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

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