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