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 boolean_t
1235 dsl_dataset_is_dirty(dsl_dataset_t *ds)
1236 {
1237         int t;
1238
1239         for (t = 0; t < TXG_SIZE; t++) {
1240                 if (txg_list_member(&ds->ds_dir->dd_pool->dp_dirty_datasets,
1241                     ds, t))
1242                         return (B_TRUE);
1243         }
1244         return (B_FALSE);
1245 }
1246
1247 /*
1248  * The unique space in the head dataset can be calculated by subtracting
1249  * the space used in the most recent snapshot, that is still being used
1250  * in this file system, from the space currently in use.  To figure out
1251  * the space in the most recent snapshot still in use, we need to take
1252  * the total space used in the snapshot and subtract out the space that
1253  * has been freed up since the snapshot was taken.
1254  */
1255 static void
1256 dsl_dataset_recalc_head_uniq(dsl_dataset_t *ds)
1257 {
1258         uint64_t mrs_used;
1259         uint64_t dlused, dlcomp, dluncomp;
1260
1261         ASSERT(!dsl_dataset_is_snapshot(ds));
1262
1263         if (ds->ds_phys->ds_prev_snap_obj != 0)
1264                 mrs_used = ds->ds_prev->ds_phys->ds_used_bytes;
1265         else
1266                 mrs_used = 0;
1267
1268         dsl_deadlist_space(&ds->ds_deadlist, &dlused, &dlcomp, &dluncomp);
1269
1270         ASSERT3U(dlused, <=, mrs_used);
1271         ds->ds_phys->ds_unique_bytes =
1272             ds->ds_phys->ds_used_bytes - (mrs_used - dlused);
1273
1274         if (spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1275             SPA_VERSION_UNIQUE_ACCURATE)
1276                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1277 }
1278
1279 struct killarg {
1280         dsl_dataset_t *ds;
1281         dmu_tx_t *tx;
1282 };
1283
1284 /* ARGSUSED */
1285 static int
1286 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
1287     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
1288 {
1289         struct killarg *ka = arg;
1290         dmu_tx_t *tx = ka->tx;
1291
1292         if (bp == NULL)
1293                 return (0);
1294
1295         if (zb->zb_level == ZB_ZIL_LEVEL) {
1296                 ASSERT(zilog != NULL);
1297                 /*
1298                  * It's a block in the intent log.  It has no
1299                  * accounting, so just free it.
1300                  */
1301                 dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
1302         } else {
1303                 ASSERT(zilog == NULL);
1304                 ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
1305                 (void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
1306         }
1307
1308         return (0);
1309 }
1310
1311 /* ARGSUSED */
1312 static int
1313 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
1314 {
1315         dsl_dataset_t *ds = arg1;
1316         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1317         uint64_t count;
1318         int err;
1319
1320         /*
1321          * Can't delete a head dataset if there are snapshots of it.
1322          * (Except if the only snapshots are from the branch we cloned
1323          * from.)
1324          */
1325         if (ds->ds_prev != NULL &&
1326             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1327                 return (EBUSY);
1328
1329         /*
1330          * This is really a dsl_dir thing, but check it here so that
1331          * we'll be less likely to leave this dataset inconsistent &
1332          * nearly destroyed.
1333          */
1334         err = zap_count(mos, ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
1335         if (err)
1336                 return (err);
1337         if (count != 0)
1338                 return (EEXIST);
1339
1340         return (0);
1341 }
1342
1343 /* ARGSUSED */
1344 static void
1345 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1346 {
1347         dsl_dataset_t *ds = arg1;
1348         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1349
1350         /* Mark it as inconsistent on-disk, in case we crash */
1351         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1352         ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1353
1354         spa_history_log_internal(LOG_DS_DESTROY_BEGIN, dp->dp_spa, tx,
1355             "dataset = %llu", ds->ds_object);
1356 }
1357
1358 static int
1359 dsl_dataset_origin_check(struct dsl_ds_destroyarg *dsda, void *tag,
1360     dmu_tx_t *tx)
1361 {
1362         dsl_dataset_t *ds = dsda->ds;
1363         dsl_dataset_t *ds_prev = ds->ds_prev;
1364
1365         if (dsl_dataset_might_destroy_origin(ds_prev)) {
1366                 struct dsl_ds_destroyarg ndsda = {0};
1367
1368                 /*
1369                  * If we're not prepared to remove the origin, don't remove
1370                  * the clone either.
1371                  */
1372                 if (dsda->rm_origin == NULL) {
1373                         dsda->need_prep = B_TRUE;
1374                         return (EBUSY);
1375                 }
1376
1377                 ndsda.ds = ds_prev;
1378                 ndsda.is_origin_rm = B_TRUE;
1379                 return (dsl_dataset_destroy_check(&ndsda, tag, tx));
1380         }
1381
1382         /*
1383          * If we're not going to remove the origin after all,
1384          * undo the open context setup.
1385          */
1386         if (dsda->rm_origin != NULL) {
1387                 dsl_dataset_disown(dsda->rm_origin, tag);
1388                 dsda->rm_origin = NULL;
1389         }
1390
1391         return (0);
1392 }
1393
1394 /*
1395  * If you add new checks here, you may need to add
1396  * additional checks to the "temporary" case in
1397  * snapshot_check() in dmu_objset.c.
1398  */
1399 /* ARGSUSED */
1400 int
1401 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1402 {
1403         struct dsl_ds_destroyarg *dsda = arg1;
1404         dsl_dataset_t *ds = dsda->ds;
1405
1406         /* we have an owner hold, so noone else can destroy us */
1407         ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1408
1409         /*
1410          * Only allow deferred destroy on pools that support it.
1411          * NOTE: deferred destroy is only supported on snapshots.
1412          */
1413         if (dsda->defer) {
1414                 if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1415                     SPA_VERSION_USERREFS)
1416                         return (ENOTSUP);
1417                 ASSERT(dsl_dataset_is_snapshot(ds));
1418                 return (0);
1419         }
1420
1421         /*
1422          * Can't delete a head dataset if there are snapshots of it.
1423          * (Except if the only snapshots are from the branch we cloned
1424          * from.)
1425          */
1426         if (ds->ds_prev != NULL &&
1427             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1428                 return (EBUSY);
1429
1430         /*
1431          * If we made changes this txg, traverse_dsl_dataset won't find
1432          * them.  Try again.
1433          */
1434         if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1435                 return (EAGAIN);
1436
1437         if (dsl_dataset_is_snapshot(ds)) {
1438                 /*
1439                  * If this snapshot has an elevated user reference count,
1440                  * we can't destroy it yet.
1441                  */
1442                 if (ds->ds_userrefs > 0 && !dsda->releasing)
1443                         return (EBUSY);
1444
1445                 mutex_enter(&ds->ds_lock);
1446                 /*
1447                  * Can't delete a branch point. However, if we're destroying
1448                  * a clone and removing its origin due to it having a user
1449                  * hold count of 0 and having been marked for deferred destroy,
1450                  * it's OK for the origin to have a single clone.
1451                  */
1452                 if (ds->ds_phys->ds_num_children >
1453                     (dsda->is_origin_rm ? 2 : 1)) {
1454                         mutex_exit(&ds->ds_lock);
1455                         return (EEXIST);
1456                 }
1457                 mutex_exit(&ds->ds_lock);
1458         } else if (dsl_dir_is_clone(ds->ds_dir)) {
1459                 return (dsl_dataset_origin_check(dsda, arg2, tx));
1460         }
1461
1462         /* XXX we should do some i/o error checking... */
1463         return (0);
1464 }
1465
1466 struct refsarg {
1467         kmutex_t lock;
1468         boolean_t gone;
1469         kcondvar_t cv;
1470 };
1471
1472 /* ARGSUSED */
1473 static void
1474 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1475 {
1476         struct refsarg *arg = argv;
1477
1478         mutex_enter(&arg->lock);
1479         arg->gone = TRUE;
1480         cv_signal(&arg->cv);
1481         mutex_exit(&arg->lock);
1482 }
1483
1484 static void
1485 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1486 {
1487         struct refsarg arg;
1488
1489         mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1490         cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1491         arg.gone = FALSE;
1492         (void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1493             dsl_dataset_refs_gone);
1494         dmu_buf_rele(ds->ds_dbuf, tag);
1495         mutex_enter(&arg.lock);
1496         while (!arg.gone)
1497                 cv_wait(&arg.cv, &arg.lock);
1498         ASSERT(arg.gone);
1499         mutex_exit(&arg.lock);
1500         ds->ds_dbuf = NULL;
1501         ds->ds_phys = NULL;
1502         mutex_destroy(&arg.lock);
1503         cv_destroy(&arg.cv);
1504 }
1505
1506 static void
1507 remove_from_next_clones(dsl_dataset_t *ds, uint64_t obj, dmu_tx_t *tx)
1508 {
1509         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1510         int err;
1511         ASSERTV(uint64_t count);
1512
1513         ASSERT(ds->ds_phys->ds_num_children >= 2);
1514         err = zap_remove_int(mos, ds->ds_phys->ds_next_clones_obj, obj, tx);
1515         /*
1516          * The err should not be ENOENT, but a bug in a previous version
1517          * of the code could cause upgrade_clones_cb() to not set
1518          * ds_next_snap_obj when it should, leading to a missing entry.
1519          * If we knew that the pool was created after
1520          * SPA_VERSION_NEXT_CLONES, we could assert that it isn't
1521          * ENOENT.  However, at least we can check that we don't have
1522          * too many entries in the next_clones_obj even after failing to
1523          * remove this one.
1524          */
1525         if (err != ENOENT) {
1526                 VERIFY3U(err, ==, 0);
1527         }
1528         ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
1529             &count));
1530         ASSERT3U(count, <=, ds->ds_phys->ds_num_children - 2);
1531 }
1532
1533 static void
1534 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
1535 {
1536         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1537         zap_cursor_t zc;
1538         zap_attribute_t za;
1539
1540         /*
1541          * If it is the old version, dd_clones doesn't exist so we can't
1542          * find the clones, but deadlist_remove_key() is a no-op so it
1543          * doesn't matter.
1544          */
1545         if (ds->ds_dir->dd_phys->dd_clones == 0)
1546                 return;
1547
1548         for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
1549             zap_cursor_retrieve(&zc, &za) == 0;
1550             zap_cursor_advance(&zc)) {
1551                 dsl_dataset_t *clone;
1552
1553                 VERIFY3U(0, ==, dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
1554                     za.za_first_integer, FTAG, &clone));
1555                 if (clone->ds_dir->dd_origin_txg > mintxg) {
1556                         dsl_deadlist_remove_key(&clone->ds_deadlist,
1557                             mintxg, tx);
1558                         dsl_dataset_remove_clones_key(clone, mintxg, tx);
1559                 }
1560                 dsl_dataset_rele(clone, FTAG);
1561         }
1562         zap_cursor_fini(&zc);
1563 }
1564
1565 struct process_old_arg {
1566         dsl_dataset_t *ds;
1567         dsl_dataset_t *ds_prev;
1568         boolean_t after_branch_point;
1569         zio_t *pio;
1570         uint64_t used, comp, uncomp;
1571 };
1572
1573 static int
1574 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1575 {
1576         struct process_old_arg *poa = arg;
1577         dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
1578
1579         if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
1580                 dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
1581                 if (poa->ds_prev && !poa->after_branch_point &&
1582                     bp->blk_birth >
1583                     poa->ds_prev->ds_phys->ds_prev_snap_txg) {
1584                         poa->ds_prev->ds_phys->ds_unique_bytes +=
1585                             bp_get_dsize_sync(dp->dp_spa, bp);
1586                 }
1587         } else {
1588                 poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
1589                 poa->comp += BP_GET_PSIZE(bp);
1590                 poa->uncomp += BP_GET_UCSIZE(bp);
1591                 dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
1592         }
1593         return (0);
1594 }
1595
1596 static void
1597 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
1598     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
1599 {
1600         struct process_old_arg poa = { 0 };
1601         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1602         objset_t *mos = dp->dp_meta_objset;
1603
1604         ASSERT(ds->ds_deadlist.dl_oldfmt);
1605         ASSERT(ds_next->ds_deadlist.dl_oldfmt);
1606
1607         poa.ds = ds;
1608         poa.ds_prev = ds_prev;
1609         poa.after_branch_point = after_branch_point;
1610         poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1611         VERIFY3U(0, ==, bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
1612             process_old_cb, &poa, tx));
1613         VERIFY3U(zio_wait(poa.pio), ==, 0);
1614         ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
1615
1616         /* change snapused */
1617         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1618             -poa.used, -poa.comp, -poa.uncomp, tx);
1619
1620         /* swap next's deadlist to our deadlist */
1621         dsl_deadlist_close(&ds->ds_deadlist);
1622         dsl_deadlist_close(&ds_next->ds_deadlist);
1623         SWITCH64(ds_next->ds_phys->ds_deadlist_obj,
1624             ds->ds_phys->ds_deadlist_obj);
1625         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
1626         dsl_deadlist_open(&ds_next->ds_deadlist, mos,
1627             ds_next->ds_phys->ds_deadlist_obj);
1628 }
1629
1630 void
1631 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1632 {
1633         struct dsl_ds_destroyarg *dsda = arg1;
1634         dsl_dataset_t *ds = dsda->ds;
1635         int err;
1636         int after_branch_point = FALSE;
1637         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1638         objset_t *mos = dp->dp_meta_objset;
1639         dsl_dataset_t *ds_prev = NULL;
1640         boolean_t wont_destroy;
1641         uint64_t obj;
1642
1643         wont_destroy = (dsda->defer &&
1644             (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1));
1645
1646         ASSERT(ds->ds_owner || wont_destroy);
1647         ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1648         ASSERT(ds->ds_prev == NULL ||
1649             ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1650         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1651
1652         if (wont_destroy) {
1653                 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1654                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1655                 ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1656                 return;
1657         }
1658
1659         /* signal any waiters that this dataset is going away */
1660         mutex_enter(&ds->ds_lock);
1661         ds->ds_owner = dsl_reaper;
1662         cv_broadcast(&ds->ds_exclusive_cv);
1663         mutex_exit(&ds->ds_lock);
1664
1665         /* Remove our reservation */
1666         if (ds->ds_reserved != 0) {
1667                 dsl_prop_setarg_t psa;
1668                 uint64_t value = 0;
1669
1670                 dsl_prop_setarg_init_uint64(&psa, "refreservation",
1671                     (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
1672                     &value);
1673                 psa.psa_effective_value = 0;    /* predict default value */
1674
1675                 dsl_dataset_set_reservation_sync(ds, &psa, tx);
1676                 ASSERT3U(ds->ds_reserved, ==, 0);
1677         }
1678
1679         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1680
1681         dsl_scan_ds_destroyed(ds, tx);
1682
1683         obj = ds->ds_object;
1684
1685         if (ds->ds_phys->ds_prev_snap_obj != 0) {
1686                 if (ds->ds_prev) {
1687                         ds_prev = ds->ds_prev;
1688                 } else {
1689                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1690                             ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1691                 }
1692                 after_branch_point =
1693                     (ds_prev->ds_phys->ds_next_snap_obj != obj);
1694
1695                 dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1696                 if (after_branch_point &&
1697                     ds_prev->ds_phys->ds_next_clones_obj != 0) {
1698                         remove_from_next_clones(ds_prev, obj, tx);
1699                         if (ds->ds_phys->ds_next_snap_obj != 0) {
1700                                 VERIFY(0 == zap_add_int(mos,
1701                                     ds_prev->ds_phys->ds_next_clones_obj,
1702                                     ds->ds_phys->ds_next_snap_obj, tx));
1703                         }
1704                 }
1705                 if (after_branch_point &&
1706                     ds->ds_phys->ds_next_snap_obj == 0) {
1707                         /* This clone is toast. */
1708                         ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1709                         ds_prev->ds_phys->ds_num_children--;
1710
1711                         /*
1712                          * If the clone's origin has no other clones, no
1713                          * user holds, and has been marked for deferred
1714                          * deletion, then we should have done the necessary
1715                          * destroy setup for it.
1716                          */
1717                         if (ds_prev->ds_phys->ds_num_children == 1 &&
1718                             ds_prev->ds_userrefs == 0 &&
1719                             DS_IS_DEFER_DESTROY(ds_prev)) {
1720                                 ASSERT3P(dsda->rm_origin, !=, NULL);
1721                         } else {
1722                                 ASSERT3P(dsda->rm_origin, ==, NULL);
1723                         }
1724                 } else if (!after_branch_point) {
1725                         ds_prev->ds_phys->ds_next_snap_obj =
1726                             ds->ds_phys->ds_next_snap_obj;
1727                 }
1728         }
1729
1730         if (dsl_dataset_is_snapshot(ds)) {
1731                 dsl_dataset_t *ds_next;
1732                 uint64_t old_unique;
1733                 uint64_t used = 0, comp = 0, uncomp = 0;
1734
1735                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1736                     ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1737                 ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1738
1739                 old_unique = ds_next->ds_phys->ds_unique_bytes;
1740
1741                 dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1742                 ds_next->ds_phys->ds_prev_snap_obj =
1743                     ds->ds_phys->ds_prev_snap_obj;
1744                 ds_next->ds_phys->ds_prev_snap_txg =
1745                     ds->ds_phys->ds_prev_snap_txg;
1746                 ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1747                     ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1748
1749
1750                 if (ds_next->ds_deadlist.dl_oldfmt) {
1751                         process_old_deadlist(ds, ds_prev, ds_next,
1752                             after_branch_point, tx);
1753                 } else {
1754                         /* Adjust prev's unique space. */
1755                         if (ds_prev && !after_branch_point) {
1756                                 dsl_deadlist_space_range(&ds_next->ds_deadlist,
1757                                     ds_prev->ds_phys->ds_prev_snap_txg,
1758                                     ds->ds_phys->ds_prev_snap_txg,
1759                                     &used, &comp, &uncomp);
1760                                 ds_prev->ds_phys->ds_unique_bytes += used;
1761                         }
1762
1763                         /* Adjust snapused. */
1764                         dsl_deadlist_space_range(&ds_next->ds_deadlist,
1765                             ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
1766                             &used, &comp, &uncomp);
1767                         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1768                             -used, -comp, -uncomp, tx);
1769
1770                         /* Move blocks to be freed to pool's free list. */
1771                         dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
1772                             &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
1773                             tx);
1774                         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
1775                             DD_USED_HEAD, used, comp, uncomp, tx);
1776                         dsl_dir_dirty(tx->tx_pool->dp_free_dir, tx);
1777
1778                         /* Merge our deadlist into next's and free it. */
1779                         dsl_deadlist_merge(&ds_next->ds_deadlist,
1780                             ds->ds_phys->ds_deadlist_obj, tx);
1781                 }
1782                 dsl_deadlist_close(&ds->ds_deadlist);
1783                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1784
1785                 /* Collapse range in clone heads */
1786                 dsl_dataset_remove_clones_key(ds,
1787                     ds->ds_phys->ds_creation_txg, tx);
1788
1789                 if (dsl_dataset_is_snapshot(ds_next)) {
1790                         dsl_dataset_t *ds_nextnext;
1791                         dsl_dataset_t *hds;
1792
1793                         /*
1794                          * Update next's unique to include blocks which
1795                          * were previously shared by only this snapshot
1796                          * and it.  Those blocks will be born after the
1797                          * prev snap and before this snap, and will have
1798                          * died after the next snap and before the one
1799                          * after that (ie. be on the snap after next's
1800                          * deadlist).
1801                          */
1802                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1803                             ds_next->ds_phys->ds_next_snap_obj,
1804                             FTAG, &ds_nextnext));
1805                         dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
1806                             ds->ds_phys->ds_prev_snap_txg,
1807                             ds->ds_phys->ds_creation_txg,
1808                             &used, &comp, &uncomp);
1809                         ds_next->ds_phys->ds_unique_bytes += used;
1810                         dsl_dataset_rele(ds_nextnext, FTAG);
1811                         ASSERT3P(ds_next->ds_prev, ==, NULL);
1812
1813                         /* Collapse range in this head. */
1814                         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
1815                             ds->ds_dir->dd_phys->dd_head_dataset_obj,
1816                             FTAG, &hds));
1817                         dsl_deadlist_remove_key(&hds->ds_deadlist,
1818                             ds->ds_phys->ds_creation_txg, tx);
1819                         dsl_dataset_rele(hds, FTAG);
1820
1821                 } else {
1822                         ASSERT3P(ds_next->ds_prev, ==, ds);
1823                         dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1824                         ds_next->ds_prev = NULL;
1825                         if (ds_prev) {
1826                                 VERIFY(0 == dsl_dataset_get_ref(dp,
1827                                     ds->ds_phys->ds_prev_snap_obj,
1828                                     ds_next, &ds_next->ds_prev));
1829                         }
1830
1831                         dsl_dataset_recalc_head_uniq(ds_next);
1832
1833                         /*
1834                          * Reduce the amount of our unconsmed refreservation
1835                          * being charged to our parent by the amount of
1836                          * new unique data we have gained.
1837                          */
1838                         if (old_unique < ds_next->ds_reserved) {
1839                                 int64_t mrsdelta;
1840                                 uint64_t new_unique =
1841                                     ds_next->ds_phys->ds_unique_bytes;
1842
1843                                 ASSERT(old_unique <= new_unique);
1844                                 mrsdelta = MIN(new_unique - old_unique,
1845                                     ds_next->ds_reserved - old_unique);
1846                                 dsl_dir_diduse_space(ds->ds_dir,
1847                                     DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1848                         }
1849                 }
1850                 dsl_dataset_rele(ds_next, FTAG);
1851         } else {
1852                 /*
1853                  * There's no next snapshot, so this is a head dataset.
1854                  * Destroy the deadlist.  Unless it's a clone, the
1855                  * deadlist should be empty.  (If it's a clone, it's
1856                  * safe to ignore the deadlist contents.)
1857                  */
1858                 struct killarg ka;
1859
1860                 dsl_deadlist_close(&ds->ds_deadlist);
1861                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1862                 ds->ds_phys->ds_deadlist_obj = 0;
1863
1864                 /*
1865                  * Free everything that we point to (that's born after
1866                  * the previous snapshot, if we are a clone)
1867                  *
1868                  * NB: this should be very quick, because we already
1869                  * freed all the objects in open context.
1870                  */
1871                 ka.ds = ds;
1872                 ka.tx = tx;
1873                 err = traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1874                     TRAVERSE_POST, kill_blkptr, &ka);
1875                 ASSERT3U(err, ==, 0);
1876                 ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1877                     ds->ds_phys->ds_unique_bytes == 0);
1878
1879                 if (ds->ds_prev != NULL) {
1880                         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
1881                                 VERIFY3U(0, ==, zap_remove_int(mos,
1882                                     ds->ds_prev->ds_dir->dd_phys->dd_clones,
1883                                     ds->ds_object, tx));
1884                         }
1885                         dsl_dataset_rele(ds->ds_prev, ds);
1886                         ds->ds_prev = ds_prev = NULL;
1887                 }
1888         }
1889
1890         /*
1891          * This must be done after the dsl_traverse(), because it will
1892          * re-open the objset.
1893          */
1894         if (ds->ds_objset) {
1895                 dmu_objset_evict(ds->ds_objset);
1896                 ds->ds_objset = NULL;
1897         }
1898
1899         if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1900                 /* Erase the link in the dir */
1901                 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1902                 ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1903                 ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1904                 err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1905                 ASSERT(err == 0);
1906         } else {
1907                 /* remove from snapshot namespace */
1908                 dsl_dataset_t *ds_head;
1909                 ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1910                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1911                     ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1912                 VERIFY(0 == dsl_dataset_get_snapname(ds));
1913 #ifdef ZFS_DEBUG
1914                 {
1915                         uint64_t val;
1916
1917                         err = dsl_dataset_snap_lookup(ds_head,
1918                             ds->ds_snapname, &val);
1919                         ASSERT3U(err, ==, 0);
1920                         ASSERT3U(val, ==, obj);
1921                 }
1922 #endif
1923                 err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1924                 ASSERT(err == 0);
1925                 dsl_dataset_rele(ds_head, FTAG);
1926         }
1927
1928         if (ds_prev && ds->ds_prev != ds_prev)
1929                 dsl_dataset_rele(ds_prev, FTAG);
1930
1931         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1932         spa_history_log_internal(LOG_DS_DESTROY, dp->dp_spa, tx,
1933             "dataset = %llu", ds->ds_object);
1934
1935         if (ds->ds_phys->ds_next_clones_obj != 0) {
1936                 ASSERTV(uint64_t count);
1937                 ASSERT(0 == zap_count(mos,
1938                     ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1939                 VERIFY(0 == dmu_object_free(mos,
1940                     ds->ds_phys->ds_next_clones_obj, tx));
1941         }
1942         if (ds->ds_phys->ds_props_obj != 0)
1943                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1944         if (ds->ds_phys->ds_userrefs_obj != 0)
1945                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1946         dsl_dir_close(ds->ds_dir, ds);
1947         ds->ds_dir = NULL;
1948         dsl_dataset_drain_refs(ds, tag);
1949         VERIFY(0 == dmu_object_free(mos, obj, tx));
1950
1951         if (dsda->rm_origin) {
1952                 /*
1953                  * Remove the origin of the clone we just destroyed.
1954                  */
1955                 struct dsl_ds_destroyarg ndsda = {0};
1956
1957                 ndsda.ds = dsda->rm_origin;
1958                 dsl_dataset_destroy_sync(&ndsda, tag, tx);
1959         }
1960 }
1961
1962 static int
1963 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
1964 {
1965         uint64_t asize;
1966
1967         if (!dmu_tx_is_syncing(tx))
1968                 return (0);
1969
1970         /*
1971          * If there's an fs-only reservation, any blocks that might become
1972          * owned by the snapshot dataset must be accommodated by space
1973          * outside of the reservation.
1974          */
1975         ASSERT(ds->ds_reserved == 0 || DS_UNIQUE_IS_ACCURATE(ds));
1976         asize = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
1977         if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
1978                 return (ENOSPC);
1979
1980         /*
1981          * Propogate any reserved space for this snapshot to other
1982          * snapshot checks in this sync group.
1983          */
1984         if (asize > 0)
1985                 dsl_dir_willuse_space(ds->ds_dir, asize, tx);
1986
1987         return (0);
1988 }
1989
1990 int
1991 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1992 {
1993         dsl_dataset_t *ds = arg1;
1994         const char *snapname = arg2;
1995         int err;
1996         uint64_t value;
1997
1998         /*
1999          * We don't allow multiple snapshots of the same txg.  If there
2000          * is already one, try again.
2001          */
2002         if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
2003                 return (EAGAIN);
2004
2005         /*
2006          * Check for conflicting name snapshot name.
2007          */
2008         err = dsl_dataset_snap_lookup(ds, snapname, &value);
2009         if (err == 0)
2010                 return (EEXIST);
2011         if (err != ENOENT)
2012                 return (err);
2013
2014         /*
2015          * Check that the dataset's name is not too long.  Name consists
2016          * of the dataset's length + 1 for the @-sign + snapshot name's length
2017          */
2018         if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
2019                 return (ENAMETOOLONG);
2020
2021         err = dsl_dataset_snapshot_reserve_space(ds, tx);
2022         if (err)
2023                 return (err);
2024
2025         ds->ds_trysnap_txg = tx->tx_txg;
2026         return (0);
2027 }
2028
2029 void
2030 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2031 {
2032         dsl_dataset_t *ds = arg1;
2033         const char *snapname = arg2;
2034         dsl_pool_t *dp = ds->ds_dir->dd_pool;
2035         dmu_buf_t *dbuf;
2036         dsl_dataset_phys_t *dsphys;
2037         uint64_t dsobj, crtxg;
2038         objset_t *mos = dp->dp_meta_objset;
2039         int err;
2040
2041         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
2042
2043         /*
2044          * The origin's ds_creation_txg has to be < TXG_INITIAL
2045          */
2046         if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2047                 crtxg = 1;
2048         else
2049                 crtxg = tx->tx_txg;
2050
2051         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2052             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2053         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2054         dmu_buf_will_dirty(dbuf, tx);
2055         dsphys = dbuf->db_data;
2056         bzero(dsphys, sizeof (dsl_dataset_phys_t));
2057         dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2058         dsphys->ds_fsid_guid = unique_create();
2059         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2060             sizeof (dsphys->ds_guid));
2061         dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2062         dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2063         dsphys->ds_next_snap_obj = ds->ds_object;
2064         dsphys->ds_num_children = 1;
2065         dsphys->ds_creation_time = gethrestime_sec();
2066         dsphys->ds_creation_txg = crtxg;
2067         dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2068         dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
2069         dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2070         dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2071         dsphys->ds_flags = ds->ds_phys->ds_flags;
2072         dsphys->ds_bp = ds->ds_phys->ds_bp;
2073         dmu_buf_rele(dbuf, FTAG);
2074
2075         ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2076         if (ds->ds_prev) {
2077                 uint64_t next_clones_obj =
2078                     ds->ds_prev->ds_phys->ds_next_clones_obj;
2079                 ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2080                     ds->ds_object ||
2081                     ds->ds_prev->ds_phys->ds_num_children > 1);
2082                 if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2083                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2084                         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2085                             ds->ds_prev->ds_phys->ds_creation_txg);
2086                         ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2087                 } else if (next_clones_obj != 0) {
2088                         remove_from_next_clones(ds->ds_prev,
2089                             dsphys->ds_next_snap_obj, tx);
2090                         VERIFY3U(0, ==, zap_add_int(mos,
2091                             next_clones_obj, dsobj, tx));
2092                 }
2093         }
2094
2095         /*
2096          * If we have a reference-reservation on this dataset, we will
2097          * need to increase the amount of refreservation being charged
2098          * since our unique space is going to zero.
2099          */
2100         if (ds->ds_reserved) {
2101                 int64_t delta;
2102                 ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
2103                 delta = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2104                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2105                     delta, 0, 0, tx);
2106         }
2107
2108         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2109         zfs_dbgmsg("taking snapshot %s@%s/%llu; newkey=%llu",
2110             ds->ds_dir->dd_myname, snapname, dsobj,
2111             ds->ds_phys->ds_prev_snap_txg);
2112         ds->ds_phys->ds_deadlist_obj = dsl_deadlist_clone(&ds->ds_deadlist,
2113             UINT64_MAX, ds->ds_phys->ds_prev_snap_obj, tx);
2114         dsl_deadlist_close(&ds->ds_deadlist);
2115         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
2116         dsl_deadlist_add_key(&ds->ds_deadlist,
2117             ds->ds_phys->ds_prev_snap_txg, tx);
2118
2119         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2120         ds->ds_phys->ds_prev_snap_obj = dsobj;
2121         ds->ds_phys->ds_prev_snap_txg = crtxg;
2122         ds->ds_phys->ds_unique_bytes = 0;
2123         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2124                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2125
2126         err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2127             snapname, 8, 1, &dsobj, tx);
2128         ASSERT(err == 0);
2129
2130         if (ds->ds_prev)
2131                 dsl_dataset_drop_ref(ds->ds_prev, ds);
2132         VERIFY(0 == dsl_dataset_get_ref(dp,
2133             ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2134
2135         dsl_scan_ds_snapshotted(ds, tx);
2136
2137         dsl_dir_snap_cmtime_update(ds->ds_dir);
2138
2139         spa_history_log_internal(LOG_DS_SNAPSHOT, dp->dp_spa, tx,
2140             "dataset = %llu", dsobj);
2141 }
2142
2143 void
2144 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2145 {
2146         ASSERT(dmu_tx_is_syncing(tx));
2147         ASSERT(ds->ds_objset != NULL);
2148         ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2149
2150         /*
2151          * in case we had to change ds_fsid_guid when we opened it,
2152          * sync it out now.
2153          */
2154         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2155         ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2156
2157         dsl_dir_dirty(ds->ds_dir, tx);
2158         dmu_objset_sync(ds->ds_objset, zio, tx);
2159 }
2160
2161 static void
2162 get_clones_stat(dsl_dataset_t *ds, nvlist_t *nv)
2163 {
2164         uint64_t count = 0;
2165         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
2166         zap_cursor_t zc;
2167         zap_attribute_t za;
2168         nvlist_t *propval;
2169         nvlist_t *val;
2170
2171         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2172         VERIFY(nvlist_alloc(&propval, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2173         VERIFY(nvlist_alloc(&val, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2174
2175         /*
2176          * There may me missing entries in ds_next_clones_obj
2177          * due to a bug in a previous version of the code.
2178          * Only trust it if it has the right number of entries.
2179          */
2180         if (ds->ds_phys->ds_next_clones_obj != 0) {
2181                 ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
2182                     &count));
2183         }
2184         if (count != ds->ds_phys->ds_num_children - 1) {
2185                 goto fail;
2186         }
2187         for (zap_cursor_init(&zc, mos, ds->ds_phys->ds_next_clones_obj);
2188             zap_cursor_retrieve(&zc, &za) == 0;
2189             zap_cursor_advance(&zc)) {
2190                 dsl_dataset_t *clone;
2191                 char buf[ZFS_MAXNAMELEN];
2192                 if (dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
2193                     za.za_first_integer, FTAG, &clone) != 0) {
2194                         goto fail;
2195                 }
2196                 dsl_dir_name(clone->ds_dir, buf);
2197                 VERIFY(nvlist_add_boolean(val, buf) == 0);
2198                 dsl_dataset_rele(clone, FTAG);
2199         }
2200         zap_cursor_fini(&zc);
2201         VERIFY(nvlist_add_nvlist(propval, ZPROP_VALUE, val) == 0);
2202         VERIFY(nvlist_add_nvlist(nv, zfs_prop_to_name(ZFS_PROP_CLONES),
2203             propval) == 0);
2204 fail:
2205         nvlist_free(val);
2206         nvlist_free(propval);
2207         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2208 }
2209
2210 void
2211 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2212 {
2213         uint64_t refd, avail, uobjs, aobjs, ratio;
2214
2215         dsl_dir_stats(ds->ds_dir, nv);
2216
2217         dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2218         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2219         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2220
2221         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2222             ds->ds_phys->ds_creation_time);
2223         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2224             ds->ds_phys->ds_creation_txg);
2225         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2226             ds->ds_quota);
2227         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2228             ds->ds_reserved);
2229         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2230             ds->ds_phys->ds_guid);
2231         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_UNIQUE,
2232             ds->ds_phys->ds_unique_bytes);
2233         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_OBJSETID,
2234             ds->ds_object);
2235         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS,
2236             ds->ds_userrefs);
2237         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2238             DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2239
2240         if (ds->ds_phys->ds_prev_snap_obj != 0) {
2241                 uint64_t written, comp, uncomp;
2242                 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2243                 dsl_dataset_t *prev;
2244                 int err;
2245
2246                 rw_enter(&dp->dp_config_rwlock, RW_READER);
2247                 err = dsl_dataset_hold_obj(dp,
2248                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
2249                 rw_exit(&dp->dp_config_rwlock);
2250                 if (err == 0) {
2251                         err = dsl_dataset_space_written(prev, ds, &written,
2252                             &comp, &uncomp);
2253                         dsl_dataset_rele(prev, FTAG);
2254                         if (err == 0) {
2255                                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_WRITTEN,
2256                                     written);
2257                         }
2258                 }
2259         }
2260
2261         ratio = ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2262             (ds->ds_phys->ds_uncompressed_bytes * 100 /
2263             ds->ds_phys->ds_compressed_bytes);
2264         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRATIO, ratio);
2265
2266         if (ds->ds_phys->ds_next_snap_obj) {
2267                 /*
2268                  * This is a snapshot; override the dd's space used with
2269                  * our unique space and compression ratio.
2270                  */
2271                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2272                     ds->ds_phys->ds_unique_bytes);
2273                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO, ratio);
2274
2275                 get_clones_stat(ds, nv);
2276         }
2277 }
2278
2279 void
2280 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2281 {
2282         stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2283         stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2284         stat->dds_guid = ds->ds_phys->ds_guid;
2285         if (ds->ds_phys->ds_next_snap_obj) {
2286                 stat->dds_is_snapshot = B_TRUE;
2287                 stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2288         } else {
2289                 stat->dds_is_snapshot = B_FALSE;
2290                 stat->dds_num_clones = 0;
2291         }
2292
2293         /* clone origin is really a dsl_dir thing... */
2294         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2295         if (dsl_dir_is_clone(ds->ds_dir)) {
2296                 dsl_dataset_t *ods;
2297
2298                 VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2299                     ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2300                 dsl_dataset_name(ods, stat->dds_origin);
2301                 dsl_dataset_drop_ref(ods, FTAG);
2302         } else {
2303                 stat->dds_origin[0] = '\0';
2304         }
2305         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2306 }
2307
2308 uint64_t
2309 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2310 {
2311         return (ds->ds_fsid_guid);
2312 }
2313
2314 void
2315 dsl_dataset_space(dsl_dataset_t *ds,
2316     uint64_t *refdbytesp, uint64_t *availbytesp,
2317     uint64_t *usedobjsp, uint64_t *availobjsp)
2318 {
2319         *refdbytesp = ds->ds_phys->ds_used_bytes;
2320         *availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2321         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2322                 *availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2323         if (ds->ds_quota != 0) {
2324                 /*
2325                  * Adjust available bytes according to refquota
2326                  */
2327                 if (*refdbytesp < ds->ds_quota)
2328                         *availbytesp = MIN(*availbytesp,
2329                             ds->ds_quota - *refdbytesp);
2330                 else
2331                         *availbytesp = 0;
2332         }
2333         *usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2334         *availobjsp = DN_MAX_OBJECT - *usedobjsp;
2335 }
2336
2337 boolean_t
2338 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2339 {
2340         ASSERTV(dsl_pool_t *dp = ds->ds_dir->dd_pool);
2341
2342         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2343             dsl_pool_sync_context(dp));
2344         if (ds->ds_prev == NULL)
2345                 return (B_FALSE);
2346         if (ds->ds_phys->ds_bp.blk_birth >
2347             ds->ds_prev->ds_phys->ds_creation_txg) {
2348                 objset_t *os, *os_prev;
2349                 /*
2350                  * It may be that only the ZIL differs, because it was
2351                  * reset in the head.  Don't count that as being
2352                  * modified.
2353                  */
2354                 if (dmu_objset_from_ds(ds, &os) != 0)
2355                         return (B_TRUE);
2356                 if (dmu_objset_from_ds(ds->ds_prev, &os_prev) != 0)
2357                         return (B_TRUE);
2358                 return (bcmp(&os->os_phys->os_meta_dnode,
2359                     &os_prev->os_phys->os_meta_dnode,
2360                     sizeof (os->os_phys->os_meta_dnode)) != 0);
2361         }
2362         return (B_FALSE);
2363 }
2364
2365 /* ARGSUSED */
2366 static int
2367 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2368 {
2369         dsl_dataset_t *ds = arg1;
2370         char *newsnapname = arg2;
2371         dsl_dir_t *dd = ds->ds_dir;
2372         dsl_dataset_t *hds;
2373         uint64_t val;
2374         int err;
2375
2376         err = dsl_dataset_hold_obj(dd->dd_pool,
2377             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2378         if (err)
2379                 return (err);
2380
2381         /* new name better not be in use */
2382         err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2383         dsl_dataset_rele(hds, FTAG);
2384
2385         if (err == 0)
2386                 err = EEXIST;
2387         else if (err == ENOENT)
2388                 err = 0;
2389
2390         /* dataset name + 1 for the "@" + the new snapshot name must fit */
2391         if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2392                 err = ENAMETOOLONG;
2393
2394         return (err);
2395 }
2396
2397 static void
2398 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2399 {
2400         dsl_dataset_t *ds = arg1;
2401         const char *newsnapname = arg2;
2402         dsl_dir_t *dd = ds->ds_dir;
2403         objset_t *mos = dd->dd_pool->dp_meta_objset;
2404         dsl_dataset_t *hds;
2405         int err;
2406
2407         ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2408
2409         VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2410             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2411
2412         VERIFY(0 == dsl_dataset_get_snapname(ds));
2413         err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2414         ASSERT3U(err, ==, 0);
2415         mutex_enter(&ds->ds_lock);
2416         (void) strcpy(ds->ds_snapname, newsnapname);
2417         mutex_exit(&ds->ds_lock);
2418         err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2419             ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2420         ASSERT3U(err, ==, 0);
2421
2422         spa_history_log_internal(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2423             "dataset = %llu", ds->ds_object);
2424         dsl_dataset_rele(hds, FTAG);
2425 }
2426
2427 struct renamesnaparg {
2428         dsl_sync_task_group_t *dstg;
2429         char failed[MAXPATHLEN];
2430         char *oldsnap;
2431         char *newsnap;
2432 };
2433
2434 static int
2435 dsl_snapshot_rename_one(const char *name, void *arg)
2436 {
2437         struct renamesnaparg *ra = arg;
2438         dsl_dataset_t *ds = NULL;
2439         char *snapname;
2440         int err;
2441
2442         snapname = kmem_asprintf("%s@%s", name, ra->oldsnap);
2443         (void) strlcpy(ra->failed, snapname, sizeof (ra->failed));
2444
2445         /*
2446          * For recursive snapshot renames the parent won't be changing
2447          * so we just pass name for both the to/from argument.
2448          */
2449         err = zfs_secpolicy_rename_perms(snapname, snapname, CRED());
2450         if (err != 0) {
2451                 strfree(snapname);
2452                 return (err == ENOENT ? 0 : err);
2453         }
2454
2455 #ifdef _KERNEL
2456         /*
2457          * For all filesystems undergoing rename, we'll need to unmount it.
2458          */
2459         (void) zfs_unmount_snap(snapname, NULL);
2460 #endif
2461         err = dsl_dataset_hold(snapname, ra->dstg, &ds);
2462         strfree(snapname);
2463         if (err != 0)
2464                 return (err == ENOENT ? 0 : err);
2465
2466         dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2467             dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2468
2469         return (0);
2470 }
2471
2472 static int
2473 dsl_recursive_rename(char *oldname, const char *newname)
2474 {
2475         int err;
2476         struct renamesnaparg *ra;
2477         dsl_sync_task_t *dst;
2478         spa_t *spa;
2479         char *cp, *fsname = spa_strdup(oldname);
2480         int len = strlen(oldname) + 1;
2481
2482         /* truncate the snapshot name to get the fsname */
2483         cp = strchr(fsname, '@');
2484         *cp = '\0';
2485
2486         err = spa_open(fsname, &spa, FTAG);
2487         if (err) {
2488                 kmem_free(fsname, len);
2489                 return (err);
2490         }
2491         ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2492         ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2493
2494         ra->oldsnap = strchr(oldname, '@') + 1;
2495         ra->newsnap = strchr(newname, '@') + 1;
2496         *ra->failed = '\0';
2497
2498         err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2499             DS_FIND_CHILDREN);
2500         kmem_free(fsname, len);
2501
2502         if (err == 0) {
2503                 err = dsl_sync_task_group_wait(ra->dstg);
2504         }
2505
2506         for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2507             dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2508                 dsl_dataset_t *ds = dst->dst_arg1;
2509                 if (dst->dst_err) {
2510                         dsl_dir_name(ds->ds_dir, ra->failed);
2511                         (void) strlcat(ra->failed, "@", sizeof (ra->failed));
2512                         (void) strlcat(ra->failed, ra->newsnap,
2513                             sizeof (ra->failed));
2514                 }
2515                 dsl_dataset_rele(ds, ra->dstg);
2516         }
2517
2518         if (err)
2519                 (void) strlcpy(oldname, ra->failed, sizeof (ra->failed));
2520
2521         dsl_sync_task_group_destroy(ra->dstg);
2522         kmem_free(ra, sizeof (struct renamesnaparg));
2523         spa_close(spa, FTAG);
2524         return (err);
2525 }
2526
2527 static int
2528 dsl_valid_rename(const char *oldname, void *arg)
2529 {
2530         int delta = *(int *)arg;
2531
2532         if (strlen(oldname) + delta >= MAXNAMELEN)
2533                 return (ENAMETOOLONG);
2534
2535         return (0);
2536 }
2537
2538 #pragma weak dmu_objset_rename = dsl_dataset_rename
2539 int
2540 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2541 {
2542         dsl_dir_t *dd;
2543         dsl_dataset_t *ds;
2544         const char *tail;
2545         int err;
2546
2547         err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2548         if (err)
2549                 return (err);
2550
2551         if (tail == NULL) {
2552                 int delta = strlen(newname) - strlen(oldname);
2553
2554                 /* if we're growing, validate child name lengths */
2555                 if (delta > 0)
2556                         err = dmu_objset_find(oldname, dsl_valid_rename,
2557                             &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2558
2559                 if (err == 0)
2560                         err = dsl_dir_rename(dd, newname);
2561                 dsl_dir_close(dd, FTAG);
2562                 return (err);
2563         }
2564
2565         if (tail[0] != '@') {
2566                 /* the name ended in a nonexistent component */
2567                 dsl_dir_close(dd, FTAG);
2568                 return (ENOENT);
2569         }
2570
2571         dsl_dir_close(dd, FTAG);
2572
2573         /* new name must be snapshot in same filesystem */
2574         tail = strchr(newname, '@');
2575         if (tail == NULL)
2576                 return (EINVAL);
2577         tail++;
2578         if (strncmp(oldname, newname, tail - newname) != 0)
2579                 return (EXDEV);
2580
2581         if (recursive) {
2582                 err = dsl_recursive_rename(oldname, newname);
2583         } else {
2584                 err = dsl_dataset_hold(oldname, FTAG, &ds);
2585                 if (err)
2586                         return (err);
2587
2588                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2589                     dsl_dataset_snapshot_rename_check,
2590                     dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2591
2592                 dsl_dataset_rele(ds, FTAG);
2593         }
2594
2595         return (err);
2596 }
2597
2598 struct promotenode {
2599         list_node_t link;
2600         dsl_dataset_t *ds;
2601 };
2602
2603 struct promotearg {
2604         list_t shared_snaps, origin_snaps, clone_snaps;
2605         dsl_dataset_t *origin_origin;
2606         uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2607         char *err_ds;
2608 };
2609
2610 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2611
2612 static int
2613 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2614 {
2615         dsl_dataset_t *hds = arg1;
2616         struct promotearg *pa = arg2;
2617         struct promotenode *snap = list_head(&pa->shared_snaps);
2618         dsl_dataset_t *origin_ds = snap->ds;
2619         int err;
2620         uint64_t unused;
2621
2622         /* Check that it is a real clone */
2623         if (!dsl_dir_is_clone(hds->ds_dir))
2624                 return (EINVAL);
2625
2626         /* Since this is so expensive, don't do the preliminary check */
2627         if (!dmu_tx_is_syncing(tx))
2628                 return (0);
2629
2630         if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2631                 return (EXDEV);
2632
2633         /* compute origin's new unique space */
2634         snap = list_tail(&pa->clone_snaps);
2635         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2636         dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2637             origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
2638             &pa->unique, &unused, &unused);
2639
2640         /*
2641          * Walk the snapshots that we are moving
2642          *
2643          * Compute space to transfer.  Consider the incremental changes
2644          * to used for each snapshot:
2645          * (my used) = (prev's used) + (blocks born) - (blocks killed)
2646          * So each snapshot gave birth to:
2647          * (blocks born) = (my used) - (prev's used) + (blocks killed)
2648          * So a sequence would look like:
2649          * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2650          * Which simplifies to:
2651          * uN + kN + kN-1 + ... + k1 + k0
2652          * Note however, if we stop before we reach the ORIGIN we get:
2653          * uN + kN + kN-1 + ... + kM - uM-1
2654          */
2655         pa->used = origin_ds->ds_phys->ds_used_bytes;
2656         pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2657         pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2658         for (snap = list_head(&pa->shared_snaps); snap;
2659             snap = list_next(&pa->shared_snaps, snap)) {
2660                 uint64_t val, dlused, dlcomp, dluncomp;
2661                 dsl_dataset_t *ds = snap->ds;
2662
2663                 /* Check that the snapshot name does not conflict */
2664                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2665                 err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2666                 if (err == 0) {
2667                         err = EEXIST;
2668                         goto out;
2669                 }
2670                 if (err != ENOENT)
2671                         goto out;
2672
2673                 /* The very first snapshot does not have a deadlist */
2674                 if (ds->ds_phys->ds_prev_snap_obj == 0)
2675                         continue;
2676
2677                 dsl_deadlist_space(&ds->ds_deadlist,
2678                     &dlused, &dlcomp, &dluncomp);
2679                 pa->used += dlused;
2680                 pa->comp += dlcomp;
2681                 pa->uncomp += dluncomp;
2682         }
2683
2684         /*
2685          * If we are a clone of a clone then we never reached ORIGIN,
2686          * so we need to subtract out the clone origin's used space.
2687          */
2688         if (pa->origin_origin) {
2689                 pa->used -= pa->origin_origin->ds_phys->ds_used_bytes;
2690                 pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2691                 pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2692         }
2693
2694         /* Check that there is enough space here */
2695         err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2696             pa->used);
2697         if (err)
2698                 return (err);
2699
2700         /*
2701          * Compute the amounts of space that will be used by snapshots
2702          * after the promotion (for both origin and clone).  For each,
2703          * it is the amount of space that will be on all of their
2704          * deadlists (that was not born before their new origin).
2705          */
2706         if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2707                 uint64_t space;
2708
2709                 /*
2710                  * Note, typically this will not be a clone of a clone,
2711                  * so dd_origin_txg will be < TXG_INITIAL, so
2712                  * these snaplist_space() -> dsl_deadlist_space_range()
2713                  * calls will be fast because they do not have to
2714                  * iterate over all bps.
2715                  */
2716                 snap = list_head(&pa->origin_snaps);
2717                 err = snaplist_space(&pa->shared_snaps,
2718                     snap->ds->ds_dir->dd_origin_txg, &pa->cloneusedsnap);
2719                 if (err)
2720                         return (err);
2721
2722                 err = snaplist_space(&pa->clone_snaps,
2723                     snap->ds->ds_dir->dd_origin_txg, &space);
2724                 if (err)
2725                         return (err);
2726                 pa->cloneusedsnap += space;
2727         }
2728         if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2729                 err = snaplist_space(&pa->origin_snaps,
2730                     origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2731                 if (err)
2732                         return (err);
2733         }
2734
2735         return (0);
2736 out:
2737         pa->err_ds =  snap->ds->ds_snapname;
2738         return (err);
2739 }
2740
2741 static void
2742 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2743 {
2744         dsl_dataset_t *hds = arg1;
2745         struct promotearg *pa = arg2;
2746         struct promotenode *snap = list_head(&pa->shared_snaps);
2747         dsl_dataset_t *origin_ds = snap->ds;
2748         dsl_dataset_t *origin_head;
2749         dsl_dir_t *dd = hds->ds_dir;
2750         dsl_pool_t *dp = hds->ds_dir->dd_pool;
2751         dsl_dir_t *odd = NULL;
2752         uint64_t oldnext_obj;
2753         int64_t delta;
2754
2755         ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2756
2757         snap = list_head(&pa->origin_snaps);
2758         origin_head = snap->ds;
2759
2760         /*
2761          * We need to explicitly open odd, since origin_ds's dd will be
2762          * changing.
2763          */
2764         VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2765             NULL, FTAG, &odd));
2766
2767         /* change origin's next snap */
2768         dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2769         oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2770         snap = list_tail(&pa->clone_snaps);
2771         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2772         origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2773
2774         /* change the origin's next clone */
2775         if (origin_ds->ds_phys->ds_next_clones_obj) {
2776                 remove_from_next_clones(origin_ds, snap->ds->ds_object, tx);
2777                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2778                     origin_ds->ds_phys->ds_next_clones_obj,
2779                     oldnext_obj, tx));
2780         }
2781
2782         /* change origin */
2783         dmu_buf_will_dirty(dd->dd_dbuf, tx);
2784         ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2785         dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2786         dd->dd_origin_txg = origin_head->ds_dir->dd_origin_txg;
2787         dmu_buf_will_dirty(odd->dd_dbuf, tx);
2788         odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2789         origin_head->ds_dir->dd_origin_txg =
2790             origin_ds->ds_phys->ds_creation_txg;
2791
2792         /* change dd_clone entries */
2793         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2794                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2795                     odd->dd_phys->dd_clones, hds->ds_object, tx));
2796                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2797                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2798                     hds->ds_object, tx));
2799
2800                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2801                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2802                     origin_head->ds_object, tx));
2803                 if (dd->dd_phys->dd_clones == 0) {
2804                         dd->dd_phys->dd_clones = zap_create(dp->dp_meta_objset,
2805                             DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
2806                 }
2807                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2808                     dd->dd_phys->dd_clones, origin_head->ds_object, tx));
2809
2810         }
2811
2812         /* move snapshots to this dir */
2813         for (snap = list_head(&pa->shared_snaps); snap;
2814             snap = list_next(&pa->shared_snaps, snap)) {
2815                 dsl_dataset_t *ds = snap->ds;
2816
2817                 /* unregister props as dsl_dir is changing */
2818                 if (ds->ds_objset) {
2819                         dmu_objset_evict(ds->ds_objset);
2820                         ds->ds_objset = NULL;
2821                 }
2822                 /* move snap name entry */
2823                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2824                 VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2825                     ds->ds_snapname, tx));
2826                 VERIFY(0 == zap_add(dp->dp_meta_objset,
2827                     hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2828                     8, 1, &ds->ds_object, tx));
2829
2830                 /* change containing dsl_dir */
2831                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2832                 ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2833                 ds->ds_phys->ds_dir_obj = dd->dd_object;
2834                 ASSERT3P(ds->ds_dir, ==, odd);
2835                 dsl_dir_close(ds->ds_dir, ds);
2836                 VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2837                     NULL, ds, &ds->ds_dir));
2838
2839                 /* move any clone references */
2840                 if (ds->ds_phys->ds_next_clones_obj &&
2841                     spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2842                         zap_cursor_t zc;
2843                         zap_attribute_t za;
2844
2845                         for (zap_cursor_init(&zc, dp->dp_meta_objset,
2846                             ds->ds_phys->ds_next_clones_obj);
2847                             zap_cursor_retrieve(&zc, &za) == 0;
2848                             zap_cursor_advance(&zc)) {
2849                                 dsl_dataset_t *cnds;
2850                                 uint64_t o;
2851
2852                                 if (za.za_first_integer == oldnext_obj) {
2853                                         /*
2854                                          * We've already moved the
2855                                          * origin's reference.
2856                                          */
2857                                         continue;
2858                                 }
2859
2860                                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
2861                                     za.za_first_integer, FTAG, &cnds));
2862                                 o = cnds->ds_dir->dd_phys->dd_head_dataset_obj;
2863
2864                                 VERIFY3U(zap_remove_int(dp->dp_meta_objset,
2865                                     odd->dd_phys->dd_clones, o, tx), ==, 0);
2866                                 VERIFY3U(zap_add_int(dp->dp_meta_objset,
2867                                     dd->dd_phys->dd_clones, o, tx), ==, 0);
2868                                 dsl_dataset_rele(cnds, FTAG);
2869                         }
2870                         zap_cursor_fini(&zc);
2871                 }
2872
2873                 ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2874         }
2875
2876         /*
2877          * Change space accounting.
2878          * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2879          * both be valid, or both be 0 (resulting in delta == 0).  This
2880          * is true for each of {clone,origin} independently.
2881          */
2882
2883         delta = pa->cloneusedsnap -
2884             dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2885         ASSERT3S(delta, >=, 0);
2886         ASSERT3U(pa->used, >=, delta);
2887         dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2888         dsl_dir_diduse_space(dd, DD_USED_HEAD,
2889             pa->used - delta, pa->comp, pa->uncomp, tx);
2890
2891         delta = pa->originusedsnap -
2892             odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2893         ASSERT3S(delta, <=, 0);
2894         ASSERT3U(pa->used, >=, -delta);
2895         dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2896         dsl_dir_diduse_space(odd, DD_USED_HEAD,
2897             -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2898
2899         origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2900
2901         /* log history record */
2902         spa_history_log_internal(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2903             "dataset = %llu", hds->ds_object);
2904
2905         dsl_dir_close(odd, FTAG);
2906 }
2907
2908 static char *snaplist_tag = "snaplist";
2909 /*
2910  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2911  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2912  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2913  * snapshots back to this dataset's origin.
2914  */
2915 static int
2916 snaplist_make(dsl_pool_t *dp, boolean_t own,
2917     uint64_t first_obj, uint64_t last_obj, list_t *l)
2918 {
2919         uint64_t obj = last_obj;
2920
2921         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2922
2923         list_create(l, sizeof (struct promotenode),
2924             offsetof(struct promotenode, link));
2925
2926         while (obj != first_obj) {
2927                 dsl_dataset_t *ds;
2928                 struct promotenode *snap;
2929                 int err;
2930
2931                 if (own) {
2932                         err = dsl_dataset_own_obj(dp, obj,
2933                             0, snaplist_tag, &ds);
2934                         if (err == 0)
2935                                 dsl_dataset_make_exclusive(ds, snaplist_tag);
2936                 } else {
2937                         err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2938                 }
2939                 if (err == ENOENT) {
2940                         /* lost race with snapshot destroy */
2941                         struct promotenode *last = list_tail(l);
2942                         ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2943                         obj = last->ds->ds_phys->ds_prev_snap_obj;
2944                         continue;
2945                 } else if (err) {
2946                         return (err);
2947                 }
2948
2949                 if (first_obj == 0)
2950                         first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2951
2952                 snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
2953                 snap->ds = ds;
2954                 list_insert_tail(l, snap);
2955                 obj = ds->ds_phys->ds_prev_snap_obj;
2956         }
2957
2958         return (0);
2959 }
2960
2961 static int
2962 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
2963 {
2964         struct promotenode *snap;
2965
2966         *spacep = 0;
2967         for (snap = list_head(l); snap; snap = list_next(l, snap)) {
2968                 uint64_t used, comp, uncomp;
2969                 dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2970                     mintxg, UINT64_MAX, &used, &comp, &uncomp);
2971                 *spacep += used;
2972         }
2973         return (0);
2974 }
2975
2976 static void
2977 snaplist_destroy(list_t *l, boolean_t own)
2978 {
2979         struct promotenode *snap;
2980
2981         if (!l || !list_link_active(&l->list_head))
2982                 return;
2983
2984         while ((snap = list_tail(l)) != NULL) {
2985                 list_remove(l, snap);
2986                 if (own)
2987                         dsl_dataset_disown(snap->ds, snaplist_tag);
2988                 else
2989                         dsl_dataset_rele(snap->ds, snaplist_tag);
2990                 kmem_free(snap, sizeof (struct promotenode));
2991         }
2992         list_destroy(l);
2993 }
2994
2995 /*
2996  * Promote a clone.  Nomenclature note:
2997  * "clone" or "cds": the original clone which is being promoted
2998  * "origin" or "ods": the snapshot which is originally clone's origin
2999  * "origin head" or "ohds": the dataset which is the head
3000  * (filesystem/volume) for the origin
3001  * "origin origin": the origin of the origin's filesystem (typically
3002  * NULL, indicating that the clone is not a clone of a clone).
3003  */
3004 int
3005 dsl_dataset_promote(const char *name, char *conflsnap)
3006 {
3007         dsl_dataset_t *ds;
3008         dsl_dir_t *dd;
3009         dsl_pool_t *dp;
3010         dmu_object_info_t doi;
3011         struct promotearg pa;
3012         struct promotenode *snap;
3013         int err;
3014
3015         bzero(&pa, sizeof(struct promotearg));
3016         err = dsl_dataset_hold(name, FTAG, &ds);
3017         if (err)
3018                 return (err);
3019         dd = ds->ds_dir;
3020         dp = dd->dd_pool;
3021
3022         err = dmu_object_info(dp->dp_meta_objset,
3023             ds->ds_phys->ds_snapnames_zapobj, &doi);
3024         if (err) {
3025                 dsl_dataset_rele(ds, FTAG);
3026                 return (err);
3027         }
3028
3029         if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
3030                 dsl_dataset_rele(ds, FTAG);
3031                 return (EINVAL);
3032         }
3033
3034         /*
3035          * We are going to inherit all the snapshots taken before our
3036          * origin (i.e., our new origin will be our parent's origin).
3037          * Take ownership of them so that we can rename them into our
3038          * namespace.
3039          */
3040         rw_enter(&dp->dp_config_rwlock, RW_READER);
3041
3042         err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
3043             &pa.shared_snaps);
3044         if (err != 0)
3045                 goto out;
3046
3047         err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
3048         if (err != 0)
3049                 goto out;
3050
3051         snap = list_head(&pa.shared_snaps);
3052         ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
3053         err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
3054             snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
3055         if (err != 0)
3056                 goto out;
3057
3058         if (snap->ds->ds_dir->dd_phys->dd_origin_obj != 0) {
3059                 err = dsl_dataset_hold_obj(dp,
3060                     snap->ds->ds_dir->dd_phys->dd_origin_obj,
3061                     FTAG, &pa.origin_origin);
3062                 if (err != 0)
3063                         goto out;
3064         }
3065
3066 out:
3067         rw_exit(&dp->dp_config_rwlock);
3068
3069         /*
3070          * Add in 128x the snapnames zapobj size, since we will be moving
3071          * a bunch of snapnames to the promoted ds, and dirtying their
3072          * bonus buffers.
3073          */
3074         if (err == 0) {
3075                 err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
3076                     dsl_dataset_promote_sync, ds, &pa,
3077                     2 + 2 * doi.doi_physical_blocks_512);
3078                 if (err && pa.err_ds && conflsnap)
3079                         (void) strncpy(conflsnap, pa.err_ds, MAXNAMELEN);
3080         }
3081
3082         snaplist_destroy(&pa.shared_snaps, B_TRUE);
3083         snaplist_destroy(&pa.clone_snaps, B_FALSE);
3084         snaplist_destroy(&pa.origin_snaps, B_FALSE);
3085         if (pa.origin_origin)
3086                 dsl_dataset_rele(pa.origin_origin, FTAG);
3087         dsl_dataset_rele(ds, FTAG);
3088         return (err);
3089 }
3090
3091 struct cloneswaparg {
3092         dsl_dataset_t *cds; /* clone dataset */
3093         dsl_dataset_t *ohds; /* origin's head dataset */
3094         boolean_t force;
3095         int64_t unused_refres_delta; /* change in unconsumed refreservation */
3096 };
3097
3098 /* ARGSUSED */
3099 static int
3100 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
3101 {
3102         struct cloneswaparg *csa = arg1;
3103
3104         /* they should both be heads */
3105         if (dsl_dataset_is_snapshot(csa->cds) ||
3106             dsl_dataset_is_snapshot(csa->ohds))
3107                 return (EINVAL);
3108
3109         /* the branch point should be just before them */
3110         if (csa->cds->ds_prev != csa->ohds->ds_prev)
3111                 return (EINVAL);
3112
3113         /* cds should be the clone (unless they are unrelated) */
3114         if (csa->cds->ds_prev != NULL &&
3115             csa->cds->ds_prev != csa->cds->ds_dir->dd_pool->dp_origin_snap &&
3116             csa->ohds->ds_object !=
3117             csa->cds->ds_prev->ds_phys->ds_next_snap_obj)
3118                 return (EINVAL);
3119
3120         /* the clone should be a child of the origin */
3121         if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
3122                 return (EINVAL);
3123
3124         /* ohds shouldn't be modified unless 'force' */
3125         if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
3126                 return (ETXTBSY);
3127
3128         /* adjust amount of any unconsumed refreservation */
3129         csa->unused_refres_delta =
3130             (int64_t)MIN(csa->ohds->ds_reserved,
3131             csa->ohds->ds_phys->ds_unique_bytes) -
3132             (int64_t)MIN(csa->ohds->ds_reserved,
3133             csa->cds->ds_phys->ds_unique_bytes);
3134
3135         if (csa->unused_refres_delta > 0 &&
3136             csa->unused_refres_delta >
3137             dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3138                 return (ENOSPC);
3139
3140         if (csa->ohds->ds_quota != 0 &&
3141             csa->cds->ds_phys->ds_unique_bytes > csa->ohds->ds_quota)
3142                 return (EDQUOT);
3143
3144         return (0);
3145 }
3146
3147 /* ARGSUSED */
3148 static void
3149 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3150 {
3151         struct cloneswaparg *csa = arg1;
3152         dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3153
3154         ASSERT(csa->cds->ds_reserved == 0);
3155         ASSERT(csa->ohds->ds_quota == 0 ||
3156             csa->cds->ds_phys->ds_unique_bytes <= csa->ohds->ds_quota);
3157
3158         dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3159         dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3160
3161         if (csa->cds->ds_objset != NULL) {
3162                 dmu_objset_evict(csa->cds->ds_objset);
3163                 csa->cds->ds_objset = NULL;
3164         }
3165
3166         if (csa->ohds->ds_objset != NULL) {
3167                 dmu_objset_evict(csa->ohds->ds_objset);
3168                 csa->ohds->ds_objset = NULL;
3169         }
3170
3171         /*
3172          * Reset origin's unique bytes, if it exists.
3173          */
3174         if (csa->cds->ds_prev) {
3175                 dsl_dataset_t *origin = csa->cds->ds_prev;
3176                 uint64_t comp, uncomp;
3177
3178                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
3179                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3180                     origin->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3181                     &origin->ds_phys->ds_unique_bytes, &comp, &uncomp);
3182         }
3183
3184         /* swap blkptrs */
3185         {
3186                 blkptr_t tmp;
3187                 tmp = csa->ohds->ds_phys->ds_bp;
3188                 csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3189                 csa->cds->ds_phys->ds_bp = tmp;
3190         }
3191
3192         /* set dd_*_bytes */
3193         {
3194                 int64_t dused, dcomp, duncomp;
3195                 uint64_t cdl_used, cdl_comp, cdl_uncomp;
3196                 uint64_t odl_used, odl_comp, odl_uncomp;
3197
3198                 ASSERT3U(csa->cds->ds_dir->dd_phys->
3199                     dd_used_breakdown[DD_USED_SNAP], ==, 0);
3200
3201                 dsl_deadlist_space(&csa->cds->ds_deadlist,
3202                     &cdl_used, &cdl_comp, &cdl_uncomp);
3203                 dsl_deadlist_space(&csa->ohds->ds_deadlist,
3204                     &odl_used, &odl_comp, &odl_uncomp);
3205
3206                 dused = csa->cds->ds_phys->ds_used_bytes + cdl_used -
3207                     (csa->ohds->ds_phys->ds_used_bytes + odl_used);
3208                 dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3209                     (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3210                 duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3211                     cdl_uncomp -
3212                     (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3213
3214                 dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3215                     dused, dcomp, duncomp, tx);
3216                 dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3217                     -dused, -dcomp, -duncomp, tx);
3218
3219                 /*
3220                  * The difference in the space used by snapshots is the
3221                  * difference in snapshot space due to the head's
3222                  * deadlist (since that's the only thing that's
3223                  * changing that affects the snapused).
3224                  */
3225                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3226                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3227                     &cdl_used, &cdl_comp, &cdl_uncomp);
3228                 dsl_deadlist_space_range(&csa->ohds->ds_deadlist,
3229                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3230                     &odl_used, &odl_comp, &odl_uncomp);
3231                 dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3232                     DD_USED_HEAD, DD_USED_SNAP, tx);
3233         }
3234
3235         /* swap ds_*_bytes */
3236         SWITCH64(csa->ohds->ds_phys->ds_used_bytes,
3237             csa->cds->ds_phys->ds_used_bytes);
3238         SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3239             csa->cds->ds_phys->ds_compressed_bytes);
3240         SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3241             csa->cds->ds_phys->ds_uncompressed_bytes);
3242         SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3243             csa->cds->ds_phys->ds_unique_bytes);
3244
3245         /* apply any parent delta for change in unconsumed refreservation */
3246         dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3247             csa->unused_refres_delta, 0, 0, tx);
3248
3249         /*
3250          * Swap deadlists.
3251          */
3252         dsl_deadlist_close(&csa->cds->ds_deadlist);
3253         dsl_deadlist_close(&csa->ohds->ds_deadlist);
3254         SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3255             csa->cds->ds_phys->ds_deadlist_obj);
3256         dsl_deadlist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3257             csa->cds->ds_phys->ds_deadlist_obj);
3258         dsl_deadlist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3259             csa->ohds->ds_phys->ds_deadlist_obj);
3260
3261         dsl_scan_ds_clone_swapped(csa->ohds, csa->cds, tx);
3262 }
3263
3264 /*
3265  * Swap 'clone' with its origin head datasets.  Used at the end of "zfs
3266  * recv" into an existing fs to swizzle the file system to the new
3267  * version, and by "zfs rollback".  Can also be used to swap two
3268  * independent head datasets if neither has any snapshots.
3269  */
3270 int
3271 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3272     boolean_t force)
3273 {
3274         struct cloneswaparg csa;
3275         int error;
3276
3277         ASSERT(clone->ds_owner);
3278         ASSERT(origin_head->ds_owner);
3279 retry:
3280         /*
3281          * Need exclusive access for the swap. If we're swapping these
3282          * datasets back after an error, we already hold the locks.
3283          */
3284         if (!RW_WRITE_HELD(&clone->ds_rwlock))
3285                 rw_enter(&clone->ds_rwlock, RW_WRITER);
3286         if (!RW_WRITE_HELD(&origin_head->ds_rwlock) &&
3287             !rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3288                 rw_exit(&clone->ds_rwlock);
3289                 rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3290                 if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3291                         rw_exit(&origin_head->ds_rwlock);
3292                         goto retry;
3293                 }
3294         }
3295         csa.cds = clone;
3296         csa.ohds = origin_head;
3297         csa.force = force;
3298         error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3299             dsl_dataset_clone_swap_check,
3300             dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3301         return (error);
3302 }
3303
3304 /*
3305  * Given a pool name and a dataset object number in that pool,
3306  * return the name of that dataset.
3307  */
3308 int
3309 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3310 {
3311         spa_t *spa;
3312         dsl_pool_t *dp;
3313         dsl_dataset_t *ds;
3314         int error;
3315
3316         if ((error = spa_open(pname, &spa, FTAG)) != 0)
3317                 return (error);
3318         dp = spa_get_dsl(spa);
3319         rw_enter(&dp->dp_config_rwlock, RW_READER);
3320         if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3321                 dsl_dataset_name(ds, buf);
3322                 dsl_dataset_rele(ds, FTAG);
3323         }
3324         rw_exit(&dp->dp_config_rwlock);
3325         spa_close(spa, FTAG);
3326
3327         return (error);
3328 }
3329
3330 int
3331 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3332     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3333 {
3334         int error = 0;
3335
3336         ASSERT3S(asize, >, 0);
3337
3338         /*
3339          * *ref_rsrv is the portion of asize that will come from any
3340          * unconsumed refreservation space.
3341          */
3342         *ref_rsrv = 0;
3343
3344         mutex_enter(&ds->ds_lock);
3345         /*
3346          * Make a space adjustment for reserved bytes.
3347          */
3348         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3349                 ASSERT3U(*used, >=,
3350                     ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3351                 *used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3352                 *ref_rsrv =
3353                     asize - MIN(asize, parent_delta(ds, asize + inflight));
3354         }
3355
3356         if (!check_quota || ds->ds_quota == 0) {
3357                 mutex_exit(&ds->ds_lock);
3358                 return (0);
3359         }
3360         /*
3361          * If they are requesting more space, and our current estimate
3362          * is over quota, they get to try again unless the actual
3363          * on-disk is over quota and there are no pending changes (which
3364          * may free up space for us).
3365          */
3366         if (ds->ds_phys->ds_used_bytes + inflight >= ds->ds_quota) {
3367                 if (inflight > 0 || ds->ds_phys->ds_used_bytes < ds->ds_quota)
3368                         error = ERESTART;
3369                 else
3370                         error = EDQUOT;
3371
3372                 DMU_TX_STAT_BUMP(dmu_tx_quota);
3373         }
3374         mutex_exit(&ds->ds_lock);
3375
3376         return (error);
3377 }
3378
3379 /* ARGSUSED */
3380 static int
3381 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3382 {
3383         dsl_dataset_t *ds = arg1;
3384         dsl_prop_setarg_t *psa = arg2;
3385         int err;
3386
3387         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3388                 return (ENOTSUP);
3389
3390         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3391                 return (err);
3392
3393         if (psa->psa_effective_value == 0)
3394                 return (0);
3395
3396         if (psa->psa_effective_value < ds->ds_phys->ds_used_bytes ||
3397             psa->psa_effective_value < ds->ds_reserved)
3398                 return (ENOSPC);
3399
3400         return (0);
3401 }
3402
3403 extern void dsl_prop_set_sync(void *, void *, dmu_tx_t *);
3404
3405 void
3406 dsl_dataset_set_quota_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3407 {
3408         dsl_dataset_t *ds = arg1;
3409         dsl_prop_setarg_t *psa = arg2;
3410         uint64_t effective_value = psa->psa_effective_value;
3411
3412         dsl_prop_set_sync(ds, psa, tx);
3413         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3414
3415         if (ds->ds_quota != effective_value) {
3416                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3417                 ds->ds_quota = effective_value;
3418         }
3419 }
3420
3421 int
3422 dsl_dataset_set_quota(const char *dsname, zprop_source_t source, uint64_t quota)
3423 {
3424         dsl_dataset_t *ds;
3425         dsl_prop_setarg_t psa;
3426         int err;
3427
3428         dsl_prop_setarg_init_uint64(&psa, "refquota", source, &quota);
3429
3430         err = dsl_dataset_hold(dsname, FTAG, &ds);
3431         if (err)
3432                 return (err);
3433
3434         /*
3435          * If someone removes a file, then tries to set the quota, we
3436          * want to make sure the file freeing takes effect.
3437          */
3438         txg_wait_open(ds->ds_dir->dd_pool, 0);
3439
3440         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3441             dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3442             ds, &psa, 0);
3443
3444         dsl_dataset_rele(ds, FTAG);
3445         return (err);
3446 }
3447
3448 static int
3449 dsl_dataset_set_reservation_check(void *arg1, void *arg2, dmu_tx_t *tx)
3450 {
3451         dsl_dataset_t *ds = arg1;
3452         dsl_prop_setarg_t *psa = arg2;
3453         uint64_t effective_value;
3454         uint64_t unique;
3455         int err;
3456
3457         if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3458             SPA_VERSION_REFRESERVATION)
3459                 return (ENOTSUP);
3460
3461         if (dsl_dataset_is_snapshot(ds))
3462                 return (EINVAL);
3463
3464         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3465                 return (err);
3466
3467         effective_value = psa->psa_effective_value;
3468
3469         /*
3470          * If we are doing the preliminary check in open context, the
3471          * space estimates may be inaccurate.
3472          */
3473         if (!dmu_tx_is_syncing(tx))
3474                 return (0);
3475
3476         mutex_enter(&ds->ds_lock);
3477         if (!DS_UNIQUE_IS_ACCURATE(ds))
3478                 dsl_dataset_recalc_head_uniq(ds);
3479         unique = ds->ds_phys->ds_unique_bytes;
3480         mutex_exit(&ds->ds_lock);
3481
3482         if (MAX(unique, effective_value) > MAX(unique, ds->ds_reserved)) {
3483                 uint64_t delta = MAX(unique, effective_value) -
3484                     MAX(unique, ds->ds_reserved);
3485
3486                 if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3487                         return (ENOSPC);
3488                 if (ds->ds_quota > 0 &&
3489                     effective_value > ds->ds_quota)
3490                         return (ENOSPC);
3491         }
3492
3493         return (0);
3494 }
3495
3496 static void
3497 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3498 {
3499         dsl_dataset_t *ds = arg1;
3500         dsl_prop_setarg_t *psa = arg2;
3501         uint64_t effective_value = psa->psa_effective_value;
3502         uint64_t unique;
3503         int64_t delta;
3504
3505         dsl_prop_set_sync(ds, psa, tx);
3506         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3507
3508         dmu_buf_will_dirty(ds->ds_dbuf, tx);
3509
3510         mutex_enter(&ds->ds_dir->dd_lock);
3511         mutex_enter(&ds->ds_lock);
3512         ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
3513         unique = ds->ds_phys->ds_unique_bytes;
3514         delta = MAX(0, (int64_t)(effective_value - unique)) -
3515             MAX(0, (int64_t)(ds->ds_reserved - unique));
3516         ds->ds_reserved = effective_value;
3517         mutex_exit(&ds->ds_lock);
3518
3519         dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3520         mutex_exit(&ds->ds_dir->dd_lock);
3521 }
3522
3523 int
3524 dsl_dataset_set_reservation(const char *dsname, zprop_source_t source,
3525     uint64_t reservation)
3526 {
3527         dsl_dataset_t *ds;
3528         dsl_prop_setarg_t psa;
3529         int err;
3530
3531         dsl_prop_setarg_init_uint64(&psa, "refreservation", source,
3532             &reservation);
3533
3534         err = dsl_dataset_hold(dsname, FTAG, &ds);
3535         if (err)
3536                 return (err);
3537
3538         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3539             dsl_dataset_set_reservation_check,
3540             dsl_dataset_set_reservation_sync, ds, &psa, 0);
3541
3542         dsl_dataset_rele(ds, FTAG);
3543         return (err);
3544 }
3545
3546 typedef struct zfs_hold_cleanup_arg {
3547         dsl_pool_t *dp;
3548         uint64_t dsobj;
3549         char htag[MAXNAMELEN];
3550 } zfs_hold_cleanup_arg_t;
3551
3552 static void
3553 dsl_dataset_user_release_onexit(void *arg)
3554 {
3555         zfs_hold_cleanup_arg_t *ca = arg;
3556
3557         (void) dsl_dataset_user_release_tmp(ca->dp, ca->dsobj, ca->htag,
3558             B_TRUE);
3559         kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t));
3560 }
3561
3562 void
3563 dsl_register_onexit_hold_cleanup(dsl_dataset_t *ds, const char *htag,
3564     minor_t minor)
3565 {
3566         zfs_hold_cleanup_arg_t *ca;
3567
3568         ca = kmem_alloc(sizeof (zfs_hold_cleanup_arg_t), KM_SLEEP);
3569         ca->dp = ds->ds_dir->dd_pool;
3570         ca->dsobj = ds->ds_object;
3571         (void) strlcpy(ca->htag, htag, sizeof (ca->htag));
3572         VERIFY3U(0, ==, zfs_onexit_add_cb(minor,
3573             dsl_dataset_user_release_onexit, ca, NULL));
3574 }
3575
3576 /*
3577  * If you add new checks here, you may need to add
3578  * additional checks to the "temporary" case in
3579  * snapshot_check() in dmu_objset.c.
3580  */
3581 static int
3582 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3583 {
3584         dsl_dataset_t *ds = arg1;
3585         struct dsl_ds_holdarg *ha = arg2;
3586         char *htag = ha->htag;
3587         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3588         int error = 0;
3589
3590         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3591                 return (ENOTSUP);
3592
3593         if (!dsl_dataset_is_snapshot(ds))
3594                 return (EINVAL);
3595
3596         /* tags must be unique */
3597         mutex_enter(&ds->ds_lock);
3598         if (ds->ds_phys->ds_userrefs_obj) {
3599                 error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3600                     8, 1, tx);
3601                 if (error == 0)
3602                         error = EEXIST;
3603                 else if (error == ENOENT)
3604                         error = 0;
3605         }
3606         mutex_exit(&ds->ds_lock);
3607
3608         if (error == 0 && ha->temphold &&
3609             strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
3610                 error = E2BIG;
3611
3612         return (error);
3613 }
3614
3615 void
3616 dsl_dataset_user_hold_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3617 {
3618         dsl_dataset_t *ds = arg1;
3619         struct dsl_ds_holdarg *ha = arg2;
3620         char *htag = ha->htag;
3621         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3622         objset_t *mos = dp->dp_meta_objset;
3623         uint64_t now = gethrestime_sec();
3624         uint64_t zapobj;
3625
3626         mutex_enter(&ds->ds_lock);
3627         if (ds->ds_phys->ds_userrefs_obj == 0) {
3628                 /*
3629                  * This is the first user hold for this dataset.  Create
3630                  * the userrefs zap object.
3631                  */
3632                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3633                 zapobj = ds->ds_phys->ds_userrefs_obj =
3634                     zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3635         } else {
3636                 zapobj = ds->ds_phys->ds_userrefs_obj;
3637         }
3638         ds->ds_userrefs++;
3639         mutex_exit(&ds->ds_lock);
3640
3641         VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3642
3643         if (ha->temphold) {
3644                 VERIFY(0 == dsl_pool_user_hold(dp, ds->ds_object,
3645                     htag, &now, tx));
3646         }
3647
3648         spa_history_log_internal(LOG_DS_USER_HOLD,
3649             dp->dp_spa, tx, "<%s> temp = %d dataset = %llu", htag,
3650             (int)ha->temphold, ds->ds_object);
3651 }
3652
3653 static int
3654 dsl_dataset_user_hold_one(const char *dsname, void *arg)
3655 {
3656         struct dsl_ds_holdarg *ha = arg;
3657         dsl_dataset_t *ds;
3658         int error;
3659         char *name;
3660
3661         /* alloc a buffer to hold dsname@snapname plus terminating NULL */
3662         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3663         error = dsl_dataset_hold(name, ha->dstg, &ds);
3664         strfree(name);
3665         if (error == 0) {
3666                 ha->gotone = B_TRUE;
3667                 dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3668                     dsl_dataset_user_hold_sync, ds, ha, 0);
3669         } else if (error == ENOENT && ha->recursive) {
3670                 error = 0;
3671         } else {
3672                 (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3673         }
3674         return (error);
3675 }
3676
3677 int
3678 dsl_dataset_user_hold_for_send(dsl_dataset_t *ds, char *htag,
3679     boolean_t temphold)
3680 {
3681         struct dsl_ds_holdarg *ha;
3682         int error;
3683
3684         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3685         ha->htag = htag;
3686         ha->temphold = temphold;
3687         error = dsl_sync_task_do(ds->ds_dir->dd_pool,
3688             dsl_dataset_user_hold_check, dsl_dataset_user_hold_sync,
3689             ds, ha, 0);
3690         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3691
3692         return (error);
3693 }
3694
3695 int
3696 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3697     boolean_t recursive, boolean_t temphold, int cleanup_fd)
3698 {
3699         struct dsl_ds_holdarg *ha;
3700         dsl_sync_task_t *dst;
3701         spa_t *spa;
3702         int error;
3703         minor_t minor = 0;
3704
3705         if (cleanup_fd != -1) {
3706                 /* Currently we only support cleanup-on-exit of tempholds. */
3707                 if (!temphold)
3708                         return (EINVAL);
3709                 error = zfs_onexit_fd_hold(cleanup_fd, &minor);
3710                 if (error)
3711                         return (error);
3712         }
3713
3714         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3715
3716         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3717
3718         error = spa_open(dsname, &spa, FTAG);
3719         if (error) {
3720                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3721                 if (cleanup_fd != -1)
3722                         zfs_onexit_fd_rele(cleanup_fd);
3723                 return (error);
3724         }
3725
3726         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3727         ha->htag = htag;
3728         ha->snapname = snapname;
3729         ha->recursive = recursive;
3730         ha->temphold = temphold;
3731
3732         if (recursive) {
3733                 error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3734                     ha, DS_FIND_CHILDREN);
3735         } else {
3736                 error = dsl_dataset_user_hold_one(dsname, ha);
3737         }
3738         if (error == 0)
3739                 error = dsl_sync_task_group_wait(ha->dstg);
3740
3741         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3742             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3743                 dsl_dataset_t *ds = dst->dst_arg1;
3744
3745                 if (dst->dst_err) {
3746                         dsl_dataset_name(ds, ha->failed);
3747                         *strchr(ha->failed, '@') = '\0';
3748                 } else if (error == 0 && minor != 0 && temphold) {
3749                         /*
3750                          * If this hold is to be released upon process exit,
3751                          * register that action now.
3752                          */
3753                         dsl_register_onexit_hold_cleanup(ds, htag, minor);
3754                 }
3755                 dsl_dataset_rele(ds, ha->dstg);
3756         }
3757
3758         if (error == 0 && recursive && !ha->gotone)
3759                 error = ENOENT;
3760
3761         if (error)
3762                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3763
3764         dsl_sync_task_group_destroy(ha->dstg);
3765
3766         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3767         spa_close(spa, FTAG);
3768         if (cleanup_fd != -1)
3769                 zfs_onexit_fd_rele(cleanup_fd);
3770         return (error);
3771 }
3772
3773 struct dsl_ds_releasearg {
3774         dsl_dataset_t *ds;
3775         const char *htag;
3776         boolean_t own;          /* do we own or just hold ds? */
3777 };
3778
3779 static int
3780 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3781     boolean_t *might_destroy)
3782 {
3783         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3784         uint64_t zapobj;
3785         uint64_t tmp;
3786         int error;
3787
3788         *might_destroy = B_FALSE;
3789
3790         mutex_enter(&ds->ds_lock);
3791         zapobj = ds->ds_phys->ds_userrefs_obj;
3792         if (zapobj == 0) {
3793                 /* The tag can't possibly exist */
3794                 mutex_exit(&ds->ds_lock);
3795                 return (ESRCH);
3796         }
3797
3798         /* Make sure the tag exists */
3799         error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3800         if (error) {
3801                 mutex_exit(&ds->ds_lock);
3802                 if (error == ENOENT)
3803                         error = ESRCH;
3804                 return (error);
3805         }
3806
3807         if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3808             DS_IS_DEFER_DESTROY(ds))
3809                 *might_destroy = B_TRUE;
3810
3811         mutex_exit(&ds->ds_lock);
3812         return (0);
3813 }
3814
3815 static int
3816 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3817 {
3818         struct dsl_ds_releasearg *ra = arg1;
3819         dsl_dataset_t *ds = ra->ds;
3820         boolean_t might_destroy;
3821         int error;
3822
3823         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3824                 return (ENOTSUP);
3825
3826         error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3827         if (error)
3828                 return (error);
3829
3830         if (might_destroy) {
3831                 struct dsl_ds_destroyarg dsda = {0};
3832
3833                 if (dmu_tx_is_syncing(tx)) {
3834                         /*
3835                          * If we're not prepared to remove the snapshot,
3836                          * we can't allow the release to happen right now.
3837                          */
3838                         if (!ra->own)
3839                                 return (EBUSY);
3840                 }
3841                 dsda.ds = ds;
3842                 dsda.releasing = B_TRUE;
3843                 return (dsl_dataset_destroy_check(&dsda, tag, tx));
3844         }
3845
3846         return (0);
3847 }
3848
3849 static void
3850 dsl_dataset_user_release_sync(void *arg1, void *tag, dmu_tx_t *tx)
3851 {
3852         struct dsl_ds_releasearg *ra = arg1;
3853         dsl_dataset_t *ds = ra->ds;
3854         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3855         objset_t *mos = dp->dp_meta_objset;
3856         uint64_t zapobj;
3857         uint64_t dsobj = ds->ds_object;
3858         uint64_t refs;
3859         int error;
3860
3861         mutex_enter(&ds->ds_lock);
3862         ds->ds_userrefs--;
3863         refs = ds->ds_userrefs;
3864         mutex_exit(&ds->ds_lock);
3865         error = dsl_pool_user_release(dp, ds->ds_object, ra->htag, tx);
3866         VERIFY(error == 0 || error == ENOENT);
3867         zapobj = ds->ds_phys->ds_userrefs_obj;
3868         VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3869
3870         spa_history_log_internal(LOG_DS_USER_RELEASE,
3871             dp->dp_spa, tx, "<%s> %lld dataset = %llu",
3872             ra->htag, (longlong_t)refs, dsobj);
3873
3874         if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3875             DS_IS_DEFER_DESTROY(ds)) {
3876                 struct dsl_ds_destroyarg dsda = {0};
3877
3878                 ASSERT(ra->own);
3879                 dsda.ds = ds;
3880                 dsda.releasing = B_TRUE;
3881                 /* We already did the destroy_check */
3882                 dsl_dataset_destroy_sync(&dsda, tag, tx);
3883         }
3884 }
3885
3886 static int
3887 dsl_dataset_user_release_one(const char *dsname, void *arg)
3888 {
3889         struct dsl_ds_holdarg *ha = arg;
3890         struct dsl_ds_releasearg *ra;
3891         dsl_dataset_t *ds;
3892         int error;
3893         void *dtag = ha->dstg;
3894         char *name;
3895         boolean_t own = B_FALSE;
3896         boolean_t might_destroy;
3897
3898         /* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3899         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3900         error = dsl_dataset_hold(name, dtag, &ds);
3901         strfree(name);
3902         if (error == ENOENT && ha->recursive)
3903                 return (0);
3904         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3905         if (error)
3906                 return (error);
3907
3908         ha->gotone = B_TRUE;
3909
3910         ASSERT(dsl_dataset_is_snapshot(ds));
3911
3912         error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3913         if (error) {
3914                 dsl_dataset_rele(ds, dtag);
3915                 return (error);
3916         }
3917
3918         if (might_destroy) {
3919 #ifdef _KERNEL
3920                 name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3921                 error = zfs_unmount_snap(name, NULL);
3922                 strfree(name);
3923                 if (error) {
3924                         dsl_dataset_rele(ds, dtag);
3925                         return (error);
3926                 }
3927 #endif
3928                 if (!dsl_dataset_tryown(ds, B_TRUE, dtag)) {
3929                         dsl_dataset_rele(ds, dtag);
3930                         return (EBUSY);
3931                 } else {
3932                         own = B_TRUE;
3933                         dsl_dataset_make_exclusive(ds, dtag);
3934                 }
3935         }
3936
3937         ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3938         ra->ds = ds;
3939         ra->htag = ha->htag;
3940         ra->own = own;
3941         dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3942             dsl_dataset_user_release_sync, ra, dtag, 0);
3943
3944         return (0);
3945 }
3946
3947 int
3948 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3949     boolean_t recursive)
3950 {
3951         struct dsl_ds_holdarg *ha;
3952         dsl_sync_task_t *dst;
3953         spa_t *spa;
3954         int error;
3955
3956 top:
3957         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3958
3959         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3960
3961         error = spa_open(dsname, &spa, FTAG);
3962         if (error) {
3963                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3964                 return (error);
3965         }
3966
3967         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3968         ha->htag = htag;
3969         ha->snapname = snapname;
3970         ha->recursive = recursive;
3971         if (recursive) {
3972                 error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
3973                     ha, DS_FIND_CHILDREN);
3974         } else {
3975                 error = dsl_dataset_user_release_one(dsname, ha);
3976         }
3977         if (error == 0)
3978                 error = dsl_sync_task_group_wait(ha->dstg);
3979
3980         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3981             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3982                 struct dsl_ds_releasearg *ra = dst->dst_arg1;
3983                 dsl_dataset_t *ds = ra->ds;
3984
3985                 if (dst->dst_err)
3986                         dsl_dataset_name(ds, ha->failed);
3987
3988                 if (ra->own)
3989                         dsl_dataset_disown(ds, ha->dstg);
3990                 else
3991                         dsl_dataset_rele(ds, ha->dstg);
3992
3993                 kmem_free(ra, sizeof (struct dsl_ds_releasearg));
3994         }
3995
3996         if (error == 0 && recursive && !ha->gotone)
3997                 error = ENOENT;
3998
3999         if (error && error != EBUSY)
4000                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
4001
4002         dsl_sync_task_group_destroy(ha->dstg);
4003         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
4004         spa_close(spa, FTAG);
4005
4006         /*
4007          * We can get EBUSY if we were racing with deferred destroy and
4008          * dsl_dataset_user_release_check() hadn't done the necessary
4009          * open context setup.  We can also get EBUSY if we're racing
4010          * with destroy and that thread is the ds_owner.  Either way
4011          * the busy condition should be transient, and we should retry
4012          * the release operation.
4013          */
4014         if (error == EBUSY)
4015                 goto top;
4016
4017         return (error);
4018 }
4019
4020 /*
4021  * Called at spa_load time (with retry == B_FALSE) to release a stale
4022  * temporary user hold. Also called by the onexit code (with retry == B_TRUE).
4023  */
4024 int
4025 dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, char *htag,
4026     boolean_t retry)
4027 {
4028         dsl_dataset_t *ds;
4029         char *snap;
4030         char *name;
4031         int namelen;
4032         int error;
4033
4034         do {
4035                 rw_enter(&dp->dp_config_rwlock, RW_READER);
4036                 error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
4037                 rw_exit(&dp->dp_config_rwlock);
4038                 if (error)
4039                         return (error);
4040                 namelen = dsl_dataset_namelen(ds)+1;
4041                 name = kmem_alloc(namelen, KM_SLEEP);
4042                 dsl_dataset_name(ds, name);
4043                 dsl_dataset_rele(ds, FTAG);
4044
4045                 snap = strchr(name, '@');
4046                 *snap = '\0';
4047                 ++snap;
4048                 error = dsl_dataset_user_release(name, snap, htag, B_FALSE);
4049                 kmem_free(name, namelen);
4050
4051                 /*
4052                  * The object can't have been destroyed because we have a hold,
4053                  * but it might have been renamed, resulting in ENOENT.  Retry
4054                  * if we've been requested to do so.
4055                  *
4056                  * It would be nice if we could use the dsobj all the way
4057                  * through and avoid ENOENT entirely.  But we might need to
4058                  * unmount the snapshot, and there's currently no way to lookup
4059                  * a vfsp using a ZFS object id.
4060                  */
4061         } while ((error == ENOENT) && retry);
4062
4063         return (error);
4064 }
4065
4066 int
4067 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
4068 {
4069         dsl_dataset_t *ds;
4070         int err;
4071
4072         err = dsl_dataset_hold(dsname, FTAG, &ds);
4073         if (err)
4074                 return (err);
4075
4076         VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
4077         if (ds->ds_phys->ds_userrefs_obj != 0) {
4078                 zap_attribute_t *za;
4079                 zap_cursor_t zc;
4080
4081                 za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
4082                 for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
4083                     ds->ds_phys->ds_userrefs_obj);
4084                     zap_cursor_retrieve(&zc, za) == 0;
4085                     zap_cursor_advance(&zc)) {
4086                         VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
4087                             za->za_first_integer));
4088                 }
4089                 zap_cursor_fini(&zc);
4090                 kmem_free(za, sizeof (zap_attribute_t));
4091         }
4092         dsl_dataset_rele(ds, FTAG);
4093         return (0);
4094 }
4095
4096 /*
4097  * Note, this function is used as the callback for dmu_objset_find().  We
4098  * always return 0 so that we will continue to find and process
4099  * inconsistent datasets, even if we encounter an error trying to
4100  * process one of them.
4101  */
4102 /* ARGSUSED */
4103 int
4104 dsl_destroy_inconsistent(const char *dsname, void *arg)
4105 {
4106         dsl_dataset_t *ds;
4107
4108         if (dsl_dataset_own(dsname, B_TRUE, FTAG, &ds) == 0) {
4109                 if (DS_IS_INCONSISTENT(ds))
4110                         (void) dsl_dataset_destroy(ds, FTAG, B_FALSE);
4111                 else
4112                         dsl_dataset_disown(ds, FTAG);
4113         }
4114         return (0);
4115 }
4116
4117
4118 /*
4119  * Return (in *usedp) the amount of space written in new that is not
4120  * present in oldsnap.  New may be a snapshot or the head.  Old must be
4121  * a snapshot before new, in new's filesystem (or its origin).  If not then
4122  * fail and return EINVAL.
4123  *
4124  * The written space is calculated by considering two components:  First, we
4125  * ignore any freed space, and calculate the written as new's used space
4126  * minus old's used space.  Next, we add in the amount of space that was freed
4127  * between the two snapshots, thus reducing new's used space relative to old's.
4128  * Specifically, this is the space that was born before old->ds_creation_txg,
4129  * and freed before new (ie. on new's deadlist or a previous deadlist).
4130  *
4131  * space freed                         [---------------------]
4132  * snapshots                       ---O-------O--------O-------O------
4133  *                                         oldsnap            new
4134  */
4135 int
4136 dsl_dataset_space_written(dsl_dataset_t *oldsnap, dsl_dataset_t *new,
4137     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
4138 {
4139         int err = 0;
4140         uint64_t snapobj;
4141         dsl_pool_t *dp = new->ds_dir->dd_pool;
4142
4143         *usedp = 0;
4144         *usedp += new->ds_phys->ds_used_bytes;
4145         *usedp -= oldsnap->ds_phys->ds_used_bytes;
4146
4147         *compp = 0;
4148         *compp += new->ds_phys->ds_compressed_bytes;
4149         *compp -= oldsnap->ds_phys->ds_compressed_bytes;
4150
4151         *uncompp = 0;
4152         *uncompp += new->ds_phys->ds_uncompressed_bytes;
4153         *uncompp -= oldsnap->ds_phys->ds_uncompressed_bytes;
4154
4155         rw_enter(&dp->dp_config_rwlock, RW_READER);
4156         snapobj = new->ds_object;
4157         while (snapobj != oldsnap->ds_object) {
4158                 dsl_dataset_t *snap;
4159                 uint64_t used, comp, uncomp;
4160
4161                 err = dsl_dataset_hold_obj(dp, snapobj, FTAG, &snap);
4162                 if (err != 0)
4163                         break;
4164
4165                 if (snap->ds_phys->ds_prev_snap_txg ==
4166                     oldsnap->ds_phys->ds_creation_txg) {
4167                         /*
4168                          * The blocks in the deadlist can not be born after
4169                          * ds_prev_snap_txg, so get the whole deadlist space,
4170                          * which is more efficient (especially for old-format
4171                          * deadlists).  Unfortunately the deadlist code
4172                          * doesn't have enough information to make this
4173                          * optimization itself.
4174                          */
4175                         dsl_deadlist_space(&snap->ds_deadlist,
4176                             &used, &comp, &uncomp);
4177                 } else {
4178                         dsl_deadlist_space_range(&snap->ds_deadlist,
4179                             0, oldsnap->ds_phys->ds_creation_txg,
4180                             &used, &comp, &uncomp);
4181                 }
4182                 *usedp += used;
4183                 *compp += comp;
4184                 *uncompp += uncomp;
4185
4186                 /*
4187                  * If we get to the beginning of the chain of snapshots
4188                  * (ds_prev_snap_obj == 0) before oldsnap, then oldsnap
4189                  * was not a snapshot of/before new.
4190                  */
4191                 snapobj = snap->ds_phys->ds_prev_snap_obj;
4192                 dsl_dataset_rele(snap, FTAG);
4193                 if (snapobj == 0) {
4194                         err = EINVAL;
4195                         break;
4196                 }
4197
4198         }
4199         rw_exit(&dp->dp_config_rwlock);
4200         return (err);
4201 }
4202
4203 /*
4204  * Return (in *usedp) the amount of space that will be reclaimed if firstsnap,
4205  * lastsnap, and all snapshots in between are deleted.
4206  *
4207  * blocks that would be freed            [---------------------------]
4208  * snapshots                       ---O-------O--------O-------O--------O
4209  *                                        firstsnap        lastsnap
4210  *
4211  * This is the set of blocks that were born after the snap before firstsnap,
4212  * (birth > firstsnap->prev_snap_txg) and died before the snap after the
4213  * last snap (ie, is on lastsnap->ds_next->ds_deadlist or an earlier deadlist).
4214  * We calculate this by iterating over the relevant deadlists (from the snap
4215  * after lastsnap, backward to the snap after firstsnap), summing up the
4216  * space on the deadlist that was born after the snap before firstsnap.
4217  */
4218 int
4219 dsl_dataset_space_wouldfree(dsl_dataset_t *firstsnap,
4220     dsl_dataset_t *lastsnap,
4221     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
4222 {
4223         int err = 0;
4224         uint64_t snapobj;
4225         dsl_pool_t *dp = firstsnap->ds_dir->dd_pool;
4226
4227         ASSERT(dsl_dataset_is_snapshot(firstsnap));
4228         ASSERT(dsl_dataset_is_snapshot(lastsnap));
4229
4230         /*
4231          * Check that the snapshots are in the same dsl_dir, and firstsnap
4232          * is before lastsnap.
4233          */
4234         if (firstsnap->ds_dir != lastsnap->ds_dir ||
4235             firstsnap->ds_phys->ds_creation_txg >
4236             lastsnap->ds_phys->ds_creation_txg)
4237                 return (EINVAL);
4238
4239         *usedp = *compp = *uncompp = 0;
4240
4241         rw_enter(&dp->dp_config_rwlock, RW_READER);
4242         snapobj = lastsnap->ds_phys->ds_next_snap_obj;
4243         while (snapobj != firstsnap->ds_object) {
4244                 dsl_dataset_t *ds;
4245                 uint64_t used, comp, uncomp;
4246
4247                 err = dsl_dataset_hold_obj(dp, snapobj, FTAG, &ds);
4248                 if (err != 0)
4249                         break;
4250
4251                 dsl_deadlist_space_range(&ds->ds_deadlist,
4252                     firstsnap->ds_phys->ds_prev_snap_txg, UINT64_MAX,
4253                     &used, &comp, &uncomp);
4254                 *usedp += used;
4255                 *compp += comp;
4256                 *uncompp += uncomp;
4257
4258                 snapobj = ds->ds_phys->ds_prev_snap_obj;
4259                 ASSERT3U(snapobj, !=, 0);
4260                 dsl_dataset_rele(ds, FTAG);
4261         }
4262         rw_exit(&dp->dp_config_rwlock);
4263         return (err);
4264 }
4265
4266 #if defined(_KERNEL) && defined(HAVE_SPL)
4267 EXPORT_SYMBOL(dmu_snapshots_destroy_nvl);
4268 EXPORT_SYMBOL(dsl_dataset_hold);
4269 EXPORT_SYMBOL(dsl_dataset_hold_obj);
4270 EXPORT_SYMBOL(dsl_dataset_own);
4271 EXPORT_SYMBOL(dsl_dataset_own_obj);
4272 EXPORT_SYMBOL(dsl_dataset_name);
4273 EXPORT_SYMBOL(dsl_dataset_rele);
4274 EXPORT_SYMBOL(dsl_dataset_disown);
4275 EXPORT_SYMBOL(dsl_dataset_drop_ref);
4276 EXPORT_SYMBOL(dsl_dataset_tryown);
4277 EXPORT_SYMBOL(dsl_dataset_make_exclusive);
4278 EXPORT_SYMBOL(dsl_dataset_create_sync);
4279 EXPORT_SYMBOL(dsl_dataset_create_sync_dd);
4280 EXPORT_SYMBOL(dsl_dataset_destroy);
4281 EXPORT_SYMBOL(dsl_dataset_destroy_check);
4282 EXPORT_SYMBOL(dsl_dataset_destroy_sync);
4283 EXPORT_SYMBOL(dsl_dataset_snapshot_check);
4284 EXPORT_SYMBOL(dsl_dataset_snapshot_sync);
4285 EXPORT_SYMBOL(dsl_dataset_rename);
4286 EXPORT_SYMBOL(dsl_dataset_promote);
4287 EXPORT_SYMBOL(dsl_dataset_clone_swap);
4288 EXPORT_SYMBOL(dsl_dataset_user_hold);
4289 EXPORT_SYMBOL(dsl_dataset_user_release);
4290 EXPORT_SYMBOL(dsl_dataset_user_release_tmp);
4291 EXPORT_SYMBOL(dsl_dataset_get_holds);
4292 EXPORT_SYMBOL(dsl_dataset_get_blkptr);
4293 EXPORT_SYMBOL(dsl_dataset_set_blkptr);
4294 EXPORT_SYMBOL(dsl_dataset_get_spa);
4295 EXPORT_SYMBOL(dsl_dataset_modified_since_lastsnap);
4296 EXPORT_SYMBOL(dsl_dataset_space_written);
4297 EXPORT_SYMBOL(dsl_dataset_space_wouldfree);
4298 EXPORT_SYMBOL(dsl_dataset_sync);
4299 EXPORT_SYMBOL(dsl_dataset_block_born);
4300 EXPORT_SYMBOL(dsl_dataset_block_kill);
4301 EXPORT_SYMBOL(dsl_dataset_block_freeable);
4302 EXPORT_SYMBOL(dsl_dataset_prev_snap_txg);
4303 EXPORT_SYMBOL(dsl_dataset_dirty);
4304 EXPORT_SYMBOL(dsl_dataset_stats);
4305 EXPORT_SYMBOL(dsl_dataset_fast_stat);
4306 EXPORT_SYMBOL(dsl_dataset_space);
4307 EXPORT_SYMBOL(dsl_dataset_fsid_guid);
4308 EXPORT_SYMBOL(dsl_dsobj_to_dsname);
4309 EXPORT_SYMBOL(dsl_dataset_check_quota);
4310 EXPORT_SYMBOL(dsl_dataset_set_quota);
4311 EXPORT_SYMBOL(dsl_dataset_set_quota_sync);
4312 EXPORT_SYMBOL(dsl_dataset_set_reservation);
4313 EXPORT_SYMBOL(dsl_destroy_inconsistent);
4314 #endif