module/zfs/dsl_dataset.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011 by Delphix. All rights reserved.
  24  * Copyright (c) 2012, Joyent, Inc. All rights reserved.
  25  */
  26
  27 #include <sys/dmu_objset.h>
  28 #include <sys/dsl_dataset.h>
  29 #include <sys/dsl_dir.h>
  30 #include <sys/dsl_prop.h>
  31 #include <sys/dsl_synctask.h>
  32 #include <sys/dmu_traverse.h>
  33 #include <sys/dmu_impl.h>
  34 #include <sys/dmu_tx.h>
  35 #include <sys/arc.h>
  36 #include <sys/zio.h>
  37 #include <sys/zap.h>
  38 #include <sys/unique.h>
  39 #include <sys/zfs_context.h>
  40 #include <sys/zfs_ioctl.h>
  41 #include <sys/spa.h>
  42 #include <sys/zfs_znode.h>
  43 #include <sys/zfs_onexit.h>
  44 #include <sys/zvol.h>
  45 #include <sys/dsl_scan.h>
  46 #include <sys/dsl_deadlist.h>
  47
  48 static char *dsl_reaper = "the grim reaper";
  49
  50 static dsl_checkfunc_t dsl_dataset_destroy_begin_check;
  51 static dsl_syncfunc_t dsl_dataset_destroy_begin_sync;
  52 static dsl_syncfunc_t dsl_dataset_set_reservation_sync;
  53
  54 #define SWITCH64(x, y) \
  55         { \
  56                 uint64_t __tmp = (x); \
  57                 (x) = (y); \
  58                 (y) = __tmp; \
  59         }
  60
  61 #define DS_REF_MAX      (1ULL << 62)
  62
  63 #define DSL_DEADLIST_BLOCKSIZE  SPA_MAXBLOCKSIZE
  64
  65 #define DSL_DATASET_IS_DESTROYED(ds)    ((ds)->ds_owner == dsl_reaper)
  66
  67
  68 /*
  69  * Figure out how much of this delta should be propogated to the dsl_dir
  70  * layer.  If there's a refreservation, that space has already been
  71  * partially accounted for in our ancestors.
  72  */
  73 static int64_t
  74 parent_delta(dsl_dataset_t *ds, int64_t delta)
  75 {
  76         uint64_t old_bytes, new_bytes;
  77
  78         if (ds->ds_reserved == 0)
  79                 return (delta);
  80
  81         old_bytes = MAX(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
  82         new_bytes = MAX(ds->ds_phys->ds_unique_bytes + delta, ds->ds_reserved);
  83
  84         ASSERT3U(ABS((int64_t)(new_bytes - old_bytes)), <=, ABS(delta));
  85         return (new_bytes - old_bytes);
  86 }
  87
  88 void
  89 dsl_dataset_block_born(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
  90 {
  91         int used, compressed, uncompressed;
  92         int64_t delta;
  93
  94         used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
  95         compressed = BP_GET_PSIZE(bp);
  96         uncompressed = BP_GET_UCSIZE(bp);
  97
  98         dprintf_bp(bp, "ds=%p", ds);
  99
 100         ASSERT(dmu_tx_is_syncing(tx));
 101         /* It could have been compressed away to nothing */
 102         if (BP_IS_HOLE(bp))
 103                 return;
 104         ASSERT(BP_GET_TYPE(bp) != DMU_OT_NONE);
 105         ASSERT3U(BP_GET_TYPE(bp), <, DMU_OT_NUMTYPES);
 106         if (ds == NULL) {
 107                 /*
 108                  * Account for the meta-objset space in its placeholder
 109                  * dsl_dir.
 110                  */
 111                 ASSERT3U(compressed, ==, uncompressed); /* it's all metadata */
 112                 dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
 113                     used, compressed, uncompressed, tx);
 114                 dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
 115                 return;
 116         }
 117         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 118
 119         mutex_enter(&ds->ds_dir->dd_lock);
 120         mutex_enter(&ds->ds_lock);
 121         delta = parent_delta(ds, used);
 122         ds->ds_phys->ds_used_bytes += used;
 123         ds->ds_phys->ds_compressed_bytes += compressed;
 124         ds->ds_phys->ds_uncompressed_bytes += uncompressed;
 125         ds->ds_phys->ds_unique_bytes += used;
 126         mutex_exit(&ds->ds_lock);
 127         dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD, delta,
 128             compressed, uncompressed, tx);
 129         dsl_dir_transfer_space(ds->ds_dir, used - delta,
 130             DD_USED_REFRSRV, DD_USED_HEAD, tx);
 131         mutex_exit(&ds->ds_dir->dd_lock);
 132 }
 133
 134 int
 135 dsl_dataset_block_kill(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx,
 136     boolean_t async)
 137 {
 138         int used, compressed, uncompressed;
 139
 140         if (BP_IS_HOLE(bp))
 141                 return (0);
 142
 143         ASSERT(dmu_tx_is_syncing(tx));
 144         ASSERT(bp->blk_birth <= tx->tx_txg);
 145
 146         used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
 147         compressed = BP_GET_PSIZE(bp);
 148         uncompressed = BP_GET_UCSIZE(bp);
 149
 150         ASSERT(used > 0);
 151         if (ds == NULL) {
 152                 /*
 153                  * Account for the meta-objset space in its placeholder
 154                  * dataset.
 155                  */
 156                 dsl_free(tx->tx_pool, tx->tx_txg, bp);
 157
 158                 dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
 159                     -used, -compressed, -uncompressed, tx);
 160                 dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
 161                 return (used);
 162         }
 163         ASSERT3P(tx->tx_pool, ==, ds->ds_dir->dd_pool);
 164
 165         ASSERT(!dsl_dataset_is_snapshot(ds));
 166         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 167
 168         if (bp->blk_birth > ds->ds_phys->ds_prev_snap_txg) {
 169                 int64_t delta;
 170
 171                 dprintf_bp(bp, "freeing ds=%llu", ds->ds_object);
 172                 dsl_free(tx->tx_pool, tx->tx_txg, bp);
 173
 174                 mutex_enter(&ds->ds_dir->dd_lock);
 175                 mutex_enter(&ds->ds_lock);
 176                 ASSERT(ds->ds_phys->ds_unique_bytes >= used ||
 177                     !DS_UNIQUE_IS_ACCURATE(ds));
 178                 delta = parent_delta(ds, -used);
 179                 ds->ds_phys->ds_unique_bytes -= used;
 180                 mutex_exit(&ds->ds_lock);
 181                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
 182                     delta, -compressed, -uncompressed, tx);
 183                 dsl_dir_transfer_space(ds->ds_dir, -used - delta,
 184                     DD_USED_REFRSRV, DD_USED_HEAD, tx);
 185                 mutex_exit(&ds->ds_dir->dd_lock);
 186         } else {
 187                 dprintf_bp(bp, "putting on dead list: %s", "");
 188                 if (async) {
 189                         /*
 190                          * We are here as part of zio's write done callback,
 191                          * which means we're a zio interrupt thread.  We can't
 192                          * call dsl_deadlist_insert() now because it may block
 193                          * waiting for I/O.  Instead, put bp on the deferred
 194                          * queue and let dsl_pool_sync() finish the job.
 195                          */
 196                         bplist_append(&ds->ds_pending_deadlist, bp);
 197                 } else {
 198                         dsl_deadlist_insert(&ds->ds_deadlist, bp, tx);
 199                 }
 200                 ASSERT3U(ds->ds_prev->ds_object, ==,
 201                     ds->ds_phys->ds_prev_snap_obj);
 202                 ASSERT(ds->ds_prev->ds_phys->ds_num_children > 0);
 203                 /* if (bp->blk_birth > prev prev snap txg) prev unique += bs */
 204                 if (ds->ds_prev->ds_phys->ds_next_snap_obj ==
 205                     ds->ds_object && bp->blk_birth >
 206                     ds->ds_prev->ds_phys->ds_prev_snap_txg) {
 207                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
 208                         mutex_enter(&ds->ds_prev->ds_lock);
 209                         ds->ds_prev->ds_phys->ds_unique_bytes += used;
 210                         mutex_exit(&ds->ds_prev->ds_lock);
 211                 }
 212                 if (bp->blk_birth > ds->ds_dir->dd_origin_txg) {
 213                         dsl_dir_transfer_space(ds->ds_dir, used,
 214                             DD_USED_HEAD, DD_USED_SNAP, tx);
 215                 }
 216         }
 217         mutex_enter(&ds->ds_lock);
 218         ASSERT3U(ds->ds_phys->ds_used_bytes, >=, used);
 219         ds->ds_phys->ds_used_bytes -= used;
 220         ASSERT3U(ds->ds_phys->ds_compressed_bytes, >=, compressed);
 221         ds->ds_phys->ds_compressed_bytes -= compressed;
 222         ASSERT3U(ds->ds_phys->ds_uncompressed_bytes, >=, uncompressed);
 223         ds->ds_phys->ds_uncompressed_bytes -= uncompressed;
 224         mutex_exit(&ds->ds_lock);
 225
 226         return (used);
 227 }
 228
 229 uint64_t
 230 dsl_dataset_prev_snap_txg(dsl_dataset_t *ds)
 231 {
 232         uint64_t trysnap = 0;
 233
 234         if (ds == NULL)
 235                 return (0);
 236         /*
 237          * The snapshot creation could fail, but that would cause an
 238          * incorrect FALSE return, which would only result in an
 239          * overestimation of the amount of space that an operation would
 240          * consume, which is OK.
 241          *
 242          * There's also a small window where we could miss a pending
 243          * snapshot, because we could set the sync task in the quiescing
 244          * phase.  So this should only be used as a guess.
 245          */
 246         if (ds->ds_trysnap_txg >
 247             spa_last_synced_txg(ds->ds_dir->dd_pool->dp_spa))
 248                 trysnap = ds->ds_trysnap_txg;
 249         return (MAX(ds->ds_phys->ds_prev_snap_txg, trysnap));
 250 }
 251
 252 boolean_t
 253 dsl_dataset_block_freeable(dsl_dataset_t *ds, const blkptr_t *bp,
 254     uint64_t blk_birth)
 255 {
 256         if (blk_birth <= dsl_dataset_prev_snap_txg(ds))
 257                 return (B_FALSE);
 258
 259         ddt_prefetch(dsl_dataset_get_spa(ds), bp);
 260
 261         return (B_TRUE);
 262 }
 263
 264 /* ARGSUSED */
 265 static void
 266 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
 267 {
 268         dsl_dataset_t *ds = dsv;
 269
 270         ASSERT(ds->ds_owner == NULL || DSL_DATASET_IS_DESTROYED(ds));
 271
 272         unique_remove(ds->ds_fsid_guid);
 273
 274         if (ds->ds_objset != NULL)
 275                 dmu_objset_evict(ds->ds_objset);
 276
 277         if (ds->ds_prev) {
 278                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 279                 ds->ds_prev = NULL;
 280         }
 281
 282         bplist_destroy(&ds->ds_pending_deadlist);
 283         if (db != NULL) {
 284                 dsl_deadlist_close(&ds->ds_deadlist);
 285         } else {
 286                 ASSERT(ds->ds_deadlist.dl_dbuf == NULL);
 287                 ASSERT(!ds->ds_deadlist.dl_oldfmt);
 288         }
 289         if (ds->ds_dir)
 290                 dsl_dir_close(ds->ds_dir, ds);
 291
 292         ASSERT(!list_link_active(&ds->ds_synced_link));
 293
 294         mutex_destroy(&ds->ds_lock);
 295         mutex_destroy(&ds->ds_recvlock);
 296         mutex_destroy(&ds->ds_opening_lock);
 297         rw_destroy(&ds->ds_rwlock);
 298         cv_destroy(&ds->ds_exclusive_cv);
 299
 300         kmem_free(ds, sizeof (dsl_dataset_t));
 301 }
 302
 303 static int
 304 dsl_dataset_get_snapname(dsl_dataset_t *ds)
 305 {
 306         dsl_dataset_phys_t *headphys;
 307         int err;
 308         dmu_buf_t *headdbuf;
 309         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 310         objset_t *mos = dp->dp_meta_objset;
 311
 312         if (ds->ds_snapname[0])
 313                 return (0);
 314         if (ds->ds_phys->ds_next_snap_obj == 0)
 315                 return (0);
 316
 317         err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
 318             FTAG, &headdbuf);
 319         if (err)
 320                 return (err);
 321         headphys = headdbuf->db_data;
 322         err = zap_value_search(dp->dp_meta_objset,
 323             headphys->ds_snapnames_zapobj, ds->ds_object, 0, ds->ds_snapname);
 324         dmu_buf_rele(headdbuf, FTAG);
 325         return (err);
 326 }
 327
 328 static int
 329 dsl_dataset_snap_lookup(dsl_dataset_t *ds, const char *name, uint64_t *value)
 330 {
 331         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 332         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 333         matchtype_t mt;
 334         int err;
 335
 336         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 337                 mt = MT_FIRST;
 338         else
 339                 mt = MT_EXACT;
 340
 341         err = zap_lookup_norm(mos, snapobj, name, 8, 1,
 342             value, mt, NULL, 0, NULL);
 343         if (err == ENOTSUP && mt == MT_FIRST)
 344                 err = zap_lookup(mos, snapobj, name, 8, 1, value);
 345         return (err);
 346 }
 347
 348 static int
 349 dsl_dataset_snap_remove(dsl_dataset_t *ds, char *name, dmu_tx_t *tx)
 350 {
 351         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 352         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 353         matchtype_t mt;
 354         int err;
 355
 356         dsl_dir_snap_cmtime_update(ds->ds_dir);
 357
 358         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 359                 mt = MT_FIRST;
 360         else
 361                 mt = MT_EXACT;
 362
 363         err = zap_remove_norm(mos, snapobj, name, mt, tx);
 364         if (err == ENOTSUP && mt == MT_FIRST)
 365                 err = zap_remove(mos, snapobj, name, tx);
 366         return (err);
 367 }
 368
 369 static int
 370 dsl_dataset_get_ref(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 371     dsl_dataset_t **dsp)
 372 {
 373         objset_t *mos = dp->dp_meta_objset;
 374         dmu_buf_t *dbuf;
 375         dsl_dataset_t *ds;
 376         int err;
 377         dmu_object_info_t doi;
 378
 379         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
 380             dsl_pool_sync_context(dp));
 381
 382         err = dmu_bonus_hold(mos, dsobj, tag, &dbuf);
 383         if (err)
 384                 return (err);
 385
 386         /* Make sure dsobj has the correct object type. */
 387         dmu_object_info_from_db(dbuf, &doi);
 388         if (doi.doi_type != DMU_OT_DSL_DATASET)
 389                 return (EINVAL);
 390
 391         ds = dmu_buf_get_user(dbuf);
 392         if (ds == NULL) {
 393                 dsl_dataset_t *winner = NULL;
 394
 395                 ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_PUSHPAGE);
 396                 ds->ds_dbuf = dbuf;
 397                 ds->ds_object = dsobj;
 398                 ds->ds_phys = dbuf->db_data;
 399                 list_link_init(&ds->ds_synced_link);
 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                 mutex_init(&ds->ds_sendstream_lock, NULL, MUTEX_DEFAULT, NULL);
 405
 406                 rw_init(&ds->ds_rwlock, NULL, RW_DEFAULT, NULL);
 407                 cv_init(&ds->ds_exclusive_cv, NULL, CV_DEFAULT, NULL);
 408
 409                 bplist_create(&ds->ds_pending_deadlist);
 410                 dsl_deadlist_open(&ds->ds_deadlist,
 411                     mos, ds->ds_phys->ds_deadlist_obj);
 412
 413                 list_create(&ds->ds_sendstreams, sizeof (dmu_sendarg_t),
 414                     offsetof(dmu_sendarg_t, dsa_link));
 415
 416                 if (err == 0) {
 417                         err = dsl_dir_open_obj(dp,
 418                             ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
 419                 }
 420                 if (err) {
 421                         mutex_destroy(&ds->ds_lock);
 422                         mutex_destroy(&ds->ds_recvlock);
 423                         mutex_destroy(&ds->ds_opening_lock);
 424                         rw_destroy(&ds->ds_rwlock);
 425                         cv_destroy(&ds->ds_exclusive_cv);
 426                         bplist_destroy(&ds->ds_pending_deadlist);
 427                         dsl_deadlist_close(&ds->ds_deadlist);
 428                         kmem_free(ds, sizeof (dsl_dataset_t));
 429                         dmu_buf_rele(dbuf, tag);
 430                         return (err);
 431                 }
 432
 433                 if (!dsl_dataset_is_snapshot(ds)) {
 434                         ds->ds_snapname[0] = '\0';
 435                         if (ds->ds_phys->ds_prev_snap_obj) {
 436                                 err = dsl_dataset_get_ref(dp,
 437                                     ds->ds_phys->ds_prev_snap_obj,
 438                                     ds, &ds->ds_prev);
 439                         }
 440                 } else {
 441                         if (zfs_flags & ZFS_DEBUG_SNAPNAMES)
 442                                 err = dsl_dataset_get_snapname(ds);
 443                         if (err == 0 && ds->ds_phys->ds_userrefs_obj != 0) {
 444                                 err = zap_count(
 445                                     ds->ds_dir->dd_pool->dp_meta_objset,
 446                                     ds->ds_phys->ds_userrefs_obj,
 447                                     &ds->ds_userrefs);
 448                         }
 449                 }
 450
 451                 if (err == 0 && !dsl_dataset_is_snapshot(ds)) {
 452                         /*
 453                          * In sync context, we're called with either no lock
 454                          * or with the write lock.  If we're not syncing,
 455                          * we're always called with the read lock held.
 456                          */
 457                         boolean_t need_lock =
 458                             !RW_WRITE_HELD(&dp->dp_config_rwlock) &&
 459                             dsl_pool_sync_context(dp);
 460
 461                         if (need_lock)
 462                                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 463
 464                         err = dsl_prop_get_ds(ds,
 465                             "refreservation", sizeof (uint64_t), 1,
 466                             &ds->ds_reserved, NULL);
 467                         if (err == 0) {
 468                                 err = dsl_prop_get_ds(ds,
 469                                     "refquota", sizeof (uint64_t), 1,
 470                                     &ds->ds_quota, NULL);
 471                         }
 472
 473                         if (need_lock)
 474                                 rw_exit(&dp->dp_config_rwlock);
 475                 } else {
 476                         ds->ds_reserved = ds->ds_quota = 0;
 477                 }
 478
 479                 if (err == 0) {
 480                         winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
 481                             dsl_dataset_evict);
 482                 }
 483                 if (err || winner) {
 484                         bplist_destroy(&ds->ds_pending_deadlist);
 485                         dsl_deadlist_close(&ds->ds_deadlist);
 486                         if (ds->ds_prev)
 487                                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 488                         dsl_dir_close(ds->ds_dir, ds);
 489                         mutex_destroy(&ds->ds_lock);
 490                         mutex_destroy(&ds->ds_recvlock);
 491                         mutex_destroy(&ds->ds_opening_lock);
 492                         rw_destroy(&ds->ds_rwlock);
 493                         cv_destroy(&ds->ds_exclusive_cv);
 494                         kmem_free(ds, sizeof (dsl_dataset_t));
 495                         if (err) {
 496                                 dmu_buf_rele(dbuf, tag);
 497                                 return (err);
 498                         }
 499                         ds = winner;
 500                 } else {
 501                         ds->ds_fsid_guid =
 502                             unique_insert(ds->ds_phys->ds_fsid_guid);
 503                 }
 504         }
 505         ASSERT3P(ds->ds_dbuf, ==, dbuf);
 506         ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
 507         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0 ||
 508             spa_version(dp->dp_spa) < SPA_VERSION_ORIGIN ||
 509             dp->dp_origin_snap == NULL || ds == dp->dp_origin_snap);
 510         mutex_enter(&ds->ds_lock);
 511         if (!dsl_pool_sync_context(dp) && DSL_DATASET_IS_DESTROYED(ds)) {
 512                 mutex_exit(&ds->ds_lock);
 513                 dmu_buf_rele(ds->ds_dbuf, tag);
 514                 return (ENOENT);
 515         }
 516         mutex_exit(&ds->ds_lock);
 517         *dsp = ds;
 518         return (0);
 519 }
 520
 521 static int
 522 dsl_dataset_hold_ref(dsl_dataset_t *ds, void *tag)
 523 {
 524         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 525
 526         /*
 527          * In syncing context we don't want the rwlock lock: there
 528          * may be an existing writer waiting for sync phase to
 529          * finish.  We don't need to worry about such writers, since
 530          * sync phase is single-threaded, so the writer can't be
 531          * doing anything while we are active.
 532          */
 533         if (dsl_pool_sync_context(dp)) {
 534                 ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
 535                 return (0);
 536         }
 537
 538         /*
 539          * Normal users will hold the ds_rwlock as a READER until they
 540          * are finished (i.e., call dsl_dataset_rele()).  "Owners" will
 541          * drop their READER lock after they set the ds_owner field.
 542          *
 543          * If the dataset is being destroyed, the destroy thread will
 544          * obtain a WRITER lock for exclusive access after it's done its
 545          * open-context work and then change the ds_owner to
 546          * dsl_reaper once destruction is assured.  So threads
 547          * may block here temporarily, until the "destructability" of
 548          * the dataset is determined.
 549          */
 550         ASSERT(!RW_WRITE_HELD(&dp->dp_config_rwlock));
 551         mutex_enter(&ds->ds_lock);
 552         while (!rw_tryenter(&ds->ds_rwlock, RW_READER)) {
 553                 rw_exit(&dp->dp_config_rwlock);
 554                 cv_wait(&ds->ds_exclusive_cv, &ds->ds_lock);
 555                 if (DSL_DATASET_IS_DESTROYED(ds)) {
 556                         mutex_exit(&ds->ds_lock);
 557                         dsl_dataset_drop_ref(ds, tag);
 558                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 559                         return (ENOENT);
 560                 }
 561                 /*
 562                  * The dp_config_rwlock lives above the ds_lock. And
 563                  * we need to check DSL_DATASET_IS_DESTROYED() while
 564                  * holding the ds_lock, so we have to drop and reacquire
 565                  * the ds_lock here.
 566                  */
 567                 mutex_exit(&ds->ds_lock);
 568                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 569                 mutex_enter(&ds->ds_lock);
 570         }
 571         mutex_exit(&ds->ds_lock);
 572         return (0);
 573 }
 574
 575 int
 576 dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 577     dsl_dataset_t **dsp)
 578 {
 579         int err = dsl_dataset_get_ref(dp, dsobj, tag, dsp);
 580
 581         if (err)
 582                 return (err);
 583         return (dsl_dataset_hold_ref(*dsp, tag));
 584 }
 585
 586 int
 587 dsl_dataset_own_obj(dsl_pool_t *dp, uint64_t dsobj, boolean_t inconsistentok,
 588     void *tag, dsl_dataset_t **dsp)
 589 {
 590         int err = dsl_dataset_hold_obj(dp, dsobj, tag, dsp);
 591         if (err)
 592                 return (err);
 593         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 594                 dsl_dataset_rele(*dsp, tag);
 595                 *dsp = NULL;
 596                 return (EBUSY);
 597         }
 598         return (0);
 599 }
 600
 601 int
 602 dsl_dataset_hold(const char *name, void *tag, dsl_dataset_t **dsp)
 603 {
 604         dsl_dir_t *dd;
 605         dsl_pool_t *dp;
 606         const char *snapname;
 607         uint64_t obj;
 608         int err = 0;
 609
 610         err = dsl_dir_open_spa(NULL, name, FTAG, &dd, &snapname);
 611         if (err)
 612                 return (err);
 613
 614         dp = dd->dd_pool;
 615         obj = dd->dd_phys->dd_head_dataset_obj;
 616         rw_enter(&dp->dp_config_rwlock, RW_READER);
 617         if (obj)
 618                 err = dsl_dataset_get_ref(dp, obj, tag, dsp);
 619         else
 620                 err = ENOENT;
 621         if (err)
 622                 goto out;
 623
 624         err = dsl_dataset_hold_ref(*dsp, tag);
 625
 626         /* we may be looking for a snapshot */
 627         if (err == 0 && snapname != NULL) {
 628                 dsl_dataset_t *ds = NULL;
 629
 630                 if (*snapname++ != '@') {
 631                         dsl_dataset_rele(*dsp, tag);
 632                         err = ENOENT;
 633                         goto out;
 634                 }
 635
 636                 dprintf("looking for snapshot '%s'\n", snapname);
 637                 err = dsl_dataset_snap_lookup(*dsp, snapname, &obj);
 638                 if (err == 0)
 639                         err = dsl_dataset_get_ref(dp, obj, tag, &ds);
 640                 dsl_dataset_rele(*dsp, tag);
 641
 642                 ASSERT3U((err == 0), ==, (ds != NULL));
 643
 644                 if (ds) {
 645                         mutex_enter(&ds->ds_lock);
 646                         if (ds->ds_snapname[0] == 0)
 647                                 (void) strlcpy(ds->ds_snapname, snapname,
 648                                     sizeof (ds->ds_snapname));
 649                         mutex_exit(&ds->ds_lock);
 650                         err = dsl_dataset_hold_ref(ds, tag);
 651                         *dsp = err ? NULL : ds;
 652                 }
 653         }
 654 out:
 655         rw_exit(&dp->dp_config_rwlock);
 656         dsl_dir_close(dd, FTAG);
 657         return (err);
 658 }
 659
 660 int
 661 dsl_dataset_own(const char *name, boolean_t inconsistentok,
 662     void *tag, dsl_dataset_t **dsp)
 663 {
 664         int err = dsl_dataset_hold(name, tag, dsp);
 665         if (err)
 666                 return (err);
 667         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 668                 dsl_dataset_rele(*dsp, tag);
 669                 return (EBUSY);
 670         }
 671         return (0);
 672 }
 673
 674 void
 675 dsl_dataset_name(dsl_dataset_t *ds, char *name)
 676 {
 677         if (ds == NULL) {
 678                 (void) strcpy(name, "mos");
 679         } else {
 680                 dsl_dir_name(ds->ds_dir, name);
 681                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 682                 if (ds->ds_snapname[0]) {
 683                         (void) strcat(name, "@");
 684                         /*
 685                          * We use a "recursive" mutex so that we
 686                          * can call dprintf_ds() with ds_lock held.
 687                          */
 688                         if (!MUTEX_HELD(&ds->ds_lock)) {
 689                                 mutex_enter(&ds->ds_lock);
 690                                 (void) strcat(name, ds->ds_snapname);
 691                                 mutex_exit(&ds->ds_lock);
 692                         } else {
 693                                 (void) strcat(name, ds->ds_snapname);
 694                         }
 695                 }
 696         }
 697 }
 698
 699 static int
 700 dsl_dataset_namelen(dsl_dataset_t *ds)
 701 {
 702         int result;
 703
 704         if (ds == NULL) {
 705                 result = 3;     /* "mos" */
 706         } else {
 707                 result = dsl_dir_namelen(ds->ds_dir);
 708                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 709                 if (ds->ds_snapname[0]) {
 710                         ++result;       /* adding one for the @-sign */
 711                         if (!MUTEX_HELD(&ds->ds_lock)) {
 712                                 mutex_enter(&ds->ds_lock);
 713                                 result += strlen(ds->ds_snapname);
 714                                 mutex_exit(&ds->ds_lock);
 715                         } else {
 716                                 result += strlen(ds->ds_snapname);
 717                         }
 718                 }
 719         }
 720
 721         return (result);
 722 }
 723
 724 void
 725 dsl_dataset_drop_ref(dsl_dataset_t *ds, void *tag)
 726 {
 727         dmu_buf_rele(ds->ds_dbuf, tag);
 728 }
 729
 730 void
 731 dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
 732 {
 733         if (!dsl_pool_sync_context(ds->ds_dir->dd_pool)) {
 734                 rw_exit(&ds->ds_rwlock);
 735         }
 736         dsl_dataset_drop_ref(ds, tag);
 737 }
 738
 739 void
 740 dsl_dataset_disown(dsl_dataset_t *ds, void *tag)
 741 {
 742         ASSERT((ds->ds_owner == tag && ds->ds_dbuf) ||
 743             (DSL_DATASET_IS_DESTROYED(ds) && ds->ds_dbuf == NULL));
 744
 745         mutex_enter(&ds->ds_lock);
 746         ds->ds_owner = NULL;
 747         if (RW_WRITE_HELD(&ds->ds_rwlock)) {
 748                 rw_exit(&ds->ds_rwlock);
 749                 cv_broadcast(&ds->ds_exclusive_cv);
 750         }
 751         mutex_exit(&ds->ds_lock);
 752         if (ds->ds_dbuf)
 753                 dsl_dataset_drop_ref(ds, tag);
 754         else
 755                 dsl_dataset_evict(NULL, ds);
 756 }
 757
 758 boolean_t
 759 dsl_dataset_tryown(dsl_dataset_t *ds, boolean_t inconsistentok, void *tag)
 760 {
 761         boolean_t gotit = FALSE;
 762
 763         mutex_enter(&ds->ds_lock);
 764         if (ds->ds_owner == NULL &&
 765             (!DS_IS_INCONSISTENT(ds) || inconsistentok)) {
 766                 ds->ds_owner = tag;
 767                 if (!dsl_pool_sync_context(ds->ds_dir->dd_pool))
 768                         rw_exit(&ds->ds_rwlock);
 769                 gotit = TRUE;
 770         }
 771         mutex_exit(&ds->ds_lock);
 772         return (gotit);
 773 }
 774
 775 void
 776 dsl_dataset_make_exclusive(dsl_dataset_t *ds, void *owner)
 777 {
 778         ASSERT3P(owner, ==, ds->ds_owner);
 779         if (!RW_WRITE_HELD(&ds->ds_rwlock))
 780                 rw_enter(&ds->ds_rwlock, RW_WRITER);
 781 }
 782
 783 uint64_t
 784 dsl_dataset_create_sync_dd(dsl_dir_t *dd, dsl_dataset_t *origin,
 785     uint64_t flags, dmu_tx_t *tx)
 786 {
 787         dsl_pool_t *dp = dd->dd_pool;
 788         dmu_buf_t *dbuf;
 789         dsl_dataset_phys_t *dsphys;
 790         uint64_t dsobj;
 791         objset_t *mos = dp->dp_meta_objset;
 792
 793         if (origin == NULL)
 794                 origin = dp->dp_origin_snap;
 795
 796         ASSERT(origin == NULL || origin->ds_dir->dd_pool == dp);
 797         ASSERT(origin == NULL || origin->ds_phys->ds_num_children > 0);
 798         ASSERT(dmu_tx_is_syncing(tx));
 799         ASSERT(dd->dd_phys->dd_head_dataset_obj == 0);
 800
 801         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
 802             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
 803         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
 804         dmu_buf_will_dirty(dbuf, tx);
 805         dsphys = dbuf->db_data;
 806         bzero(dsphys, sizeof (dsl_dataset_phys_t));
 807         dsphys->ds_dir_obj = dd->dd_object;
 808         dsphys->ds_flags = flags;
 809         dsphys->ds_fsid_guid = unique_create();
 810         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
 811             sizeof (dsphys->ds_guid));
 812         dsphys->ds_snapnames_zapobj =
 813             zap_create_norm(mos, U8_TEXTPREP_TOUPPER, DMU_OT_DSL_DS_SNAP_MAP,
 814             DMU_OT_NONE, 0, tx);
 815         dsphys->ds_creation_time = gethrestime_sec();
 816         dsphys->ds_creation_txg = tx->tx_txg == TXG_INITIAL ? 1 : tx->tx_txg;
 817
 818         if (origin == NULL) {
 819                 dsphys->ds_deadlist_obj = dsl_deadlist_alloc(mos, tx);
 820         } else {
 821                 dsl_dataset_t *ohds;
 822
 823                 dsphys->ds_prev_snap_obj = origin->ds_object;
 824                 dsphys->ds_prev_snap_txg =
 825                     origin->ds_phys->ds_creation_txg;
 826                 dsphys->ds_used_bytes =
 827                     origin->ds_phys->ds_used_bytes;
 828                 dsphys->ds_compressed_bytes =
 829                     origin->ds_phys->ds_compressed_bytes;
 830                 dsphys->ds_uncompressed_bytes =
 831                     origin->ds_phys->ds_uncompressed_bytes;
 832                 dsphys->ds_bp = origin->ds_phys->ds_bp;
 833                 dsphys->ds_flags |= origin->ds_phys->ds_flags;
 834
 835                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
 836                 origin->ds_phys->ds_num_children++;
 837
 838                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
 839                     origin->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ohds));
 840                 dsphys->ds_deadlist_obj = dsl_deadlist_clone(&ohds->ds_deadlist,
 841                     dsphys->ds_prev_snap_txg, dsphys->ds_prev_snap_obj, tx);
 842                 dsl_dataset_rele(ohds, FTAG);
 843
 844                 if (spa_version(dp->dp_spa) >= SPA_VERSION_NEXT_CLONES) {
 845                         if (origin->ds_phys->ds_next_clones_obj == 0) {
 846                                 origin->ds_phys->ds_next_clones_obj =
 847                                     zap_create(mos,
 848                                     DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
 849                         }
 850                         VERIFY(0 == zap_add_int(mos,
 851                             origin->ds_phys->ds_next_clones_obj,
 852                             dsobj, tx));
 853                 }
 854
 855                 dmu_buf_will_dirty(dd->dd_dbuf, tx);
 856                 dd->dd_phys->dd_origin_obj = origin->ds_object;
 857                 if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
 858                         if (origin->ds_dir->dd_phys->dd_clones == 0) {
 859                                 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
 860                                 origin->ds_dir->dd_phys->dd_clones =
 861                                     zap_create(mos,
 862                                     DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
 863                         }
 864                         VERIFY3U(0, ==, zap_add_int(mos,
 865                             origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
 866                 }
 867         }
 868
 869         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
 870                 dsphys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
 871
 872         dmu_buf_rele(dbuf, FTAG);
 873
 874         dmu_buf_will_dirty(dd->dd_dbuf, tx);
 875         dd->dd_phys->dd_head_dataset_obj = dsobj;
 876
 877         return (dsobj);
 878 }
 879
 880 uint64_t
 881 dsl_dataset_create_sync(dsl_dir_t *pdd, const char *lastname,
 882     dsl_dataset_t *origin, uint64_t flags, cred_t *cr, dmu_tx_t *tx)
 883 {
 884         dsl_pool_t *dp = pdd->dd_pool;
 885         uint64_t dsobj, ddobj;
 886         dsl_dir_t *dd;
 887
 888         ASSERT(lastname[0] != '@');
 889
 890         ddobj = dsl_dir_create_sync(dp, pdd, lastname, tx);
 891         VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
 892
 893         dsobj = dsl_dataset_create_sync_dd(dd, origin, flags, tx);
 894
 895         dsl_deleg_set_create_perms(dd, tx, cr);
 896
 897         dsl_dir_close(dd, FTAG);
 898
 899         /*
 900          * If we are creating a clone, make sure we zero out any stale
 901          * data from the origin snapshots zil header.
 902          */
 903         if (origin != NULL) {
 904                 dsl_dataset_t *ds;
 905                 objset_t *os;
 906
 907                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 908                 VERIFY3U(0, ==, dmu_objset_from_ds(ds, &os));
 909                 bzero(&os->os_zil_header, sizeof (os->os_zil_header));
 910                 dsl_dataset_dirty(ds, tx);
 911                 dsl_dataset_rele(ds, FTAG);
 912         }
 913
 914         return (dsobj);
 915 }
 916
 917 /*
 918  * The snapshots must all be in the same pool.
 919  */
 920 int
 921 dmu_snapshots_destroy_nvl(nvlist_t *snaps, boolean_t defer, char *failed)
 922 {
 923         int err;
 924         dsl_sync_task_t *dst;
 925         spa_t *spa;
 926         nvpair_t *pair;
 927         dsl_sync_task_group_t *dstg;
 928
 929         pair = nvlist_next_nvpair(snaps, NULL);
 930         if (pair == NULL)
 931                 return (0);
 932
 933         err = spa_open(nvpair_name(pair), &spa, FTAG);
 934         if (err)
 935                 return (err);
 936         dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
 937
 938         for (pair = nvlist_next_nvpair(snaps, NULL); pair != NULL;
 939             pair = nvlist_next_nvpair(snaps, pair)) {
 940                 dsl_dataset_t *ds;
 941                 int err;
 942
 943                 err = dsl_dataset_own(nvpair_name(pair), B_TRUE, dstg, &ds);
 944                 if (err == 0) {
 945                         struct dsl_ds_destroyarg *dsda;
 946
 947                         dsl_dataset_make_exclusive(ds, dstg);
 948                         dsda = kmem_zalloc(sizeof (struct dsl_ds_destroyarg),
 949                             KM_SLEEP);
 950                         dsda->ds = ds;
 951                         dsda->defer = defer;
 952                         dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
 953                             dsl_dataset_destroy_sync, dsda, dstg, 0);
 954                 } else if (err == ENOENT) {
 955                         err = 0;
 956                 } else {
 957                         (void) strcpy(failed, nvpair_name(pair));
 958                         break;
 959                 }
 960         }
 961
 962         if (err == 0)
 963                 err = dsl_sync_task_group_wait(dstg);
 964
 965         for (dst = list_head(&dstg->dstg_tasks); dst;
 966             dst = list_next(&dstg->dstg_tasks, dst)) {
 967                 struct dsl_ds_destroyarg *dsda = dst->dst_arg1;
 968                 dsl_dataset_t *ds = dsda->ds;
 969
 970                 /*
 971                  * Return the file system name that triggered the error
 972                  */
 973                 if (dst->dst_err) {
 974                         dsl_dataset_name(ds, failed);
 975                 }
 976                 ASSERT3P(dsda->rm_origin, ==, NULL);
 977                 dsl_dataset_disown(ds, dstg);
 978                 kmem_free(dsda, sizeof (struct dsl_ds_destroyarg));
 979         }
 980
 981         dsl_sync_task_group_destroy(dstg);
 982         spa_close(spa, FTAG);
 983         return (err);
 984
 985 }
 986
 987 static boolean_t
 988 dsl_dataset_might_destroy_origin(dsl_dataset_t *ds)
 989 {
 990         boolean_t might_destroy = B_FALSE;
 991
 992         mutex_enter(&ds->ds_lock);
 993         if (ds->ds_phys->ds_num_children == 2 && ds->ds_userrefs == 0 &&
 994             DS_IS_DEFER_DESTROY(ds))
 995                 might_destroy = B_TRUE;
 996         mutex_exit(&ds->ds_lock);
 997
 998         return (might_destroy);
 999 }
1000
1001 /*
1002  * If we're removing a clone, and these three conditions are true:
1003  *      1) the clone's origin has no other children
1004  *      2) the clone's origin has no user references
1005  *      3) the clone's origin has been marked for deferred destruction
1006  * Then, prepare to remove the origin as part of this sync task group.
1007  */
1008 static int
1009 dsl_dataset_origin_rm_prep(struct dsl_ds_destroyarg *dsda, void *tag)
1010 {
1011         dsl_dataset_t *ds = dsda->ds;
1012         dsl_dataset_t *origin = ds->ds_prev;
1013
1014         if (dsl_dataset_might_destroy_origin(origin)) {
1015                 char *name;
1016                 int namelen;
1017                 int error;
1018
1019                 namelen = dsl_dataset_namelen(origin) + 1;
1020                 name = kmem_alloc(namelen, KM_SLEEP);
1021                 dsl_dataset_name(origin, name);
1022 #ifdef _KERNEL
1023                 error = zfs_unmount_snap(name, NULL);
1024                 if (error) {
1025                         kmem_free(name, namelen);
1026                         return (error);
1027                 }
1028 #endif
1029                 error = dsl_dataset_own(name, B_TRUE, tag, &origin);
1030                 kmem_free(name, namelen);
1031                 if (error)
1032                         return (error);
1033                 dsda->rm_origin = origin;
1034                 dsl_dataset_make_exclusive(origin, tag);
1035         }
1036
1037         return (0);
1038 }
1039
1040 /*
1041  * ds must be opened as OWNER.  On return (whether successful or not),
1042  * ds will be closed and caller can no longer dereference it.
1043  */
1044 int
1045 dsl_dataset_destroy(dsl_dataset_t *ds, void *tag, boolean_t defer)
1046 {
1047         int err;
1048         dsl_sync_task_group_t *dstg;
1049         objset_t *os;
1050         dsl_dir_t *dd;
1051         uint64_t obj;
1052         struct dsl_ds_destroyarg dsda = { 0 };
1053         dsl_dataset_t *dummy_ds;
1054
1055         dsda.ds = ds;
1056
1057         if (dsl_dataset_is_snapshot(ds)) {
1058                 /* Destroying a snapshot is simpler */
1059                 dsl_dataset_make_exclusive(ds, tag);
1060
1061                 dsda.defer = defer;
1062                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1063                     dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
1064                     &dsda, tag, 0);
1065                 ASSERT3P(dsda.rm_origin, ==, NULL);
1066                 goto out;
1067         } else if (defer) {
1068                 err = EINVAL;
1069                 goto out;
1070         }
1071
1072         dd = ds->ds_dir;
1073         dummy_ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
1074         dummy_ds->ds_dir = dd;
1075         dummy_ds->ds_object = ds->ds_object;
1076
1077         /*
1078          * Check for errors and mark this ds as inconsistent, in
1079          * case we crash while freeing the objects.
1080          */
1081         err = dsl_sync_task_do(dd->dd_pool, dsl_dataset_destroy_begin_check,
1082             dsl_dataset_destroy_begin_sync, ds, NULL, 0);
1083         if (err)
1084                 goto out_free;
1085
1086         err = dmu_objset_from_ds(ds, &os);
1087         if (err)
1088                 goto out_free;
1089
1090         /*
1091          * remove the objects in open context, so that we won't
1092          * have too much to do in syncing context.
1093          */
1094         for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE,
1095             ds->ds_phys->ds_prev_snap_txg)) {
1096                 /*
1097                  * Ignore errors, if there is not enough disk space
1098                  * we will deal with it in dsl_dataset_destroy_sync().
1099                  */
1100                 (void) dmu_free_object(os, obj);
1101         }
1102         if (err != ESRCH)
1103                 goto out_free;
1104
1105         /*
1106          * Only the ZIL knows how to free log blocks.
1107          */
1108         zil_destroy(dmu_objset_zil(os), B_FALSE);
1109
1110         /*
1111          * Sync out all in-flight IO.
1112          */
1113         txg_wait_synced(dd->dd_pool, 0);
1114
1115         /*
1116          * If we managed to free all the objects in open
1117          * context, the user space accounting should be zero.
1118          */
1119         if (ds->ds_phys->ds_bp.blk_fill == 0 &&
1120             dmu_objset_userused_enabled(os)) {
1121                 ASSERTV(uint64_t count);
1122                 ASSERT(zap_count(os, DMU_USERUSED_OBJECT, &count) != 0 ||
1123                     count == 0);
1124                 ASSERT(zap_count(os, DMU_GROUPUSED_OBJECT, &count) != 0 ||
1125                     count == 0);
1126         }
1127
1128         rw_enter(&dd->dd_pool->dp_config_rwlock, RW_READER);
1129         err = dsl_dir_open_obj(dd->dd_pool, dd->dd_object, NULL, FTAG, &dd);
1130         rw_exit(&dd->dd_pool->dp_config_rwlock);
1131
1132         if (err)
1133                 goto out_free;
1134
1135         /*
1136          * Blow away the dsl_dir + head dataset.
1137          */
1138         dsl_dataset_make_exclusive(ds, tag);
1139         /*
1140          * If we're removing a clone, we might also need to remove its
1141          * origin.
1142          */
1143         do {
1144                 dsda.need_prep = B_FALSE;
1145                 if (dsl_dir_is_clone(dd)) {
1146                         err = dsl_dataset_origin_rm_prep(&dsda, tag);
1147                         if (err) {
1148                                 dsl_dir_close(dd, FTAG);
1149                                 goto out_free;
1150                         }
1151                 }
1152
1153                 dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
1154                 dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
1155                     dsl_dataset_destroy_sync, &dsda, tag, 0);
1156                 dsl_sync_task_create(dstg, dsl_dir_destroy_check,
1157                     dsl_dir_destroy_sync, dummy_ds, FTAG, 0);
1158                 err = dsl_sync_task_group_wait(dstg);
1159                 dsl_sync_task_group_destroy(dstg);
1160
1161                 /*
1162                  * We could be racing against 'zfs release' or 'zfs destroy -d'
1163                  * on the origin snap, in which case we can get EBUSY if we
1164                  * needed to destroy the origin snap but were not ready to
1165                  * do so.
1166                  */
1167                 if (dsda.need_prep) {
1168                         ASSERT(err == EBUSY);
1169                         ASSERT(dsl_dir_is_clone(dd));
1170                         ASSERT(dsda.rm_origin == NULL);
1171                 }
1172         } while (dsda.need_prep);
1173
1174         if (dsda.rm_origin != NULL)
1175                 dsl_dataset_disown(dsda.rm_origin, tag);
1176
1177         /* if it is successful, dsl_dir_destroy_sync will close the dd */
1178         if (err)
1179                 dsl_dir_close(dd, FTAG);
1180
1181 out_free:
1182         kmem_free(dummy_ds, sizeof (dsl_dataset_t));
1183 out:
1184         dsl_dataset_disown(ds, tag);
1185         return (err);
1186 }
1187
1188 blkptr_t *
1189 dsl_dataset_get_blkptr(dsl_dataset_t *ds)
1190 {
1191         return (&ds->ds_phys->ds_bp);
1192 }
1193
1194 void
1195 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
1196 {
1197         ASSERT(dmu_tx_is_syncing(tx));
1198         /* If it's the meta-objset, set dp_meta_rootbp */
1199         if (ds == NULL) {
1200                 tx->tx_pool->dp_meta_rootbp = *bp;
1201         } else {
1202                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1203                 ds->ds_phys->ds_bp = *bp;
1204         }
1205 }
1206
1207 spa_t *
1208 dsl_dataset_get_spa(dsl_dataset_t *ds)
1209 {
1210         return (ds->ds_dir->dd_pool->dp_spa);
1211 }
1212
1213 void
1214 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
1215 {
1216         dsl_pool_t *dp;
1217
1218         if (ds == NULL) /* this is the meta-objset */
1219                 return;
1220
1221         ASSERT(ds->ds_objset != NULL);
1222
1223         if (ds->ds_phys->ds_next_snap_obj != 0)
1224                 panic("dirtying snapshot!");
1225
1226         dp = ds->ds_dir->dd_pool;
1227
1228         if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
1229                 /* up the hold count until we can be written out */
1230                 dmu_buf_add_ref(ds->ds_dbuf, ds);
1231         }
1232 }
1233
1234 /*
1235  * The unique space in the head dataset can be calculated by subtracting
1236  * the space used in the most recent snapshot, that is still being used
1237  * in this file system, from the space currently in use.  To figure out
1238  * the space in the most recent snapshot still in use, we need to take
1239  * the total space used in the snapshot and subtract out the space that
1240  * has been freed up since the snapshot was taken.
1241  */
1242 static void
1243 dsl_dataset_recalc_head_uniq(dsl_dataset_t *ds)
1244 {
1245         uint64_t mrs_used;
1246         uint64_t dlused, dlcomp, dluncomp;
1247
1248         ASSERT(!dsl_dataset_is_snapshot(ds));
1249
1250         if (ds->ds_phys->ds_prev_snap_obj != 0)
1251                 mrs_used = ds->ds_prev->ds_phys->ds_used_bytes;
1252         else
1253                 mrs_used = 0;
1254
1255         dsl_deadlist_space(&ds->ds_deadlist, &dlused, &dlcomp, &dluncomp);
1256
1257         ASSERT3U(dlused, <=, mrs_used);
1258         ds->ds_phys->ds_unique_bytes =
1259             ds->ds_phys->ds_used_bytes - (mrs_used - dlused);
1260
1261         if (spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1262             SPA_VERSION_UNIQUE_ACCURATE)
1263                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1264 }
1265
1266 struct killarg {
1267         dsl_dataset_t *ds;
1268         dmu_tx_t *tx;
1269 };
1270
1271 /* ARGSUSED */
1272 static int
1273 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
1274     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
1275 {
1276         struct killarg *ka = arg;
1277         dmu_tx_t *tx = ka->tx;
1278
1279         if (bp == NULL)
1280                 return (0);
1281
1282         if (zb->zb_level == ZB_ZIL_LEVEL) {
1283                 ASSERT(zilog != NULL);
1284                 /*
1285                  * It's a block in the intent log.  It has no
1286                  * accounting, so just free it.
1287                  */
1288                 dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
1289         } else {
1290                 ASSERT(zilog == NULL);
1291                 ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
1292                 (void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
1293         }
1294
1295         return (0);
1296 }
1297
1298 /* ARGSUSED */
1299 static int
1300 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
1301 {
1302         dsl_dataset_t *ds = arg1;
1303         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1304         uint64_t count;
1305         int err;
1306
1307         /*
1308          * Can't delete a head dataset if there are snapshots of it.
1309          * (Except if the only snapshots are from the branch we cloned
1310          * from.)
1311          */
1312         if (ds->ds_prev != NULL &&
1313             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1314                 return (EBUSY);
1315
1316         /*
1317          * This is really a dsl_dir thing, but check it here so that
1318          * we'll be less likely to leave this dataset inconsistent &
1319          * nearly destroyed.
1320          */
1321         err = zap_count(mos, ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
1322         if (err)
1323                 return (err);
1324         if (count != 0)
1325                 return (EEXIST);
1326
1327         return (0);
1328 }
1329
1330 /* ARGSUSED */
1331 static void
1332 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1333 {
1334         dsl_dataset_t *ds = arg1;
1335         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1336
1337         /* Mark it as inconsistent on-disk, in case we crash */
1338         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1339         ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1340
1341         spa_history_log_internal(LOG_DS_DESTROY_BEGIN, dp->dp_spa, tx,
1342             "dataset = %llu", ds->ds_object);
1343 }
1344
1345 static int
1346 dsl_dataset_origin_check(struct dsl_ds_destroyarg *dsda, void *tag,
1347     dmu_tx_t *tx)
1348 {
1349         dsl_dataset_t *ds = dsda->ds;
1350         dsl_dataset_t *ds_prev = ds->ds_prev;
1351
1352         if (dsl_dataset_might_destroy_origin(ds_prev)) {
1353                 struct dsl_ds_destroyarg ndsda = {0};
1354
1355                 /*
1356                  * If we're not prepared to remove the origin, don't remove
1357                  * the clone either.
1358                  */
1359                 if (dsda->rm_origin == NULL) {
1360                         dsda->need_prep = B_TRUE;
1361                         return (EBUSY);
1362                 }
1363
1364                 ndsda.ds = ds_prev;
1365                 ndsda.is_origin_rm = B_TRUE;
1366                 return (dsl_dataset_destroy_check(&ndsda, tag, tx));
1367         }
1368
1369         /*
1370          * If we're not going to remove the origin after all,
1371          * undo the open context setup.
1372          */
1373         if (dsda->rm_origin != NULL) {
1374                 dsl_dataset_disown(dsda->rm_origin, tag);
1375                 dsda->rm_origin = NULL;
1376         }
1377
1378         return (0);
1379 }
1380
1381 /*
1382  * If you add new checks here, you may need to add
1383  * additional checks to the "temporary" case in
1384  * snapshot_check() in dmu_objset.c.
1385  */
1386 /* ARGSUSED */
1387 int
1388 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1389 {
1390         struct dsl_ds_destroyarg *dsda = arg1;
1391         dsl_dataset_t *ds = dsda->ds;
1392
1393         /* we have an owner hold, so noone else can destroy us */
1394         ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1395
1396         /*
1397          * Only allow deferred destroy on pools that support it.
1398          * NOTE: deferred destroy is only supported on snapshots.
1399          */
1400         if (dsda->defer) {
1401                 if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1402                     SPA_VERSION_USERREFS)
1403                         return (ENOTSUP);
1404                 ASSERT(dsl_dataset_is_snapshot(ds));
1405                 return (0);
1406         }
1407
1408         /*
1409          * Can't delete a head dataset if there are snapshots of it.
1410          * (Except if the only snapshots are from the branch we cloned
1411          * from.)
1412          */
1413         if (ds->ds_prev != NULL &&
1414             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1415                 return (EBUSY);
1416
1417         /*
1418          * If we made changes this txg, traverse_dsl_dataset won't find
1419          * them.  Try again.
1420          */
1421         if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1422                 return (EAGAIN);
1423
1424         if (dsl_dataset_is_snapshot(ds)) {
1425                 /*
1426                  * If this snapshot has an elevated user reference count,
1427                  * we can't destroy it yet.
1428                  */
1429                 if (ds->ds_userrefs > 0 && !dsda->releasing)
1430                         return (EBUSY);
1431
1432                 mutex_enter(&ds->ds_lock);
1433                 /*
1434                  * Can't delete a branch point. However, if we're destroying
1435                  * a clone and removing its origin due to it having a user
1436                  * hold count of 0 and having been marked for deferred destroy,
1437                  * it's OK for the origin to have a single clone.
1438                  */
1439                 if (ds->ds_phys->ds_num_children >
1440                     (dsda->is_origin_rm ? 2 : 1)) {
1441                         mutex_exit(&ds->ds_lock);
1442                         return (EEXIST);
1443                 }
1444                 mutex_exit(&ds->ds_lock);
1445         } else if (dsl_dir_is_clone(ds->ds_dir)) {
1446                 return (dsl_dataset_origin_check(dsda, arg2, tx));
1447         }
1448
1449         /* XXX we should do some i/o error checking... */
1450         return (0);
1451 }
1452
1453 struct refsarg {
1454         kmutex_t lock;
1455         boolean_t gone;
1456         kcondvar_t cv;
1457 };
1458
1459 /* ARGSUSED */
1460 static void
1461 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1462 {
1463         struct refsarg *arg = argv;
1464
1465         mutex_enter(&arg->lock);
1466         arg->gone = TRUE;
1467         cv_signal(&arg->cv);
1468         mutex_exit(&arg->lock);
1469 }
1470
1471 static void
1472 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1473 {
1474         struct refsarg arg;
1475
1476         mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1477         cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1478         arg.gone = FALSE;
1479         (void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1480             dsl_dataset_refs_gone);
1481         dmu_buf_rele(ds->ds_dbuf, tag);
1482         mutex_enter(&arg.lock);
1483         while (!arg.gone)
1484                 cv_wait(&arg.cv, &arg.lock);
1485         ASSERT(arg.gone);
1486         mutex_exit(&arg.lock);
1487         ds->ds_dbuf = NULL;
1488         ds->ds_phys = NULL;
1489         mutex_destroy(&arg.lock);
1490         cv_destroy(&arg.cv);
1491 }
1492
1493 static void
1494 remove_from_next_clones(dsl_dataset_t *ds, uint64_t obj, dmu_tx_t *tx)
1495 {
1496         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1497         int err;
1498         ASSERTV(uint64_t count);
1499
1500         ASSERT(ds->ds_phys->ds_num_children >= 2);
1501         err = zap_remove_int(mos, ds->ds_phys->ds_next_clones_obj, obj, tx);
1502         /*
1503          * The err should not be ENOENT, but a bug in a previous version
1504          * of the code could cause upgrade_clones_cb() to not set
1505          * ds_next_snap_obj when it should, leading to a missing entry.
1506          * If we knew that the pool was created after
1507          * SPA_VERSION_NEXT_CLONES, we could assert that it isn't
1508          * ENOENT.  However, at least we can check that we don't have
1509          * too many entries in the next_clones_obj even after failing to
1510          * remove this one.
1511          */
1512         if (err != ENOENT) {
1513                 VERIFY3U(err, ==, 0);
1514         }
1515         ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
1516             &count));
1517         ASSERT3U(count, <=, ds->ds_phys->ds_num_children - 2);
1518 }
1519
1520 static void
1521 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
1522 {
1523         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1524         zap_cursor_t zc;
1525         zap_attribute_t za;
1526
1527         /*
1528          * If it is the old version, dd_clones doesn't exist so we can't
1529          * find the clones, but deadlist_remove_key() is a no-op so it
1530          * doesn't matter.
1531          */
1532         if (ds->ds_dir->dd_phys->dd_clones == 0)
1533                 return;
1534
1535         for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
1536             zap_cursor_retrieve(&zc, &za) == 0;
1537             zap_cursor_advance(&zc)) {
1538                 dsl_dataset_t *clone;
1539
1540                 VERIFY3U(0, ==, dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
1541                     za.za_first_integer, FTAG, &clone));
1542                 if (clone->ds_dir->dd_origin_txg > mintxg) {
1543                         dsl_deadlist_remove_key(&clone->ds_deadlist,
1544                             mintxg, tx);
1545                         dsl_dataset_remove_clones_key(clone, mintxg, tx);
1546                 }
1547                 dsl_dataset_rele(clone, FTAG);
1548         }
1549         zap_cursor_fini(&zc);
1550 }
1551
1552 struct process_old_arg {
1553         dsl_dataset_t *ds;
1554         dsl_dataset_t *ds_prev;
1555         boolean_t after_branch_point;
1556         zio_t *pio;
1557         uint64_t used, comp, uncomp;
1558 };
1559
1560 static int
1561 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1562 {
1563         struct process_old_arg *poa = arg;
1564         dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
1565
1566         if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
1567                 dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
1568                 if (poa->ds_prev && !poa->after_branch_point &&
1569                     bp->blk_birth >
1570                     poa->ds_prev->ds_phys->ds_prev_snap_txg) {
1571                         poa->ds_prev->ds_phys->ds_unique_bytes +=
1572                             bp_get_dsize_sync(dp->dp_spa, bp);
1573                 }
1574         } else {
1575                 poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
1576                 poa->comp += BP_GET_PSIZE(bp);
1577                 poa->uncomp += BP_GET_UCSIZE(bp);
1578                 dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
1579         }
1580         return (0);
1581 }
1582
1583 static void
1584 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
1585     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
1586 {
1587         struct process_old_arg poa = { 0 };
1588         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1589         objset_t *mos = dp->dp_meta_objset;
1590
1591         ASSERT(ds->ds_deadlist.dl_oldfmt);
1592         ASSERT(ds_next->ds_deadlist.dl_oldfmt);
1593
1594         poa.ds = ds;
1595         poa.ds_prev = ds_prev;
1596         poa.after_branch_point = after_branch_point;
1597         poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1598         VERIFY3U(0, ==, bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
1599             process_old_cb, &poa, tx));
1600         VERIFY3U(zio_wait(poa.pio), ==, 0);
1601         ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
1602
1603         /* change snapused */
1604         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1605             -poa.used, -poa.comp, -poa.uncomp, tx);
1606
1607         /* swap next's deadlist to our deadlist */
1608         dsl_deadlist_close(&ds->ds_deadlist);
1609         dsl_deadlist_close(&ds_next->ds_deadlist);
1610         SWITCH64(ds_next->ds_phys->ds_deadlist_obj,
1611             ds->ds_phys->ds_deadlist_obj);
1612         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
1613         dsl_deadlist_open(&ds_next->ds_deadlist, mos,
1614             ds_next->ds_phys->ds_deadlist_obj);
1615 }
1616
1617 void
1618 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1619 {
1620         struct dsl_ds_destroyarg *dsda = arg1;
1621         dsl_dataset_t *ds = dsda->ds;
1622         int err;
1623         int after_branch_point = FALSE;
1624         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1625         objset_t *mos = dp->dp_meta_objset;
1626         dsl_dataset_t *ds_prev = NULL;
1627         boolean_t wont_destroy;
1628         uint64_t obj;
1629
1630         wont_destroy = (dsda->defer &&
1631             (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1));
1632
1633         ASSERT(ds->ds_owner || wont_destroy);
1634         ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1635         ASSERT(ds->ds_prev == NULL ||
1636             ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1637         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1638
1639         if (wont_destroy) {
1640                 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1641                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1642                 ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1643                 return;
1644         }
1645
1646         /* signal any waiters that this dataset is going away */
1647         mutex_enter(&ds->ds_lock);
1648         ds->ds_owner = dsl_reaper;
1649         cv_broadcast(&ds->ds_exclusive_cv);
1650         mutex_exit(&ds->ds_lock);
1651
1652         /* Remove our reservation */
1653         if (ds->ds_reserved != 0) {
1654                 dsl_prop_setarg_t psa;
1655                 uint64_t value = 0;
1656
1657                 dsl_prop_setarg_init_uint64(&psa, "refreservation",
1658                     (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
1659                     &value);
1660                 psa.psa_effective_value = 0;    /* predict default value */
1661
1662                 dsl_dataset_set_reservation_sync(ds, &psa, tx);
1663                 ASSERT3U(ds->ds_reserved, ==, 0);
1664         }
1665
1666         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1667
1668         dsl_scan_ds_destroyed(ds, tx);
1669
1670         obj = ds->ds_object;
1671
1672         if (ds->ds_phys->ds_prev_snap_obj != 0) {
1673                 if (ds->ds_prev) {
1674                         ds_prev = ds->ds_prev;
1675                 } else {
1676                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1677                             ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1678                 }
1679                 after_branch_point =
1680                     (ds_prev->ds_phys->ds_next_snap_obj != obj);
1681
1682                 dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1683                 if (after_branch_point &&
1684                     ds_prev->ds_phys->ds_next_clones_obj != 0) {
1685                         remove_from_next_clones(ds_prev, obj, tx);
1686                         if (ds->ds_phys->ds_next_snap_obj != 0) {
1687                                 VERIFY(0 == zap_add_int(mos,
1688                                     ds_prev->ds_phys->ds_next_clones_obj,
1689                                     ds->ds_phys->ds_next_snap_obj, tx));
1690                         }
1691                 }
1692                 if (after_branch_point &&
1693                     ds->ds_phys->ds_next_snap_obj == 0) {
1694                         /* This clone is toast. */
1695                         ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1696                         ds_prev->ds_phys->ds_num_children--;
1697
1698                         /*
1699                          * If the clone's origin has no other clones, no
1700                          * user holds, and has been marked for deferred
1701                          * deletion, then we should have done the necessary
1702                          * destroy setup for it.
1703                          */
1704                         if (ds_prev->ds_phys->ds_num_children == 1 &&
1705                             ds_prev->ds_userrefs == 0 &&
1706                             DS_IS_DEFER_DESTROY(ds_prev)) {
1707                                 ASSERT3P(dsda->rm_origin, !=, NULL);
1708                         } else {
1709                                 ASSERT3P(dsda->rm_origin, ==, NULL);
1710                         }
1711                 } else if (!after_branch_point) {
1712                         ds_prev->ds_phys->ds_next_snap_obj =
1713                             ds->ds_phys->ds_next_snap_obj;
1714                 }
1715         }
1716
1717         if (dsl_dataset_is_snapshot(ds)) {
1718                 dsl_dataset_t *ds_next;
1719                 uint64_t old_unique;
1720                 uint64_t used = 0, comp = 0, uncomp = 0;
1721
1722                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1723                     ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1724                 ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1725
1726                 old_unique = ds_next->ds_phys->ds_unique_bytes;
1727
1728                 dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1729                 ds_next->ds_phys->ds_prev_snap_obj =
1730                     ds->ds_phys->ds_prev_snap_obj;
1731                 ds_next->ds_phys->ds_prev_snap_txg =
1732                     ds->ds_phys->ds_prev_snap_txg;
1733                 ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1734                     ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1735
1736
1737                 if (ds_next->ds_deadlist.dl_oldfmt) {
1738                         process_old_deadlist(ds, ds_prev, ds_next,
1739                             after_branch_point, tx);
1740                 } else {
1741                         /* Adjust prev's unique space. */
1742                         if (ds_prev && !after_branch_point) {
1743                                 dsl_deadlist_space_range(&ds_next->ds_deadlist,
1744                                     ds_prev->ds_phys->ds_prev_snap_txg,
1745                                     ds->ds_phys->ds_prev_snap_txg,
1746                                     &used, &comp, &uncomp);
1747                                 ds_prev->ds_phys->ds_unique_bytes += used;
1748                         }
1749
1750                         /* Adjust snapused. */
1751                         dsl_deadlist_space_range(&ds_next->ds_deadlist,
1752                             ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
1753                             &used, &comp, &uncomp);
1754                         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1755                             -used, -comp, -uncomp, tx);
1756
1757                         /* Move blocks to be freed to pool's free list. */
1758                         dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
1759                             &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
1760                             tx);
1761                         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
1762                             DD_USED_HEAD, used, comp, uncomp, tx);
1763                         dsl_dir_dirty(tx->tx_pool->dp_free_dir, tx);
1764
1765                         /* Merge our deadlist into next's and free it. */
1766                         dsl_deadlist_merge(&ds_next->ds_deadlist,
1767                             ds->ds_phys->ds_deadlist_obj, tx);
1768                 }
1769                 dsl_deadlist_close(&ds->ds_deadlist);
1770                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1771
1772                 /* Collapse range in clone heads */
1773                 dsl_dataset_remove_clones_key(ds,
1774                     ds->ds_phys->ds_creation_txg, tx);
1775
1776                 if (dsl_dataset_is_snapshot(ds_next)) {
1777                         dsl_dataset_t *ds_nextnext;
1778                         dsl_dataset_t *hds;
1779
1780                         /*
1781                          * Update next's unique to include blocks which
1782                          * were previously shared by only this snapshot
1783                          * and it.  Those blocks will be born after the
1784                          * prev snap and before this snap, and will have
1785                          * died after the next snap and before the one
1786                          * after that (ie. be on the snap after next's
1787                          * deadlist).
1788                          */
1789                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1790                             ds_next->ds_phys->ds_next_snap_obj,
1791                             FTAG, &ds_nextnext));
1792                         dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
1793                             ds->ds_phys->ds_prev_snap_txg,
1794                             ds->ds_phys->ds_creation_txg,
1795                             &used, &comp, &uncomp);
1796                         ds_next->ds_phys->ds_unique_bytes += used;
1797                         dsl_dataset_rele(ds_nextnext, FTAG);
1798                         ASSERT3P(ds_next->ds_prev, ==, NULL);
1799
1800                         /* Collapse range in this head. */
1801                         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
1802                             ds->ds_dir->dd_phys->dd_head_dataset_obj,
1803                             FTAG, &hds));
1804                         dsl_deadlist_remove_key(&hds->ds_deadlist,
1805                             ds->ds_phys->ds_creation_txg, tx);
1806                         dsl_dataset_rele(hds, FTAG);
1807
1808                 } else {
1809                         ASSERT3P(ds_next->ds_prev, ==, ds);
1810                         dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1811                         ds_next->ds_prev = NULL;
1812                         if (ds_prev) {
1813                                 VERIFY(0 == dsl_dataset_get_ref(dp,
1814                                     ds->ds_phys->ds_prev_snap_obj,
1815                                     ds_next, &ds_next->ds_prev));
1816                         }
1817
1818                         dsl_dataset_recalc_head_uniq(ds_next);
1819
1820                         /*
1821                          * Reduce the amount of our unconsmed refreservation
1822                          * being charged to our parent by the amount of
1823                          * new unique data we have gained.
1824                          */
1825                         if (old_unique < ds_next->ds_reserved) {
1826                                 int64_t mrsdelta;
1827                                 uint64_t new_unique =
1828                                     ds_next->ds_phys->ds_unique_bytes;
1829
1830                                 ASSERT(old_unique <= new_unique);
1831                                 mrsdelta = MIN(new_unique - old_unique,
1832                                     ds_next->ds_reserved - old_unique);
1833                                 dsl_dir_diduse_space(ds->ds_dir,
1834                                     DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1835                         }
1836                 }
1837                 dsl_dataset_rele(ds_next, FTAG);
1838         } else {
1839                 /*
1840                  * There's no next snapshot, so this is a head dataset.
1841                  * Destroy the deadlist.  Unless it's a clone, the
1842                  * deadlist should be empty.  (If it's a clone, it's
1843                  * safe to ignore the deadlist contents.)
1844                  */
1845                 struct killarg ka;
1846
1847                 dsl_deadlist_close(&ds->ds_deadlist);
1848                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1849                 ds->ds_phys->ds_deadlist_obj = 0;
1850
1851                 /*
1852                  * Free everything that we point to (that's born after
1853                  * the previous snapshot, if we are a clone)
1854                  *
1855                  * NB: this should be very quick, because we already
1856                  * freed all the objects in open context.
1857                  */
1858                 ka.ds = ds;
1859                 ka.tx = tx;
1860                 err = traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1861                     TRAVERSE_POST, kill_blkptr, &ka);
1862                 ASSERT3U(err, ==, 0);
1863                 ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1864                     ds->ds_phys->ds_unique_bytes == 0);
1865
1866                 if (ds->ds_prev != NULL) {
1867                         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
1868                                 VERIFY3U(0, ==, zap_remove_int(mos,
1869                                     ds->ds_prev->ds_dir->dd_phys->dd_clones,
1870                                     ds->ds_object, tx));
1871                         }
1872                         dsl_dataset_rele(ds->ds_prev, ds);
1873                         ds->ds_prev = ds_prev = NULL;
1874                 }
1875         }
1876
1877         /*
1878          * This must be done after the dsl_traverse(), because it will
1879          * re-open the objset.
1880          */
1881         if (ds->ds_objset) {
1882                 dmu_objset_evict(ds->ds_objset);
1883                 ds->ds_objset = NULL;
1884         }
1885
1886         if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1887                 /* Erase the link in the dir */
1888                 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1889                 ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1890                 ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1891                 err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1892                 ASSERT(err == 0);
1893         } else {
1894                 /* remove from snapshot namespace */
1895                 dsl_dataset_t *ds_head;
1896                 ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1897                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1898                     ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1899                 VERIFY(0 == dsl_dataset_get_snapname(ds));
1900 #ifdef ZFS_DEBUG
1901                 {
1902                         uint64_t val;
1903
1904                         err = dsl_dataset_snap_lookup(ds_head,
1905                             ds->ds_snapname, &val);
1906                         ASSERT3U(err, ==, 0);
1907                         ASSERT3U(val, ==, obj);
1908                 }
1909 #endif
1910                 err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1911                 ASSERT(err == 0);
1912                 dsl_dataset_rele(ds_head, FTAG);
1913         }
1914
1915         if (ds_prev && ds->ds_prev != ds_prev)
1916                 dsl_dataset_rele(ds_prev, FTAG);
1917
1918         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1919         spa_history_log_internal(LOG_DS_DESTROY, dp->dp_spa, tx,
1920             "dataset = %llu", ds->ds_object);
1921
1922         if (ds->ds_phys->ds_next_clones_obj != 0) {
1923                 ASSERTV(uint64_t count);
1924                 ASSERT(0 == zap_count(mos,
1925                     ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1926                 VERIFY(0 == dmu_object_free(mos,
1927                     ds->ds_phys->ds_next_clones_obj, tx));
1928         }
1929         if (ds->ds_phys->ds_props_obj != 0)
1930                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1931         if (ds->ds_phys->ds_userrefs_obj != 0)
1932                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1933         dsl_dir_close(ds->ds_dir, ds);
1934         ds->ds_dir = NULL;
1935         dsl_dataset_drain_refs(ds, tag);
1936         VERIFY(0 == dmu_object_free(mos, obj, tx));
1937
1938         if (dsda->rm_origin) {
1939                 /*
1940                  * Remove the origin of the clone we just destroyed.
1941                  */
1942                 struct dsl_ds_destroyarg ndsda = {0};
1943
1944                 ndsda.ds = dsda->rm_origin;
1945                 dsl_dataset_destroy_sync(&ndsda, tag, tx);
1946         }
1947 }
1948
1949 static int
1950 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
1951 {
1952         uint64_t asize;
1953
1954         if (!dmu_tx_is_syncing(tx))
1955                 return (0);
1956
1957         /*
1958          * If there's an fs-only reservation, any blocks that might become
1959          * owned by the snapshot dataset must be accommodated by space
1960          * outside of the reservation.
1961          */
1962         ASSERT(ds->ds_reserved == 0 || DS_UNIQUE_IS_ACCURATE(ds));
1963         asize = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
1964         if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
1965                 return (ENOSPC);
1966
1967         /*
1968          * Propogate any reserved space for this snapshot to other
1969          * snapshot checks in this sync group.
1970          */
1971         if (asize > 0)
1972                 dsl_dir_willuse_space(ds->ds_dir, asize, tx);
1973
1974         return (0);
1975 }
1976
1977 int
1978 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1979 {
1980         dsl_dataset_t *ds = arg1;
1981         const char *snapname = arg2;
1982         int err;
1983         uint64_t value;
1984
1985         /*
1986          * We don't allow multiple snapshots of the same txg.  If there
1987          * is already one, try again.
1988          */
1989         if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
1990                 return (EAGAIN);
1991
1992         /*
1993          * Check for conflicting name snapshot name.
1994          */
1995         err = dsl_dataset_snap_lookup(ds, snapname, &value);
1996         if (err == 0)
1997                 return (EEXIST);
1998         if (err != ENOENT)
1999                 return (err);
2000
2001         /*
2002          * Check that the dataset's name is not too long.  Name consists
2003          * of the dataset's length + 1 for the @-sign + snapshot name's length
2004          */
2005         if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
2006                 return (ENAMETOOLONG);
2007
2008         err = dsl_dataset_snapshot_reserve_space(ds, tx);
2009         if (err)
2010                 return (err);
2011
2012         ds->ds_trysnap_txg = tx->tx_txg;
2013         return (0);
2014 }
2015
2016 void
2017 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2018 {
2019         dsl_dataset_t *ds = arg1;
2020         const char *snapname = arg2;
2021         dsl_pool_t *dp = ds->ds_dir->dd_pool;
2022         dmu_buf_t *dbuf;
2023         dsl_dataset_phys_t *dsphys;
2024         uint64_t dsobj, crtxg;
2025         objset_t *mos = dp->dp_meta_objset;
2026         int err;
2027
2028         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
2029
2030         /*
2031          * The origin's ds_creation_txg has to be < TXG_INITIAL
2032          */
2033         if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2034                 crtxg = 1;
2035         else
2036                 crtxg = tx->tx_txg;
2037
2038         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2039             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2040         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2041         dmu_buf_will_dirty(dbuf, tx);
2042         dsphys = dbuf->db_data;
2043         bzero(dsphys, sizeof (dsl_dataset_phys_t));
2044         dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2045         dsphys->ds_fsid_guid = unique_create();
2046         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2047             sizeof (dsphys->ds_guid));
2048         dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2049         dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2050         dsphys->ds_next_snap_obj = ds->ds_object;
2051         dsphys->ds_num_children = 1;
2052         dsphys->ds_creation_time = gethrestime_sec();
2053         dsphys->ds_creation_txg = crtxg;
2054         dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2055         dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
2056         dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2057         dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2058         dsphys->ds_flags = ds->ds_phys->ds_flags;
2059         dsphys->ds_bp = ds->ds_phys->ds_bp;
2060         dmu_buf_rele(dbuf, FTAG);
2061
2062         ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2063         if (ds->ds_prev) {
2064                 uint64_t next_clones_obj =
2065                     ds->ds_prev->ds_phys->ds_next_clones_obj;
2066                 ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2067                     ds->ds_object ||
2068                     ds->ds_prev->ds_phys->ds_num_children > 1);
2069                 if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2070                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2071                         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2072                             ds->ds_prev->ds_phys->ds_creation_txg);
2073                         ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2074                 } else if (next_clones_obj != 0) {
2075                         remove_from_next_clones(ds->ds_prev,
2076                             dsphys->ds_next_snap_obj, tx);
2077                         VERIFY3U(0, ==, zap_add_int(mos,
2078                             next_clones_obj, dsobj, tx));
2079                 }
2080         }
2081
2082         /*
2083          * If we have a reference-reservation on this dataset, we will
2084          * need to increase the amount of refreservation being charged
2085          * since our unique space is going to zero.
2086          */
2087         if (ds->ds_reserved) {
2088                 int64_t delta;
2089                 ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
2090                 delta = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2091                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2092                     delta, 0, 0, tx);
2093         }
2094
2095         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2096         zfs_dbgmsg("taking snapshot %s@%s/%llu; newkey=%llu",
2097             ds->ds_dir->dd_myname, snapname, dsobj,
2098             ds->ds_phys->ds_prev_snap_txg);
2099         ds->ds_phys->ds_deadlist_obj = dsl_deadlist_clone(&ds->ds_deadlist,
2100             UINT64_MAX, ds->ds_phys->ds_prev_snap_obj, tx);
2101         dsl_deadlist_close(&ds->ds_deadlist);
2102         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
2103         dsl_deadlist_add_key(&ds->ds_deadlist,
2104             ds->ds_phys->ds_prev_snap_txg, tx);
2105
2106         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2107         ds->ds_phys->ds_prev_snap_obj = dsobj;
2108         ds->ds_phys->ds_prev_snap_txg = crtxg;
2109         ds->ds_phys->ds_unique_bytes = 0;
2110         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2111                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2112
2113         err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2114             snapname, 8, 1, &dsobj, tx);
2115         ASSERT(err == 0);
2116
2117         if (ds->ds_prev)
2118                 dsl_dataset_drop_ref(ds->ds_prev, ds);
2119         VERIFY(0 == dsl_dataset_get_ref(dp,
2120             ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2121
2122         dsl_scan_ds_snapshotted(ds, tx);
2123
2124         dsl_dir_snap_cmtime_update(ds->ds_dir);
2125
2126         spa_history_log_internal(LOG_DS_SNAPSHOT, dp->dp_spa, tx,
2127             "dataset = %llu", dsobj);
2128 }
2129
2130 void
2131 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2132 {
2133         ASSERT(dmu_tx_is_syncing(tx));
2134         ASSERT(ds->ds_objset != NULL);
2135         ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2136
2137         /*
2138          * in case we had to change ds_fsid_guid when we opened it,
2139          * sync it out now.
2140          */
2141         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2142         ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2143
2144         dsl_dir_dirty(ds->ds_dir, tx);
2145         dmu_objset_sync(ds->ds_objset, zio, tx);
2146 }
2147
2148 static void
2149 get_clones_stat(dsl_dataset_t *ds, nvlist_t *nv)
2150 {
2151         uint64_t count = 0;
2152         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
2153         zap_cursor_t zc;
2154         zap_attribute_t za;
2155         nvlist_t *propval;
2156         nvlist_t *val;
2157
2158         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2159         VERIFY(nvlist_alloc(&propval, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2160         VERIFY(nvlist_alloc(&val, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2161
2162         /*
2163          * There may me missing entries in ds_next_clones_obj
2164          * due to a bug in a previous version of the code.
2165          * Only trust it if it has the right number of entries.
2166          */
2167         if (ds->ds_phys->ds_next_clones_obj != 0) {
2168                 ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
2169                     &count));
2170         }
2171         if (count != ds->ds_phys->ds_num_children - 1) {
2172                 goto fail;
2173         }
2174         for (zap_cursor_init(&zc, mos, ds->ds_phys->ds_next_clones_obj);
2175             zap_cursor_retrieve(&zc, &za) == 0;
2176             zap_cursor_advance(&zc)) {
2177                 dsl_dataset_t *clone;
2178                 char buf[ZFS_MAXNAMELEN];
2179                 if (dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
2180                     za.za_first_integer, FTAG, &clone) != 0) {
2181                         goto fail;
2182                 }
2183                 dsl_dir_name(clone->ds_dir, buf);
2184                 VERIFY(nvlist_add_boolean(val, buf) == 0);
2185                 dsl_dataset_rele(clone, FTAG);
2186         }
2187         zap_cursor_fini(&zc);
2188         VERIFY(nvlist_add_nvlist(propval, ZPROP_VALUE, val) == 0);
2189         VERIFY(nvlist_add_nvlist(nv, zfs_prop_to_name(ZFS_PROP_CLONES),
2190             propval) == 0);
2191 fail:
2192         nvlist_free(val);
2193         nvlist_free(propval);
2194         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2195 }
2196
2197 void
2198 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2199 {
2200         uint64_t refd, avail, uobjs, aobjs, ratio;
2201
2202         dsl_dir_stats(ds->ds_dir, nv);
2203
2204         dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2205         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2206         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2207
2208         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2209             ds->ds_phys->ds_creation_time);
2210         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2211             ds->ds_phys->ds_creation_txg);
2212         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2213             ds->ds_quota);
2214         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2215             ds->ds_reserved);
2216         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2217             ds->ds_phys->ds_guid);
2218         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_UNIQUE,
2219             ds->ds_phys->ds_unique_bytes);
2220         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_OBJSETID,
2221             ds->ds_object);
2222         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS,
2223             ds->ds_userrefs);
2224         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2225             DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2226
2227         if (ds->ds_phys->ds_prev_snap_obj != 0) {
2228                 uint64_t written, comp, uncomp;
2229                 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2230                 dsl_dataset_t *prev;
2231                 int err;
2232
2233                 rw_enter(&dp->dp_config_rwlock, RW_READER);
2234                 err = dsl_dataset_hold_obj(dp,
2235                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
2236                 rw_exit(&dp->dp_config_rwlock);
2237                 if (err == 0) {
2238                         err = dsl_dataset_space_written(prev, ds, &written,
2239                             &comp, &uncomp);
2240                         dsl_dataset_rele(prev, FTAG);
2241                         if (err == 0) {
2242                                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_WRITTEN,
2243                                     written);
2244                         }
2245                 }
2246         }
2247
2248         ratio = ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2249             (ds->ds_phys->ds_uncompressed_bytes * 100 /
2250             ds->ds_phys->ds_compressed_bytes);
2251         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRATIO, ratio);
2252
2253         if (ds->ds_phys->ds_next_snap_obj) {
2254                 /*
2255                  * This is a snapshot; override the dd's space used with
2256                  * our unique space and compression ratio.
2257                  */
2258                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2259                     ds->ds_phys->ds_unique_bytes);
2260                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO, ratio);
2261
2262                 get_clones_stat(ds, nv);
2263         }
2264 }
2265
2266 void
2267 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2268 {
2269         stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2270         stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2271         stat->dds_guid = ds->ds_phys->ds_guid;
2272         if (ds->ds_phys->ds_next_snap_obj) {
2273                 stat->dds_is_snapshot = B_TRUE;
2274                 stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2275         } else {
2276                 stat->dds_is_snapshot = B_FALSE;
2277                 stat->dds_num_clones = 0;
2278         }
2279
2280         /* clone origin is really a dsl_dir thing... */
2281         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2282         if (dsl_dir_is_clone(ds->ds_dir)) {
2283                 dsl_dataset_t *ods;
2284
2285                 VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2286                     ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2287                 dsl_dataset_name(ods, stat->dds_origin);
2288                 dsl_dataset_drop_ref(ods, FTAG);
2289         } else {
2290                 stat->dds_origin[0] = '\0';
2291         }
2292         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2293 }
2294
2295 uint64_t
2296 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2297 {
2298         return (ds->ds_fsid_guid);
2299 }
2300
2301 void
2302 dsl_dataset_space(dsl_dataset_t *ds,
2303     uint64_t *refdbytesp, uint64_t *availbytesp,
2304     uint64_t *usedobjsp, uint64_t *availobjsp)
2305 {
2306         *refdbytesp = ds->ds_phys->ds_used_bytes;
2307         *availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2308         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2309                 *availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2310         if (ds->ds_quota != 0) {
2311                 /*
2312                  * Adjust available bytes according to refquota
2313                  */
2314                 if (*refdbytesp < ds->ds_quota)
2315                         *availbytesp = MIN(*availbytesp,
2316                             ds->ds_quota - *refdbytesp);
2317                 else
2318                         *availbytesp = 0;
2319         }
2320         *usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2321         *availobjsp = DN_MAX_OBJECT - *usedobjsp;
2322 }
2323
2324 boolean_t
2325 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2326 {
2327         ASSERTV(dsl_pool_t *dp = ds->ds_dir->dd_pool);
2328
2329         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2330             dsl_pool_sync_context(dp));
2331         if (ds->ds_prev == NULL)
2332                 return (B_FALSE);
2333         if (ds->ds_phys->ds_bp.blk_birth >
2334             ds->ds_prev->ds_phys->ds_creation_txg) {
2335                 objset_t *os, *os_prev;
2336                 /*
2337                  * It may be that only the ZIL differs, because it was
2338                  * reset in the head.  Don't count that as being
2339                  * modified.
2340                  */
2341                 if (dmu_objset_from_ds(ds, &os) != 0)
2342                         return (B_TRUE);
2343                 if (dmu_objset_from_ds(ds->ds_prev, &os_prev) != 0)
2344                         return (B_TRUE);
2345                 return (bcmp(&os->os_phys->os_meta_dnode,
2346                     &os_prev->os_phys->os_meta_dnode,
2347                     sizeof (os->os_phys->os_meta_dnode)) != 0);
2348         }
2349         return (B_FALSE);
2350 }
2351
2352 /* ARGSUSED */
2353 static int
2354 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2355 {
2356         dsl_dataset_t *ds = arg1;
2357         char *newsnapname = arg2;
2358         dsl_dir_t *dd = ds->ds_dir;
2359         dsl_dataset_t *hds;
2360         uint64_t val;
2361         int err;
2362
2363         err = dsl_dataset_hold_obj(dd->dd_pool,
2364             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2365         if (err)
2366                 return (err);
2367
2368         /* new name better not be in use */
2369         err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2370         dsl_dataset_rele(hds, FTAG);
2371
2372         if (err == 0)
2373                 err = EEXIST;
2374         else if (err == ENOENT)
2375                 err = 0;
2376
2377         /* dataset name + 1 for the "@" + the new snapshot name must fit */
2378         if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2379                 err = ENAMETOOLONG;
2380
2381         return (err);
2382 }
2383
2384 static void
2385 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2386 {
2387         dsl_dataset_t *ds = arg1;
2388         const char *newsnapname = arg2;
2389         dsl_dir_t *dd = ds->ds_dir;
2390         objset_t *mos = dd->dd_pool->dp_meta_objset;
2391         dsl_dataset_t *hds;
2392         int err;
2393
2394         ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2395
2396         VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2397             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2398
2399         VERIFY(0 == dsl_dataset_get_snapname(ds));
2400         err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2401         ASSERT3U(err, ==, 0);
2402         mutex_enter(&ds->ds_lock);
2403         (void) strcpy(ds->ds_snapname, newsnapname);
2404         mutex_exit(&ds->ds_lock);
2405         err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2406             ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2407         ASSERT3U(err, ==, 0);
2408
2409         spa_history_log_internal(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2410             "dataset = %llu", ds->ds_object);
2411         dsl_dataset_rele(hds, FTAG);
2412 }
2413
2414 struct renamesnaparg {
2415         dsl_sync_task_group_t *dstg;
2416         char failed[MAXPATHLEN];
2417         char *oldsnap;
2418         char *newsnap;
2419 };
2420
2421 static int
2422 dsl_snapshot_rename_one(const char *name, void *arg)
2423 {
2424         struct renamesnaparg *ra = arg;
2425         dsl_dataset_t *ds = NULL;
2426         char *snapname;
2427         int err;
2428
2429         snapname = kmem_asprintf("%s@%s", name, ra->oldsnap);
2430         (void) strlcpy(ra->failed, snapname, sizeof (ra->failed));
2431
2432         /*
2433          * For recursive snapshot renames the parent won't be changing
2434          * so we just pass name for both the to/from argument.
2435          */
2436         err = zfs_secpolicy_rename_perms(snapname, snapname, CRED());
2437         if (err != 0) {
2438                 strfree(snapname);
2439                 return (err == ENOENT ? 0 : err);
2440         }
2441
2442 #ifdef _KERNEL
2443         /*
2444          * For all filesystems undergoing rename, we'll need to unmount it.
2445          */
2446         (void) zfs_unmount_snap(snapname, NULL);
2447 #endif
2448         err = dsl_dataset_hold(snapname, ra->dstg, &ds);
2449         strfree(snapname);
2450         if (err != 0)
2451                 return (err == ENOENT ? 0 : err);
2452
2453         dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2454             dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2455
2456         return (0);
2457 }
2458
2459 static int
2460 dsl_recursive_rename(char *oldname, const char *newname)
2461 {
2462         int err;
2463         struct renamesnaparg *ra;
2464         dsl_sync_task_t *dst;
2465         spa_t *spa;
2466         char *cp, *fsname = spa_strdup(oldname);
2467         int len = strlen(oldname) + 1;
2468
2469         /* truncate the snapshot name to get the fsname */
2470         cp = strchr(fsname, '@');
2471         *cp = '\0';
2472
2473         err = spa_open(fsname, &spa, FTAG);
2474         if (err) {
2475                 kmem_free(fsname, len);
2476                 return (err);
2477         }
2478         ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2479         ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2480
2481         ra->oldsnap = strchr(oldname, '@') + 1;
2482         ra->newsnap = strchr(newname, '@') + 1;
2483         *ra->failed = '\0';
2484
2485         err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2486             DS_FIND_CHILDREN);
2487         kmem_free(fsname, len);
2488
2489         if (err == 0) {
2490                 err = dsl_sync_task_group_wait(ra->dstg);
2491         }
2492
2493         for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2494             dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2495                 dsl_dataset_t *ds = dst->dst_arg1;
2496                 if (dst->dst_err) {
2497                         dsl_dir_name(ds->ds_dir, ra->failed);
2498                         (void) strlcat(ra->failed, "@", sizeof (ra->failed));
2499                         (void) strlcat(ra->failed, ra->newsnap,
2500                             sizeof (ra->failed));
2501                 }
2502                 dsl_dataset_rele(ds, ra->dstg);
2503         }
2504
2505         if (err)
2506                 (void) strlcpy(oldname, ra->failed, sizeof (ra->failed));
2507
2508         dsl_sync_task_group_destroy(ra->dstg);
2509         kmem_free(ra, sizeof (struct renamesnaparg));
2510         spa_close(spa, FTAG);
2511         return (err);
2512 }
2513
2514 static int
2515 dsl_valid_rename(const char *oldname, void *arg)
2516 {
2517         int delta = *(int *)arg;
2518
2519         if (strlen(oldname) + delta >= MAXNAMELEN)
2520                 return (ENAMETOOLONG);
2521
2522         return (0);
2523 }
2524
2525 #pragma weak dmu_objset_rename = dsl_dataset_rename
2526 int
2527 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2528 {
2529         dsl_dir_t *dd;
2530         dsl_dataset_t *ds;
2531         const char *tail;
2532         int err;
2533
2534         err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2535         if (err)
2536                 return (err);
2537
2538         if (tail == NULL) {
2539                 int delta = strlen(newname) - strlen(oldname);
2540
2541                 /* if we're growing, validate child name lengths */
2542                 if (delta > 0)
2543                         err = dmu_objset_find(oldname, dsl_valid_rename,
2544                             &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2545
2546                 if (err == 0)
2547                         err = dsl_dir_rename(dd, newname);
2548                 dsl_dir_close(dd, FTAG);
2549                 return (err);
2550         }
2551
2552         if (tail[0] != '@') {
2553                 /* the name ended in a nonexistent component */
2554                 dsl_dir_close(dd, FTAG);
2555                 return (ENOENT);
2556         }
2557
2558         dsl_dir_close(dd, FTAG);
2559
2560         /* new name must be snapshot in same filesystem */
2561         tail = strchr(newname, '@');
2562         if (tail == NULL)
2563                 return (EINVAL);
2564         tail++;
2565         if (strncmp(oldname, newname, tail - newname) != 0)
2566                 return (EXDEV);
2567
2568         if (recursive) {
2569                 err = dsl_recursive_rename(oldname, newname);
2570         } else {
2571                 err = dsl_dataset_hold(oldname, FTAG, &ds);
2572                 if (err)
2573                         return (err);
2574
2575                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2576                     dsl_dataset_snapshot_rename_check,
2577                     dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2578
2579                 dsl_dataset_rele(ds, FTAG);
2580         }
2581
2582         return (err);
2583 }
2584
2585 struct promotenode {
2586         list_node_t link;
2587         dsl_dataset_t *ds;
2588 };
2589
2590 struct promotearg {
2591         list_t shared_snaps, origin_snaps, clone_snaps;
2592         dsl_dataset_t *origin_origin;
2593         uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2594         char *err_ds;
2595 };
2596
2597 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2598
2599 static int
2600 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2601 {
2602         dsl_dataset_t *hds = arg1;
2603         struct promotearg *pa = arg2;
2604         struct promotenode *snap = list_head(&pa->shared_snaps);
2605         dsl_dataset_t *origin_ds = snap->ds;
2606         int err;
2607         uint64_t unused;
2608
2609         /* Check that it is a real clone */
2610         if (!dsl_dir_is_clone(hds->ds_dir))
2611                 return (EINVAL);
2612
2613         /* Since this is so expensive, don't do the preliminary check */
2614         if (!dmu_tx_is_syncing(tx))
2615                 return (0);
2616
2617         if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2618                 return (EXDEV);
2619
2620         /* compute origin's new unique space */
2621         snap = list_tail(&pa->clone_snaps);
2622         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2623         dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2624             origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
2625             &pa->unique, &unused, &unused);
2626
2627         /*
2628          * Walk the snapshots that we are moving
2629          *
2630          * Compute space to transfer.  Consider the incremental changes
2631          * to used for each snapshot:
2632          * (my used) = (prev's used) + (blocks born) - (blocks killed)
2633          * So each snapshot gave birth to:
2634          * (blocks born) = (my used) - (prev's used) + (blocks killed)
2635          * So a sequence would look like:
2636          * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2637          * Which simplifies to:
2638          * uN + kN + kN-1 + ... + k1 + k0
2639          * Note however, if we stop before we reach the ORIGIN we get:
2640          * uN + kN + kN-1 + ... + kM - uM-1
2641          */
2642         pa->used = origin_ds->ds_phys->ds_used_bytes;
2643         pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2644         pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2645         for (snap = list_head(&pa->shared_snaps); snap;
2646             snap = list_next(&pa->shared_snaps, snap)) {
2647                 uint64_t val, dlused, dlcomp, dluncomp;
2648                 dsl_dataset_t *ds = snap->ds;
2649
2650                 /* Check that the snapshot name does not conflict */
2651                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2652                 err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2653                 if (err == 0) {
2654                         err = EEXIST;
2655                         goto out;
2656                 }
2657                 if (err != ENOENT)
2658                         goto out;
2659
2660                 /* The very first snapshot does not have a deadlist */
2661                 if (ds->ds_phys->ds_prev_snap_obj == 0)
2662                         continue;
2663
2664                 dsl_deadlist_space(&ds->ds_deadlist,
2665                     &dlused, &dlcomp, &dluncomp);
2666                 pa->used += dlused;
2667                 pa->comp += dlcomp;
2668                 pa->uncomp += dluncomp;
2669         }
2670
2671         /*
2672          * If we are a clone of a clone then we never reached ORIGIN,
2673          * so we need to subtract out the clone origin's used space.
2674          */
2675         if (pa->origin_origin) {
2676                 pa->used -= pa->origin_origin->ds_phys->ds_used_bytes;
2677                 pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2678                 pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2679         }
2680
2681         /* Check that there is enough space here */
2682         err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2683             pa->used);
2684         if (err)
2685                 return (err);
2686
2687         /*
2688          * Compute the amounts of space that will be used by snapshots
2689          * after the promotion (for both origin and clone).  For each,
2690          * it is the amount of space that will be on all of their
2691          * deadlists (that was not born before their new origin).
2692          */
2693         if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2694                 uint64_t space;
2695
2696                 /*
2697                  * Note, typically this will not be a clone of a clone,
2698                  * so dd_origin_txg will be < TXG_INITIAL, so
2699                  * these snaplist_space() -> dsl_deadlist_space_range()
2700                  * calls will be fast because they do not have to
2701                  * iterate over all bps.
2702                  */
2703                 snap = list_head(&pa->origin_snaps);
2704                 err = snaplist_space(&pa->shared_snaps,
2705                     snap->ds->ds_dir->dd_origin_txg, &pa->cloneusedsnap);
2706                 if (err)
2707                         return (err);
2708
2709                 err = snaplist_space(&pa->clone_snaps,
2710                     snap->ds->ds_dir->dd_origin_txg, &space);
2711                 if (err)
2712                         return (err);
2713                 pa->cloneusedsnap += space;
2714         }
2715         if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2716                 err = snaplist_space(&pa->origin_snaps,
2717                     origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2718                 if (err)
2719                         return (err);
2720         }
2721
2722         return (0);
2723 out:
2724         pa->err_ds =  snap->ds->ds_snapname;
2725         return (err);
2726 }
2727
2728 static void
2729 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2730 {
2731         dsl_dataset_t *hds = arg1;
2732         struct promotearg *pa = arg2;
2733         struct promotenode *snap = list_head(&pa->shared_snaps);
2734         dsl_dataset_t *origin_ds = snap->ds;
2735         dsl_dataset_t *origin_head;
2736         dsl_dir_t *dd = hds->ds_dir;
2737         dsl_pool_t *dp = hds->ds_dir->dd_pool;
2738         dsl_dir_t *odd = NULL;
2739         uint64_t oldnext_obj;
2740         int64_t delta;
2741
2742         ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2743
2744         snap = list_head(&pa->origin_snaps);
2745         origin_head = snap->ds;
2746
2747         /*
2748          * We need to explicitly open odd, since origin_ds's dd will be
2749          * changing.
2750          */
2751         VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2752             NULL, FTAG, &odd));
2753
2754         /* change origin's next snap */
2755         dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2756         oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2757         snap = list_tail(&pa->clone_snaps);
2758         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2759         origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2760
2761         /* change the origin's next clone */
2762         if (origin_ds->ds_phys->ds_next_clones_obj) {
2763                 remove_from_next_clones(origin_ds, snap->ds->ds_object, tx);
2764                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2765                     origin_ds->ds_phys->ds_next_clones_obj,
2766                     oldnext_obj, tx));
2767         }
2768
2769         /* change origin */
2770         dmu_buf_will_dirty(dd->dd_dbuf, tx);
2771         ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2772         dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2773         dd->dd_origin_txg = origin_head->ds_dir->dd_origin_txg;
2774         dmu_buf_will_dirty(odd->dd_dbuf, tx);
2775         odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2776         origin_head->ds_dir->dd_origin_txg =
2777             origin_ds->ds_phys->ds_creation_txg;
2778
2779         /* change dd_clone entries */
2780         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2781                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2782                     odd->dd_phys->dd_clones, hds->ds_object, tx));
2783                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2784                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2785                     hds->ds_object, tx));
2786
2787                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2788                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2789                     origin_head->ds_object, tx));
2790                 if (dd->dd_phys->dd_clones == 0) {
2791                         dd->dd_phys->dd_clones = zap_create(dp->dp_meta_objset,
2792                             DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
2793                 }
2794                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2795                     dd->dd_phys->dd_clones, origin_head->ds_object, tx));
2796
2797         }
2798
2799         /* move snapshots to this dir */
2800         for (snap = list_head(&pa->shared_snaps); snap;
2801             snap = list_next(&pa->shared_snaps, snap)) {
2802                 dsl_dataset_t *ds = snap->ds;
2803
2804                 /* unregister props as dsl_dir is changing */
2805                 if (ds->ds_objset) {
2806                         dmu_objset_evict(ds->ds_objset);
2807                         ds->ds_objset = NULL;
2808                 }
2809                 /* move snap name entry */
2810                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2811                 VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2812                     ds->ds_snapname, tx));
2813                 VERIFY(0 == zap_add(dp->dp_meta_objset,
2814                     hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2815                     8, 1, &ds->ds_object, tx));
2816
2817                 /* change containing dsl_dir */
2818                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2819                 ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2820                 ds->ds_phys->ds_dir_obj = dd->dd_object;
2821                 ASSERT3P(ds->ds_dir, ==, odd);
2822                 dsl_dir_close(ds->ds_dir, ds);
2823                 VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2824                     NULL, ds, &ds->ds_dir));
2825
2826                 /* move any clone references */
2827                 if (ds->ds_phys->ds_next_clones_obj &&
2828                     spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2829                         zap_cursor_t zc;
2830                         zap_attribute_t za;
2831
2832                         for (zap_cursor_init(&zc, dp->dp_meta_objset,
2833                             ds->ds_phys->ds_next_clones_obj);
2834                             zap_cursor_retrieve(&zc, &za) == 0;
2835                             zap_cursor_advance(&zc)) {
2836                                 dsl_dataset_t *cnds;
2837                                 uint64_t o;
2838
2839                                 if (za.za_first_integer == oldnext_obj) {
2840                                         /*
2841                                          * We've already moved the
2842                                          * origin's reference.
2843                                          */
2844                                         continue;
2845                                 }
2846
2847                                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
2848                                     za.za_first_integer, FTAG, &cnds));
2849                                 o = cnds->ds_dir->dd_phys->dd_head_dataset_obj;
2850
2851                                 VERIFY3U(zap_remove_int(dp->dp_meta_objset,
2852                                     odd->dd_phys->dd_clones, o, tx), ==, 0);
2853                                 VERIFY3U(zap_add_int(dp->dp_meta_objset,
2854                                     dd->dd_phys->dd_clones, o, tx), ==, 0);
2855                                 dsl_dataset_rele(cnds, FTAG);
2856                         }
2857                         zap_cursor_fini(&zc);
2858                 }
2859
2860                 ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2861         }
2862
2863         /*
2864          * Change space accounting.
2865          * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2866          * both be valid, or both be 0 (resulting in delta == 0).  This
2867          * is true for each of {clone,origin} independently.
2868          */
2869
2870         delta = pa->cloneusedsnap -
2871             dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2872         ASSERT3S(delta, >=, 0);
2873         ASSERT3U(pa->used, >=, delta);
2874         dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2875         dsl_dir_diduse_space(dd, DD_USED_HEAD,
2876             pa->used - delta, pa->comp, pa->uncomp, tx);
2877
2878         delta = pa->originusedsnap -
2879             odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2880         ASSERT3S(delta, <=, 0);
2881         ASSERT3U(pa->used, >=, -delta);
2882         dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2883         dsl_dir_diduse_space(odd, DD_USED_HEAD,
2884             -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2885
2886         origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2887
2888         /* log history record */
2889         spa_history_log_internal(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2890             "dataset = %llu", hds->ds_object);
2891
2892         dsl_dir_close(odd, FTAG);
2893 }
2894
2895 static char *snaplist_tag = "snaplist";
2896 /*
2897  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2898  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2899  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2900  * snapshots back to this dataset's origin.
2901  */
2902 static int
2903 snaplist_make(dsl_pool_t *dp, boolean_t own,
2904     uint64_t first_obj, uint64_t last_obj, list_t *l)
2905 {
2906         uint64_t obj = last_obj;
2907
2908         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2909
2910         list_create(l, sizeof (struct promotenode),
2911             offsetof(struct promotenode, link));
2912
2913         while (obj != first_obj) {
2914                 dsl_dataset_t *ds;
2915                 struct promotenode *snap;
2916                 int err;
2917
2918                 if (own) {
2919                         err = dsl_dataset_own_obj(dp, obj,
2920                             0, snaplist_tag, &ds);
2921                         if (err == 0)
2922                                 dsl_dataset_make_exclusive(ds, snaplist_tag);
2923                 } else {
2924                         err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2925                 }
2926                 if (err == ENOENT) {
2927                         /* lost race with snapshot destroy */
2928                         struct promotenode *last = list_tail(l);
2929                         ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2930                         obj = last->ds->ds_phys->ds_prev_snap_obj;
2931                         continue;
2932                 } else if (err) {
2933                         return (err);
2934                 }
2935
2936                 if (first_obj == 0)
2937                         first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2938
2939                 snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
2940                 snap->ds = ds;
2941                 list_insert_tail(l, snap);
2942                 obj = ds->ds_phys->ds_prev_snap_obj;
2943         }
2944
2945         return (0);
2946 }
2947
2948 static int
2949 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
2950 {
2951         struct promotenode *snap;
2952
2953         *spacep = 0;
2954         for (snap = list_head(l); snap; snap = list_next(l, snap)) {
2955                 uint64_t used, comp, uncomp;
2956                 dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2957                     mintxg, UINT64_MAX, &used, &comp, &uncomp);
2958                 *spacep += used;
2959         }
2960         return (0);
2961 }
2962
2963 static void
2964 snaplist_destroy(list_t *l, boolean_t own)
2965 {
2966         struct promotenode *snap;
2967
2968         if (!l || !list_link_active(&l->list_head))
2969                 return;
2970
2971         while ((snap = list_tail(l)) != NULL) {
2972                 list_remove(l, snap);
2973                 if (own)
2974                         dsl_dataset_disown(snap->ds, snaplist_tag);
2975                 else
2976                         dsl_dataset_rele(snap->ds, snaplist_tag);
2977                 kmem_free(snap, sizeof (struct promotenode));
2978         }
2979         list_destroy(l);
2980 }
2981
2982 /*
2983  * Promote a clone.  Nomenclature note:
2984  * "clone" or "cds": the original clone which is being promoted
2985  * "origin" or "ods": the snapshot which is originally clone's origin
2986  * "origin head" or "ohds": the dataset which is the head
2987  * (filesystem/volume) for the origin
2988  * "origin origin": the origin of the origin's filesystem (typically
2989  * NULL, indicating that the clone is not a clone of a clone).
2990  */
2991 int
2992 dsl_dataset_promote(const char *name, char *conflsnap)
2993 {
2994         dsl_dataset_t *ds;
2995         dsl_dir_t *dd;
2996         dsl_pool_t *dp;
2997         dmu_object_info_t doi;
2998         struct promotearg pa;
2999         struct promotenode *snap;
3000         int err;
3001
3002         bzero(&pa, sizeof(struct promotearg));
3003         err = dsl_dataset_hold(name, FTAG, &ds);
3004         if (err)
3005                 return (err);
3006         dd = ds->ds_dir;
3007         dp = dd->dd_pool;
3008
3009         err = dmu_object_info(dp->dp_meta_objset,
3010             ds->ds_phys->ds_snapnames_zapobj, &doi);
3011         if (err) {
3012                 dsl_dataset_rele(ds, FTAG);
3013                 return (err);
3014         }
3015
3016         if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
3017                 dsl_dataset_rele(ds, FTAG);
3018                 return (EINVAL);
3019         }
3020
3021         /*
3022          * We are going to inherit all the snapshots taken before our
3023          * origin (i.e., our new origin will be our parent's origin).
3024          * Take ownership of them so that we can rename them into our
3025          * namespace.
3026          */
3027         rw_enter(&dp->dp_config_rwlock, RW_READER);
3028
3029         err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
3030             &pa.shared_snaps);
3031         if (err != 0)
3032                 goto out;
3033
3034         err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
3035         if (err != 0)
3036                 goto out;
3037
3038         snap = list_head(&pa.shared_snaps);
3039         ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
3040         err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
3041             snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
3042         if (err != 0)
3043                 goto out;
3044
3045         if (snap->ds->ds_dir->dd_phys->dd_origin_obj != 0) {
3046                 err = dsl_dataset_hold_obj(dp,
3047                     snap->ds->ds_dir->dd_phys->dd_origin_obj,
3048                     FTAG, &pa.origin_origin);
3049                 if (err != 0)
3050                         goto out;
3051         }
3052
3053 out:
3054         rw_exit(&dp->dp_config_rwlock);
3055
3056         /*
3057          * Add in 128x the snapnames zapobj size, since we will be moving
3058          * a bunch of snapnames to the promoted ds, and dirtying their
3059          * bonus buffers.
3060          */
3061         if (err == 0) {
3062                 err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
3063                     dsl_dataset_promote_sync, ds, &pa,
3064                     2 + 2 * doi.doi_physical_blocks_512);
3065                 if (err && pa.err_ds && conflsnap)
3066                         (void) strncpy(conflsnap, pa.err_ds, MAXNAMELEN);
3067         }
3068
3069         snaplist_destroy(&pa.shared_snaps, B_TRUE);
3070         snaplist_destroy(&pa.clone_snaps, B_FALSE);
3071         snaplist_destroy(&pa.origin_snaps, B_FALSE);
3072         if (pa.origin_origin)
3073                 dsl_dataset_rele(pa.origin_origin, FTAG);
3074         dsl_dataset_rele(ds, FTAG);
3075         return (err);
3076 }
3077
3078 struct cloneswaparg {
3079         dsl_dataset_t *cds; /* clone dataset */
3080         dsl_dataset_t *ohds; /* origin's head dataset */
3081         boolean_t force;
3082         int64_t unused_refres_delta; /* change in unconsumed refreservation */
3083 };
3084
3085 /* ARGSUSED */
3086 static int
3087 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
3088 {
3089         struct cloneswaparg *csa = arg1;
3090
3091         /* they should both be heads */
3092         if (dsl_dataset_is_snapshot(csa->cds) ||
3093             dsl_dataset_is_snapshot(csa->ohds))
3094                 return (EINVAL);
3095
3096         /* the branch point should be just before them */
3097         if (csa->cds->ds_prev != csa->ohds->ds_prev)
3098                 return (EINVAL);
3099
3100         /* cds should be the clone (unless they are unrelated) */
3101         if (csa->cds->ds_prev != NULL &&
3102             csa->cds->ds_prev != csa->cds->ds_dir->dd_pool->dp_origin_snap &&
3103             csa->ohds->ds_object !=
3104             csa->cds->ds_prev->ds_phys->ds_next_snap_obj)
3105                 return (EINVAL);
3106
3107         /* the clone should be a child of the origin */
3108         if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
3109                 return (EINVAL);
3110
3111         /* ohds shouldn't be modified unless 'force' */
3112         if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
3113                 return (ETXTBSY);
3114
3115         /* adjust amount of any unconsumed refreservation */
3116         csa->unused_refres_delta =
3117             (int64_t)MIN(csa->ohds->ds_reserved,
3118             csa->ohds->ds_phys->ds_unique_bytes) -
3119             (int64_t)MIN(csa->ohds->ds_reserved,
3120             csa->cds->ds_phys->ds_unique_bytes);
3121
3122         if (csa->unused_refres_delta > 0 &&
3123             csa->unused_refres_delta >
3124             dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3125                 return (ENOSPC);
3126
3127         if (csa->ohds->ds_quota != 0 &&
3128             csa->cds->ds_phys->ds_unique_bytes > csa->ohds->ds_quota)
3129                 return (EDQUOT);
3130
3131         return (0);
3132 }
3133
3134 /* ARGSUSED */
3135 static void
3136 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3137 {
3138         struct cloneswaparg *csa = arg1;
3139         dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3140
3141         ASSERT(csa->cds->ds_reserved == 0);
3142         ASSERT(csa->ohds->ds_quota == 0 ||
3143             csa->cds->ds_phys->ds_unique_bytes <= csa->ohds->ds_quota);
3144
3145         dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3146         dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3147
3148         if (csa->cds->ds_objset != NULL) {
3149                 dmu_objset_evict(csa->cds->ds_objset);
3150                 csa->cds->ds_objset = NULL;
3151         }
3152
3153         if (csa->ohds->ds_objset != NULL) {
3154                 dmu_objset_evict(csa->ohds->ds_objset);
3155                 csa->ohds->ds_objset = NULL;
3156         }
3157
3158         /*
3159          * Reset origin's unique bytes, if it exists.
3160          */
3161         if (csa->cds->ds_prev) {
3162                 dsl_dataset_t *origin = csa->cds->ds_prev;
3163                 uint64_t comp, uncomp;
3164
3165                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
3166                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3167                     origin->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3168                     &origin->ds_phys->ds_unique_bytes, &comp, &uncomp);
3169         }
3170
3171         /* swap blkptrs */
3172         {
3173                 blkptr_t tmp;
3174                 tmp = csa->ohds->ds_phys->ds_bp;
3175                 csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3176                 csa->cds->ds_phys->ds_bp = tmp;
3177         }
3178
3179         /* set dd_*_bytes */
3180         {
3181                 int64_t dused, dcomp, duncomp;
3182                 uint64_t cdl_used, cdl_comp, cdl_uncomp;
3183                 uint64_t odl_used, odl_comp, odl_uncomp;
3184
3185                 ASSERT3U(csa->cds->ds_dir->dd_phys->
3186                     dd_used_breakdown[DD_USED_SNAP], ==, 0);
3187
3188                 dsl_deadlist_space(&csa->cds->ds_deadlist,
3189                     &cdl_used, &cdl_comp, &cdl_uncomp);
3190                 dsl_deadlist_space(&csa->ohds->ds_deadlist,
3191                     &odl_used, &odl_comp, &odl_uncomp);
3192
3193                 dused = csa->cds->ds_phys->ds_used_bytes + cdl_used -
3194                     (csa->ohds->ds_phys->ds_used_bytes + odl_used);
3195                 dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3196                     (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3197                 duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3198                     cdl_uncomp -
3199                     (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3200
3201                 dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3202                     dused, dcomp, duncomp, tx);
3203                 dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3204                     -dused, -dcomp, -duncomp, tx);
3205
3206                 /*
3207                  * The difference in the space used by snapshots is the
3208                  * difference in snapshot space due to the head's
3209                  * deadlist (since that's the only thing that's
3210                  * changing that affects the snapused).
3211                  */
3212                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3213                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3214                     &cdl_used, &cdl_comp, &cdl_uncomp);
3215                 dsl_deadlist_space_range(&csa->ohds->ds_deadlist,
3216                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3217                     &odl_used, &odl_comp, &odl_uncomp);
3218                 dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3219                     DD_USED_HEAD, DD_USED_SNAP, tx);
3220         }
3221
3222         /* swap ds_*_bytes */
3223         SWITCH64(csa->ohds->ds_phys->ds_used_bytes,
3224             csa->cds->ds_phys->ds_used_bytes);
3225         SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3226             csa->cds->ds_phys->ds_compressed_bytes);
3227         SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3228             csa->cds->ds_phys->ds_uncompressed_bytes);
3229         SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3230             csa->cds->ds_phys->ds_unique_bytes);
3231
3232         /* apply any parent delta for change in unconsumed refreservation */
3233         dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3234             csa->unused_refres_delta, 0, 0, tx);
3235
3236         /*
3237          * Swap deadlists.
3238          */
3239         dsl_deadlist_close(&csa->cds->ds_deadlist);
3240         dsl_deadlist_close(&csa->ohds->ds_deadlist);
3241         SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3242             csa->cds->ds_phys->ds_deadlist_obj);
3243         dsl_deadlist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3244             csa->cds->ds_phys->ds_deadlist_obj);
3245         dsl_deadlist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3246             csa->ohds->ds_phys->ds_deadlist_obj);
3247
3248         dsl_scan_ds_clone_swapped(csa->ohds, csa->cds, tx);
3249 }
3250
3251 /*
3252  * Swap 'clone' with its origin head datasets.  Used at the end of "zfs
3253  * recv" into an existing fs to swizzle the file system to the new
3254  * version, and by "zfs rollback".  Can also be used to swap two
3255  * independent head datasets if neither has any snapshots.
3256  */
3257 int
3258 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3259     boolean_t force)
3260 {
3261         struct cloneswaparg csa;
3262         int error;
3263
3264         ASSERT(clone->ds_owner);
3265         ASSERT(origin_head->ds_owner);
3266 retry:
3267         /*
3268          * Need exclusive access for the swap. If we're swapping these
3269          * datasets back after an error, we already hold the locks.
3270          */
3271         if (!RW_WRITE_HELD(&clone->ds_rwlock))
3272                 rw_enter(&clone->ds_rwlock, RW_WRITER);
3273         if (!RW_WRITE_HELD(&origin_head->ds_rwlock) &&
3274             !rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3275                 rw_exit(&clone->ds_rwlock);
3276                 rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3277                 if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3278                         rw_exit(&origin_head->ds_rwlock);
3279                         goto retry;
3280                 }
3281         }
3282         csa.cds = clone;
3283         csa.ohds = origin_head;
3284         csa.force = force;
3285         error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3286             dsl_dataset_clone_swap_check,
3287             dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3288         return (error);
3289 }
3290
3291 /*
3292  * Given a pool name and a dataset object number in that pool,
3293  * return the name of that dataset.
3294  */
3295 int
3296 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3297 {
3298         spa_t *spa;
3299         dsl_pool_t *dp;
3300         dsl_dataset_t *ds;
3301         int error;
3302
3303         if ((error = spa_open(pname, &spa, FTAG)) != 0)
3304                 return (error);
3305         dp = spa_get_dsl(spa);
3306         rw_enter(&dp->dp_config_rwlock, RW_READER);
3307         if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3308                 dsl_dataset_name(ds, buf);
3309                 dsl_dataset_rele(ds, FTAG);
3310         }
3311         rw_exit(&dp->dp_config_rwlock);
3312         spa_close(spa, FTAG);
3313
3314         return (error);
3315 }
3316
3317 int
3318 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3319     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3320 {
3321         int error = 0;
3322
3323         ASSERT3S(asize, >, 0);
3324
3325         /*
3326          * *ref_rsrv is the portion of asize that will come from any
3327          * unconsumed refreservation space.
3328          */
3329         *ref_rsrv = 0;
3330
3331         mutex_enter(&ds->ds_lock);
3332         /*
3333          * Make a space adjustment for reserved bytes.
3334          */
3335         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3336                 ASSERT3U(*used, >=,
3337                     ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3338                 *used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3339                 *ref_rsrv =
3340                     asize - MIN(asize, parent_delta(ds, asize + inflight));
3341         }
3342
3343         if (!check_quota || ds->ds_quota == 0) {
3344                 mutex_exit(&ds->ds_lock);
3345                 return (0);
3346         }
3347         /*
3348          * If they are requesting more space, and our current estimate
3349          * is over quota, they get to try again unless the actual
3350          * on-disk is over quota and there are no pending changes (which
3351          * may free up space for us).
3352          */
3353         if (ds->ds_phys->ds_used_bytes + inflight >= ds->ds_quota) {
3354                 if (inflight > 0 || ds->ds_phys->ds_used_bytes < ds->ds_quota)
3355                         error = ERESTART;
3356                 else
3357                         error = EDQUOT;
3358
3359                 DMU_TX_STAT_BUMP(dmu_tx_quota);
3360         }
3361         mutex_exit(&ds->ds_lock);
3362
3363         return (error);
3364 }
3365
3366 /* ARGSUSED */
3367 static int
3368 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3369 {
3370         dsl_dataset_t *ds = arg1;
3371         dsl_prop_setarg_t *psa = arg2;
3372         int err;
3373
3374         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3375                 return (ENOTSUP);
3376
3377         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3378                 return (err);
3379
3380         if (psa->psa_effective_value == 0)
3381                 return (0);
3382
3383         if (psa->psa_effective_value < ds->ds_phys->ds_used_bytes ||
3384             psa->psa_effective_value < ds->ds_reserved)
3385                 return (ENOSPC);
3386
3387         return (0);
3388 }
3389
3390 extern void dsl_prop_set_sync(void *, void *, dmu_tx_t *);
3391
3392 void
3393 dsl_dataset_set_quota_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3394 {
3395         dsl_dataset_t *ds = arg1;
3396         dsl_prop_setarg_t *psa = arg2;
3397         uint64_t effective_value = psa->psa_effective_value;
3398
3399         dsl_prop_set_sync(ds, psa, tx);
3400         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3401
3402         if (ds->ds_quota != effective_value) {
3403                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3404                 ds->ds_quota = effective_value;
3405
3406                 spa_history_log_internal(LOG_DS_REFQUOTA,
3407                     ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu ",
3408                     (longlong_t)ds->ds_quota, ds->ds_object);
3409         }
3410 }
3411
3412 int
3413 dsl_dataset_set_quota(const char *dsname, zprop_source_t source, uint64_t quota)
3414 {
3415         dsl_dataset_t *ds;
3416         dsl_prop_setarg_t psa;
3417         int err;
3418
3419         dsl_prop_setarg_init_uint64(&psa, "refquota", source, &quota);
3420
3421         err = dsl_dataset_hold(dsname, FTAG, &ds);
3422         if (err)
3423                 return (err);
3424
3425         /*
3426          * If someone removes a file, then tries to set the quota, we
3427          * want to make sure the file freeing takes effect.
3428          */
3429         txg_wait_open(ds->ds_dir->dd_pool, 0);
3430
3431         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3432             dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3433             ds, &psa, 0);
3434
3435         dsl_dataset_rele(ds, FTAG);
3436         return (err);
3437 }
3438
3439 static int
3440 dsl_dataset_set_reservation_check(void *arg1, void *arg2, dmu_tx_t *tx)
3441 {
3442         dsl_dataset_t *ds = arg1;
3443         dsl_prop_setarg_t *psa = arg2;
3444         uint64_t effective_value;
3445         uint64_t unique;
3446         int err;
3447
3448         if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3449             SPA_VERSION_REFRESERVATION)
3450                 return (ENOTSUP);
3451
3452         if (dsl_dataset_is_snapshot(ds))
3453                 return (EINVAL);
3454
3455         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3456                 return (err);
3457
3458         effective_value = psa->psa_effective_value;
3459
3460         /*
3461          * If we are doing the preliminary check in open context, the
3462          * space estimates may be inaccurate.
3463          */
3464         if (!dmu_tx_is_syncing(tx))
3465                 return (0);
3466
3467         mutex_enter(&ds->ds_lock);
3468         if (!DS_UNIQUE_IS_ACCURATE(ds))
3469                 dsl_dataset_recalc_head_uniq(ds);
3470         unique = ds->ds_phys->ds_unique_bytes;
3471         mutex_exit(&ds->ds_lock);
3472
3473         if (MAX(unique, effective_value) > MAX(unique, ds->ds_reserved)) {
3474                 uint64_t delta = MAX(unique, effective_value) -
3475                     MAX(unique, ds->ds_reserved);
3476
3477                 if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3478                         return (ENOSPC);
3479                 if (ds->ds_quota > 0 &&
3480                     effective_value > ds->ds_quota)
3481                         return (ENOSPC);
3482         }
3483
3484         return (0);
3485 }
3486
3487 static void
3488 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3489 {
3490         dsl_dataset_t *ds = arg1;
3491         dsl_prop_setarg_t *psa = arg2;
3492         uint64_t effective_value = psa->psa_effective_value;
3493         uint64_t unique;
3494         int64_t delta;
3495
3496         dsl_prop_set_sync(ds, psa, tx);
3497         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3498
3499         dmu_buf_will_dirty(ds->ds_dbuf, tx);
3500
3501         mutex_enter(&ds->ds_dir->dd_lock);
3502         mutex_enter(&ds->ds_lock);
3503         ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
3504         unique = ds->ds_phys->ds_unique_bytes;
3505         delta = MAX(0, (int64_t)(effective_value - unique)) -
3506             MAX(0, (int64_t)(ds->ds_reserved - unique));
3507         ds->ds_reserved = effective_value;
3508         mutex_exit(&ds->ds_lock);
3509
3510         dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3511         mutex_exit(&ds->ds_dir->dd_lock);
3512
3513         spa_history_log_internal(LOG_DS_REFRESERV,
3514             ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu",
3515             (longlong_t)effective_value, ds->ds_object);
3516 }
3517
3518 int
3519 dsl_dataset_set_reservation(const char *dsname, zprop_source_t source,
3520     uint64_t reservation)
3521 {
3522         dsl_dataset_t *ds;
3523         dsl_prop_setarg_t psa;
3524         int err;
3525
3526         dsl_prop_setarg_init_uint64(&psa, "refreservation", source,
3527             &reservation);
3528
3529         err = dsl_dataset_hold(dsname, FTAG, &ds);
3530         if (err)
3531                 return (err);
3532
3533         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3534             dsl_dataset_set_reservation_check,
3535             dsl_dataset_set_reservation_sync, ds, &psa, 0);
3536
3537         dsl_dataset_rele(ds, FTAG);
3538         return (err);
3539 }
3540
3541 typedef struct zfs_hold_cleanup_arg {
3542         dsl_pool_t *dp;
3543         uint64_t dsobj;
3544         char htag[MAXNAMELEN];
3545 } zfs_hold_cleanup_arg_t;
3546
3547 static void
3548 dsl_dataset_user_release_onexit(void *arg)
3549 {
3550         zfs_hold_cleanup_arg_t *ca = arg;
3551
3552         (void) dsl_dataset_user_release_tmp(ca->dp, ca->dsobj, ca->htag,
3553             B_TRUE);
3554         kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t));
3555 }
3556
3557 void
3558 dsl_register_onexit_hold_cleanup(dsl_dataset_t *ds, const char *htag,
3559     minor_t minor)
3560 {
3561         zfs_hold_cleanup_arg_t *ca;
3562
3563         ca = kmem_alloc(sizeof (zfs_hold_cleanup_arg_t), KM_SLEEP);
3564         ca->dp = ds->ds_dir->dd_pool;
3565         ca->dsobj = ds->ds_object;
3566         (void) strlcpy(ca->htag, htag, sizeof (ca->htag));
3567         VERIFY3U(0, ==, zfs_onexit_add_cb(minor,
3568             dsl_dataset_user_release_onexit, ca, NULL));
3569 }
3570
3571 /*
3572  * If you add new checks here, you may need to add
3573  * additional checks to the "temporary" case in
3574  * snapshot_check() in dmu_objset.c.
3575  */
3576 static int
3577 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3578 {
3579         dsl_dataset_t *ds = arg1;
3580         struct dsl_ds_holdarg *ha = arg2;
3581         char *htag = ha->htag;
3582         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3583         int error = 0;
3584
3585         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3586                 return (ENOTSUP);
3587
3588         if (!dsl_dataset_is_snapshot(ds))
3589                 return (EINVAL);
3590
3591         /* tags must be unique */
3592         mutex_enter(&ds->ds_lock);
3593         if (ds->ds_phys->ds_userrefs_obj) {
3594                 error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3595                     8, 1, tx);
3596                 if (error == 0)
3597                         error = EEXIST;
3598                 else if (error == ENOENT)
3599                         error = 0;
3600         }
3601         mutex_exit(&ds->ds_lock);
3602
3603         if (error == 0 && ha->temphold &&
3604             strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
3605                 error = E2BIG;
3606
3607         return (error);
3608 }
3609
3610 void
3611 dsl_dataset_user_hold_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3612 {
3613         dsl_dataset_t *ds = arg1;
3614         struct dsl_ds_holdarg *ha = arg2;
3615         char *htag = ha->htag;
3616         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3617         objset_t *mos = dp->dp_meta_objset;
3618         uint64_t now = gethrestime_sec();
3619         uint64_t zapobj;
3620
3621         mutex_enter(&ds->ds_lock);
3622         if (ds->ds_phys->ds_userrefs_obj == 0) {
3623                 /*
3624                  * This is the first user hold for this dataset.  Create
3625                  * the userrefs zap object.
3626                  */
3627                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3628                 zapobj = ds->ds_phys->ds_userrefs_obj =
3629                     zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3630         } else {
3631                 zapobj = ds->ds_phys->ds_userrefs_obj;
3632         }
3633         ds->ds_userrefs++;
3634         mutex_exit(&ds->ds_lock);
3635
3636         VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3637
3638         if (ha->temphold) {
3639                 VERIFY(0 == dsl_pool_user_hold(dp, ds->ds_object,
3640                     htag, &now, tx));
3641         }
3642
3643         spa_history_log_internal(LOG_DS_USER_HOLD,
3644             dp->dp_spa, tx, "<%s> temp = %d dataset = %llu", htag,
3645             (int)ha->temphold, ds->ds_object);
3646 }
3647
3648 static int
3649 dsl_dataset_user_hold_one(const char *dsname, void *arg)
3650 {
3651         struct dsl_ds_holdarg *ha = arg;
3652         dsl_dataset_t *ds;
3653         int error;
3654         char *name;
3655
3656         /* alloc a buffer to hold dsname@snapname plus terminating NULL */
3657         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3658         error = dsl_dataset_hold(name, ha->dstg, &ds);
3659         strfree(name);
3660         if (error == 0) {
3661                 ha->gotone = B_TRUE;
3662                 dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3663                     dsl_dataset_user_hold_sync, ds, ha, 0);
3664         } else if (error == ENOENT && ha->recursive) {
3665                 error = 0;
3666         } else {
3667                 (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3668         }
3669         return (error);
3670 }
3671
3672 int
3673 dsl_dataset_user_hold_for_send(dsl_dataset_t *ds, char *htag,
3674     boolean_t temphold)
3675 {
3676         struct dsl_ds_holdarg *ha;
3677         int error;
3678
3679         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3680         ha->htag = htag;
3681         ha->temphold = temphold;
3682         error = dsl_sync_task_do(ds->ds_dir->dd_pool,
3683             dsl_dataset_user_hold_check, dsl_dataset_user_hold_sync,
3684             ds, ha, 0);
3685         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3686
3687         return (error);
3688 }
3689
3690 int
3691 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3692     boolean_t recursive, boolean_t temphold, int cleanup_fd)
3693 {
3694         struct dsl_ds_holdarg *ha;
3695         dsl_sync_task_t *dst;
3696         spa_t *spa;
3697         int error;
3698         minor_t minor = 0;
3699
3700         if (cleanup_fd != -1) {
3701                 /* Currently we only support cleanup-on-exit of tempholds. */
3702                 if (!temphold)
3703                         return (EINVAL);
3704                 error = zfs_onexit_fd_hold(cleanup_fd, &minor);
3705                 if (error)
3706                         return (error);
3707         }
3708
3709         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3710
3711         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3712
3713         error = spa_open(dsname, &spa, FTAG);
3714         if (error) {
3715                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3716                 if (cleanup_fd != -1)
3717                         zfs_onexit_fd_rele(cleanup_fd);
3718                 return (error);
3719         }
3720
3721         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3722         ha->htag = htag;
3723         ha->snapname = snapname;
3724         ha->recursive = recursive;
3725         ha->temphold = temphold;
3726
3727         if (recursive) {
3728                 error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3729                     ha, DS_FIND_CHILDREN);
3730         } else {
3731                 error = dsl_dataset_user_hold_one(dsname, ha);
3732         }
3733         if (error == 0)
3734                 error = dsl_sync_task_group_wait(ha->dstg);
3735
3736         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3737             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3738                 dsl_dataset_t *ds = dst->dst_arg1;
3739
3740                 if (dst->dst_err) {
3741                         dsl_dataset_name(ds, ha->failed);
3742                         *strchr(ha->failed, '@') = '\0';
3743                 } else if (error == 0 && minor != 0 && temphold) {
3744                         /*
3745                          * If this hold is to be released upon process exit,
3746                          * register that action now.
3747                          */
3748                         dsl_register_onexit_hold_cleanup(ds, htag, minor);
3749                 }
3750                 dsl_dataset_rele(ds, ha->dstg);
3751         }
3752
3753         if (error == 0 && recursive && !ha->gotone)
3754                 error = ENOENT;
3755
3756         if (error)
3757                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3758
3759         dsl_sync_task_group_destroy(ha->dstg);
3760
3761         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3762         spa_close(spa, FTAG);
3763         if (cleanup_fd != -1)
3764                 zfs_onexit_fd_rele(cleanup_fd);
3765         return (error);
3766 }
3767
3768 struct dsl_ds_releasearg {
3769         dsl_dataset_t *ds;
3770         const char *htag;
3771         boolean_t own;          /* do we own or just hold ds? */
3772 };
3773
3774 static int
3775 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3776     boolean_t *might_destroy)
3777 {
3778         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3779         uint64_t zapobj;
3780         uint64_t tmp;
3781         int error;
3782
3783         *might_destroy = B_FALSE;
3784
3785         mutex_enter(&ds->ds_lock);
3786         zapobj = ds->ds_phys->ds_userrefs_obj;
3787         if (zapobj == 0) {
3788                 /* The tag can't possibly exist */
3789                 mutex_exit(&ds->ds_lock);
3790                 return (ESRCH);
3791         }
3792
3793         /* Make sure the tag exists */
3794         error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3795         if (error) {
3796                 mutex_exit(&ds->ds_lock);
3797                 if (error == ENOENT)
3798                         error = ESRCH;
3799                 return (error);
3800         }
3801
3802         if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3803             DS_IS_DEFER_DESTROY(ds))
3804                 *might_destroy = B_TRUE;
3805
3806         mutex_exit(&ds->ds_lock);
3807         return (0);
3808 }
3809
3810 static int
3811 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3812 {
3813         struct dsl_ds_releasearg *ra = arg1;
3814         dsl_dataset_t *ds = ra->ds;
3815         boolean_t might_destroy;
3816         int error;
3817
3818         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3819                 return (ENOTSUP);
3820
3821         error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3822         if (error)
3823                 return (error);
3824
3825         if (might_destroy) {
3826                 struct dsl_ds_destroyarg dsda = {0};
3827
3828                 if (dmu_tx_is_syncing(tx)) {
3829                         /*
3830                          * If we're not prepared to remove the snapshot,
3831                          * we can't allow the release to happen right now.
3832                          */
3833                         if (!ra->own)
3834                                 return (EBUSY);
3835                 }
3836                 dsda.ds = ds;
3837                 dsda.releasing = B_TRUE;
3838                 return (dsl_dataset_destroy_check(&dsda, tag, tx));
3839         }
3840
3841         return (0);
3842 }
3843
3844 static void
3845 dsl_dataset_user_release_sync(void *arg1, void *tag, dmu_tx_t *tx)
3846 {
3847         struct dsl_ds_releasearg *ra = arg1;
3848         dsl_dataset_t *ds = ra->ds;
3849         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3850         objset_t *mos = dp->dp_meta_objset;
3851         uint64_t zapobj;
3852         uint64_t dsobj = ds->ds_object;
3853         uint64_t refs;
3854         int error;
3855
3856         mutex_enter(&ds->ds_lock);
3857         ds->ds_userrefs--;
3858         refs = ds->ds_userrefs;
3859         mutex_exit(&ds->ds_lock);
3860         error = dsl_pool_user_release(dp, ds->ds_object, ra->htag, tx);
3861         VERIFY(error == 0 || error == ENOENT);
3862         zapobj = ds->ds_phys->ds_userrefs_obj;
3863         VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3864
3865         spa_history_log_internal(LOG_DS_USER_RELEASE,
3866             dp->dp_spa, tx, "<%s> %lld dataset = %llu",
3867             ra->htag, (longlong_t)refs, dsobj);
3868
3869         if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3870             DS_IS_DEFER_DESTROY(ds)) {
3871                 struct dsl_ds_destroyarg dsda = {0};
3872
3873                 ASSERT(ra->own);
3874                 dsda.ds = ds;
3875                 dsda.releasing = B_TRUE;
3876                 /* We already did the destroy_check */
3877                 dsl_dataset_destroy_sync(&dsda, tag, tx);
3878         }
3879 }
3880
3881 static int
3882 dsl_dataset_user_release_one(const char *dsname, void *arg)
3883 {
3884         struct dsl_ds_holdarg *ha = arg;
3885         struct dsl_ds_releasearg *ra;
3886         dsl_dataset_t *ds;
3887         int error;
3888         void *dtag = ha->dstg;
3889         char *name;
3890         boolean_t own = B_FALSE;
3891         boolean_t might_destroy;
3892
3893         /* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3894         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3895         error = dsl_dataset_hold(name, dtag, &ds);
3896         strfree(name);
3897         if (error == ENOENT && ha->recursive)
3898                 return (0);
3899         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3900         if (error)
3901                 return (error);
3902
3903         ha->gotone = B_TRUE;
3904
3905         ASSERT(dsl_dataset_is_snapshot(ds));
3906
3907         error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3908         if (error) {
3909                 dsl_dataset_rele(ds, dtag);
3910                 return (error);
3911         }
3912
3913         if (might_destroy) {
3914 #ifdef _KERNEL
3915                 name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3916                 error = zfs_unmount_snap(name, NULL);
3917                 strfree(name);
3918                 if (error) {
3919                         dsl_dataset_rele(ds, dtag);
3920                         return (error);
3921                 }
3922 #endif
3923                 if (!dsl_dataset_tryown(ds, B_TRUE, dtag)) {
3924                         dsl_dataset_rele(ds, dtag);
3925                         return (EBUSY);
3926                 } else {
3927                         own = B_TRUE;
3928                         dsl_dataset_make_exclusive(ds, dtag);
3929                 }
3930         }
3931
3932         ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3933         ra->ds = ds;
3934         ra->htag = ha->htag;
3935         ra->own = own;
3936         dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3937             dsl_dataset_user_release_sync, ra, dtag, 0);
3938
3939         return (0);
3940 }
3941
3942 int
3943 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3944     boolean_t recursive)
3945 {
3946         struct dsl_ds_holdarg *ha;
3947         dsl_sync_task_t *dst;
3948         spa_t *spa;
3949         int error;
3950
3951 top:
3952         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3953
3954         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3955
3956         error = spa_open(dsname, &spa, FTAG);
3957         if (error) {
3958                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3959                 return (error);
3960         }
3961
3962         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3963         ha->htag = htag;
3964         ha->snapname = snapname;
3965         ha->recursive = recursive;
3966         if (recursive) {
3967                 error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
3968                     ha, DS_FIND_CHILDREN);
3969         } else {
3970                 error = dsl_dataset_user_release_one(dsname, ha);
3971         }
3972         if (error == 0)
3973                 error = dsl_sync_task_group_wait(ha->dstg);
3974
3975         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3976             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3977                 struct dsl_ds_releasearg *ra = dst->dst_arg1;
3978                 dsl_dataset_t *ds = ra->ds;
3979
3980                 if (dst->dst_err)
3981                         dsl_dataset_name(ds, ha->failed);
3982
3983                 if (ra->own)
3984                         dsl_dataset_disown(ds, ha->dstg);
3985                 else
3986                         dsl_dataset_rele(ds, ha->dstg);
3987
3988                 kmem_free(ra, sizeof (struct dsl_ds_releasearg));
3989         }
3990
3991         if (error == 0 && recursive && !ha->gotone)
3992                 error = ENOENT;
3993
3994         if (error && error != EBUSY)
3995                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3996
3997         dsl_sync_task_group_destroy(ha->dstg);
3998         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3999         spa_close(spa, FTAG);
4000
4001         /*
4002          * We can get EBUSY if we were racing with deferred destroy and
4003          * dsl_dataset_user_release_check() hadn't done the necessary
4004          * open context setup.  We can also get EBUSY if we're racing
4005          * with destroy and that thread is the ds_owner.  Either way
4006          * the busy condition should be transient, and we should retry
4007          * the release operation.
4008          */
4009         if (error == EBUSY)
4010                 goto top;
4011
4012         return (error);
4013 }
4014
4015 /*
4016  * Called at spa_load time (with retry == B_FALSE) to release a stale
4017  * temporary user hold. Also called by the onexit code (with retry == B_TRUE).
4018  */
4019 int
4020 dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, char *htag,
4021     boolean_t retry)
4022 {
4023         dsl_dataset_t *ds;
4024         char *snap;
4025         char *name;
4026         int namelen;
4027         int error;
4028
4029         do {
4030                 rw_enter(&dp->dp_config_rwlock, RW_READER);
4031                 error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
4032                 rw_exit(&dp->dp_config_rwlock);
4033                 if (error)
4034                         return (error);
4035                 namelen = dsl_dataset_namelen(ds)+1;
4036                 name = kmem_alloc(namelen, KM_SLEEP);
4037                 dsl_dataset_name(ds, name);
4038                 dsl_dataset_rele(ds, FTAG);
4039
4040                 snap = strchr(name, '@');
4041                 *snap = '\0';
4042                 ++snap;
4043                 error = dsl_dataset_user_release(name, snap, htag, B_FALSE);
4044                 kmem_free(name, namelen);
4045
4046                 /*
4047                  * The object can't have been destroyed because we have a hold,
4048                  * but it might have been renamed, resulting in ENOENT.  Retry
4049                  * if we've been requested to do so.
4050                  *
4051                  * It would be nice if we could use the dsobj all the way
4052                  * through and avoid ENOENT entirely.  But we might need to
4053                  * unmount the snapshot, and there's currently no way to lookup
4054                  * a vfsp using a ZFS object id.
4055                  */
4056         } while ((error == ENOENT) && retry);
4057
4058         return (error);
4059 }
4060
4061 int
4062 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
4063 {
4064         dsl_dataset_t *ds;
4065         int err;
4066
4067         err = dsl_dataset_hold(dsname, FTAG, &ds);
4068         if (err)
4069                 return (err);
4070
4071         VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
4072         if (ds->ds_phys->ds_userrefs_obj != 0) {
4073                 zap_attribute_t *za;
4074                 zap_cursor_t zc;
4075
4076                 za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
4077                 for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
4078                     ds->ds_phys->ds_userrefs_obj);
4079                     zap_cursor_retrieve(&zc, za) == 0;
4080                     zap_cursor_advance(&zc)) {
4081                         VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
4082                             za->za_first_integer));
4083                 }
4084                 zap_cursor_fini(&zc);
4085                 kmem_free(za, sizeof (zap_attribute_t));
4086         }
4087         dsl_dataset_rele(ds, FTAG);
4088         return (0);
4089 }
4090
4091 /*
4092  * Note, this function is used as the callback for dmu_objset_find().  We
4093  * always return 0 so that we will continue to find and process
4094  * inconsistent datasets, even if we encounter an error trying to
4095  * process one of them.
4096  */
4097 /* ARGSUSED */
4098 int
4099 dsl_destroy_inconsistent(const char *dsname, void *arg)
4100 {
4101         dsl_dataset_t *ds;
4102
4103         if (dsl_dataset_own(dsname, B_TRUE, FTAG, &ds) == 0) {
4104                 if (DS_IS_INCONSISTENT(ds))
4105                         (void) dsl_dataset_destroy(ds, FTAG, B_FALSE);
4106                 else
4107                         dsl_dataset_disown(ds, FTAG);
4108         }
4109         return (0);
4110 }
4111
4112
4113 /*
4114  * Return (in *usedp) the amount of space written in new that is not
4115  * present in oldsnap.  New may be a snapshot or the head.  Old must be
4116  * a snapshot before new, in new's filesystem (or its origin).  If not then
4117  * fail and return EINVAL.
4118  *
4119  * The written space is calculated by considering two components:  First, we
4120  * ignore any freed space, and calculate the written as new's used space
4121  * minus old's used space.  Next, we add in the amount of space that was freed
4122  * between the two snapshots, thus reducing new's used space relative to old's.
4123  * Specifically, this is the space that was born before old->ds_creation_txg,
4124  * and freed before new (ie. on new's deadlist or a previous deadlist).
4125  *
4126  * space freed                         [---------------------]
4127  * snapshots                       ---O-------O--------O-------O------
4128  *                                         oldsnap            new
4129  */
4130 int
4131 dsl_dataset_space_written(dsl_dataset_t *oldsnap, dsl_dataset_t *new,
4132     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
4133 {
4134         int err = 0;
4135         uint64_t snapobj;
4136         dsl_pool_t *dp = new->ds_dir->dd_pool;
4137
4138         *usedp = 0;
4139         *usedp += new->ds_phys->ds_used_bytes;
4140         *usedp -= oldsnap->ds_phys->ds_used_bytes;
4141
4142         *compp = 0;
4143         *compp += new->ds_phys->ds_compressed_bytes;
4144         *compp -= oldsnap->ds_phys->ds_compressed_bytes;
4145
4146         *uncompp = 0;
4147         *uncompp += new->ds_phys->ds_uncompressed_bytes;
4148         *uncompp -= oldsnap->ds_phys->ds_uncompressed_bytes;
4149
4150         rw_enter(&dp->dp_config_rwlock, RW_READER);
4151         snapobj = new->ds_object;
4152         while (snapobj != oldsnap->ds_object) {
4153                 dsl_dataset_t *snap;
4154                 uint64_t used, comp, uncomp;
4155
4156                 err = dsl_dataset_hold_obj(dp, snapobj, FTAG, &snap);
4157                 if (err != 0)
4158                         break;
4159
4160                 if (snap->ds_phys->ds_prev_snap_txg ==
4161                     oldsnap->ds_phys->ds_creation_txg) {
4162                         /*
4163                          * The blocks in the deadlist can not be born after
4164                          * ds_prev_snap_txg, so get the whole deadlist space,
4165                          * which is more efficient (especially for old-format
4166                          * deadlists).  Unfortunately the deadlist code
4167                          * doesn't have enough information to make this
4168                          * optimization itself.
4169                          */
4170                         dsl_deadlist_space(&snap->ds_deadlist,
4171                             &used, &comp, &uncomp);
4172                 } else {
4173                         dsl_deadlist_space_range(&snap->ds_deadlist,
4174                             0, oldsnap->ds_phys->ds_creation_txg,
4175                             &used, &comp, &uncomp);
4176                 }
4177                 *usedp += used;
4178                 *compp += comp;
4179                 *uncompp += uncomp;
4180
4181                 /*
4182                  * If we get to the beginning of the chain of snapshots
4183                  * (ds_prev_snap_obj == 0) before oldsnap, then oldsnap
4184                  * was not a snapshot of/before new.
4185                  */
4186                 snapobj = snap->ds_phys->ds_prev_snap_obj;
4187                 dsl_dataset_rele(snap, FTAG);
4188                 if (snapobj == 0) {
4189                         err = EINVAL;
4190                         break;
4191                 }
4192
4193         }
4194         rw_exit(&dp->dp_config_rwlock);
4195         return (err);
4196 }
4197
4198 /*
4199  * Return (in *usedp) the amount of space that will be reclaimed if firstsnap,
4200  * lastsnap, and all snapshots in between are deleted.
4201  *
4202  * blocks that would be freed            [---------------------------]
4203  * snapshots                       ---O-------O--------O-------O--------O
4204  *                                        firstsnap        lastsnap
4205  *
4206  * This is the set of blocks that were born after the snap before firstsnap,
4207  * (birth > firstsnap->prev_snap_txg) and died before the snap after the
4208  * last snap (ie, is on lastsnap->ds_next->ds_deadlist or an earlier deadlist).
4209  * We calculate this by iterating over the relevant deadlists (from the snap
4210  * after lastsnap, backward to the snap after firstsnap), summing up the
4211  * space on the deadlist that was born after the snap before firstsnap.
4212  */
4213 int
4214 dsl_dataset_space_wouldfree(dsl_dataset_t *firstsnap,
4215     dsl_dataset_t *lastsnap,
4216     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
4217 {
4218         int err = 0;
4219         uint64_t snapobj;
4220         dsl_pool_t *dp = firstsnap->ds_dir->dd_pool;
4221
4222         ASSERT(dsl_dataset_is_snapshot(firstsnap));
4223         ASSERT(dsl_dataset_is_snapshot(lastsnap));
4224
4225         /*
4226          * Check that the snapshots are in the same dsl_dir, and firstsnap
4227          * is before lastsnap.
4228          */
4229         if (firstsnap->ds_dir != lastsnap->ds_dir ||
4230             firstsnap->ds_phys->ds_creation_txg >
4231             lastsnap->ds_phys->ds_creation_txg)
4232                 return (EINVAL);
4233
4234         *usedp = *compp = *uncompp = 0;
4235
4236         rw_enter(&dp->dp_config_rwlock, RW_READER);
4237         snapobj = lastsnap->ds_phys->ds_next_snap_obj;
4238         while (snapobj != firstsnap->ds_object) {
4239                 dsl_dataset_t *ds;
4240                 uint64_t used, comp, uncomp;
4241
4242                 err = dsl_dataset_hold_obj(dp, snapobj, FTAG, &ds);
4243                 if (err != 0)
4244                         break;
4245
4246                 dsl_deadlist_space_range(&ds->ds_deadlist,
4247                     firstsnap->ds_phys->ds_prev_snap_txg, UINT64_MAX,
4248                     &used, &comp, &uncomp);
4249                 *usedp += used;
4250                 *compp += comp;
4251                 *uncompp += uncomp;
4252
4253                 snapobj = ds->ds_phys->ds_prev_snap_obj;
4254                 ASSERT3U(snapobj, !=, 0);
4255                 dsl_dataset_rele(ds, FTAG);
4256         }
4257         rw_exit(&dp->dp_config_rwlock);
4258         return (err);
4259 }
4260
4261 #if defined(_KERNEL) && defined(HAVE_SPL)
4262 EXPORT_SYMBOL(dmu_snapshots_destroy_nvl);
4263 EXPORT_SYMBOL(dsl_dataset_hold);
4264 EXPORT_SYMBOL(dsl_dataset_hold_obj);
4265 EXPORT_SYMBOL(dsl_dataset_own);
4266 EXPORT_SYMBOL(dsl_dataset_own_obj);
4267 EXPORT_SYMBOL(dsl_dataset_name);
4268 EXPORT_SYMBOL(dsl_dataset_rele);
4269 EXPORT_SYMBOL(dsl_dataset_disown);
4270 EXPORT_SYMBOL(dsl_dataset_drop_ref);
4271 EXPORT_SYMBOL(dsl_dataset_tryown);
4272 EXPORT_SYMBOL(dsl_dataset_make_exclusive);
4273 EXPORT_SYMBOL(dsl_dataset_create_sync);
4274 EXPORT_SYMBOL(dsl_dataset_create_sync_dd);
4275 EXPORT_SYMBOL(dsl_dataset_destroy);
4276 EXPORT_SYMBOL(dsl_dataset_destroy_check);
4277 EXPORT_SYMBOL(dsl_dataset_destroy_sync);
4278 EXPORT_SYMBOL(dsl_dataset_snapshot_check);
4279 EXPORT_SYMBOL(dsl_dataset_snapshot_sync);
4280 EXPORT_SYMBOL(dsl_dataset_rename);
4281 EXPORT_SYMBOL(dsl_dataset_promote);
4282 EXPORT_SYMBOL(dsl_dataset_clone_swap);
4283 EXPORT_SYMBOL(dsl_dataset_user_hold);
4284 EXPORT_SYMBOL(dsl_dataset_user_release);
4285 EXPORT_SYMBOL(dsl_dataset_user_release_tmp);
4286 EXPORT_SYMBOL(dsl_dataset_get_holds);
4287 EXPORT_SYMBOL(dsl_dataset_get_blkptr);
4288 EXPORT_SYMBOL(dsl_dataset_set_blkptr);
4289 EXPORT_SYMBOL(dsl_dataset_get_spa);
4290 EXPORT_SYMBOL(dsl_dataset_modified_since_lastsnap);
4291 EXPORT_SYMBOL(dsl_dataset_space_written);
4292 EXPORT_SYMBOL(dsl_dataset_space_wouldfree);
4293 EXPORT_SYMBOL(dsl_dataset_sync);
4294 EXPORT_SYMBOL(dsl_dataset_block_born);
4295 EXPORT_SYMBOL(dsl_dataset_block_kill);
4296 EXPORT_SYMBOL(dsl_dataset_block_freeable);
4297 EXPORT_SYMBOL(dsl_dataset_prev_snap_txg);
4298 EXPORT_SYMBOL(dsl_dataset_dirty);
4299 EXPORT_SYMBOL(dsl_dataset_stats);
4300 EXPORT_SYMBOL(dsl_dataset_fast_stat);
4301 EXPORT_SYMBOL(dsl_dataset_space);
4302 EXPORT_SYMBOL(dsl_dataset_fsid_guid);
4303 EXPORT_SYMBOL(dsl_dsobj_to_dsname);
4304 EXPORT_SYMBOL(dsl_dataset_check_quota);
4305 EXPORT_SYMBOL(dsl_dataset_set_quota);
4306 EXPORT_SYMBOL(dsl_dataset_set_quota_sync);
4307 EXPORT_SYMBOL(dsl_dataset_set_reservation);
4308 EXPORT_SYMBOL(dsl_destroy_inconsistent);
4309 #endif