t *ds;
3955         char *snap;
3956         char *name;
3957         int namelen;
3958         int error;
3959
3960         do {
3961                 rw_enter(&dp->dp_config_rwlock, RW_READER);
3962                 error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
3963                 rw_exit(&dp->dp_config_rwlock);
3964                 if (error)
3965                         return (error);
3966                 namelen = dsl_dataset_namelen(ds)+1;
3967                 name = kmem_alloc(namelen, KM_SLEEP);
3968                 dsl_dataset_name(ds, name);
3969                 dsl_dataset_rele(ds, FTAG);
3970
3971                 snap = strchr(name, '@');
3972                 *snap = '\0';
3973                 ++snap;
3974                 error = dsl_dataset_user_release(name, snap, htag, B_FALSE);
3975                 kmem_free(name, namelen);
3976
3977                 /*
3978                  * The object can't have been destroyed because we have a hold,
3979                  * but it might have been renamed, resulting in ENOENT.  Retry
3980                  * if we've been requested to do so.
3981                  *
3982                  * It would be nice if we could use the dsobj all the way
3983                  * through and avoid ENOENT entirely.  But we might need to
3984                  * unmount the snapshot, and there's currently no way to lookup
3985                  * a vfsp using a ZFS object id.
3986                  */
3987         } while ((error == ENOENT) && retry);
3988
3989         return (error);
3990 }
3991
3992 int
3993 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
3994 {
3995         dsl_dataset_t *ds;
3996         int err;
3997
3998         err = dsl_dataset_hold(dsname, FTAG, &ds);
3999         if (err)
4000                 return (err);
4001
4002         VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
4003         if (ds->ds_phys->ds_userrefs_obj != 0) {
4004                 zap_attribute_t *za;
4005                 zap_cursor_t zc;
4006
4007                 za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
4008                 for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
4009                     ds->ds_phys->ds_userrefs_obj);
4010                     zap_cursor_retrieve(&zc, za) == 0;
4011                     zap_cursor_advance(&zc)) {
4012                         VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
4013                             za->za_first_integer));
4014                 }
4015                 zap_cursor_fini(&zc);
4016                 kmem_free(za, sizeof (zap_attribute_t));
4017         }
4018         dsl_dataset_rele(ds, FTAG);
4019         return (0);
4020 }
4021
4022 /*
4023  * Note, this fuction is used as the callback for dmu_objset_find().  We
4024  * always return 0 so that we will continue to find and process
4025  * inconsistent datasets, even if we encounter an error trying to
4026  * process one of them.
4027  */
4028 /* ARGSUSED */
4029 int
4030 dsl_destroy_inconsistent(const char *dsname, void *arg)
4031 {
4032         dsl_dataset_t *ds;
4033
4034         if (dsl_dataset_own(dsname, B_TRUE, FTAG, &ds) == 0) {
4035                 if (DS_IS_INCONSISTENT(ds))
4036                         (void) dsl_dataset_destroy(ds, FTAG, B_FALSE);
4037                 else
4038                         dsl_dataset_disown(ds, FTAG);
4039         }
4040         return (0);
4041 }