module/zfs/dsl_dataset.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011 by Delphix. All rights reserved.
  24  */
  25
  26 #include <sys/dmu_objset.h>
  27 #include <sys/dsl_dataset.h>
  28 #include <sys/dsl_dir.h>
  29 #include <sys/dsl_prop.h>
  30 #include <sys/dsl_synctask.h>
  31 #include <sys/dmu_traverse.h>
  32 #include <sys/dmu_tx.h>
  33 #include <sys/arc.h>
  34 #include <sys/zio.h>
  35 #include <sys/zap.h>
  36 #include <sys/unique.h>
  37 #include <sys/zfs_context.h>
  38 #include <sys/zfs_ioctl.h>
  39 #include <sys/spa.h>
  40 #include <sys/zfs_znode.h>
  41 #include <sys/zfs_onexit.h>
  42 #include <sys/zvol.h>
  43 #include <sys/dsl_scan.h>
  44 #include <sys/dsl_deadlist.h>
  45
  46 static char *dsl_reaper = "the grim reaper";
  47
  48 static dsl_checkfunc_t dsl_dataset_destroy_begin_check;
  49 static dsl_syncfunc_t dsl_dataset_destroy_begin_sync;
  50 static dsl_syncfunc_t dsl_dataset_set_reservation_sync;
  51
  52 #define SWITCH64(x, y) \
  53         { \
  54                 uint64_t __tmp = (x); \
  55                 (x) = (y); \
  56                 (y) = __tmp; \
  57         }
  58
  59 #define DS_REF_MAX      (1ULL << 62)
  60
  61 #define DSL_DEADLIST_BLOCKSIZE  SPA_MAXBLOCKSIZE
  62
  63 #define DSL_DATASET_IS_DESTROYED(ds)    ((ds)->ds_owner == dsl_reaper)
  64
  65
  66 /*
  67  * Figure out how much of this delta should be propogated to the dsl_dir
  68  * layer.  If there's a refreservation, that space has already been
  69  * partially accounted for in our ancestors.
  70  */
  71 static int64_t
  72 parent_delta(dsl_dataset_t *ds, int64_t delta)
  73 {
  74         uint64_t old_bytes, new_bytes;
  75
  76         if (ds->ds_reserved == 0)
  77                 return (delta);
  78
  79         old_bytes = MAX(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
  80         new_bytes = MAX(ds->ds_phys->ds_unique_bytes + delta, ds->ds_reserved);
  81
  82         ASSERT3U(ABS((int64_t)(new_bytes - old_bytes)), <=, ABS(delta));
  83         return (new_bytes - old_bytes);
  84 }
  85
  86 void
  87 dsl_dataset_block_born(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
  88 {
  89         int used, compressed, uncompressed;
  90         int64_t delta;
  91
  92         used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
  93         compressed = BP_GET_PSIZE(bp);
  94         uncompressed = BP_GET_UCSIZE(bp);
  95
  96         dprintf_bp(bp, "ds=%p", ds);
  97
  98         ASSERT(dmu_tx_is_syncing(tx));
  99         /* It could have been compressed away to nothing */
 100         if (BP_IS_HOLE(bp))
 101                 return;
 102         ASSERT(BP_GET_TYPE(bp) != DMU_OT_NONE);
 103         ASSERT3U(BP_GET_TYPE(bp), <, DMU_OT_NUMTYPES);
 104         if (ds == NULL) {
 105                 /*
 106                  * Account for the meta-objset space in its placeholder
 107                  * dsl_dir.
 108                  */
 109                 ASSERT3U(compressed, ==, uncompressed); /* it's all metadata */
 110                 dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
 111                     used, compressed, uncompressed, tx);
 112                 dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
 113                 return;
 114         }
 115         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 116
 117         mutex_enter(&ds->ds_dir->dd_lock);
 118         mutex_enter(&ds->ds_lock);
 119         delta = parent_delta(ds, used);
 120         ds->ds_phys->ds_used_bytes += used;
 121         ds->ds_phys->ds_compressed_bytes += compressed;
 122         ds->ds_phys->ds_uncompressed_bytes += uncompressed;
 123         ds->ds_phys->ds_unique_bytes += used;
 124         mutex_exit(&ds->ds_lock);
 125         dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD, delta,
 126             compressed, uncompressed, tx);
 127         dsl_dir_transfer_space(ds->ds_dir, used - delta,
 128             DD_USED_REFRSRV, DD_USED_HEAD, tx);
 129         mutex_exit(&ds->ds_dir->dd_lock);
 130 }
 131
 132 int
 133 dsl_dataset_block_kill(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx,
 134     boolean_t async)
 135 {
 136         int used, compressed, uncompressed;
 137
 138         if (BP_IS_HOLE(bp))
 139                 return (0);
 140
 141         ASSERT(dmu_tx_is_syncing(tx));
 142         ASSERT(bp->blk_birth <= tx->tx_txg);
 143
 144         used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
 145         compressed = BP_GET_PSIZE(bp);
 146         uncompressed = BP_GET_UCSIZE(bp);
 147
 148         ASSERT(used > 0);
 149         if (ds == NULL) {
 150                 /*
 151                  * Account for the meta-objset space in its placeholder
 152                  * dataset.
 153                  */
 154                 dsl_free(tx->tx_pool, tx->tx_txg, bp);
 155
 156                 dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
 157                     -used, -compressed, -uncompressed, tx);
 158                 dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
 159                 return (used);
 160         }
 161         ASSERT3P(tx->tx_pool, ==, ds->ds_dir->dd_pool);
 162
 163         ASSERT(!dsl_dataset_is_snapshot(ds));
 164         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 165
 166         if (bp->blk_birth > ds->ds_phys->ds_prev_snap_txg) {
 167                 int64_t delta;
 168
 169                 dprintf_bp(bp, "freeing ds=%llu", ds->ds_object);
 170                 dsl_free(tx->tx_pool, tx->tx_txg, bp);
 171
 172                 mutex_enter(&ds->ds_dir->dd_lock);
 173                 mutex_enter(&ds->ds_lock);
 174                 ASSERT(ds->ds_phys->ds_unique_bytes >= used ||
 175                     !DS_UNIQUE_IS_ACCURATE(ds));
 176                 delta = parent_delta(ds, -used);
 177                 ds->ds_phys->ds_unique_bytes -= used;
 178                 mutex_exit(&ds->ds_lock);
 179                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
 180                     delta, -compressed, -uncompressed, tx);
 181                 dsl_dir_transfer_space(ds->ds_dir, -used - delta,
 182                     DD_USED_REFRSRV, DD_USED_HEAD, tx);
 183                 mutex_exit(&ds->ds_dir->dd_lock);
 184         } else {
 185                 dprintf_bp(bp, "putting on dead list: %s", "");
 186                 if (async) {
 187                         /*
 188                          * We are here as part of zio's write done callback,
 189                          * which means we're a zio interrupt thread.  We can't
 190                          * call dsl_deadlist_insert() now because it may block
 191                          * waiting for I/O.  Instead, put bp on the deferred
 192                          * queue and let dsl_pool_sync() finish the job.
 193                          */
 194                         bplist_append(&ds->ds_pending_deadlist, bp);
 195                 } else {
 196                         dsl_deadlist_insert(&ds->ds_deadlist, bp, tx);
 197                 }
 198                 ASSERT3U(ds->ds_prev->ds_object, ==,
 199                     ds->ds_phys->ds_prev_snap_obj);
 200                 ASSERT(ds->ds_prev->ds_phys->ds_num_children > 0);
 201                 /* if (bp->blk_birth > prev prev snap txg) prev unique += bs */
 202                 if (ds->ds_prev->ds_phys->ds_next_snap_obj ==
 203                     ds->ds_object && bp->blk_birth >
 204                     ds->ds_prev->ds_phys->ds_prev_snap_txg) {
 205                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
 206                         mutex_enter(&ds->ds_prev->ds_lock);
 207                         ds->ds_prev->ds_phys->ds_unique_bytes += used;
 208                         mutex_exit(&ds->ds_prev->ds_lock);
 209                 }
 210                 if (bp->blk_birth > ds->ds_dir->dd_origin_txg) {
 211                         dsl_dir_transfer_space(ds->ds_dir, used,
 212                             DD_USED_HEAD, DD_USED_SNAP, tx);
 213                 }
 214         }
 215         mutex_enter(&ds->ds_lock);
 216         ASSERT3U(ds->ds_phys->ds_used_bytes, >=, used);
 217         ds->ds_phys->ds_used_bytes -= used;
 218         ASSERT3U(ds->ds_phys->ds_compressed_bytes, >=, compressed);
 219         ds->ds_phys->ds_compressed_bytes -= compressed;
 220         ASSERT3U(ds->ds_phys->ds_uncompressed_bytes, >=, uncompressed);
 221         ds->ds_phys->ds_uncompressed_bytes -= uncompressed;
 222         mutex_exit(&ds->ds_lock);
 223
 224         return (used);
 225 }
 226
 227 uint64_t
 228 dsl_dataset_prev_snap_txg(dsl_dataset_t *ds)
 229 {
 230         uint64_t trysnap = 0;
 231
 232         if (ds == NULL)
 233                 return (0);
 234         /*
 235          * The snapshot creation could fail, but that would cause an
 236          * incorrect FALSE return, which would only result in an
 237          * overestimation of the amount of space that an operation would
 238          * consume, which is OK.
 239          *
 240          * There's also a small window where we could miss a pending
 241          * snapshot, because we could set the sync task in the quiescing
 242          * phase.  So this should only be used as a guess.
 243          */
 244         if (ds->ds_trysnap_txg >
 245             spa_last_synced_txg(ds->ds_dir->dd_pool->dp_spa))
 246                 trysnap = ds->ds_trysnap_txg;
 247         return (MAX(ds->ds_phys->ds_prev_snap_txg, trysnap));
 248 }
 249
 250 boolean_t
 251 dsl_dataset_block_freeable(dsl_dataset_t *ds, const blkptr_t *bp,
 252     uint64_t blk_birth)
 253 {
 254         if (blk_birth <= dsl_dataset_prev_snap_txg(ds))
 255                 return (B_FALSE);
 256
 257         ddt_prefetch(dsl_dataset_get_spa(ds), bp);
 258
 259         return (B_TRUE);
 260 }
 261
 262 /* ARGSUSED */
 263 static void
 264 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
 265 {
 266         dsl_dataset_t *ds = dsv;
 267
 268         ASSERT(ds->ds_owner == NULL || DSL_DATASET_IS_DESTROYED(ds));
 269
 270         unique_remove(ds->ds_fsid_guid);
 271
 272         if (ds->ds_objset != NULL)
 273                 dmu_objset_evict(ds->ds_objset);
 274
 275         if (ds->ds_prev) {
 276                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 277                 ds->ds_prev = NULL;
 278         }
 279
 280         bplist_destroy(&ds->ds_pending_deadlist);
 281         if (db != NULL) {
 282                 dsl_deadlist_close(&ds->ds_deadlist);
 283         } else {
 284                 ASSERT(ds->ds_deadlist.dl_dbuf == NULL);
 285                 ASSERT(!ds->ds_deadlist.dl_oldfmt);
 286         }
 287         if (ds->ds_dir)
 288                 dsl_dir_close(ds->ds_dir, ds);
 289
 290         ASSERT(!list_link_active(&ds->ds_synced_link));
 291
 292         mutex_destroy(&ds->ds_lock);
 293         mutex_destroy(&ds->ds_recvlock);
 294         mutex_destroy(&ds->ds_opening_lock);
 295         rw_destroy(&ds->ds_rwlock);
 296         cv_destroy(&ds->ds_exclusive_cv);
 297
 298         kmem_free(ds, sizeof (dsl_dataset_t));
 299 }
 300
 301 static int
 302 dsl_dataset_get_snapname(dsl_dataset_t *ds)
 303 {
 304         dsl_dataset_phys_t *headphys;
 305         int err;
 306         dmu_buf_t *headdbuf;
 307         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 308         objset_t *mos = dp->dp_meta_objset;
 309
 310         if (ds->ds_snapname[0])
 311                 return (0);
 312         if (ds->ds_phys->ds_next_snap_obj == 0)
 313                 return (0);
 314
 315         err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
 316             FTAG, &headdbuf);
 317         if (err)
 318                 return (err);
 319         headphys = headdbuf->db_data;
 320         err = zap_value_search(dp->dp_meta_objset,
 321             headphys->ds_snapnames_zapobj, ds->ds_object, 0, ds->ds_snapname);
 322         dmu_buf_rele(headdbuf, FTAG);
 323         return (err);
 324 }
 325
 326 static int
 327 dsl_dataset_snap_lookup(dsl_dataset_t *ds, const char *name, uint64_t *value)
 328 {
 329         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 330         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 331         matchtype_t mt;
 332         int err;
 333
 334         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 335                 mt = MT_FIRST;
 336         else
 337                 mt = MT_EXACT;
 338
 339         err = zap_lookup_norm(mos, snapobj, name, 8, 1,
 340             value, mt, NULL, 0, NULL);
 341         if (err == ENOTSUP && mt == MT_FIRST)
 342                 err = zap_lookup(mos, snapobj, name, 8, 1, value);
 343         return (err);
 344 }
 345
 346 static int
 347 dsl_dataset_snap_remove(dsl_dataset_t *ds, char *name, dmu_tx_t *tx)
 348 {
 349         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 350         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 351         matchtype_t mt;
 352         int err;
 353
 354         dsl_dir_snap_cmtime_update(ds->ds_dir);
 355
 356         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 357                 mt = MT_FIRST;
 358         else
 359                 mt = MT_EXACT;
 360
 361         err = zap_remove_norm(mos, snapobj, name, mt, tx);
 362         if (err == ENOTSUP && mt == MT_FIRST)
 363                 err = zap_remove(mos, snapobj, name, tx);
 364         return (err);
 365 }
 366
 367 static int
 368 dsl_dataset_get_ref(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 369     dsl_dataset_t **dsp)
 370 {
 371         objset_t *mos = dp->dp_meta_objset;
 372         dmu_buf_t *dbuf;
 373         dsl_dataset_t *ds;
 374         int err;
 375         dmu_object_info_t doi;
 376
 377         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
 378             dsl_pool_sync_context(dp));
 379
 380         err = dmu_bonus_hold(mos, dsobj, tag, &dbuf);
 381         if (err)
 382                 return (err);
 383
 384         /* Make sure dsobj has the correct object type. */
 385         dmu_object_info_from_db(dbuf, &doi);
 386         if (doi.doi_type != DMU_OT_DSL_DATASET)
 387                 return (EINVAL);
 388
 389         ds = dmu_buf_get_user(dbuf);
 390         if (ds == NULL) {
 391                 dsl_dataset_t *winner = NULL;
 392
 393                 ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_PUSHPAGE);
 394                 ds->ds_dbuf = dbuf;
 395                 ds->ds_object = dsobj;
 396                 ds->ds_phys = dbuf->db_data;
 397                 list_link_init(&ds->ds_synced_link);
 398
 399                 mutex_init(&ds->ds_lock, NULL, MUTEX_DEFAULT, NULL);
 400                 mutex_init(&ds->ds_recvlock, NULL, MUTEX_DEFAULT, NULL);
 401                 mutex_init(&ds->ds_opening_lock, NULL, MUTEX_DEFAULT, NULL);
 402                 rw_init(&ds->ds_rwlock, NULL, RW_DEFAULT, NULL);
 403                 cv_init(&ds->ds_exclusive_cv, NULL, CV_DEFAULT, NULL);
 404
 405                 bplist_create(&ds->ds_pending_deadlist);
 406                 dsl_deadlist_open(&ds->ds_deadlist,
 407                     mos, ds->ds_phys->ds_deadlist_obj);
 408
 409                 if (err == 0) {
 410                         err = dsl_dir_open_obj(dp,
 411                             ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
 412                 }
 413                 if (err) {
 414                         mutex_destroy(&ds->ds_lock);
 415                         mutex_destroy(&ds->ds_recvlock);
 416                         mutex_destroy(&ds->ds_opening_lock);
 417                         rw_destroy(&ds->ds_rwlock);
 418                         cv_destroy(&ds->ds_exclusive_cv);
 419                         bplist_destroy(&ds->ds_pending_deadlist);
 420                         dsl_deadlist_close(&ds->ds_deadlist);
 421                         kmem_free(ds, sizeof (dsl_dataset_t));
 422                         dmu_buf_rele(dbuf, tag);
 423                         return (err);
 424                 }
 425
 426                 if (!dsl_dataset_is_snapshot(ds)) {
 427                         ds->ds_snapname[0] = '\0';
 428                         if (ds->ds_phys->ds_prev_snap_obj) {
 429                                 err = dsl_dataset_get_ref(dp,
 430                                     ds->ds_phys->ds_prev_snap_obj,
 431                                     ds, &ds->ds_prev);
 432                         }
 433                 } else {
 434                         if (zfs_flags & ZFS_DEBUG_SNAPNAMES)
 435                                 err = dsl_dataset_get_snapname(ds);
 436                         if (err == 0 && ds->ds_phys->ds_userrefs_obj != 0) {
 437                                 err = zap_count(
 438                                     ds->ds_dir->dd_pool->dp_meta_objset,
 439                                     ds->ds_phys->ds_userrefs_obj,
 440                                     &ds->ds_userrefs);
 441                         }
 442                 }
 443
 444                 if (err == 0 && !dsl_dataset_is_snapshot(ds)) {
 445                         /*
 446                          * In sync context, we're called with either no lock
 447                          * or with the write lock.  If we're not syncing,
 448                          * we're always called with the read lock held.
 449                          */
 450                         boolean_t need_lock =
 451                             !RW_WRITE_HELD(&dp->dp_config_rwlock) &&
 452                             dsl_pool_sync_context(dp);
 453
 454                         if (need_lock)
 455                                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 456
 457                         err = dsl_prop_get_ds(ds,
 458                             "refreservation", sizeof (uint64_t), 1,
 459                             &ds->ds_reserved, NULL);
 460                         if (err == 0) {
 461                                 err = dsl_prop_get_ds(ds,
 462                                     "refquota", sizeof (uint64_t), 1,
 463                                     &ds->ds_quota, NULL);
 464                         }
 465
 466                         if (need_lock)
 467                                 rw_exit(&dp->dp_config_rwlock);
 468                 } else {
 469                         ds->ds_reserved = ds->ds_quota = 0;
 470                 }
 471
 472                 if (err == 0) {
 473                         winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
 474                             dsl_dataset_evict);
 475                 }
 476                 if (err || winner) {
 477                         bplist_destroy(&ds->ds_pending_deadlist);
 478                         dsl_deadlist_close(&ds->ds_deadlist);
 479                         if (ds->ds_prev)
 480                                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 481                         dsl_dir_close(ds->ds_dir, ds);
 482                         mutex_destroy(&ds->ds_lock);
 483                         mutex_destroy(&ds->ds_recvlock);
 484                         mutex_destroy(&ds->ds_opening_lock);
 485                         rw_destroy(&ds->ds_rwlock);
 486                         cv_destroy(&ds->ds_exclusive_cv);
 487                         kmem_free(ds, sizeof (dsl_dataset_t));
 488                         if (err) {
 489                                 dmu_buf_rele(dbuf, tag);
 490                                 return (err);
 491                         }
 492                         ds = winner;
 493                 } else {
 494                         ds->ds_fsid_guid =
 495                             unique_insert(ds->ds_phys->ds_fsid_guid);
 496                 }
 497         }
 498         ASSERT3P(ds->ds_dbuf, ==, dbuf);
 499         ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
 500         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0 ||
 501             spa_version(dp->dp_spa) < SPA_VERSION_ORIGIN ||
 502             dp->dp_origin_snap == NULL || ds == dp->dp_origin_snap);
 503         mutex_enter(&ds->ds_lock);
 504         if (!dsl_pool_sync_context(dp) && DSL_DATASET_IS_DESTROYED(ds)) {
 505                 mutex_exit(&ds->ds_lock);
 506                 dmu_buf_rele(ds->ds_dbuf, tag);
 507                 return (ENOENT);
 508         }
 509         mutex_exit(&ds->ds_lock);
 510         *dsp = ds;
 511         return (0);
 512 }
 513
 514 static int
 515 dsl_dataset_hold_ref(dsl_dataset_t *ds, void *tag)
 516 {
 517         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 518
 519         /*
 520          * In syncing context we don't want the rwlock lock: there
 521          * may be an existing writer waiting for sync phase to
 522          * finish.  We don't need to worry about such writers, since
 523          * sync phase is single-threaded, so the writer can't be
 524          * doing anything while we are active.
 525          */
 526         if (dsl_pool_sync_context(dp)) {
 527                 ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
 528                 return (0);
 529         }
 530
 531         /*
 532          * Normal users will hold the ds_rwlock as a READER until they
 533          * are finished (i.e., call dsl_dataset_rele()).  "Owners" will
 534          * drop their READER lock after they set the ds_owner field.
 535          *
 536          * If the dataset is being destroyed, the destroy thread will
 537          * obtain a WRITER lock for exclusive access after it's done its
 538          * open-context work and then change the ds_owner to
 539          * dsl_reaper once destruction is assured.  So threads
 540          * may block here temporarily, until the "destructability" of
 541          * the dataset is determined.
 542          */
 543         ASSERT(!RW_WRITE_HELD(&dp->dp_config_rwlock));
 544         mutex_enter(&ds->ds_lock);
 545         while (!rw_tryenter(&ds->ds_rwlock, RW_READER)) {
 546                 rw_exit(&dp->dp_config_rwlock);
 547                 cv_wait(&ds->ds_exclusive_cv, &ds->ds_lock);
 548                 if (DSL_DATASET_IS_DESTROYED(ds)) {
 549                         mutex_exit(&ds->ds_lock);
 550                         dsl_dataset_drop_ref(ds, tag);
 551                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 552                         return (ENOENT);
 553                 }
 554                 /*
 555                  * The dp_config_rwlock lives above the ds_lock. And
 556                  * we need to check DSL_DATASET_IS_DESTROYED() while
 557                  * holding the ds_lock, so we have to drop and reacquire
 558                  * the ds_lock here.
 559                  */
 560                 mutex_exit(&ds->ds_lock);
 561                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 562                 mutex_enter(&ds->ds_lock);
 563         }
 564         mutex_exit(&ds->ds_lock);
 565         return (0);
 566 }
 567
 568 int
 569 dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 570     dsl_dataset_t **dsp)
 571 {
 572         int err = dsl_dataset_get_ref(dp, dsobj, tag, dsp);
 573
 574         if (err)
 575                 return (err);
 576         return (dsl_dataset_hold_ref(*dsp, tag));
 577 }
 578
 579 int
 580 dsl_dataset_own_obj(dsl_pool_t *dp, uint64_t dsobj, boolean_t inconsistentok,
 581     void *tag, dsl_dataset_t **dsp)
 582 {
 583         int err = dsl_dataset_hold_obj(dp, dsobj, tag, dsp);
 584         if (err)
 585                 return (err);
 586         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 587                 dsl_dataset_rele(*dsp, tag);
 588                 *dsp = NULL;
 589                 return (EBUSY);
 590         }
 591         return (0);
 592 }
 593
 594 int
 595 dsl_dataset_hold(const char *name, void *tag, dsl_dataset_t **dsp)
 596 {
 597         dsl_dir_t *dd;
 598         dsl_pool_t *dp;
 599         const char *snapname;
 600         uint64_t obj;
 601         int err = 0;
 602
 603         err = dsl_dir_open_spa(NULL, name, FTAG, &dd, &snapname);
 604         if (err)
 605                 return (err);
 606
 607         dp = dd->dd_pool;
 608         obj = dd->dd_phys->dd_head_dataset_obj;
 609         rw_enter(&dp->dp_config_rwlock, RW_READER);
 610         if (obj)
 611                 err = dsl_dataset_get_ref(dp, obj, tag, dsp);
 612         else
 613                 err = ENOENT;
 614         if (err)
 615                 goto out;
 616
 617         err = dsl_dataset_hold_ref(*dsp, tag);
 618
 619         /* we may be looking for a snapshot */
 620         if (err == 0 && snapname != NULL) {
 621                 dsl_dataset_t *ds = NULL;
 622
 623                 if (*snapname++ != '@') {
 624                         dsl_dataset_rele(*dsp, tag);
 625                         err = ENOENT;
 626                         goto out;
 627                 }
 628
 629                 dprintf("looking for snapshot '%s'\n", snapname);
 630                 err = dsl_dataset_snap_lookup(*dsp, snapname, &obj);
 631                 if (err == 0)
 632                         err = dsl_dataset_get_ref(dp, obj, tag, &ds);
 633                 dsl_dataset_rele(*dsp, tag);
 634
 635                 ASSERT3U((err == 0), ==, (ds != NULL));
 636
 637                 if (ds) {
 638                         mutex_enter(&ds->ds_lock);
 639                         if (ds->ds_snapname[0] == 0)
 640                                 (void) strlcpy(ds->ds_snapname, snapname,
 641                                     sizeof (ds->ds_snapname));
 642                         mutex_exit(&ds->ds_lock);
 643                         err = dsl_dataset_hold_ref(ds, tag);
 644                         *dsp = err ? NULL : ds;
 645                 }
 646         }
 647 out:
 648         rw_exit(&dp->dp_config_rwlock);
 649         dsl_dir_close(dd, FTAG);
 650         return (err);
 651 }
 652
 653 int
 654 dsl_dataset_own(const char *name, boolean_t inconsistentok,
 655     void *tag, dsl_dataset_t **dsp)
 656 {
 657         int err = dsl_dataset_hold(name, tag, dsp);
 658         if (err)
 659                 return (err);
 660         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 661                 dsl_dataset_rele(*dsp, tag);
 662                 return (EBUSY);
 663         }
 664         return (0);
 665 }
 666
 667 void
 668 dsl_dataset_name(dsl_dataset_t *ds, char *name)
 669 {
 670         if (ds == NULL) {
 671                 (void) strcpy(name, "mos");
 672         } else {
 673                 dsl_dir_name(ds->ds_dir, name);
 674                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 675                 if (ds->ds_snapname[0]) {
 676                         (void) strcat(name, "@");
 677                         /*
 678                          * We use a "recursive" mutex so that we
 679                          * can call dprintf_ds() with ds_lock held.
 680                          */
 681                         if (!MUTEX_HELD(&ds->ds_lock)) {
 682                                 mutex_enter(&ds->ds_lock);
 683                                 (void) strcat(name, ds->ds_snapname);
 684                                 mutex_exit(&ds->ds_lock);
 685                         } else {
 686                                 (void) strcat(name, ds->ds_snapname);
 687                         }
 688                 }
 689         }
 690 }
 691
 692 static int
 693 dsl_dataset_namelen(dsl_dataset_t *ds)
 694 {
 695         int result;
 696
 697         if (ds == NULL) {
 698                 result = 3;     /* "mos" */
 699         } else {
 700                 result = dsl_dir_namelen(ds->ds_dir);
 701                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 702                 if (ds->ds_snapname[0]) {
 703                         ++result;       /* adding one for the @-sign */
 704                         if (!MUTEX_HELD(&ds->ds_lock)) {
 705                                 mutex_enter(&ds->ds_lock);
 706                                 result += strlen(ds->ds_snapname);
 707                                 mutex_exit(&ds->ds_lock);
 708                         } else {
 709                                 result += strlen(ds->ds_snapname);
 710                         }
 711                 }
 712         }
 713
 714         return (result);
 715 }
 716
 717 void
 718 dsl_dataset_drop_ref(dsl_dataset_t *ds, void *tag)
 719 {
 720         dmu_buf_rele(ds->ds_dbuf, tag);
 721 }
 722
 723 void
 724 dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
 725 {
 726         if (!dsl_pool_sync_context(ds->ds_dir->dd_pool)) {
 727                 rw_exit(&ds->ds_rwlock);
 728         }
 729         dsl_dataset_drop_ref(ds, tag);
 730 }
 731
 732 void
 733 dsl_dataset_disown(dsl_dataset_t *ds, void *tag)
 734 {
 735         ASSERT((ds->ds_owner == tag && ds->ds_dbuf) ||
 736             (DSL_DATASET_IS_DESTROYED(ds) && ds->ds_dbuf == NULL));
 737
 738         mutex_enter(&ds->ds_lock);
 739         ds->ds_owner = NULL;
 740         if (RW_WRITE_HELD(&ds->ds_rwlock)) {
 741                 rw_exit(&ds->ds_rwlock);
 742                 cv_broadcast(&ds->ds_exclusive_cv);
 743         }
 744         mutex_exit(&ds->ds_lock);
 745         if (ds->ds_dbuf)
 746                 dsl_dataset_drop_ref(ds, tag);
 747         else
 748                 dsl_dataset_evict(NULL, ds);
 749 }
 750
 751 boolean_t
 752 dsl_dataset_tryown(dsl_dataset_t *ds, boolean_t inconsistentok, void *tag)
 753 {
 754         boolean_t gotit = FALSE;
 755
 756         mutex_enter(&ds->ds_lock);
 757         if (ds->ds_owner == NULL &&
 758             (!DS_IS_INCONSISTENT(ds) || inconsistentok)) {
 759                 ds->ds_owner = tag;
 760                 if (!dsl_pool_sync_context(ds->ds_dir->dd_pool))
 761                         rw_exit(&ds->ds_rwlock);
 762                 gotit = TRUE;
 763         }
 764         mutex_exit(&ds->ds_lock);
 765         return (gotit);
 766 }
 767
 768 void
 769 dsl_dataset_make_exclusive(dsl_dataset_t *ds, void *owner)
 770 {
 771         ASSERT3P(owner, ==, ds->ds_owner);
 772         if (!RW_WRITE_HELD(&ds->ds_rwlock))
 773                 rw_enter(&ds->ds_rwlock, RW_WRITER);
 774 }
 775
 776 uint64_t
 777 dsl_dataset_create_sync_dd(dsl_dir_t *dd, dsl_dataset_t *origin,
 778     uint64_t flags, dmu_tx_t *tx)
 779 {
 780         dsl_pool_t *dp = dd->dd_pool;
 781         dmu_buf_t *dbuf;
 782         dsl_dataset_phys_t *dsphys;
 783         uint64_t dsobj;
 784         objset_t *mos = dp->dp_meta_objset;
 785
 786         if (origin == NULL)
 787                 origin = dp->dp_origin_snap;
 788
 789         ASSERT(origin == NULL || origin->ds_dir->dd_pool == dp);
 790         ASSERT(origin == NULL || origin->ds_phys->ds_num_children > 0);
 791         ASSERT(dmu_tx_is_syncing(tx));
 792         ASSERT(dd->dd_phys->dd_head_dataset_obj == 0);
 793
 794         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
 795             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
 796         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
 797         dmu_buf_will_dirty(dbuf, tx);
 798         dsphys = dbuf->db_data;
 799         bzero(dsphys, sizeof (dsl_dataset_phys_t));
 800         dsphys->ds_dir_obj = dd->dd_object;
 801         dsphys->ds_flags = flags;
 802         dsphys->ds_fsid_guid = unique_create();
 803         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
 804             sizeof (dsphys->ds_guid));
 805         dsphys->ds_snapnames_zapobj =
 806             zap_create_norm(mos, U8_TEXTPREP_TOUPPER, DMU_OT_DSL_DS_SNAP_MAP,
 807             DMU_OT_NONE, 0, tx);
 808         dsphys->ds_creation_time = gethrestime_sec();
 809         dsphys->ds_creation_txg = tx->tx_txg == TXG_INITIAL ? 1 : tx->tx_txg;
 810
 811         if (origin == NULL) {
 812                 dsphys->ds_deadlist_obj = dsl_deadlist_alloc(mos, tx);
 813         } else {
 814                 dsl_dataset_t *ohds;
 815
 816                 dsphys->ds_prev_snap_obj = origin->ds_object;
 817                 dsphys->ds_prev_snap_txg =
 818                     origin->ds_phys->ds_creation_txg;
 819                 dsphys->ds_used_bytes =
 820                     origin->ds_phys->ds_used_bytes;
 821                 dsphys->ds_compressed_bytes =
 822                     origin->ds_phys->ds_compressed_bytes;
 823                 dsphys->ds_uncompressed_bytes =
 824                     origin->ds_phys->ds_uncompressed_bytes;
 825                 dsphys->ds_bp = origin->ds_phys->ds_bp;
 826                 dsphys->ds_flags |= origin->ds_phys->ds_flags;
 827
 828                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
 829                 origin->ds_phys->ds_num_children++;
 830
 831                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
 832                     origin->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ohds));
 833                 dsphys->ds_deadlist_obj = dsl_deadlist_clone(&ohds->ds_deadlist,
 834                     dsphys->ds_prev_snap_txg, dsphys->ds_prev_snap_obj, tx);
 835                 dsl_dataset_rele(ohds, FTAG);
 836
 837                 if (spa_version(dp->dp_spa) >= SPA_VERSION_NEXT_CLONES) {
 838                         if (origin->ds_phys->ds_next_clones_obj == 0) {
 839                                 origin->ds_phys->ds_next_clones_obj =
 840                                     zap_create(mos,
 841                                     DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
 842                         }
 843                         VERIFY(0 == zap_add_int(mos,
 844                             origin->ds_phys->ds_next_clones_obj,
 845                             dsobj, tx));
 846                 }
 847
 848                 dmu_buf_will_dirty(dd->dd_dbuf, tx);
 849                 dd->dd_phys->dd_origin_obj = origin->ds_object;
 850                 if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
 851                         if (origin->ds_dir->dd_phys->dd_clones == 0) {
 852                                 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
 853                                 origin->ds_dir->dd_phys->dd_clones =
 854                                     zap_create(mos,
 855                                     DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
 856                         }
 857                         VERIFY3U(0, ==, zap_add_int(mos,
 858                             origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
 859                 }
 860         }
 861
 862         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
 863                 dsphys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
 864
 865         dmu_buf_rele(dbuf, FTAG);
 866
 867         dmu_buf_will_dirty(dd->dd_dbuf, tx);
 868         dd->dd_phys->dd_head_dataset_obj = dsobj;
 869
 870         return (dsobj);
 871 }
 872
 873 uint64_t
 874 dsl_dataset_create_sync(dsl_dir_t *pdd, const char *lastname,
 875     dsl_dataset_t *origin, uint64_t flags, cred_t *cr, dmu_tx_t *tx)
 876 {
 877         dsl_pool_t *dp = pdd->dd_pool;
 878         uint64_t dsobj, ddobj;
 879         dsl_dir_t *dd;
 880
 881         ASSERT(lastname[0] != '@');
 882
 883         ddobj = dsl_dir_create_sync(dp, pdd, lastname, tx);
 884         VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
 885
 886         dsobj = dsl_dataset_create_sync_dd(dd, origin, flags, tx);
 887
 888         dsl_deleg_set_create_perms(dd, tx, cr);
 889
 890         dsl_dir_close(dd, FTAG);
 891
 892         /*
 893          * If we are creating a clone, make sure we zero out any stale
 894          * data from the origin snapshots zil header.
 895          */
 896         if (origin != NULL) {
 897                 dsl_dataset_t *ds;
 898                 objset_t *os;
 899
 900                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 901                 VERIFY3U(0, ==, dmu_objset_from_ds(ds, &os));
 902                 bzero(&os->os_zil_header, sizeof (os->os_zil_header));
 903                 dsl_dataset_dirty(ds, tx);
 904                 dsl_dataset_rele(ds, FTAG);
 905         }
 906
 907         return (dsobj);
 908 }
 909
 910 /*
 911  * The snapshots must all be in the same pool.
 912  */
 913 int
 914 dmu_snapshots_destroy_nvl(nvlist_t *snaps, boolean_t defer, char *failed)
 915 {
 916         int err;
 917         dsl_sync_task_t *dst;
 918         spa_t *spa;
 919         nvpair_t *pair;
 920         dsl_sync_task_group_t *dstg;
 921
 922         pair = nvlist_next_nvpair(snaps, NULL);
 923         if (pair == NULL)
 924                 return (0);
 925
 926         err = spa_open(nvpair_name(pair), &spa, FTAG);
 927         if (err)
 928                 return (err);
 929         dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
 930
 931         for (pair = nvlist_next_nvpair(snaps, NULL); pair != NULL;
 932             pair = nvlist_next_nvpair(snaps, pair)) {
 933                 dsl_dataset_t *ds;
 934                 int err;
 935
 936                 err = dsl_dataset_own(nvpair_name(pair), B_TRUE, dstg, &ds);
 937                 if (err == 0) {
 938                         struct dsl_ds_destroyarg *dsda;
 939
 940                         dsl_dataset_make_exclusive(ds, dstg);
 941                         dsda = kmem_zalloc(sizeof (struct dsl_ds_destroyarg),
 942                             KM_SLEEP);
 943                         dsda->ds = ds;
 944                         dsda->defer = defer;
 945                         dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
 946                             dsl_dataset_destroy_sync, dsda, dstg, 0);
 947                 } else if (err == ENOENT) {
 948                         err = 0;
 949                 } else {
 950                         (void) strcpy(failed, nvpair_name(pair));
 951                         break;
 952                 }
 953         }
 954
 955         if (err == 0)
 956                 err = dsl_sync_task_group_wait(dstg);
 957
 958         for (dst = list_head(&dstg->dstg_tasks); dst;
 959             dst = list_next(&dstg->dstg_tasks, dst)) {
 960                 struct dsl_ds_destroyarg *dsda = dst->dst_arg1;
 961                 dsl_dataset_t *ds = dsda->ds;
 962
 963                 /*
 964                  * Return the file system name that triggered the error
 965                  */
 966                 if (dst->dst_err) {
 967                         dsl_dataset_name(ds, failed);
 968                 }
 969                 ASSERT3P(dsda->rm_origin, ==, NULL);
 970                 dsl_dataset_disown(ds, dstg);
 971                 kmem_free(dsda, sizeof (struct dsl_ds_destroyarg));
 972         }
 973
 974         dsl_sync_task_group_destroy(dstg);
 975         spa_close(spa, FTAG);
 976         return (err);
 977
 978 }
 979
 980 static boolean_t
 981 dsl_dataset_might_destroy_origin(dsl_dataset_t *ds)
 982 {
 983         boolean_t might_destroy = B_FALSE;
 984
 985         mutex_enter(&ds->ds_lock);
 986         if (ds->ds_phys->ds_num_children == 2 && ds->ds_userrefs == 0 &&
 987             DS_IS_DEFER_DESTROY(ds))
 988                 might_destroy = B_TRUE;
 989         mutex_exit(&ds->ds_lock);
 990
 991         return (might_destroy);
 992 }
 993
 994 /*
 995  * If we're removing a clone, and these three conditions are true:
 996  *      1) the clone's origin has no other children
 997  *      2) the clone's origin has no user references
 998  *      3) the clone's origin has been marked for deferred destruction
 999  * Then, prepare to remove the origin as part of this sync task group.
1000  */
1001 static int
1002 dsl_dataset_origin_rm_prep(struct dsl_ds_destroyarg *dsda, void *tag)
1003 {
1004         dsl_dataset_t *ds = dsda->ds;
1005         dsl_dataset_t *origin = ds->ds_prev;
1006
1007         if (dsl_dataset_might_destroy_origin(origin)) {
1008                 char *name;
1009                 int namelen;
1010                 int error;
1011
1012                 namelen = dsl_dataset_namelen(origin) + 1;
1013                 name = kmem_alloc(namelen, KM_SLEEP);
1014                 dsl_dataset_name(origin, name);
1015 #ifdef _KERNEL
1016                 error = zfs_unmount_snap(name, NULL);
1017                 if (error) {
1018                         kmem_free(name, namelen);
1019                         return (error);
1020                 }
1021 #endif
1022                 error = dsl_dataset_own(name, B_TRUE, tag, &origin);
1023                 kmem_free(name, namelen);
1024                 if (error)
1025                         return (error);
1026                 dsda->rm_origin = origin;
1027                 dsl_dataset_make_exclusive(origin, tag);
1028         }
1029
1030         return (0);
1031 }
1032
1033 /*
1034  * ds must be opened as OWNER.  On return (whether successful or not),
1035  * ds will be closed and caller can no longer dereference it.
1036  */
1037 int
1038 dsl_dataset_destroy(dsl_dataset_t *ds, void *tag, boolean_t defer)
1039 {
1040         int err;
1041         dsl_sync_task_group_t *dstg;
1042         objset_t *os;
1043         dsl_dir_t *dd;
1044         uint64_t obj;
1045         struct dsl_ds_destroyarg dsda = { 0 };
1046         dsl_dataset_t *dummy_ds;
1047
1048         dsda.ds = ds;
1049
1050         if (dsl_dataset_is_snapshot(ds)) {
1051                 /* Destroying a snapshot is simpler */
1052                 dsl_dataset_make_exclusive(ds, tag);
1053
1054                 dsda.defer = defer;
1055                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1056                     dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
1057                     &dsda, tag, 0);
1058                 ASSERT3P(dsda.rm_origin, ==, NULL);
1059                 goto out;
1060         } else if (defer) {
1061                 err = EINVAL;
1062                 goto out;
1063         }
1064
1065         dd = ds->ds_dir;
1066         dummy_ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
1067         dummy_ds->ds_dir = dd;
1068         dummy_ds->ds_object = ds->ds_object;
1069
1070         /*
1071          * Check for errors and mark this ds as inconsistent, in
1072          * case we crash while freeing the objects.
1073          */
1074         err = dsl_sync_task_do(dd->dd_pool, dsl_dataset_destroy_begin_check,
1075             dsl_dataset_destroy_begin_sync, ds, NULL, 0);
1076         if (err)
1077                 goto out_free;
1078
1079         err = dmu_objset_from_ds(ds, &os);
1080         if (err)
1081                 goto out_free;
1082
1083         /*
1084          * remove the objects in open context, so that we won't
1085          * have too much to do in syncing context.
1086          */
1087         for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE,
1088             ds->ds_phys->ds_prev_snap_txg)) {
1089                 /*
1090                  * Ignore errors, if there is not enough disk space
1091                  * we will deal with it in dsl_dataset_destroy_sync().
1092                  */
1093                 (void) dmu_free_object(os, obj);
1094         }
1095         if (err != ESRCH)
1096                 goto out_free;
1097
1098         /*
1099          * Only the ZIL knows how to free log blocks.
1100          */
1101         zil_destroy(dmu_objset_zil(os), B_FALSE);
1102
1103         /*
1104          * Sync out all in-flight IO.
1105          */
1106         txg_wait_synced(dd->dd_pool, 0);
1107
1108         /*
1109          * If we managed to free all the objects in open
1110          * context, the user space accounting should be zero.
1111          */
1112         if (ds->ds_phys->ds_bp.blk_fill == 0 &&
1113             dmu_objset_userused_enabled(os)) {
1114                 ASSERTV(uint64_t count);
1115                 ASSERT(zap_count(os, DMU_USERUSED_OBJECT, &count) != 0 ||
1116                     count == 0);
1117                 ASSERT(zap_count(os, DMU_GROUPUSED_OBJECT, &count) != 0 ||
1118                     count == 0);
1119         }
1120
1121         rw_enter(&dd->dd_pool->dp_config_rwlock, RW_READER);
1122         err = dsl_dir_open_obj(dd->dd_pool, dd->dd_object, NULL, FTAG, &dd);
1123         rw_exit(&dd->dd_pool->dp_config_rwlock);
1124
1125         if (err)
1126                 goto out_free;
1127
1128         /*
1129          * Blow away the dsl_dir + head dataset.
1130          */
1131         dsl_dataset_make_exclusive(ds, tag);
1132         /*
1133          * If we're removing a clone, we might also need to remove its
1134          * origin.
1135          */
1136         do {
1137                 dsda.need_prep = B_FALSE;
1138                 if (dsl_dir_is_clone(dd)) {
1139                         err = dsl_dataset_origin_rm_prep(&dsda, tag);
1140                         if (err) {
1141                                 dsl_dir_close(dd, FTAG);
1142                                 goto out_free;
1143                         }
1144                 }
1145
1146                 dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
1147                 dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
1148                     dsl_dataset_destroy_sync, &dsda, tag, 0);
1149                 dsl_sync_task_create(dstg, dsl_dir_destroy_check,
1150                     dsl_dir_destroy_sync, dummy_ds, FTAG, 0);
1151                 err = dsl_sync_task_group_wait(dstg);
1152                 dsl_sync_task_group_destroy(dstg);
1153
1154                 /*
1155                  * We could be racing against 'zfs release' or 'zfs destroy -d'
1156                  * on the origin snap, in which case we can get EBUSY if we
1157                  * needed to destroy the origin snap but were not ready to
1158                  * do so.
1159                  */
1160                 if (dsda.need_prep) {
1161                         ASSERT(err == EBUSY);
1162                         ASSERT(dsl_dir_is_clone(dd));
1163                         ASSERT(dsda.rm_origin == NULL);
1164                 }
1165         } while (dsda.need_prep);
1166
1167         if (dsda.rm_origin != NULL)
1168                 dsl_dataset_disown(dsda.rm_origin, tag);
1169
1170         /* if it is successful, dsl_dir_destroy_sync will close the dd */
1171         if (err)
1172                 dsl_dir_close(dd, FTAG);
1173
1174 out_free:
1175         kmem_free(dummy_ds, sizeof (dsl_dataset_t));
1176 out:
1177         dsl_dataset_disown(ds, tag);
1178         return (err);
1179 }
1180
1181 blkptr_t *
1182 dsl_dataset_get_blkptr(dsl_dataset_t *ds)
1183 {
1184         return (&ds->ds_phys->ds_bp);
1185 }
1186
1187 void
1188 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
1189 {
1190         ASSERT(dmu_tx_is_syncing(tx));
1191         /* If it's the meta-objset, set dp_meta_rootbp */
1192         if (ds == NULL) {
1193                 tx->tx_pool->dp_meta_rootbp = *bp;
1194         } else {
1195                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1196                 ds->ds_phys->ds_bp = *bp;
1197         }
1198 }
1199
1200 spa_t *
1201 dsl_dataset_get_spa(dsl_dataset_t *ds)
1202 {
1203         return (ds->ds_dir->dd_pool->dp_spa);
1204 }
1205
1206 void
1207 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
1208 {
1209         dsl_pool_t *dp;
1210
1211         if (ds == NULL) /* this is the meta-objset */
1212                 return;
1213
1214         ASSERT(ds->ds_objset != NULL);
1215
1216         if (ds->ds_phys->ds_next_snap_obj != 0)
1217                 panic("dirtying snapshot!");
1218
1219         dp = ds->ds_dir->dd_pool;
1220
1221         if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
1222                 /* up the hold count until we can be written out */
1223                 dmu_buf_add_ref(ds->ds_dbuf, ds);
1224         }
1225 }
1226
1227 /*
1228  * The unique space in the head dataset can be calculated by subtracting
1229  * the space used in the most recent snapshot, that is still being used
1230  * in this file system, from the space currently in use.  To figure out
1231  * the space in the most recent snapshot still in use, we need to take
1232  * the total space used in the snapshot and subtract out the space that
1233  * has been freed up since the snapshot was taken.
1234  */
1235 static void
1236 dsl_dataset_recalc_head_uniq(dsl_dataset_t *ds)
1237 {
1238         uint64_t mrs_used;
1239         uint64_t dlused, dlcomp, dluncomp;
1240
1241         ASSERT(!dsl_dataset_is_snapshot(ds));
1242
1243         if (ds->ds_phys->ds_prev_snap_obj != 0)
1244                 mrs_used = ds->ds_prev->ds_phys->ds_used_bytes;
1245         else
1246                 mrs_used = 0;
1247
1248         dsl_deadlist_space(&ds->ds_deadlist, &dlused, &dlcomp, &dluncomp);
1249
1250         ASSERT3U(dlused, <=, mrs_used);
1251         ds->ds_phys->ds_unique_bytes =
1252             ds->ds_phys->ds_used_bytes - (mrs_used - dlused);
1253
1254         if (spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1255             SPA_VERSION_UNIQUE_ACCURATE)
1256                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1257 }
1258
1259 struct killarg {
1260         dsl_dataset_t *ds;
1261         dmu_tx_t *tx;
1262 };
1263
1264 /* ARGSUSED */
1265 static int
1266 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
1267     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
1268 {
1269         struct killarg *ka = arg;
1270         dmu_tx_t *tx = ka->tx;
1271
1272         if (bp == NULL)
1273                 return (0);
1274
1275         if (zb->zb_level == ZB_ZIL_LEVEL) {
1276                 ASSERT(zilog != NULL);
1277                 /*
1278                  * It's a block in the intent log.  It has no
1279                  * accounting, so just free it.
1280                  */
1281                 dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
1282         } else {
1283                 ASSERT(zilog == NULL);
1284                 ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
1285                 (void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
1286         }
1287
1288         return (0);
1289 }
1290
1291 /* ARGSUSED */
1292 static int
1293 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
1294 {
1295         dsl_dataset_t *ds = arg1;
1296         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1297         uint64_t count;
1298         int err;
1299
1300         /*
1301          * Can't delete a head dataset if there are snapshots of it.
1302          * (Except if the only snapshots are from the branch we cloned
1303          * from.)
1304          */
1305         if (ds->ds_prev != NULL &&
1306             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1307                 return (EBUSY);
1308
1309         /*
1310          * This is really a dsl_dir thing, but check it here so that
1311          * we'll be less likely to leave this dataset inconsistent &
1312          * nearly destroyed.
1313          */
1314         err = zap_count(mos, ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
1315         if (err)
1316                 return (err);
1317         if (count != 0)
1318                 return (EEXIST);
1319
1320         return (0);
1321 }
1322
1323 /* ARGSUSED */
1324 static void
1325 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1326 {
1327         dsl_dataset_t *ds = arg1;
1328         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1329
1330         /* Mark it as inconsistent on-disk, in case we crash */
1331         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1332         ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1333
1334         spa_history_log_internal(LOG_DS_DESTROY_BEGIN, dp->dp_spa, tx,
1335             "dataset = %llu", ds->ds_object);
1336 }
1337
1338 static int
1339 dsl_dataset_origin_check(struct dsl_ds_destroyarg *dsda, void *tag,
1340     dmu_tx_t *tx)
1341 {
1342         dsl_dataset_t *ds = dsda->ds;
1343         dsl_dataset_t *ds_prev = ds->ds_prev;
1344
1345         if (dsl_dataset_might_destroy_origin(ds_prev)) {
1346                 struct dsl_ds_destroyarg ndsda = {0};
1347
1348                 /*
1349                  * If we're not prepared to remove the origin, don't remove
1350                  * the clone either.
1351                  */
1352                 if (dsda->rm_origin == NULL) {
1353                         dsda->need_prep = B_TRUE;
1354                         return (EBUSY);
1355                 }
1356
1357                 ndsda.ds = ds_prev;
1358                 ndsda.is_origin_rm = B_TRUE;
1359                 return (dsl_dataset_destroy_check(&ndsda, tag, tx));
1360         }
1361
1362         /*
1363          * If we're not going to remove the origin after all,
1364          * undo the open context setup.
1365          */
1366         if (dsda->rm_origin != NULL) {
1367                 dsl_dataset_disown(dsda->rm_origin, tag);
1368                 dsda->rm_origin = NULL;
1369         }
1370
1371         return (0);
1372 }
1373
1374 /*
1375  * If you add new checks here, you may need to add
1376  * additional checks to the "temporary" case in
1377  * snapshot_check() in dmu_objset.c.
1378  */
1379 /* ARGSUSED */
1380 int
1381 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1382 {
1383         struct dsl_ds_destroyarg *dsda = arg1;
1384         dsl_dataset_t *ds = dsda->ds;
1385
1386         /* we have an owner hold, so noone else can destroy us */
1387         ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1388
1389         /*
1390          * Only allow deferred destroy on pools that support it.
1391          * NOTE: deferred destroy is only supported on snapshots.
1392          */
1393         if (dsda->defer) {
1394                 if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1395                     SPA_VERSION_USERREFS)
1396                         return (ENOTSUP);
1397                 ASSERT(dsl_dataset_is_snapshot(ds));
1398                 return (0);
1399         }
1400
1401         /*
1402          * Can't delete a head dataset if there are snapshots of it.
1403          * (Except if the only snapshots are from the branch we cloned
1404          * from.)
1405          */
1406         if (ds->ds_prev != NULL &&
1407             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1408                 return (EBUSY);
1409
1410         /*
1411          * If we made changes this txg, traverse_dsl_dataset won't find
1412          * them.  Try again.
1413          */
1414         if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1415                 return (EAGAIN);
1416
1417         if (dsl_dataset_is_snapshot(ds)) {
1418                 /*
1419                  * If this snapshot has an elevated user reference count,
1420                  * we can't destroy it yet.
1421                  */
1422                 if (ds->ds_userrefs > 0 && !dsda->releasing)
1423                         return (EBUSY);
1424
1425                 mutex_enter(&ds->ds_lock);
1426                 /*
1427                  * Can't delete a branch point. However, if we're destroying
1428                  * a clone and removing its origin due to it having a user
1429                  * hold count of 0 and having been marked for deferred destroy,
1430                  * it's OK for the origin to have a single clone.
1431                  */
1432                 if (ds->ds_phys->ds_num_children >
1433                     (dsda->is_origin_rm ? 2 : 1)) {
1434                         mutex_exit(&ds->ds_lock);
1435                         return (EEXIST);
1436                 }
1437                 mutex_exit(&ds->ds_lock);
1438         } else if (dsl_dir_is_clone(ds->ds_dir)) {
1439                 return (dsl_dataset_origin_check(dsda, arg2, tx));
1440         }
1441
1442         /* XXX we should do some i/o error checking... */
1443         return (0);
1444 }
1445
1446 struct refsarg {
1447         kmutex_t lock;
1448         boolean_t gone;
1449         kcondvar_t cv;
1450 };
1451
1452 /* ARGSUSED */
1453 static void
1454 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1455 {
1456         struct refsarg *arg = argv;
1457
1458         mutex_enter(&arg->lock);
1459         arg->gone = TRUE;
1460         cv_signal(&arg->cv);
1461         mutex_exit(&arg->lock);
1462 }
1463
1464 static void
1465 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1466 {
1467         struct refsarg arg;
1468
1469         mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1470         cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1471         arg.gone = FALSE;
1472         (void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1473             dsl_dataset_refs_gone);
1474         dmu_buf_rele(ds->ds_dbuf, tag);
1475         mutex_enter(&arg.lock);
1476         while (!arg.gone)
1477                 cv_wait(&arg.cv, &arg.lock);
1478         ASSERT(arg.gone);
1479         mutex_exit(&arg.lock);
1480         ds->ds_dbuf = NULL;
1481         ds->ds_phys = NULL;
1482         mutex_destroy(&arg.lock);
1483         cv_destroy(&arg.cv);
1484 }
1485
1486 static void
1487 remove_from_next_clones(dsl_dataset_t *ds, uint64_t obj, dmu_tx_t *tx)
1488 {
1489         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1490         int err;
1491         ASSERTV(uint64_t count);
1492
1493         ASSERT(ds->ds_phys->ds_num_children >= 2);
1494         err = zap_remove_int(mos, ds->ds_phys->ds_next_clones_obj, obj, tx);
1495         /*
1496          * The err should not be ENOENT, but a bug in a previous version
1497          * of the code could cause upgrade_clones_cb() to not set
1498          * ds_next_snap_obj when it should, leading to a missing entry.
1499          * If we knew that the pool was created after
1500          * SPA_VERSION_NEXT_CLONES, we could assert that it isn't
1501          * ENOENT.  However, at least we can check that we don't have
1502          * too many entries in the next_clones_obj even after failing to
1503          * remove this one.
1504          */
1505         if (err != ENOENT) {
1506                 VERIFY3U(err, ==, 0);
1507         }
1508         ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
1509             &count));
1510         ASSERT3U(count, <=, ds->ds_phys->ds_num_children - 2);
1511 }
1512
1513 static void
1514 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
1515 {
1516         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1517         zap_cursor_t zc;
1518         zap_attribute_t za;
1519
1520         /*
1521          * If it is the old version, dd_clones doesn't exist so we can't
1522          * find the clones, but deadlist_remove_key() is a no-op so it
1523          * doesn't matter.
1524          */
1525         if (ds->ds_dir->dd_phys->dd_clones == 0)
1526                 return;
1527
1528         for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
1529             zap_cursor_retrieve(&zc, &za) == 0;
1530             zap_cursor_advance(&zc)) {
1531                 dsl_dataset_t *clone;
1532
1533                 VERIFY3U(0, ==, dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
1534                     za.za_first_integer, FTAG, &clone));
1535                 if (clone->ds_dir->dd_origin_txg > mintxg) {
1536                         dsl_deadlist_remove_key(&clone->ds_deadlist,
1537                             mintxg, tx);
1538                         dsl_dataset_remove_clones_key(clone, mintxg, tx);
1539                 }
1540                 dsl_dataset_rele(clone, FTAG);
1541         }
1542         zap_cursor_fini(&zc);
1543 }
1544
1545 struct process_old_arg {
1546         dsl_dataset_t *ds;
1547         dsl_dataset_t *ds_prev;
1548         boolean_t after_branch_point;
1549         zio_t *pio;
1550         uint64_t used, comp, uncomp;
1551 };
1552
1553 static int
1554 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1555 {
1556         struct process_old_arg *poa = arg;
1557         dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
1558
1559         if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
1560                 dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
1561                 if (poa->ds_prev && !poa->after_branch_point &&
1562                     bp->blk_birth >
1563                     poa->ds_prev->ds_phys->ds_prev_snap_txg) {
1564                         poa->ds_prev->ds_phys->ds_unique_bytes +=
1565                             bp_get_dsize_sync(dp->dp_spa, bp);
1566                 }
1567         } else {
1568                 poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
1569                 poa->comp += BP_GET_PSIZE(bp);
1570                 poa->uncomp += BP_GET_UCSIZE(bp);
1571                 dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
1572         }
1573         return (0);
1574 }
1575
1576 static void
1577 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
1578     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
1579 {
1580         struct process_old_arg poa = { 0 };
1581         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1582         objset_t *mos = dp->dp_meta_objset;
1583
1584         ASSERT(ds->ds_deadlist.dl_oldfmt);
1585         ASSERT(ds_next->ds_deadlist.dl_oldfmt);
1586
1587         poa.ds = ds;
1588         poa.ds_prev = ds_prev;
1589         poa.after_branch_point = after_branch_point;
1590         poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1591         VERIFY3U(0, ==, bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
1592             process_old_cb, &poa, tx));
1593         VERIFY3U(zio_wait(poa.pio), ==, 0);
1594         ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
1595
1596         /* change snapused */
1597         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1598             -poa.used, -poa.comp, -poa.uncomp, tx);
1599
1600         /* swap next's deadlist to our deadlist */
1601         dsl_deadlist_close(&ds->ds_deadlist);
1602         dsl_deadlist_close(&ds_next->ds_deadlist);
1603         SWITCH64(ds_next->ds_phys->ds_deadlist_obj,
1604             ds->ds_phys->ds_deadlist_obj);
1605         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
1606         dsl_deadlist_open(&ds_next->ds_deadlist, mos,
1607             ds_next->ds_phys->ds_deadlist_obj);
1608 }
1609
1610 void
1611 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1612 {
1613         struct dsl_ds_destroyarg *dsda = arg1;
1614         dsl_dataset_t *ds = dsda->ds;
1615         int err;
1616         int after_branch_point = FALSE;
1617         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1618         objset_t *mos = dp->dp_meta_objset;
1619         dsl_dataset_t *ds_prev = NULL;
1620         boolean_t wont_destroy;
1621         uint64_t obj;
1622
1623         wont_destroy = (dsda->defer &&
1624             (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1));
1625
1626         ASSERT(ds->ds_owner || wont_destroy);
1627         ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1628         ASSERT(ds->ds_prev == NULL ||
1629             ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1630         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1631
1632         if (wont_destroy) {
1633                 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1634                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1635                 ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1636                 return;
1637         }
1638
1639         /* signal any waiters that this dataset is going away */
1640         mutex_enter(&ds->ds_lock);
1641         ds->ds_owner = dsl_reaper;
1642         cv_broadcast(&ds->ds_exclusive_cv);
1643         mutex_exit(&ds->ds_lock);
1644
1645         /* Remove our reservation */
1646         if (ds->ds_reserved != 0) {
1647                 dsl_prop_setarg_t psa;
1648                 uint64_t value = 0;
1649
1650                 dsl_prop_setarg_init_uint64(&psa, "refreservation",
1651                     (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
1652                     &value);
1653                 psa.psa_effective_value = 0;    /* predict default value */
1654
1655                 dsl_dataset_set_reservation_sync(ds, &psa, tx);
1656                 ASSERT3U(ds->ds_reserved, ==, 0);
1657         }
1658
1659         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1660
1661         dsl_scan_ds_destroyed(ds, tx);
1662
1663         obj = ds->ds_object;
1664
1665         if (ds->ds_phys->ds_prev_snap_obj != 0) {
1666                 if (ds->ds_prev) {
1667                         ds_prev = ds->ds_prev;
1668                 } else {
1669                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1670                             ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1671                 }
1672                 after_branch_point =
1673                     (ds_prev->ds_phys->ds_next_snap_obj != obj);
1674
1675                 dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1676                 if (after_branch_point &&
1677                     ds_prev->ds_phys->ds_next_clones_obj != 0) {
1678                         remove_from_next_clones(ds_prev, obj, tx);
1679                         if (ds->ds_phys->ds_next_snap_obj != 0) {
1680                                 VERIFY(0 == zap_add_int(mos,
1681                                     ds_prev->ds_phys->ds_next_clones_obj,
1682                                     ds->ds_phys->ds_next_snap_obj, tx));
1683                         }
1684                 }
1685                 if (after_branch_point &&
1686                     ds->ds_phys->ds_next_snap_obj == 0) {
1687                         /* This clone is toast. */
1688                         ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1689                         ds_prev->ds_phys->ds_num_children--;
1690
1691                         /*
1692                          * If the clone's origin has no other clones, no
1693                          * user holds, and has been marked for deferred
1694                          * deletion, then we should have done the necessary
1695                          * destroy setup for it.
1696                          */
1697                         if (ds_prev->ds_phys->ds_num_children == 1 &&
1698                             ds_prev->ds_userrefs == 0 &&
1699                             DS_IS_DEFER_DESTROY(ds_prev)) {
1700                                 ASSERT3P(dsda->rm_origin, !=, NULL);
1701                         } else {
1702                                 ASSERT3P(dsda->rm_origin, ==, NULL);
1703                         }
1704                 } else if (!after_branch_point) {
1705                         ds_prev->ds_phys->ds_next_snap_obj =
1706                             ds->ds_phys->ds_next_snap_obj;
1707                 }
1708         }
1709
1710         if (dsl_dataset_is_snapshot(ds)) {
1711                 dsl_dataset_t *ds_next;
1712                 uint64_t old_unique;
1713                 uint64_t used = 0, comp = 0, uncomp = 0;
1714
1715                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1716                     ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1717                 ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1718
1719                 old_unique = ds_next->ds_phys->ds_unique_bytes;
1720
1721                 dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1722                 ds_next->ds_phys->ds_prev_snap_obj =
1723                     ds->ds_phys->ds_prev_snap_obj;
1724                 ds_next->ds_phys->ds_prev_snap_txg =
1725                     ds->ds_phys->ds_prev_snap_txg;
1726                 ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1727                     ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1728
1729
1730                 if (ds_next->ds_deadlist.dl_oldfmt) {
1731                         process_old_deadlist(ds, ds_prev, ds_next,
1732                             after_branch_point, tx);
1733                 } else {
1734                         /* Adjust prev's unique space. */
1735                         if (ds_prev && !after_branch_point) {
1736                                 dsl_deadlist_space_range(&ds_next->ds_deadlist,
1737                                     ds_prev->ds_phys->ds_prev_snap_txg,
1738                                     ds->ds_phys->ds_prev_snap_txg,
1739                                     &used, &comp, &uncomp);
1740                                 ds_prev->ds_phys->ds_unique_bytes += used;
1741                         }
1742
1743                         /* Adjust snapused. */
1744                         dsl_deadlist_space_range(&ds_next->ds_deadlist,
1745                             ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
1746                             &used, &comp, &uncomp);
1747                         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1748                             -used, -comp, -uncomp, tx);
1749
1750                         /* Move blocks to be freed to pool's free list. */
1751                         dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
1752                             &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
1753                             tx);
1754                         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
1755                             DD_USED_HEAD, used, comp, uncomp, tx);
1756                         dsl_dir_dirty(tx->tx_pool->dp_free_dir, tx);
1757
1758                         /* Merge our deadlist into next's and free it. */
1759                         dsl_deadlist_merge(&ds_next->ds_deadlist,
1760                             ds->ds_phys->ds_deadlist_obj, tx);
1761                 }
1762                 dsl_deadlist_close(&ds->ds_deadlist);
1763                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1764
1765                 /* Collapse range in clone heads */
1766                 dsl_dataset_remove_clones_key(ds,
1767                     ds->ds_phys->ds_creation_txg, tx);
1768
1769                 if (dsl_dataset_is_snapshot(ds_next)) {
1770                         dsl_dataset_t *ds_nextnext;
1771                         dsl_dataset_t *hds;
1772
1773                         /*
1774                          * Update next's unique to include blocks which
1775                          * were previously shared by only this snapshot
1776                          * and it.  Those blocks will be born after the
1777                          * prev snap and before this snap, and will have
1778                          * died after the next snap and before the one
1779                          * after that (ie. be on the snap after next's
1780                          * deadlist).
1781                          */
1782                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1783                             ds_next->ds_phys->ds_next_snap_obj,
1784                             FTAG, &ds_nextnext));
1785                         dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
1786                             ds->ds_phys->ds_prev_snap_txg,
1787                             ds->ds_phys->ds_creation_txg,
1788                             &used, &comp, &uncomp);
1789                         ds_next->ds_phys->ds_unique_bytes += used;
1790                         dsl_dataset_rele(ds_nextnext, FTAG);
1791                         ASSERT3P(ds_next->ds_prev, ==, NULL);
1792
1793                         /* Collapse range in this head. */
1794                         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
1795                             ds->ds_dir->dd_phys->dd_head_dataset_obj,
1796                             FTAG, &hds));
1797                         dsl_deadlist_remove_key(&hds->ds_deadlist,
1798                             ds->ds_phys->ds_creation_txg, tx);
1799                         dsl_dataset_rele(hds, FTAG);
1800
1801                 } else {
1802                         ASSERT3P(ds_next->ds_prev, ==, ds);
1803                         dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1804                         ds_next->ds_prev = NULL;
1805                         if (ds_prev) {
1806                                 VERIFY(0 == dsl_dataset_get_ref(dp,
1807                                     ds->ds_phys->ds_prev_snap_obj,
1808                                     ds_next, &ds_next->ds_prev));
1809                         }
1810
1811                         dsl_dataset_recalc_head_uniq(ds_next);
1812
1813                         /*
1814                          * Reduce the amount of our unconsmed refreservation
1815                          * being charged to our parent by the amount of
1816                          * new unique data we have gained.
1817                          */
1818                         if (old_unique < ds_next->ds_reserved) {
1819                                 int64_t mrsdelta;
1820                                 uint64_t new_unique =
1821                                     ds_next->ds_phys->ds_unique_bytes;
1822
1823                                 ASSERT(old_unique <= new_unique);
1824                                 mrsdelta = MIN(new_unique - old_unique,
1825                                     ds_next->ds_reserved - old_unique);
1826                                 dsl_dir_diduse_space(ds->ds_dir,
1827                                     DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1828                         }
1829                 }
1830                 dsl_dataset_rele(ds_next, FTAG);
1831         } else {
1832                 /*
1833                  * There's no next snapshot, so this is a head dataset.
1834                  * Destroy the deadlist.  Unless it's a clone, the
1835                  * deadlist should be empty.  (If it's a clone, it's
1836                  * safe to ignore the deadlist contents.)
1837                  */
1838                 struct killarg ka;
1839
1840                 dsl_deadlist_close(&ds->ds_deadlist);
1841                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1842                 ds->ds_phys->ds_deadlist_obj = 0;
1843
1844                 /*
1845                  * Free everything that we point to (that's born after
1846                  * the previous snapshot, if we are a clone)
1847                  *
1848                  * NB: this should be very quick, because we already
1849                  * freed all the objects in open context.
1850                  */
1851                 ka.ds = ds;
1852                 ka.tx = tx;
1853                 err = traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1854                     TRAVERSE_POST, kill_blkptr, &ka);
1855                 ASSERT3U(err, ==, 0);
1856                 ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1857                     ds->ds_phys->ds_unique_bytes == 0);
1858
1859                 if (ds->ds_prev != NULL) {
1860                         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
1861                                 VERIFY3U(0, ==, zap_remove_int(mos,
1862                                     ds->ds_prev->ds_dir->dd_phys->dd_clones,
1863                                     ds->ds_object, tx));
1864                         }
1865                         dsl_dataset_rele(ds->ds_prev, ds);
1866                         ds->ds_prev = ds_prev = NULL;
1867                 }
1868         }
1869
1870         /*
1871          * This must be done after the dsl_traverse(), because it will
1872          * re-open the objset.
1873          */
1874         if (ds->ds_objset) {
1875                 dmu_objset_evict(ds->ds_objset);
1876                 ds->ds_objset = NULL;
1877         }
1878
1879         if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1880                 /* Erase the link in the dir */
1881                 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1882                 ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1883                 ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1884                 err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1885                 ASSERT(err == 0);
1886         } else {
1887                 /* remove from snapshot namespace */
1888                 dsl_dataset_t *ds_head;
1889                 ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1890                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1891                     ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1892                 VERIFY(0 == dsl_dataset_get_snapname(ds));
1893 #ifdef ZFS_DEBUG
1894                 {
1895                         uint64_t val;
1896
1897                         err = dsl_dataset_snap_lookup(ds_head,
1898                             ds->ds_snapname, &val);
1899                         ASSERT3U(err, ==, 0);
1900                         ASSERT3U(val, ==, obj);
1901                 }
1902 #endif
1903                 err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1904                 ASSERT(err == 0);
1905                 dsl_dataset_rele(ds_head, FTAG);
1906         }
1907
1908         if (ds_prev && ds->ds_prev != ds_prev)
1909                 dsl_dataset_rele(ds_prev, FTAG);
1910
1911         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1912         spa_history_log_internal(LOG_DS_DESTROY, dp->dp_spa, tx,
1913             "dataset = %llu", ds->ds_object);
1914
1915         if (ds->ds_phys->ds_next_clones_obj != 0) {
1916                 ASSERTV(uint64_t count);
1917                 ASSERT(0 == zap_count(mos,
1918                     ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1919                 VERIFY(0 == dmu_object_free(mos,
1920                     ds->ds_phys->ds_next_clones_obj, tx));
1921         }
1922         if (ds->ds_phys->ds_props_obj != 0)
1923                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1924         if (ds->ds_phys->ds_userrefs_obj != 0)
1925                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1926         dsl_dir_close(ds->ds_dir, ds);
1927         ds->ds_dir = NULL;
1928         dsl_dataset_drain_refs(ds, tag);
1929         VERIFY(0 == dmu_object_free(mos, obj, tx));
1930
1931         if (dsda->rm_origin) {
1932                 /*
1933                  * Remove the origin of the clone we just destroyed.
1934                  */
1935                 struct dsl_ds_destroyarg ndsda = {0};
1936
1937                 ndsda.ds = dsda->rm_origin;
1938                 dsl_dataset_destroy_sync(&ndsda, tag, tx);
1939         }
1940 }
1941
1942 static int
1943 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
1944 {
1945         uint64_t asize;
1946
1947         if (!dmu_tx_is_syncing(tx))
1948                 return (0);
1949
1950         /*
1951          * If there's an fs-only reservation, any blocks that might become
1952          * owned by the snapshot dataset must be accommodated by space
1953          * outside of the reservation.
1954          */
1955         ASSERT(ds->ds_reserved == 0 || DS_UNIQUE_IS_ACCURATE(ds));
1956         asize = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
1957         if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
1958                 return (ENOSPC);
1959
1960         /*
1961          * Propogate any reserved space for this snapshot to other
1962          * snapshot checks in this sync group.
1963          */
1964         if (asize > 0)
1965                 dsl_dir_willuse_space(ds->ds_dir, asize, tx);
1966
1967         return (0);
1968 }
1969
1970 int
1971 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1972 {
1973         dsl_dataset_t *ds = arg1;
1974         const char *snapname = arg2;
1975         int err;
1976         uint64_t value;
1977
1978         /*
1979          * We don't allow multiple snapshots of the same txg.  If there
1980          * is already one, try again.
1981          */
1982         if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
1983                 return (EAGAIN);
1984
1985         /*
1986          * Check for conflicting name snapshot name.
1987          */
1988         err = dsl_dataset_snap_lookup(ds, snapname, &value);
1989         if (err == 0)
1990                 return (EEXIST);
1991         if (err != ENOENT)
1992                 return (err);
1993
1994         /*
1995          * Check that the dataset's name is not too long.  Name consists
1996          * of the dataset's length + 1 for the @-sign + snapshot name's length
1997          */
1998         if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
1999                 return (ENAMETOOLONG);
2000
2001         err = dsl_dataset_snapshot_reserve_space(ds, tx);
2002         if (err)
2003                 return (err);
2004
2005         ds->ds_trysnap_txg = tx->tx_txg;
2006         return (0);
2007 }
2008
2009 void
2010 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2011 {
2012         dsl_dataset_t *ds = arg1;
2013         const char *snapname = arg2;
2014         dsl_pool_t *dp = ds->ds_dir->dd_pool;
2015         dmu_buf_t *dbuf;
2016         dsl_dataset_phys_t *dsphys;
2017         uint64_t dsobj, crtxg;
2018         objset_t *mos = dp->dp_meta_objset;
2019         int err;
2020
2021         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
2022
2023         /*
2024          * The origin's ds_creation_txg has to be < TXG_INITIAL
2025          */
2026         if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2027                 crtxg = 1;
2028         else
2029                 crtxg = tx->tx_txg;
2030
2031         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2032             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2033         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2034         dmu_buf_will_dirty(dbuf, tx);
2035         dsphys = dbuf->db_data;
2036         bzero(dsphys, sizeof (dsl_dataset_phys_t));
2037         dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2038         dsphys->ds_fsid_guid = unique_create();
2039         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2040             sizeof (dsphys->ds_guid));
2041         dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2042         dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2043         dsphys->ds_next_snap_obj = ds->ds_object;
2044         dsphys->ds_num_children = 1;
2045         dsphys->ds_creation_time = gethrestime_sec();
2046         dsphys->ds_creation_txg = crtxg;
2047         dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2048         dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
2049         dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2050         dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2051         dsphys->ds_flags = ds->ds_phys->ds_flags;
2052         dsphys->ds_bp = ds->ds_phys->ds_bp;
2053         dmu_buf_rele(dbuf, FTAG);
2054
2055         ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2056         if (ds->ds_prev) {
2057                 uint64_t next_clones_obj =
2058                     ds->ds_prev->ds_phys->ds_next_clones_obj;
2059                 ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2060                     ds->ds_object ||
2061                     ds->ds_prev->ds_phys->ds_num_children > 1);
2062                 if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2063                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2064                         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2065                             ds->ds_prev->ds_phys->ds_creation_txg);
2066                         ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2067                 } else if (next_clones_obj != 0) {
2068                         remove_from_next_clones(ds->ds_prev,
2069                             dsphys->ds_next_snap_obj, tx);
2070                         VERIFY3U(0, ==, zap_add_int(mos,
2071                             next_clones_obj, dsobj, tx));
2072                 }
2073         }
2074
2075         /*
2076          * If we have a reference-reservation on this dataset, we will
2077          * need to increase the amount of refreservation being charged
2078          * since our unique space is going to zero.
2079          */
2080         if (ds->ds_reserved) {
2081                 int64_t delta;
2082                 ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
2083                 delta = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2084                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2085                     delta, 0, 0, tx);
2086         }
2087
2088         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2089         zfs_dbgmsg("taking snapshot %s@%s/%llu; newkey=%llu",
2090             ds->ds_dir->dd_myname, snapname, dsobj,
2091             ds->ds_phys->ds_prev_snap_txg);
2092         ds->ds_phys->ds_deadlist_obj = dsl_deadlist_clone(&ds->ds_deadlist,
2093             UINT64_MAX, ds->ds_phys->ds_prev_snap_obj, tx);
2094         dsl_deadlist_close(&ds->ds_deadlist);
2095         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
2096         dsl_deadlist_add_key(&ds->ds_deadlist,
2097             ds->ds_phys->ds_prev_snap_txg, tx);
2098
2099         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2100         ds->ds_phys->ds_prev_snap_obj = dsobj;
2101         ds->ds_phys->ds_prev_snap_txg = crtxg;
2102         ds->ds_phys->ds_unique_bytes = 0;
2103         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2104                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2105
2106         err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2107             snapname, 8, 1, &dsobj, tx);
2108         ASSERT(err == 0);
2109
2110         if (ds->ds_prev)
2111                 dsl_dataset_drop_ref(ds->ds_prev, ds);
2112         VERIFY(0 == dsl_dataset_get_ref(dp,
2113             ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2114
2115         dsl_scan_ds_snapshotted(ds, tx);
2116
2117         dsl_dir_snap_cmtime_update(ds->ds_dir);
2118
2119         spa_history_log_internal(LOG_DS_SNAPSHOT, dp->dp_spa, tx,
2120             "dataset = %llu", dsobj);
2121 }
2122
2123 void
2124 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2125 {
2126         ASSERT(dmu_tx_is_syncing(tx));
2127         ASSERT(ds->ds_objset != NULL);
2128         ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2129
2130         /*
2131          * in case we had to change ds_fsid_guid when we opened it,
2132          * sync it out now.
2133          */
2134         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2135         ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2136
2137         dsl_dir_dirty(ds->ds_dir, tx);
2138         dmu_objset_sync(ds->ds_objset, zio, tx);
2139 }
2140
2141 static void
2142 get_clones_stat(dsl_dataset_t *ds, nvlist_t *nv)
2143 {
2144         uint64_t count = 0;
2145         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
2146         zap_cursor_t zc;
2147         zap_attribute_t za;
2148         nvlist_t *propval;
2149         nvlist_t *val;
2150
2151         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2152         VERIFY(nvlist_alloc(&propval, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2153         VERIFY(nvlist_alloc(&val, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2154
2155         /*
2156          * There may me missing entries in ds_next_clones_obj
2157          * due to a bug in a previous version of the code.
2158          * Only trust it if it has the right number of entries.
2159          */
2160         if (ds->ds_phys->ds_next_clones_obj != 0) {
2161                 ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
2162                     &count));
2163         }
2164         if (count != ds->ds_phys->ds_num_children - 1) {
2165                 goto fail;
2166         }
2167         for (zap_cursor_init(&zc, mos, ds->ds_phys->ds_next_clones_obj);
2168             zap_cursor_retrieve(&zc, &za) == 0;
2169             zap_cursor_advance(&zc)) {
2170                 dsl_dataset_t *clone;
2171                 char buf[ZFS_MAXNAMELEN];
2172                 if (dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
2173                     za.za_first_integer, FTAG, &clone) != 0) {
2174                         goto fail;
2175                 }
2176                 dsl_dir_name(clone->ds_dir, buf);
2177                 VERIFY(nvlist_add_boolean(val, buf) == 0);
2178                 dsl_dataset_rele(clone, FTAG);
2179         }
2180         zap_cursor_fini(&zc);
2181         VERIFY(nvlist_add_nvlist(propval, ZPROP_VALUE, val) == 0);
2182         VERIFY(nvlist_add_nvlist(nv, zfs_prop_to_name(ZFS_PROP_CLONES),
2183             propval) == 0);
2184 fail:
2185         nvlist_free(val);
2186         nvlist_free(propval);
2187         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2188 }
2189
2190 void
2191 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2192 {
2193         uint64_t refd, avail, uobjs, aobjs, ratio;
2194
2195         dsl_dir_stats(ds->ds_dir, nv);
2196
2197         dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2198         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2199         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2200
2201         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2202             ds->ds_phys->ds_creation_time);
2203         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2204             ds->ds_phys->ds_creation_txg);
2205         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2206             ds->ds_quota);
2207         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2208             ds->ds_reserved);
2209         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2210             ds->ds_phys->ds_guid);
2211         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_UNIQUE,
2212             ds->ds_phys->ds_unique_bytes);
2213         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_OBJSETID,
2214             ds->ds_object);
2215         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS,
2216             ds->ds_userrefs);
2217         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2218             DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2219
2220         if (ds->ds_phys->ds_prev_snap_obj != 0) {
2221                 uint64_t written, comp, uncomp;
2222                 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2223                 dsl_dataset_t *prev;
2224                 int err;
2225
2226                 rw_enter(&dp->dp_config_rwlock, RW_READER);
2227                 err = dsl_dataset_hold_obj(dp,
2228                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
2229                 rw_exit(&dp->dp_config_rwlock);
2230                 if (err == 0) {
2231                         err = dsl_dataset_space_written(prev, ds, &written,
2232                             &comp, &uncomp);
2233                         dsl_dataset_rele(prev, FTAG);
2234                         if (err == 0) {
2235                                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_WRITTEN,
2236                                     written);
2237                         }
2238                 }
2239         }
2240
2241         ratio = ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2242             (ds->ds_phys->ds_uncompressed_bytes * 100 /
2243             ds->ds_phys->ds_compressed_bytes);
2244         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRATIO, ratio);
2245
2246         if (ds->ds_phys->ds_next_snap_obj) {
2247                 /*
2248                  * This is a snapshot; override the dd's space used with
2249                  * our unique space and compression ratio.
2250                  */
2251                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2252                     ds->ds_phys->ds_unique_bytes);
2253                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO, ratio);
2254
2255                 get_clones_stat(ds, nv);
2256         }
2257 }
2258
2259 void
2260 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2261 {
2262         stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2263         stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2264         stat->dds_guid = ds->ds_phys->ds_guid;
2265         if (ds->ds_phys->ds_next_snap_obj) {
2266                 stat->dds_is_snapshot = B_TRUE;
2267                 stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2268         } else {
2269                 stat->dds_is_snapshot = B_FALSE;
2270                 stat->dds_num_clones = 0;
2271         }
2272
2273         /* clone origin is really a dsl_dir thing... */
2274         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2275         if (dsl_dir_is_clone(ds->ds_dir)) {
2276                 dsl_dataset_t *ods;
2277
2278                 VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2279                     ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2280                 dsl_dataset_name(ods, stat->dds_origin);
2281                 dsl_dataset_drop_ref(ods, FTAG);
2282         } else {
2283                 stat->dds_origin[0] = '\0';
2284         }
2285         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2286 }
2287
2288 uint64_t
2289 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2290 {
2291         return (ds->ds_fsid_guid);
2292 }
2293
2294 void
2295 dsl_dataset_space(dsl_dataset_t *ds,
2296     uint64_t *refdbytesp, uint64_t *availbytesp,
2297     uint64_t *usedobjsp, uint64_t *availobjsp)
2298 {
2299         *refdbytesp = ds->ds_phys->ds_used_bytes;
2300         *availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2301         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2302                 *availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2303         if (ds->ds_quota != 0) {
2304                 /*
2305                  * Adjust available bytes according to refquota
2306                  */
2307                 if (*refdbytesp < ds->ds_quota)
2308                         *availbytesp = MIN(*availbytesp,
2309                             ds->ds_quota - *refdbytesp);
2310                 else
2311                         *availbytesp = 0;
2312         }
2313         *usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2314         *availobjsp = DN_MAX_OBJECT - *usedobjsp;
2315 }
2316
2317 boolean_t
2318 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2319 {
2320         ASSERTV(dsl_pool_t *dp = ds->ds_dir->dd_pool);
2321
2322         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2323             dsl_pool_sync_context(dp));
2324         if (ds->ds_prev == NULL)
2325                 return (B_FALSE);
2326         if (ds->ds_phys->ds_bp.blk_birth >
2327             ds->ds_prev->ds_phys->ds_creation_txg) {
2328                 objset_t *os, *os_prev;
2329                 /*
2330                  * It may be that only the ZIL differs, because it was
2331                  * reset in the head.  Don't count that as being
2332                  * modified.
2333                  */
2334                 if (dmu_objset_from_ds(ds, &os) != 0)
2335                         return (B_TRUE);
2336                 if (dmu_objset_from_ds(ds->ds_prev, &os_prev) != 0)
2337                         return (B_TRUE);
2338                 return (bcmp(&os->os_phys->os_meta_dnode,
2339                     &os_prev->os_phys->os_meta_dnode,
2340                     sizeof (os->os_phys->os_meta_dnode)) != 0);
2341         }
2342         return (B_FALSE);
2343 }
2344
2345 /* ARGSUSED */
2346 static int
2347 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2348 {
2349         dsl_dataset_t *ds = arg1;
2350         char *newsnapname = arg2;
2351         dsl_dir_t *dd = ds->ds_dir;
2352         dsl_dataset_t *hds;
2353         uint64_t val;
2354         int err;
2355
2356         err = dsl_dataset_hold_obj(dd->dd_pool,
2357             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2358         if (err)
2359                 return (err);
2360
2361         /* new name better not be in use */
2362         err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2363         dsl_dataset_rele(hds, FTAG);
2364
2365         if (err == 0)
2366                 err = EEXIST;
2367         else if (err == ENOENT)
2368                 err = 0;
2369
2370         /* dataset name + 1 for the "@" + the new snapshot name must fit */
2371         if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2372                 err = ENAMETOOLONG;
2373
2374         return (err);
2375 }
2376
2377 static void
2378 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2379 {
2380         dsl_dataset_t *ds = arg1;
2381         const char *newsnapname = arg2;
2382         dsl_dir_t *dd = ds->ds_dir;
2383         objset_t *mos = dd->dd_pool->dp_meta_objset;
2384         dsl_dataset_t *hds;
2385         int err;
2386
2387         ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2388
2389         VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2390             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2391
2392         VERIFY(0 == dsl_dataset_get_snapname(ds));
2393         err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2394         ASSERT3U(err, ==, 0);
2395         mutex_enter(&ds->ds_lock);
2396         (void) strcpy(ds->ds_snapname, newsnapname);
2397         mutex_exit(&ds->ds_lock);
2398         err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2399             ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2400         ASSERT3U(err, ==, 0);
2401
2402         spa_history_log_internal(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2403             "dataset = %llu", ds->ds_object);
2404         dsl_dataset_rele(hds, FTAG);
2405 }
2406
2407 struct renamesnaparg {
2408         dsl_sync_task_group_t *dstg;
2409         char failed[MAXPATHLEN];
2410         char *oldsnap;
2411         char *newsnap;
2412 };
2413
2414 static int
2415 dsl_snapshot_rename_one(const char *name, void *arg)
2416 {
2417         struct renamesnaparg *ra = arg;
2418         dsl_dataset_t *ds = NULL;
2419         char *snapname;
2420         int err;
2421
2422         snapname = kmem_asprintf("%s@%s", name, ra->oldsnap);
2423         (void) strlcpy(ra->failed, snapname, sizeof (ra->failed));
2424
2425         /*
2426          * For recursive snapshot renames the parent won't be changing
2427          * so we just pass name for both the to/from argument.
2428          */
2429         err = zfs_secpolicy_rename_perms(snapname, snapname, CRED());
2430         if (err != 0) {
2431                 strfree(snapname);
2432                 return (err == ENOENT ? 0 : err);
2433         }
2434
2435 #ifdef _KERNEL
2436         /*
2437          * For all filesystems undergoing rename, we'll need to unmount it.
2438          */
2439         (void) zfs_unmount_snap(snapname, NULL);
2440 #endif
2441         err = dsl_dataset_hold(snapname, ra->dstg, &ds);
2442         strfree(snapname);
2443         if (err != 0)
2444                 return (err == ENOENT ? 0 : err);
2445
2446         dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2447             dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2448
2449         return (0);
2450 }
2451
2452 static int
2453 dsl_recursive_rename(char *oldname, const char *newname)
2454 {
2455         int err;
2456         struct renamesnaparg *ra;
2457         dsl_sync_task_t *dst;
2458         spa_t *spa;
2459         char *cp, *fsname = spa_strdup(oldname);
2460         int len = strlen(oldname) + 1;
2461
2462         /* truncate the snapshot name to get the fsname */
2463         cp = strchr(fsname, '@');
2464         *cp = '\0';
2465
2466         err = spa_open(fsname, &spa, FTAG);
2467         if (err) {
2468                 kmem_free(fsname, len);
2469                 return (err);
2470         }
2471         ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2472         ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2473
2474         ra->oldsnap = strchr(oldname, '@') + 1;
2475         ra->newsnap = strchr(newname, '@') + 1;
2476         *ra->failed = '\0';
2477
2478         err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2479             DS_FIND_CHILDREN);
2480         kmem_free(fsname, len);
2481
2482         if (err == 0) {
2483                 err = dsl_sync_task_group_wait(ra->dstg);
2484         }
2485
2486         for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2487             dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2488                 dsl_dataset_t *ds = dst->dst_arg1;
2489                 if (dst->dst_err) {
2490                         dsl_dir_name(ds->ds_dir, ra->failed);
2491                         (void) strlcat(ra->failed, "@", sizeof (ra->failed));
2492                         (void) strlcat(ra->failed, ra->newsnap,
2493                             sizeof (ra->failed));
2494                 }
2495                 dsl_dataset_rele(ds, ra->dstg);
2496         }
2497
2498         if (err)
2499                 (void) strlcpy(oldname, ra->failed, sizeof (ra->failed));
2500
2501         dsl_sync_task_group_destroy(ra->dstg);
2502         kmem_free(ra, sizeof (struct renamesnaparg));
2503         spa_close(spa, FTAG);
2504         return (err);
2505 }
2506
2507 static int
2508 dsl_valid_rename(const char *oldname, void *arg)
2509 {
2510         int delta = *(int *)arg;
2511
2512         if (strlen(oldname) + delta >= MAXNAMELEN)
2513                 return (ENAMETOOLONG);
2514
2515         return (0);
2516 }
2517
2518 #pragma weak dmu_objset_rename = dsl_dataset_rename
2519 int
2520 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2521 {
2522         dsl_dir_t *dd;
2523         dsl_dataset_t *ds;
2524         const char *tail;
2525         int err;
2526
2527         err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2528         if (err)
2529                 return (err);
2530
2531         if (tail == NULL) {
2532                 int delta = strlen(newname) - strlen(oldname);
2533
2534                 /* if we're growing, validate child name lengths */
2535                 if (delta > 0)
2536                         err = dmu_objset_find(oldname, dsl_valid_rename,
2537                             &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2538
2539                 if (err == 0)
2540                         err = dsl_dir_rename(dd, newname);
2541                 dsl_dir_close(dd, FTAG);
2542                 return (err);
2543         }
2544
2545         if (tail[0] != '@') {
2546                 /* the name ended in a nonexistent component */
2547                 dsl_dir_close(dd, FTAG);
2548                 return (ENOENT);
2549         }
2550
2551         dsl_dir_close(dd, FTAG);
2552
2553         /* new name must be snapshot in same filesystem */
2554         tail = strchr(newname, '@');
2555         if (tail == NULL)
2556                 return (EINVAL);
2557         tail++;
2558         if (strncmp(oldname, newname, tail - newname) != 0)
2559                 return (EXDEV);
2560
2561         if (recursive) {
2562                 err = dsl_recursive_rename(oldname, newname);
2563         } else {
2564                 err = dsl_dataset_hold(oldname, FTAG, &ds);
2565                 if (err)
2566                         return (err);
2567
2568                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2569                     dsl_dataset_snapshot_rename_check,
2570                     dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2571
2572                 dsl_dataset_rele(ds, FTAG);
2573         }
2574
2575         return (err);
2576 }
2577
2578 struct promotenode {
2579         list_node_t link;
2580         dsl_dataset_t *ds;
2581 };
2582
2583 struct promotearg {
2584         list_t shared_snaps, origin_snaps, clone_snaps;
2585         dsl_dataset_t *origin_origin;
2586         uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2587         char *err_ds;
2588 };
2589
2590 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2591
2592 static int
2593 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2594 {
2595         dsl_dataset_t *hds = arg1;
2596         struct promotearg *pa = arg2;
2597         struct promotenode *snap = list_head(&pa->shared_snaps);
2598         dsl_dataset_t *origin_ds = snap->ds;
2599         int err;
2600         uint64_t unused;
2601
2602         /* Check that it is a real clone */
2603         if (!dsl_dir_is_clone(hds->ds_dir))
2604                 return (EINVAL);
2605
2606         /* Since this is so expensive, don't do the preliminary check */
2607         if (!dmu_tx_is_syncing(tx))
2608                 return (0);
2609
2610         if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2611                 return (EXDEV);
2612
2613         /* compute origin's new unique space */
2614         snap = list_tail(&pa->clone_snaps);
2615         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2616         dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2617             origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
2618             &pa->unique, &unused, &unused);
2619
2620         /*
2621          * Walk the snapshots that we are moving
2622          *
2623          * Compute space to transfer.  Consider the incremental changes
2624          * to used for each snapshot:
2625          * (my used) = (prev's used) + (blocks born) - (blocks killed)
2626          * So each snapshot gave birth to:
2627          * (blocks born) = (my used) - (prev's used) + (blocks killed)
2628          * So a sequence would look like:
2629          * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2630          * Which simplifies to:
2631          * uN + kN + kN-1 + ... + k1 + k0
2632          * Note however, if we stop before we reach the ORIGIN we get:
2633          * uN + kN + kN-1 + ... + kM - uM-1
2634          */
2635         pa->used = origin_ds->ds_phys->ds_used_bytes;
2636         pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2637         pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2638         for (snap = list_head(&pa->shared_snaps); snap;
2639             snap = list_next(&pa->shared_snaps, snap)) {
2640                 uint64_t val, dlused, dlcomp, dluncomp;
2641                 dsl_dataset_t *ds = snap->ds;
2642
2643                 /* Check that the snapshot name does not conflict */
2644                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2645                 err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2646                 if (err == 0) {
2647                         err = EEXIST;
2648                         goto out;
2649                 }
2650                 if (err != ENOENT)
2651                         goto out;
2652
2653                 /* The very first snapshot does not have a deadlist */
2654                 if (ds->ds_phys->ds_prev_snap_obj == 0)
2655                         continue;
2656
2657                 dsl_deadlist_space(&ds->ds_deadlist,
2658                     &dlused, &dlcomp, &dluncomp);
2659                 pa->used += dlused;
2660                 pa->comp += dlcomp;
2661                 pa->uncomp += dluncomp;
2662         }
2663
2664         /*
2665          * If we are a clone of a clone then we never reached ORIGIN,
2666          * so we need to subtract out the clone origin's used space.
2667          */
2668         if (pa->origin_origin) {
2669                 pa->used -= pa->origin_origin->ds_phys->ds_used_bytes;
2670                 pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2671                 pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2672         }
2673
2674         /* Check that there is enough space here */
2675         err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2676             pa->used);
2677         if (err)
2678                 return (err);
2679
2680         /*
2681          * Compute the amounts of space that will be used by snapshots
2682          * after the promotion (for both origin and clone).  For each,
2683          * it is the amount of space that will be on all of their
2684          * deadlists (that was not born before their new origin).
2685          */
2686         if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2687                 uint64_t space;
2688
2689                 /*
2690                  * Note, typically this will not be a clone of a clone,
2691                  * so dd_origin_txg will be < TXG_INITIAL, so
2692                  * these snaplist_space() -> dsl_deadlist_space_range()
2693                  * calls will be fast because they do not have to
2694                  * iterate over all bps.
2695                  */
2696                 snap = list_head(&pa->origin_snaps);
2697                 err = snaplist_space(&pa->shared_snaps,
2698                     snap->ds->ds_dir->dd_origin_txg, &pa->cloneusedsnap);
2699                 if (err)
2700                         return (err);
2701
2702                 err = snaplist_space(&pa->clone_snaps,
2703                     snap->ds->ds_dir->dd_origin_txg, &space);
2704                 if (err)
2705                         return (err);
2706                 pa->cloneusedsnap += space;
2707         }
2708         if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2709                 err = snaplist_space(&pa->origin_snaps,
2710                     origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2711                 if (err)
2712                         return (err);
2713         }
2714
2715         return (0);
2716 out:
2717         pa->err_ds =  snap->ds->ds_snapname;
2718         return (err);
2719 }
2720
2721 static void
2722 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2723 {
2724         dsl_dataset_t *hds = arg1;
2725         struct promotearg *pa = arg2;
2726         struct promotenode *snap = list_head(&pa->shared_snaps);
2727         dsl_dataset_t *origin_ds = snap->ds;
2728         dsl_dataset_t *origin_head;
2729         dsl_dir_t *dd = hds->ds_dir;
2730         dsl_pool_t *dp = hds->ds_dir->dd_pool;
2731         dsl_dir_t *odd = NULL;
2732         uint64_t oldnext_obj;
2733         int64_t delta;
2734
2735         ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2736
2737         snap = list_head(&pa->origin_snaps);
2738         origin_head = snap->ds;
2739
2740         /*
2741          * We need to explicitly open odd, since origin_ds's dd will be
2742          * changing.
2743          */
2744         VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2745             NULL, FTAG, &odd));
2746
2747         /* change origin's next snap */
2748         dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2749         oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2750         snap = list_tail(&pa->clone_snaps);
2751         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2752         origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2753
2754         /* change the origin's next clone */
2755         if (origin_ds->ds_phys->ds_next_clones_obj) {
2756                 remove_from_next_clones(origin_ds, snap->ds->ds_object, tx);
2757                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2758                     origin_ds->ds_phys->ds_next_clones_obj,
2759                     oldnext_obj, tx));
2760         }
2761
2762         /* change origin */
2763         dmu_buf_will_dirty(dd->dd_dbuf, tx);
2764         ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2765         dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2766         dd->dd_origin_txg = origin_head->ds_dir->dd_origin_txg;
2767         dmu_buf_will_dirty(odd->dd_dbuf, tx);
2768         odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2769         origin_head->ds_dir->dd_origin_txg =
2770             origin_ds->ds_phys->ds_creation_txg;
2771
2772         /* change dd_clone entries */
2773         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2774                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2775                     odd->dd_phys->dd_clones, hds->ds_object, tx));
2776                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2777                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2778                     hds->ds_object, tx));
2779
2780                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2781                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2782                     origin_head->ds_object, tx));
2783                 if (dd->dd_phys->dd_clones == 0) {
2784                         dd->dd_phys->dd_clones = zap_create(dp->dp_meta_objset,
2785                             DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
2786                 }
2787                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2788                     dd->dd_phys->dd_clones, origin_head->ds_object, tx));
2789
2790         }
2791
2792         /* move snapshots to this dir */
2793         for (snap = list_head(&pa->shared_snaps); snap;
2794             snap = list_next(&pa->shared_snaps, snap)) {
2795                 dsl_dataset_t *ds = snap->ds;
2796
2797                 /* unregister props as dsl_dir is changing */
2798                 if (ds->ds_objset) {
2799                         dmu_objset_evict(ds->ds_objset);
2800                         ds->ds_objset = NULL;
2801                 }
2802                 /* move snap name entry */
2803                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2804                 VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2805                     ds->ds_snapname, tx));
2806                 VERIFY(0 == zap_add(dp->dp_meta_objset,
2807                     hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2808                     8, 1, &ds->ds_object, tx));
2809
2810                 /* change containing dsl_dir */
2811                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2812                 ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2813                 ds->ds_phys->ds_dir_obj = dd->dd_object;
2814                 ASSERT3P(ds->ds_dir, ==, odd);
2815                 dsl_dir_close(ds->ds_dir, ds);
2816                 VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2817                     NULL, ds, &ds->ds_dir));
2818
2819                 /* move any clone references */
2820                 if (ds->ds_phys->ds_next_clones_obj &&
2821                     spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2822                         zap_cursor_t zc;
2823                         zap_attribute_t za;
2824
2825                         for (zap_cursor_init(&zc, dp->dp_meta_objset,
2826                             ds->ds_phys->ds_next_clones_obj);
2827                             zap_cursor_retrieve(&zc, &za) == 0;
2828                             zap_cursor_advance(&zc)) {
2829                                 dsl_dataset_t *cnds;
2830                                 uint64_t o;
2831
2832                                 if (za.za_first_integer == oldnext_obj) {
2833                                         /*
2834                                          * We've already moved the
2835                                          * origin's reference.
2836                                          */
2837                                         continue;
2838                                 }
2839
2840                                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
2841                                     za.za_first_integer, FTAG, &cnds));
2842                                 o = cnds->ds_dir->dd_phys->dd_head_dataset_obj;
2843
2844                                 VERIFY3U(zap_remove_int(dp->dp_meta_objset,
2845                                     odd->dd_phys->dd_clones, o, tx), ==, 0);
2846                                 VERIFY3U(zap_add_int(dp->dp_meta_objset,
2847                                     dd->dd_phys->dd_clones, o, tx), ==, 0);
2848                                 dsl_dataset_rele(cnds, FTAG);
2849                         }
2850                         zap_cursor_fini(&zc);
2851                 }
2852
2853                 ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2854         }
2855
2856         /*
2857          * Change space accounting.
2858          * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2859          * both be valid, or both be 0 (resulting in delta == 0).  This
2860          * is true for each of {clone,origin} independently.
2861          */
2862
2863         delta = pa->cloneusedsnap -
2864             dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2865         ASSERT3S(delta, >=, 0);
2866         ASSERT3U(pa->used, >=, delta);
2867         dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2868         dsl_dir_diduse_space(dd, DD_USED_HEAD,
2869             pa->used - delta, pa->comp, pa->uncomp, tx);
2870
2871         delta = pa->originusedsnap -
2872             odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2873         ASSERT3S(delta, <=, 0);
2874         ASSERT3U(pa->used, >=, -delta);
2875         dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2876         dsl_dir_diduse_space(odd, DD_USED_HEAD,
2877             -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2878
2879         origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2880
2881         /* log history record */
2882         spa_history_log_internal(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2883             "dataset = %llu", hds->ds_object);
2884
2885         dsl_dir_close(odd, FTAG);
2886 }
2887
2888 static char *snaplist_tag = "snaplist";
2889 /*
2890  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2891  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2892  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2893  * snapshots back to this dataset's origin.
2894  */
2895 static int
2896 snaplist_make(dsl_pool_t *dp, boolean_t own,
2897     uint64_t first_obj, uint64_t last_obj, list_t *l)
2898 {
2899         uint64_t obj = last_obj;
2900
2901         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2902
2903         list_create(l, sizeof (struct promotenode),
2904             offsetof(struct promotenode, link));
2905
2906         while (obj != first_obj) {
2907                 dsl_dataset_t *ds;
2908                 struct promotenode *snap;
2909                 int err;
2910
2911                 if (own) {
2912                         err = dsl_dataset_own_obj(dp, obj,
2913                             0, snaplist_tag, &ds);
2914                         if (err == 0)
2915                                 dsl_dataset_make_exclusive(ds, snaplist_tag);
2916                 } else {
2917                         err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2918                 }
2919                 if (err == ENOENT) {
2920                         /* lost race with snapshot destroy */
2921                         struct promotenode *last = list_tail(l);
2922                         ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2923                         obj = last->ds->ds_phys->ds_prev_snap_obj;
2924                         continue;
2925                 } else if (err) {
2926                         return (err);
2927                 }
2928
2929                 if (first_obj == 0)
2930                         first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2931
2932                 snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
2933                 snap->ds = ds;
2934                 list_insert_tail(l, snap);
2935                 obj = ds->ds_phys->ds_prev_snap_obj;
2936         }
2937
2938         return (0);
2939 }
2940
2941 static int
2942 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
2943 {
2944         struct promotenode *snap;
2945
2946         *spacep = 0;
2947         for (snap = list_head(l); snap; snap = list_next(l, snap)) {
2948                 uint64_t used, comp, uncomp;
2949                 dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2950                     mintxg, UINT64_MAX, &used, &comp, &uncomp);
2951                 *spacep += used;
2952         }
2953         return (0);
2954 }
2955
2956 static void
2957 snaplist_destroy(list_t *l, boolean_t own)
2958 {
2959         struct promotenode *snap;
2960
2961         if (!l || !list_link_active(&l->list_head))
2962                 return;
2963
2964         while ((snap = list_tail(l)) != NULL) {
2965                 list_remove(l, snap);
2966                 if (own)
2967                         dsl_dataset_disown(snap->ds, snaplist_tag);
2968                 else
2969                         dsl_dataset_rele(snap->ds, snaplist_tag);
2970                 kmem_free(snap, sizeof (struct promotenode));
2971         }
2972         list_destroy(l);
2973 }
2974
2975 /*
2976  * Promote a clone.  Nomenclature note:
2977  * "clone" or "cds": the original clone which is being promoted
2978  * "origin" or "ods": the snapshot which is originally clone's origin
2979  * "origin head" or "ohds": the dataset which is the head
2980  * (filesystem/volume) for the origin
2981  * "origin origin": the origin of the origin's filesystem (typically
2982  * NULL, indicating that the clone is not a clone of a clone).
2983  */
2984 int
2985 dsl_dataset_promote(const char *name, char *conflsnap)
2986 {
2987         dsl_dataset_t *ds;
2988         dsl_dir_t *dd;
2989         dsl_pool_t *dp;
2990         dmu_object_info_t doi;
2991         struct promotearg pa;
2992         struct promotenode *snap;
2993         int err;
2994
2995         bzero(&pa, sizeof(struct promotearg));
2996         err = dsl_dataset_hold(name, FTAG, &ds);
2997         if (err)
2998                 return (err);
2999         dd = ds->ds_dir;
3000         dp = dd->dd_pool;
3001
3002         err = dmu_object_info(dp->dp_meta_objset,
3003             ds->ds_phys->ds_snapnames_zapobj, &doi);
3004         if (err) {
3005                 dsl_dataset_rele(ds, FTAG);
3006                 return (err);
3007         }
3008
3009         if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
3010                 dsl_dataset_rele(ds, FTAG);
3011                 return (EINVAL);
3012         }
3013
3014         /*
3015          * We are going to inherit all the snapshots taken before our
3016          * origin (i.e., our new origin will be our parent's origin).
3017          * Take ownership of them so that we can rename them into our
3018          * namespace.
3019          */
3020         rw_enter(&dp->dp_config_rwlock, RW_READER);
3021
3022         err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
3023             &pa.shared_snaps);
3024         if (err != 0)
3025                 goto out;
3026
3027         err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
3028         if (err != 0)
3029                 goto out;
3030
3031         snap = list_head(&pa.shared_snaps);
3032         ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
3033         err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
3034             snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
3035         if (err != 0)
3036                 goto out;
3037
3038         if (snap->ds->ds_dir->dd_phys->dd_origin_obj != 0) {
3039                 err = dsl_dataset_hold_obj(dp,
3040                     snap->ds->ds_dir->dd_phys->dd_origin_obj,
3041                     FTAG, &pa.origin_origin);
3042                 if (err != 0)
3043                         goto out;
3044         }
3045
3046 out:
3047         rw_exit(&dp->dp_config_rwlock);
3048
3049         /*
3050          * Add in 128x the snapnames zapobj size, since we will be moving
3051          * a bunch of snapnames to the promoted ds, and dirtying their
3052          * bonus buffers.
3053          */
3054         if (err == 0) {
3055                 err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
3056                     dsl_dataset_promote_sync, ds, &pa,
3057                     2 + 2 * doi.doi_physical_blocks_512);
3058                 if (err && pa.err_ds && conflsnap)
3059                         (void) strncpy(conflsnap, pa.err_ds, MAXNAMELEN);
3060         }
3061
3062         snaplist_destroy(&pa.shared_snaps, B_TRUE);
3063         snaplist_destroy(&pa.clone_snaps, B_FALSE);
3064         snaplist_destroy(&pa.origin_snaps, B_FALSE);
3065         if (pa.origin_origin)
3066                 dsl_dataset_rele(pa.origin_origin, FTAG);
3067         dsl_dataset_rele(ds, FTAG);
3068         return (err);
3069 }
3070
3071 struct cloneswaparg {
3072         dsl_dataset_t *cds; /* clone dataset */
3073         dsl_dataset_t *ohds; /* origin's head dataset */
3074         boolean_t force;
3075         int64_t unused_refres_delta; /* change in unconsumed refreservation */
3076 };
3077
3078 /* ARGSUSED */
3079 static int
3080 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
3081 {
3082         struct cloneswaparg *csa = arg1;
3083
3084         /* they should both be heads */
3085         if (dsl_dataset_is_snapshot(csa->cds) ||
3086             dsl_dataset_is_snapshot(csa->ohds))
3087                 return (EINVAL);
3088
3089         /* the branch point should be just before them */
3090         if (csa->cds->ds_prev != csa->ohds->ds_prev)
3091                 return (EINVAL);
3092
3093         /* cds should be the clone (unless they are unrelated) */
3094         if (csa->cds->ds_prev != NULL &&
3095             csa->cds->ds_prev != csa->cds->ds_dir->dd_pool->dp_origin_snap &&
3096             csa->ohds->ds_object !=
3097             csa->cds->ds_prev->ds_phys->ds_next_snap_obj)
3098                 return (EINVAL);
3099
3100         /* the clone should be a child of the origin */
3101         if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
3102                 return (EINVAL);
3103
3104         /* ohds shouldn't be modified unless 'force' */
3105         if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
3106                 return (ETXTBSY);
3107
3108         /* adjust amount of any unconsumed refreservation */
3109         csa->unused_refres_delta =
3110             (int64_t)MIN(csa->ohds->ds_reserved,
3111             csa->ohds->ds_phys->ds_unique_bytes) -
3112             (int64_t)MIN(csa->ohds->ds_reserved,
3113             csa->cds->ds_phys->ds_unique_bytes);
3114
3115         if (csa->unused_refres_delta > 0 &&
3116             csa->unused_refres_delta >
3117             dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3118                 return (ENOSPC);
3119
3120         if (csa->ohds->ds_quota != 0 &&
3121             csa->cds->ds_phys->ds_unique_bytes > csa->ohds->ds_quota)
3122                 return (EDQUOT);
3123
3124         return (0);
3125 }
3126
3127 /* ARGSUSED */
3128 static void
3129 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3130 {
3131         struct cloneswaparg *csa = arg1;
3132         dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3133
3134         ASSERT(csa->cds->ds_reserved == 0);
3135         ASSERT(csa->ohds->ds_quota == 0 ||
3136             csa->cds->ds_phys->ds_unique_bytes <= csa->ohds->ds_quota);
3137
3138         dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3139         dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3140
3141         if (csa->cds->ds_objset != NULL) {
3142                 dmu_objset_evict(csa->cds->ds_objset);
3143                 csa->cds->ds_objset = NULL;
3144         }
3145
3146         if (csa->ohds->ds_objset != NULL) {
3147                 dmu_objset_evict(csa->ohds->ds_objset);
3148                 csa->ohds->ds_objset = NULL;
3149         }
3150
3151         /*
3152          * Reset origin's unique bytes, if it exists.
3153          */
3154         if (csa->cds->ds_prev) {
3155                 dsl_dataset_t *origin = csa->cds->ds_prev;
3156                 uint64_t comp, uncomp;
3157
3158                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
3159                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3160                     origin->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3161                     &origin->ds_phys->ds_unique_bytes, &comp, &uncomp);
3162         }
3163
3164         /* swap blkptrs */
3165         {
3166                 blkptr_t tmp;
3167                 tmp = csa->ohds->ds_phys->ds_bp;
3168                 csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3169                 csa->cds->ds_phys->ds_bp = tmp;
3170         }
3171
3172         /* set dd_*_bytes */
3173         {
3174                 int64_t dused, dcomp, duncomp;
3175                 uint64_t cdl_used, cdl_comp, cdl_uncomp;
3176                 uint64_t odl_used, odl_comp, odl_uncomp;
3177
3178                 ASSERT3U(csa->cds->ds_dir->dd_phys->
3179                     dd_used_breakdown[DD_USED_SNAP], ==, 0);
3180
3181                 dsl_deadlist_space(&csa->cds->ds_deadlist,
3182                     &cdl_used, &cdl_comp, &cdl_uncomp);
3183                 dsl_deadlist_space(&csa->ohds->ds_deadlist,
3184                     &odl_used, &odl_comp, &odl_uncomp);
3185
3186                 dused = csa->cds->ds_phys->ds_used_bytes + cdl_used -
3187                     (csa->ohds->ds_phys->ds_used_bytes + odl_used);
3188                 dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3189                     (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3190                 duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3191                     cdl_uncomp -
3192                     (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3193
3194                 dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3195                     dused, dcomp, duncomp, tx);
3196                 dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3197                     -dused, -dcomp, -duncomp, tx);
3198
3199                 /*
3200                  * The difference in the space used by snapshots is the
3201                  * difference in snapshot space due to the head's
3202                  * deadlist (since that's the only thing that's
3203                  * changing that affects the snapused).
3204                  */
3205                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3206                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3207                     &cdl_used, &cdl_comp, &cdl_uncomp);
3208                 dsl_deadlist_space_range(&csa->ohds->ds_deadlist,
3209                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3210                     &odl_used, &odl_comp, &odl_uncomp);
3211                 dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3212                     DD_USED_HEAD, DD_USED_SNAP, tx);
3213         }
3214
3215         /* swap ds_*_bytes */
3216         SWITCH64(csa->ohds->ds_phys->ds_used_bytes,
3217             csa->cds->ds_phys->ds_used_bytes);
3218         SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3219             csa->cds->ds_phys->ds_compressed_bytes);
3220         SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3221             csa->cds->ds_phys->ds_uncompressed_bytes);
3222         SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3223             csa->cds->ds_phys->ds_unique_bytes);
3224
3225         /* apply any parent delta for change in unconsumed refreservation */
3226         dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3227             csa->unused_refres_delta, 0, 0, tx);
3228
3229         /*
3230          * Swap deadlists.
3231          */
3232         dsl_deadlist_close(&csa->cds->ds_deadlist);
3233         dsl_deadlist_close(&csa->ohds->ds_deadlist);
3234         SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3235             csa->cds->ds_phys->ds_deadlist_obj);
3236         dsl_deadlist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3237             csa->cds->ds_phys->ds_deadlist_obj);
3238         dsl_deadlist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3239             csa->ohds->ds_phys->ds_deadlist_obj);
3240
3241         dsl_scan_ds_clone_swapped(csa->ohds, csa->cds, tx);
3242 }
3243
3244 /*
3245  * Swap 'clone' with its origin head datasets.  Used at the end of "zfs
3246  * recv" into an existing fs to swizzle the file system to the new
3247  * version, and by "zfs rollback".  Can also be used to swap two
3248  * independent head datasets if neither has any snapshots.
3249  */
3250 int
3251 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3252     boolean_t force)
3253 {
3254         struct cloneswaparg csa;
3255         int error;
3256
3257         ASSERT(clone->ds_owner);
3258         ASSERT(origin_head->ds_owner);
3259 retry:
3260         /*
3261          * Need exclusive access for the swap. If we're swapping these
3262          * datasets back after an error, we already hold the locks.
3263          */
3264         if (!RW_WRITE_HELD(&clone->ds_rwlock))
3265                 rw_enter(&clone->ds_rwlock, RW_WRITER);
3266         if (!RW_WRITE_HELD(&origin_head->ds_rwlock) &&
3267             !rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3268                 rw_exit(&clone->ds_rwlock);
3269                 rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3270                 if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3271                         rw_exit(&origin_head->ds_rwlock);
3272                         goto retry;
3273                 }
3274         }
3275         csa.cds = clone;
3276         csa.ohds = origin_head;
3277         csa.force = force;
3278         error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3279             dsl_dataset_clone_swap_check,
3280             dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3281         return (error);
3282 }
3283
3284 /*
3285  * Given a pool name and a dataset object number in that pool,
3286  * return the name of that dataset.
3287  */
3288 int
3289 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3290 {
3291         spa_t *spa;
3292         dsl_pool_t *dp;
3293         dsl_dataset_t *ds;
3294         int error;
3295
3296         if ((error = spa_open(pname, &spa, FTAG)) != 0)
3297                 return (error);
3298         dp = spa_get_dsl(spa);
3299         rw_enter(&dp->dp_config_rwlock, RW_READER);
3300         if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3301                 dsl_dataset_name(ds, buf);
3302                 dsl_dataset_rele(ds, FTAG);
3303         }
3304         rw_exit(&dp->dp_config_rwlock);
3305         spa_close(spa, FTAG);
3306
3307         return (error);
3308 }
3309
3310 int
3311 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3312     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3313 {
3314         int error = 0;
3315
3316         ASSERT3S(asize, >, 0);
3317
3318         /*
3319          * *ref_rsrv is the portion of asize that will come from any
3320          * unconsumed refreservation space.
3321          */
3322         *ref_rsrv = 0;
3323
3324         mutex_enter(&ds->ds_lock);
3325         /*
3326          * Make a space adjustment for reserved bytes.
3327          */
3328         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3329                 ASSERT3U(*used, >=,
3330                     ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3331                 *used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3332                 *ref_rsrv =
3333                     asize - MIN(asize, parent_delta(ds, asize + inflight));
3334         }
3335
3336         if (!check_quota || ds->ds_quota == 0) {
3337                 mutex_exit(&ds->ds_lock);
3338                 return (0);
3339         }
3340         /*
3341          * If they are requesting more space, and our current estimate
3342          * is over quota, they get to try again unless the actual
3343          * on-disk is over quota and there are no pending changes (which
3344          * may free up space for us).
3345          */
3346         if (ds->ds_phys->ds_used_bytes + inflight >= ds->ds_quota) {
3347                 if (inflight > 0 || ds->ds_phys->ds_used_bytes < ds->ds_quota)
3348                         error = ERESTART;
3349                 else
3350                         error = EDQUOT;
3351
3352                 DMU_TX_STAT_BUMP(dmu_tx_quota);
3353         }
3354         mutex_exit(&ds->ds_lock);
3355
3356         return (error);
3357 }
3358
3359 /* ARGSUSED */
3360 static int
3361 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3362 {
3363         dsl_dataset_t *ds = arg1;
3364         dsl_prop_setarg_t *psa = arg2;
3365         int err;
3366
3367         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3368                 return (ENOTSUP);
3369
3370         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3371                 return (err);
3372
3373         if (psa->psa_effective_value == 0)
3374                 return (0);
3375
3376         if (psa->psa_effective_value < ds->ds_phys->ds_used_bytes ||
3377             psa->psa_effective_value < ds->ds_reserved)
3378                 return (ENOSPC);
3379
3380         return (0);
3381 }
3382
3383 extern void dsl_prop_set_sync(void *, void *, dmu_tx_t *);
3384
3385 void
3386 dsl_dataset_set_quota_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3387 {
3388         dsl_dataset_t *ds = arg1;
3389         dsl_prop_setarg_t *psa = arg2;
3390         uint64_t effective_value = psa->psa_effective_value;
3391
3392         dsl_prop_set_sync(ds, psa, tx);
3393         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3394
3395         if (ds->ds_quota != effective_value) {
3396                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3397                 ds->ds_quota = effective_value;
3398
3399                 spa_history_log_internal(LOG_DS_REFQUOTA,
3400                     ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu ",
3401                     (longlong_t)ds->ds_quota, ds->ds_object);
3402         }
3403 }
3404
3405 int
3406 dsl_dataset_set_quota(const char *dsname, zprop_source_t source, uint64_t quota)
3407 {
3408         dsl_dataset_t *ds;
3409         dsl_prop_setarg_t psa;
3410         int err;
3411
3412         dsl_prop_setarg_init_uint64(&psa, "refquota", source, &quota);
3413
3414         err = dsl_dataset_hold(dsname, FTAG, &ds);
3415         if (err)
3416                 return (err);
3417
3418         /*
3419          * If someone removes a file, then tries to set the quota, we
3420          * want to make sure the file freeing takes effect.
3421          */
3422         txg_wait_open(ds->ds_dir->dd_pool, 0);
3423
3424         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3425             dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3426             ds, &psa, 0);
3427
3428         dsl_dataset_rele(ds, FTAG);
3429         return (err);
3430 }
3431
3432 static int
3433 dsl_dataset_set_reservation_check(void *arg1, void *arg2, dmu_tx_t *tx)
3434 {
3435         dsl_dataset_t *ds = arg1;
3436         dsl_prop_setarg_t *psa = arg2;
3437         uint64_t effective_value;
3438         uint64_t unique;
3439         int err;
3440
3441         if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3442             SPA_VERSION_REFRESERVATION)
3443                 return (ENOTSUP);
3444
3445         if (dsl_dataset_is_snapshot(ds))
3446                 return (EINVAL);
3447
3448         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3449                 return (err);
3450
3451         effective_value = psa->psa_effective_value;
3452
3453         /*
3454          * If we are doing the preliminary check in open context, the
3455          * space estimates may be inaccurate.
3456          */
3457         if (!dmu_tx_is_syncing(tx))
3458                 return (0);
3459
3460         mutex_enter(&ds->ds_lock);
3461         if (!DS_UNIQUE_IS_ACCURATE(ds))
3462                 dsl_dataset_recalc_head_uniq(ds);
3463         unique = ds->ds_phys->ds_unique_bytes;
3464         mutex_exit(&ds->ds_lock);
3465
3466         if (MAX(unique, effective_value) > MAX(unique, ds->ds_reserved)) {
3467                 uint64_t delta = MAX(unique, effective_value) -
3468                     MAX(unique, ds->ds_reserved);
3469
3470                 if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3471                         return (ENOSPC);
3472                 if (ds->ds_quota > 0 &&
3473                     effective_value > ds->ds_quota)
3474                         return (ENOSPC);
3475         }
3476
3477         return (0);
3478 }
3479
3480 static void
3481 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3482 {
3483         dsl_dataset_t *ds = arg1;
3484         dsl_prop_setarg_t *psa = arg2;
3485         uint64_t effective_value = psa->psa_effective_value;
3486         uint64_t unique;
3487         int64_t delta;
3488
3489         dsl_prop_set_sync(ds, psa, tx);
3490         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3491
3492         dmu_buf_will_dirty(ds->ds_dbuf, tx);
3493
3494         mutex_enter(&ds->ds_dir->dd_lock);
3495         mutex_enter(&ds->ds_lock);
3496         ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
3497         unique = ds->ds_phys->ds_unique_bytes;
3498         delta = MAX(0, (int64_t)(effective_value - unique)) -
3499             MAX(0, (int64_t)(ds->ds_reserved - unique));
3500         ds->ds_reserved = effective_value;
3501         mutex_exit(&ds->ds_lock);
3502
3503         dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3504         mutex_exit(&ds->ds_dir->dd_lock);
3505
3506         spa_history_log_internal(LOG_DS_REFRESERV,
3507             ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu",
3508             (longlong_t)effective_value, ds->ds_object);
3509 }
3510
3511 int
3512 dsl_dataset_set_reservation(const char *dsname, zprop_source_t source,
3513     uint64_t reservation)
3514 {
3515         dsl_dataset_t *ds;
3516         dsl_prop_setarg_t psa;
3517         int err;
3518
3519         dsl_prop_setarg_init_uint64(&psa, "refreservation", source,
3520             &reservation);
3521
3522         err = dsl_dataset_hold(dsname, FTAG, &ds);
3523         if (err)
3524                 return (err);
3525
3526         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3527             dsl_dataset_set_reservation_check,
3528             dsl_dataset_set_reservation_sync, ds, &psa, 0);
3529
3530         dsl_dataset_rele(ds, FTAG);
3531         return (err);
3532 }
3533
3534 typedef struct zfs_hold_cleanup_arg {
3535         dsl_pool_t *dp;
3536         uint64_t dsobj;
3537         char htag[MAXNAMELEN];
3538 } zfs_hold_cleanup_arg_t;
3539
3540 static void
3541 dsl_dataset_user_release_onexit(void *arg)
3542 {
3543         zfs_hold_cleanup_arg_t *ca = arg;
3544
3545         (void) dsl_dataset_user_release_tmp(ca->dp, ca->dsobj, ca->htag,
3546             B_TRUE);
3547         kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t));
3548 }
3549
3550 void
3551 dsl_register_onexit_hold_cleanup(dsl_dataset_t *ds, const char *htag,
3552     minor_t minor)
3553 {
3554         zfs_hold_cleanup_arg_t *ca;
3555
3556         ca = kmem_alloc(sizeof (zfs_hold_cleanup_arg_t), KM_SLEEP);
3557         ca->dp = ds->ds_dir->dd_pool;
3558         ca->dsobj = ds->ds_object;
3559         (void) strlcpy(ca->htag, htag, sizeof (ca->htag));
3560         VERIFY3U(0, ==, zfs_onexit_add_cb(minor,
3561             dsl_dataset_user_release_onexit, ca, NULL));
3562 }
3563
3564 /*
3565  * If you add new checks here, you may need to add
3566  * additional checks to the "temporary" case in
3567  * snapshot_check() in dmu_objset.c.
3568  */
3569 static int
3570 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3571 {
3572         dsl_dataset_t *ds = arg1;
3573         struct dsl_ds_holdarg *ha = arg2;
3574         char *htag = ha->htag;
3575         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3576         int error = 0;
3577
3578         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3579                 return (ENOTSUP);
3580
3581         if (!dsl_dataset_is_snapshot(ds))
3582                 return (EINVAL);
3583
3584         /* tags must be unique */
3585         mutex_enter(&ds->ds_lock);
3586         if (ds->ds_phys->ds_userrefs_obj) {
3587                 error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3588                     8, 1, tx);
3589                 if (error == 0)
3590                         error = EEXIST;
3591                 else if (error == ENOENT)
3592                         error = 0;
3593         }
3594         mutex_exit(&ds->ds_lock);
3595
3596         if (error == 0 && ha->temphold &&
3597             strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
3598                 error = E2BIG;
3599
3600         return (error);
3601 }
3602
3603 void
3604 dsl_dataset_user_hold_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3605 {
3606         dsl_dataset_t *ds = arg1;
3607         struct dsl_ds_holdarg *ha = arg2;
3608         char *htag = ha->htag;
3609         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3610         objset_t *mos = dp->dp_meta_objset;
3611         uint64_t now = gethrestime_sec();
3612         uint64_t zapobj;
3613
3614         mutex_enter(&ds->ds_lock);
3615         if (ds->ds_phys->ds_userrefs_obj == 0) {
3616                 /*
3617                  * This is the first user hold for this dataset.  Create
3618                  * the userrefs zap object.
3619                  */
3620                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3621                 zapobj = ds->ds_phys->ds_userrefs_obj =
3622                     zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3623         } else {
3624                 zapobj = ds->ds_phys->ds_userrefs_obj;
3625         }
3626         ds->ds_userrefs++;
3627         mutex_exit(&ds->ds_lock);
3628
3629         VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3630
3631         if (ha->temphold) {
3632                 VERIFY(0 == dsl_pool_user_hold(dp, ds->ds_object,
3633                     htag, &now, tx));
3634         }
3635
3636         spa_history_log_internal(LOG_DS_USER_HOLD,
3637             dp->dp_spa, tx, "<%s> temp = %d dataset = %llu", htag,
3638             (int)ha->temphold, ds->ds_object);
3639 }
3640
3641 static int
3642 dsl_dataset_user_hold_one(const char *dsname, void *arg)
3643 {
3644         struct dsl_ds_holdarg *ha = arg;
3645         dsl_dataset_t *ds;
3646         int error;
3647         char *name;
3648
3649         /* alloc a buffer to hold dsname@snapname plus terminating NULL */
3650         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3651         error = dsl_dataset_hold(name, ha->dstg, &ds);
3652         strfree(name);
3653         if (error == 0) {
3654                 ha->gotone = B_TRUE;
3655                 dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3656                     dsl_dataset_user_hold_sync, ds, ha, 0);
3657         } else if (error == ENOENT && ha->recursive) {
3658                 error = 0;
3659         } else {
3660                 (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3661         }
3662         return (error);
3663 }
3664
3665 int
3666 dsl_dataset_user_hold_for_send(dsl_dataset_t *ds, char *htag,
3667     boolean_t temphold)
3668 {
3669         struct dsl_ds_holdarg *ha;
3670         int error;
3671
3672         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3673         ha->htag = htag;
3674         ha->temphold = temphold;
3675         error = dsl_sync_task_do(ds->ds_dir->dd_pool,
3676             dsl_dataset_user_hold_check, dsl_dataset_user_hold_sync,
3677             ds, ha, 0);
3678         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3679
3680         return (error);
3681 }
3682
3683 int
3684 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3685     boolean_t recursive, boolean_t temphold, int cleanup_fd)
3686 {
3687         struct dsl_ds_holdarg *ha;
3688         dsl_sync_task_t *dst;
3689         spa_t *spa;
3690         int error;
3691         minor_t minor = 0;
3692
3693         if (cleanup_fd != -1) {
3694                 /* Currently we only support cleanup-on-exit of tempholds. */
3695                 if (!temphold)
3696                         return (EINVAL);
3697                 error = zfs_onexit_fd_hold(cleanup_fd, &minor);
3698                 if (error)
3699                         return (error);
3700         }
3701
3702         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3703
3704         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3705
3706         error = spa_open(dsname, &spa, FTAG);
3707         if (error) {
3708                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3709                 if (cleanup_fd != -1)
3710                         zfs_onexit_fd_rele(cleanup_fd);
3711                 return (error);
3712         }
3713
3714         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3715         ha->htag = htag;
3716         ha->snapname = snapname;
3717         ha->recursive = recursive;
3718         ha->temphold = temphold;
3719
3720         if (recursive) {
3721                 error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3722                     ha, DS_FIND_CHILDREN);
3723         } else {
3724                 error = dsl_dataset_user_hold_one(dsname, ha);
3725         }
3726         if (error == 0)
3727                 error = dsl_sync_task_group_wait(ha->dstg);
3728
3729         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3730             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3731                 dsl_dataset_t *ds = dst->dst_arg1;
3732
3733                 if (dst->dst_err) {
3734                         dsl_dataset_name(ds, ha->failed);
3735                         *strchr(ha->failed, '@') = '\0';
3736                 } else if (error == 0 && minor != 0 && temphold) {
3737                         /*
3738                          * If this hold is to be released upon process exit,
3739                          * register that action now.
3740                          */
3741                         dsl_register_onexit_hold_cleanup(ds, htag, minor);
3742                 }
3743                 dsl_dataset_rele(ds, ha->dstg);
3744         }
3745
3746         if (error == 0 && recursive && !ha->gotone)
3747                 error = ENOENT;
3748
3749         if (error)
3750                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3751
3752         dsl_sync_task_group_destroy(ha->dstg);
3753
3754         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3755         spa_close(spa, FTAG);
3756         if (cleanup_fd != -1)
3757                 zfs_onexit_fd_rele(cleanup_fd);
3758         return (error);
3759 }
3760
3761 struct dsl_ds_releasearg {
3762         dsl_dataset_t *ds;
3763         const char *htag;
3764         boolean_t own;          /* do we own or just hold ds? */
3765 };
3766
3767 static int
3768 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3769     boolean_t *might_destroy)
3770 {
3771         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3772         uint64_t zapobj;
3773         uint64_t tmp;
3774         int error;
3775
3776         *might_destroy = B_FALSE;
3777
3778         mutex_enter(&ds->ds_lock);
3779         zapobj = ds->ds_phys->ds_userrefs_obj;
3780         if (zapobj == 0) {
3781                 /* The tag can't possibly exist */
3782                 mutex_exit(&ds->ds_lock);
3783                 return (ESRCH);
3784         }
3785
3786         /* Make sure the tag exists */
3787         error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3788         if (error) {
3789                 mutex_exit(&ds->ds_lock);
3790                 if (error == ENOENT)
3791                         error = ESRCH;
3792                 return (error);
3793         }
3794
3795         if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3796             DS_IS_DEFER_DESTROY(ds))
3797                 *might_destroy = B_TRUE;
3798
3799         mutex_exit(&ds->ds_lock);
3800         return (0);
3801 }
3802
3803 static int
3804 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3805 {
3806         struct dsl_ds_releasearg *ra = arg1;
3807         dsl_dataset_t *ds = ra->ds;
3808         boolean_t might_destroy;
3809         int error;
3810
3811         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3812                 return (ENOTSUP);
3813
3814         error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3815         if (error)
3816                 return (error);
3817
3818         if (might_destroy) {
3819                 struct dsl_ds_destroyarg dsda = {0};
3820
3821                 if (dmu_tx_is_syncing(tx)) {
3822                         /*
3823                          * If we're not prepared to remove the snapshot,
3824                          * we can't allow the release to happen right now.
3825                          */
3826                         if (!ra->own)
3827                                 return (EBUSY);
3828                 }
3829                 dsda.ds = ds;
3830                 dsda.releasing = B_TRUE;
3831                 return (dsl_dataset_destroy_check(&dsda, tag, tx));
3832         }
3833
3834         return (0);
3835 }
3836
3837 static void
3838 dsl_dataset_user_release_sync(void *arg1, void *tag, dmu_tx_t *tx)
3839 {
3840         struct dsl_ds_releasearg *ra = arg1;
3841         dsl_dataset_t *ds = ra->ds;
3842         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3843         objset_t *mos = dp->dp_meta_objset;
3844         uint64_t zapobj;
3845         uint64_t dsobj = ds->ds_object;
3846         uint64_t refs;
3847         int error;
3848
3849         mutex_enter(&ds->ds_lock);
3850         ds->ds_userrefs--;
3851         refs = ds->ds_userrefs;
3852         mutex_exit(&ds->ds_lock);
3853         error = dsl_pool_user_release(dp, ds->ds_object, ra->htag, tx);
3854         VERIFY(error == 0 || error == ENOENT);
3855         zapobj = ds->ds_phys->ds_userrefs_obj;
3856         VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3857
3858         spa_history_log_internal(LOG_DS_USER_RELEASE,
3859             dp->dp_spa, tx, "<%s> %lld dataset = %llu",
3860             ra->htag, (longlong_t)refs, dsobj);
3861
3862         if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3863             DS_IS_DEFER_DESTROY(ds)) {
3864                 struct dsl_ds_destroyarg dsda = {0};
3865
3866                 ASSERT(ra->own);
3867                 dsda.ds = ds;
3868                 dsda.releasing = B_TRUE;
3869                 /* We already did the destroy_check */
3870                 dsl_dataset_destroy_sync(&dsda, tag, tx);
3871         }
3872 }
3873
3874 static int
3875 dsl_dataset_user_release_one(const char *dsname, void *arg)
3876 {
3877         struct dsl_ds_holdarg *ha = arg;
3878         struct dsl_ds_releasearg *ra;
3879         dsl_dataset_t *ds;
3880         int error;
3881         void *dtag = ha->dstg;
3882         char *name;
3883         boolean_t own = B_FALSE;
3884         boolean_t might_destroy;
3885
3886         /* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3887         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3888         error = dsl_dataset_hold(name, dtag, &ds);
3889         strfree(name);
3890         if (error == ENOENT && ha->recursive)
3891                 return (0);
3892         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3893         if (error)
3894                 return (error);
3895
3896         ha->gotone = B_TRUE;
3897
3898         ASSERT(dsl_dataset_is_snapshot(ds));
3899
3900         error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3901         if (error) {
3902                 dsl_dataset_rele(ds, dtag);
3903                 return (error);
3904         }
3905
3906         if (might_destroy) {
3907 #ifdef _KERNEL
3908                 name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3909                 error = zfs_unmount_snap(name, NULL);
3910                 strfree(name);
3911                 if (error) {
3912                         dsl_dataset_rele(ds, dtag);
3913                         return (error);
3914                 }
3915 #endif
3916                 if (!dsl_dataset_tryown(ds, B_TRUE, dtag)) {
3917                         dsl_dataset_rele(ds, dtag);
3918                         return (EBUSY);
3919                 } else {
3920                         own = B_TRUE;
3921                         dsl_dataset_make_exclusive(ds, dtag);
3922                 }
3923         }
3924
3925         ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3926         ra->ds = ds;
3927         ra->htag = ha->htag;
3928         ra->own = own;
3929         dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3930             dsl_dataset_user_release_sync, ra, dtag, 0);
3931
3932         return (0);
3933 }
3934
3935 int
3936 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3937     boolean_t recursive)
3938 {
3939         struct dsl_ds_holdarg *ha;
3940         dsl_sync_task_t *dst;
3941         spa_t *spa;
3942         int error;
3943
3944 top:
3945         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3946
3947         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3948
3949         error = spa_open(dsname, &spa, FTAG);
3950         if (error) {
3951                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3952                 return (error);
3953         }
3954
3955         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3956         ha->htag = htag;
3957         ha->snapname = snapname;
3958         ha->recursive = recursive;
3959         if (recursive) {
3960                 error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
3961                     ha, DS_FIND_CHILDREN);
3962         } else {
3963                 error = dsl_dataset_user_release_one(dsname, ha);
3964         }
3965         if (error == 0)
3966                 error = dsl_sync_task_group_wait(ha->dstg);
3967
3968         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3969             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3970                 struct dsl_ds_releasearg *ra = dst->dst_arg1;
3971                 dsl_dataset_t *ds = ra->ds;
3972
3973                 if (dst->dst_err)
3974                         dsl_dataset_name(ds, ha->failed);
3975
3976                 if (ra->own)
3977                         dsl_dataset_disown(ds, ha->dstg);
3978                 else
3979                         dsl_dataset_rele(ds, ha->dstg);
3980
3981                 kmem_free(ra, sizeof (struct dsl_ds_releasearg));
3982         }
3983
3984         if (error == 0 && recursive && !ha->gotone)
3985                 error = ENOENT;
3986
3987         if (error && error != EBUSY)
3988                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3989
3990         dsl_sync_task_group_destroy(ha->dstg);
3991         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3992         spa_close(spa, FTAG);
3993
3994         /*
3995          * We can get EBUSY if we were racing with deferred destroy and
3996          * dsl_dataset_user_release_check() hadn't done the necessary
3997          * open context setup.  We can also get EBUSY if we're racing
3998          * with destroy and that thread is the ds_owner.  Either way
3999          * the busy condition should be transient, and we should retry
4000          * the release operation.
4001          */
4002         if (error == EBUSY)
4003                 goto top;
4004
4005         return (error);
4006 }
4007
4008 /*
4009  * Called at spa_load time (with retry == B_FALSE) to release a stale
4010  * temporary user hold. Also called by the onexit code (with retry == B_TRUE).
4011  */
4012 int
4013 dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, char *htag,
4014     boolean_t retry)
4015 {
4016         dsl_dataset_t *ds;
4017         char *snap;
4018         char *name;
4019         int namelen;
4020         int error;
4021
4022         do {
4023                 rw_enter(&dp->dp_config_rwlock, RW_READER);
4024                 error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
4025                 rw_exit(&dp->dp_config_rwlock);
4026                 if (error)
4027                         return (error);
4028                 namelen = dsl_dataset_namelen(ds)+1;
4029                 name = kmem_alloc(namelen, KM_SLEEP);
4030                 dsl_dataset_name(ds, name);
4031                 dsl_dataset_rele(ds, FTAG);
4032
4033                 snap = strchr(name, '@');
4034                 *snap = '\0';
4035                 ++snap;
4036                 error = dsl_dataset_user_release(name, snap, htag, B_FALSE);
4037                 kmem_free(name, namelen);
4038
4039                 /*
4040                  * The object can't have been destroyed because we have a hold,
4041                  * but it might have been renamed, resulting in ENOENT.  Retry
4042                  * if we've been requested to do so.
4043                  *
4044                  * It would be nice if we could use the dsobj all the way
4045                  * through and avoid ENOENT entirely.  But we might need to
4046                  * unmount the snapshot, and there's currently no way to lookup
4047                  * a vfsp using a ZFS object id.
4048                  */
4049         } while ((error == ENOENT) && retry);
4050
4051         return (error);
4052 }
4053
4054 int
4055 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
4056 {
4057         dsl_dataset_t *ds;
4058         int err;
4059
4060         err = dsl_dataset_hold(dsname, FTAG, &ds);
4061         if (err)
4062                 return (err);
4063
4064         VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
4065         if (ds->ds_phys->ds_userrefs_obj != 0) {
4066                 zap_attribute_t *za;
4067                 zap_cursor_t zc;
4068
4069                 za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
4070                 for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
4071                     ds->ds_phys->ds_userrefs_obj);
4072                     zap_cursor_retrieve(&zc, za) == 0;
4073                     zap_cursor_advance(&zc)) {
4074                         VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
4075                             za->za_first_integer));
4076                 }
4077                 zap_cursor_fini(&zc);
4078                 kmem_free(za, sizeof (zap_attribute_t));
4079         }
4080         dsl_dataset_rele(ds, FTAG);
4081         return (0);
4082 }
4083
4084 /*
4085  * Note, this function is used as the callback for dmu_objset_find().  We
4086  * always return 0 so that we will continue to find and process
4087  * inconsistent datasets, even if we encounter an error trying to
4088  * process one of them.
4089  */
4090 /* ARGSUSED */
4091 int
4092 dsl_destroy_inconsistent(const char *dsname, void *arg)
4093 {
4094         dsl_dataset_t *ds;
4095
4096         if (dsl_dataset_own(dsname, B_TRUE, FTAG, &ds) == 0) {
4097                 if (DS_IS_INCONSISTENT(ds))
4098                         (void) dsl_dataset_destroy(ds, FTAG, B_FALSE);
4099                 else
4100                         dsl_dataset_disown(ds, FTAG);
4101         }
4102         return (0);
4103 }
4104
4105
4106 /*
4107  * Return (in *usedp) the amount of space written in new that is not
4108  * present in oldsnap.  New may be a snapshot or the head.  Old must be
4109  * a snapshot before new, in new's filesystem (or its origin).  If not then
4110  * fail and return EINVAL.
4111  *
4112  * The written space is calculated by considering two components:  First, we
4113  * ignore any freed space, and calculate the written as new's used space
4114  * minus old's used space.  Next, we add in the amount of space that was freed
4115  * between the two snapshots, thus reducing new's used space relative to old's.
4116  * Specifically, this is the space that was born before old->ds_creation_txg,
4117  * and freed before new (ie. on new's deadlist or a previous deadlist).
4118  *
4119  * space freed                         [---------------------]
4120  * snapshots                       ---O-------O--------O-------O------
4121  *                                         oldsnap            new
4122  */
4123 int
4124 dsl_dataset_space_written(dsl_dataset_t *oldsnap, dsl_dataset_t *new,
4125     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
4126 {
4127         int err = 0;
4128         uint64_t snapobj;
4129         dsl_pool_t *dp = new->ds_dir->dd_pool;
4130
4131         *usedp = 0;
4132         *usedp += new->ds_phys->ds_used_bytes;
4133         *usedp -= oldsnap->ds_phys->ds_used_bytes;
4134
4135         *compp = 0;
4136         *compp += new->ds_phys->ds_compressed_bytes;
4137         *compp -= oldsnap->ds_phys->ds_compressed_bytes;
4138
4139         *uncompp = 0;
4140         *uncompp += new->ds_phys->ds_uncompressed_bytes;
4141         *uncompp -= oldsnap->ds_phys->ds_uncompressed_bytes;
4142
4143         rw_enter(&dp->dp_config_rwlock, RW_READER);
4144         snapobj = new->ds_object;
4145         while (snapobj != oldsnap->ds_object) {
4146                 dsl_dataset_t *snap;
4147                 uint64_t used, comp, uncomp;
4148
4149                 err = dsl_dataset_hold_obj(dp, snapobj, FTAG, &snap);
4150                 if (err != 0)
4151                         break;
4152
4153                 if (snap->ds_phys->ds_prev_snap_txg ==
4154                     oldsnap->ds_phys->ds_creation_txg) {
4155                         /*
4156                          * The blocks in the deadlist can not be born after
4157                          * ds_prev_snap_txg, so get the whole deadlist space,
4158                          * which is more efficient (especially for old-format
4159                          * deadlists).  Unfortunately the deadlist code
4160                          * doesn't have enough information to make this
4161                          * optimization itself.
4162                          */
4163                         dsl_deadlist_space(&snap->ds_deadlist,
4164                             &used, &comp, &uncomp);
4165                 } else {
4166                         dsl_deadlist_space_range(&snap->ds_deadlist,
4167                             0, oldsnap->ds_phys->ds_creation_txg,
4168                             &used, &comp, &uncomp);
4169                 }
4170                 *usedp += used;
4171                 *compp += comp;
4172                 *uncompp += uncomp;
4173
4174                 /*
4175                  * If we get to the beginning of the chain of snapshots
4176                  * (ds_prev_snap_obj == 0) before oldsnap, then oldsnap
4177                  * was not a snapshot of/before new.
4178                  */
4179                 snapobj = snap->ds_phys->ds_prev_snap_obj;
4180                 dsl_dataset_rele(snap, FTAG);
4181                 if (snapobj == 0) {
4182                         err = EINVAL;
4183                         break;
4184                 }
4185
4186         }
4187         rw_exit(&dp->dp_config_rwlock);
4188         return (err);
4189 }
4190
4191 /*
4192  * Return (in *usedp) the amount of space that will be reclaimed if firstsnap,
4193  * lastsnap, and all snapshots in between are deleted.
4194  *
4195  * blocks that would be freed            [---------------------------]
4196  * snapshots                       ---O-------O--------O-------O--------O
4197  *                                        firstsnap        lastsnap
4198  *
4199  * This is the set of blocks that were born after the snap before firstsnap,
4200  * (birth > firstsnap->prev_snap_txg) and died before the snap after the
4201  * last snap (ie, is on lastsnap->ds_next->ds_deadlist or an earlier deadlist).
4202  * We calculate this by iterating over the relevant deadlists (from the snap
4203  * after lastsnap, backward to the snap after firstsnap), summing up the
4204  * space on the deadlist that was born after the snap before firstsnap.
4205  */
4206 int
4207 dsl_dataset_space_wouldfree(dsl_dataset_t *firstsnap,
4208     dsl_dataset_t *lastsnap,
4209     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
4210 {
4211         int err = 0;
4212         uint64_t snapobj;
4213         dsl_pool_t *dp = firstsnap->ds_dir->dd_pool;
4214
4215         ASSERT(dsl_dataset_is_snapshot(firstsnap));
4216         ASSERT(dsl_dataset_is_snapshot(lastsnap));
4217
4218         /*
4219          * Check that the snapshots are in the same dsl_dir, and firstsnap
4220          * is before lastsnap.
4221          */
4222         if (firstsnap->ds_dir != lastsnap->ds_dir ||
4223             firstsnap->ds_phys->ds_creation_txg >
4224             lastsnap->ds_phys->ds_creation_txg)
4225                 return (EINVAL);
4226
4227         *usedp = *compp = *uncompp = 0;
4228
4229         rw_enter(&dp->dp_config_rwlock, RW_READER);
4230         snapobj = lastsnap->ds_phys->ds_next_snap_obj;
4231         while (snapobj != firstsnap->ds_object) {
4232                 dsl_dataset_t *ds;
4233                 uint64_t used, comp, uncomp;
4234
4235                 err = dsl_dataset_hold_obj(dp, snapobj, FTAG, &ds);
4236                 if (err != 0)
4237                         break;
4238
4239                 dsl_deadlist_space_range(&ds->ds_deadlist,
4240                     firstsnap->ds_phys->ds_prev_snap_txg, UINT64_MAX,
4241                     &used, &comp, &uncomp);
4242                 *usedp += used;
4243                 *compp += comp;
4244                 *uncompp += uncomp;
4245
4246                 snapobj = ds->ds_phys->ds_prev_snap_obj;
4247                 ASSERT3U(snapobj, !=, 0);
4248                 dsl_dataset_rele(ds, FTAG);
4249         }
4250         rw_exit(&dp->dp_config_rwlock);
4251         return (err);
4252 }
4253
4254 #if defined(_KERNEL) && defined(HAVE_SPL)
4255 EXPORT_SYMBOL(dmu_snapshots_destroy_nvl);
4256 EXPORT_SYMBOL(dsl_dataset_hold);
4257 EXPORT_SYMBOL(dsl_dataset_hold_obj);
4258 EXPORT_SYMBOL(dsl_dataset_own);
4259 EXPORT_SYMBOL(dsl_dataset_own_obj);
4260 EXPORT_SYMBOL(dsl_dataset_name);
4261 EXPORT_SYMBOL(dsl_dataset_rele);
4262 EXPORT_SYMBOL(dsl_dataset_disown);
4263 EXPORT_SYMBOL(dsl_dataset_drop_ref);
4264 EXPORT_SYMBOL(dsl_dataset_tryown);
4265 EXPORT_SYMBOL(dsl_dataset_make_exclusive);
4266 EXPORT_SYMBOL(dsl_dataset_create_sync);
4267 EXPORT_SYMBOL(dsl_dataset_create_sync_dd);
4268 EXPORT_SYMBOL(dsl_dataset_destroy);
4269 EXPORT_SYMBOL(dsl_dataset_destroy_check);
4270 EXPORT_SYMBOL(dsl_dataset_destroy_sync);
4271 EXPORT_SYMBOL(dsl_dataset_snapshot_check);
4272 EXPORT_SYMBOL(dsl_dataset_snapshot_sync);
4273 EXPORT_SYMBOL(dsl_dataset_rename);
4274 EXPORT_SYMBOL(dsl_dataset_promote);
4275 EXPORT_SYMBOL(dsl_dataset_clone_swap);
4276 EXPORT_SYMBOL(dsl_dataset_user_hold);
4277 EXPORT_SYMBOL(dsl_dataset_user_release);
4278 EXPORT_SYMBOL(dsl_dataset_user_release_tmp);
4279 EXPORT_SYMBOL(dsl_dataset_get_holds);
4280 EXPORT_SYMBOL(dsl_dataset_get_blkptr);
4281 EXPORT_SYMBOL(dsl_dataset_set_blkptr);
4282 EXPORT_SYMBOL(dsl_dataset_get_spa);
4283 EXPORT_SYMBOL(dsl_dataset_modified_since_lastsnap);
4284 EXPORT_SYMBOL(dsl_dataset_space_written);
4285 EXPORT_SYMBOL(dsl_dataset_space_wouldfree);
4286 EXPORT_SYMBOL(dsl_dataset_sync);
4287 EXPORT_SYMBOL(dsl_dataset_block_born);
4288 EXPORT_SYMBOL(dsl_dataset_block_kill);
4289 EXPORT_SYMBOL(dsl_dataset_block_freeable);
4290 EXPORT_SYMBOL(dsl_dataset_prev_snap_txg);
4291 EXPORT_SYMBOL(dsl_dataset_dirty);
4292 EXPORT_SYMBOL(dsl_dataset_stats);
4293 EXPORT_SYMBOL(dsl_dataset_fast_stat);
4294 EXPORT_SYMBOL(dsl_dataset_space);
4295 EXPORT_SYMBOL(dsl_dataset_fsid_guid);
4296 EXPORT_SYMBOL(dsl_dsobj_to_dsname);
4297 EXPORT_SYMBOL(dsl_dataset_check_quota);
4298 EXPORT_SYMBOL(dsl_dataset_set_quota);
4299 EXPORT_SYMBOL(dsl_dataset_set_quota_sync);
4300 EXPORT_SYMBOL(dsl_dataset_set_reservation);
4301 EXPORT_SYMBOL(dsl_destroy_inconsistent);
4302 #endif