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) 2012 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/zfeature.h>
  39 #include <sys/unique.h>
  40 #include <sys/zfs_context.h>
  41 #include <sys/zfs_ioctl.h>
  42 #include <sys/spa.h>
  43 #include <sys/zfs_znode.h>
  44 #include <sys/zfs_onexit.h>
  45 #include <sys/zvol.h>
  46 #include <sys/dsl_scan.h>
  47 #include <sys/dsl_deadlist.h>
  48
  49 static char *dsl_reaper = "the grim reaper";
  50
  51 static dsl_checkfunc_t dsl_dataset_destroy_begin_check;
  52 static dsl_syncfunc_t dsl_dataset_destroy_begin_sync;
  53 static dsl_syncfunc_t dsl_dataset_set_reservation_sync;
  54
  55 #define SWITCH64(x, y) \
  56         { \
  57                 uint64_t __tmp = (x); \
  58                 (x) = (y); \
  59                 (y) = __tmp; \
  60         }
  61
  62 #define DS_REF_MAX      (1ULL << 62)
  63
  64 #define DSL_DEADLIST_BLOCKSIZE  SPA_MAXBLOCKSIZE
  65
  66 #define DSL_DATASET_IS_DESTROYED(ds)    ((ds)->ds_owner == dsl_reaper)
  67
  68
  69 /*
  70  * Figure out how much of this delta should be propogated to the dsl_dir
  71  * layer.  If there's a refreservation, that space has already been
  72  * partially accounted for in our ancestors.
  73  */
  74 static int64_t
  75 parent_delta(dsl_dataset_t *ds, int64_t delta)
  76 {
  77         uint64_t old_bytes, new_bytes;
  78
  79         if (ds->ds_reserved == 0)
  80                 return (delta);
  81
  82         old_bytes = MAX(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
  83         new_bytes = MAX(ds->ds_phys->ds_unique_bytes + delta, ds->ds_reserved);
  84
  85         ASSERT3U(ABS((int64_t)(new_bytes - old_bytes)), <=, ABS(delta));
  86         return (new_bytes - old_bytes);
  87 }
  88
  89 void
  90 dsl_dataset_block_born(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
  91 {
  92         int used, compressed, uncompressed;
  93         int64_t delta;
  94
  95         used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
  96         compressed = BP_GET_PSIZE(bp);
  97         uncompressed = BP_GET_UCSIZE(bp);
  98
  99         dprintf_bp(bp, "ds=%p", ds);
 100
 101         ASSERT(dmu_tx_is_syncing(tx));
 102         /* It could have been compressed away to nothing */
 103         if (BP_IS_HOLE(bp))
 104                 return;
 105         ASSERT(BP_GET_TYPE(bp) != DMU_OT_NONE);
 106         ASSERT(DMU_OT_IS_VALID(BP_GET_TYPE(bp)));
 107         if (ds == NULL) {
 108                 dsl_pool_mos_diduse_space(tx->tx_pool,
 109                     used, compressed, uncompressed);
 110                 return;
 111         }
 112         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 113
 114         mutex_enter(&ds->ds_dir->dd_lock);
 115         mutex_enter(&ds->ds_lock);
 116         delta = parent_delta(ds, used);
 117         ds->ds_phys->ds_referenced_bytes += used;
 118         ds->ds_phys->ds_compressed_bytes += compressed;
 119         ds->ds_phys->ds_uncompressed_bytes += uncompressed;
 120         ds->ds_phys->ds_unique_bytes += used;
 121         mutex_exit(&ds->ds_lock);
 122         dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD, delta,
 123             compressed, uncompressed, tx);
 124         dsl_dir_transfer_space(ds->ds_dir, used - delta,
 125             DD_USED_REFRSRV, DD_USED_HEAD, tx);
 126         mutex_exit(&ds->ds_dir->dd_lock);
 127 }
 128
 129 int
 130 dsl_dataset_block_kill(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx,
 131     boolean_t async)
 132 {
 133         int used, compressed, uncompressed;
 134
 135         if (BP_IS_HOLE(bp))
 136                 return (0);
 137
 138         ASSERT(dmu_tx_is_syncing(tx));
 139         ASSERT(bp->blk_birth <= tx->tx_txg);
 140
 141         used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
 142         compressed = BP_GET_PSIZE(bp);
 143         uncompressed = BP_GET_UCSIZE(bp);
 144
 145         ASSERT(used > 0);
 146         if (ds == NULL) {
 147                 dsl_free(tx->tx_pool, tx->tx_txg, bp);
 148                 dsl_pool_mos_diduse_space(tx->tx_pool,
 149                     -used, -compressed, -uncompressed);
 150                 return (used);
 151         }
 152         ASSERT3P(tx->tx_pool, ==, ds->ds_dir->dd_pool);
 153
 154         ASSERT(!dsl_dataset_is_snapshot(ds));
 155         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 156
 157         if (bp->blk_birth > ds->ds_phys->ds_prev_snap_txg) {
 158                 int64_t delta;
 159
 160                 dprintf_bp(bp, "freeing ds=%llu", ds->ds_object);
 161                 dsl_free(tx->tx_pool, tx->tx_txg, bp);
 162
 163                 mutex_enter(&ds->ds_dir->dd_lock);
 164                 mutex_enter(&ds->ds_lock);
 165                 ASSERT(ds->ds_phys->ds_unique_bytes >= used ||
 166                     !DS_UNIQUE_IS_ACCURATE(ds));
 167                 delta = parent_delta(ds, -used);
 168                 ds->ds_phys->ds_unique_bytes -= used;
 169                 mutex_exit(&ds->ds_lock);
 170                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
 171                     delta, -compressed, -uncompressed, tx);
 172                 dsl_dir_transfer_space(ds->ds_dir, -used - delta,
 173                     DD_USED_REFRSRV, DD_USED_HEAD, tx);
 174                 mutex_exit(&ds->ds_dir->dd_lock);
 175         } else {
 176                 dprintf_bp(bp, "putting on dead list: %s", "");
 177                 if (async) {
 178                         /*
 179                          * We are here as part of zio's write done callback,
 180                          * which means we're a zio interrupt thread.  We can't
 181                          * call dsl_deadlist_insert() now because it may block
 182                          * waiting for I/O.  Instead, put bp on the deferred
 183                          * queue and let dsl_pool_sync() finish the job.
 184                          */
 185                         bplist_append(&ds->ds_pending_deadlist, bp);
 186                 } else {
 187                         dsl_deadlist_insert(&ds->ds_deadlist, bp, tx);
 188                 }
 189                 ASSERT3U(ds->ds_prev->ds_object, ==,
 190                     ds->ds_phys->ds_prev_snap_obj);
 191                 ASSERT(ds->ds_prev->ds_phys->ds_num_children > 0);
 192                 /* if (bp->blk_birth > prev prev snap txg) prev unique += bs */
 193                 if (ds->ds_prev->ds_phys->ds_next_snap_obj ==
 194                     ds->ds_object && bp->blk_birth >
 195                     ds->ds_prev->ds_phys->ds_prev_snap_txg) {
 196                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
 197                         mutex_enter(&ds->ds_prev->ds_lock);
 198                         ds->ds_prev->ds_phys->ds_unique_bytes += used;
 199                         mutex_exit(&ds->ds_prev->ds_lock);
 200                 }
 201                 if (bp->blk_birth > ds->ds_dir->dd_origin_txg) {
 202                         dsl_dir_transfer_space(ds->ds_dir, used,
 203                             DD_USED_HEAD, DD_USED_SNAP, tx);
 204                 }
 205         }
 206         mutex_enter(&ds->ds_lock);
 207         ASSERT3U(ds->ds_phys->ds_referenced_bytes, >=, used);
 208         ds->ds_phys->ds_referenced_bytes -= used;
 209         ASSERT3U(ds->ds_phys->ds_compressed_bytes, >=, compressed);
 210         ds->ds_phys->ds_compressed_bytes -= compressed;
 211         ASSERT3U(ds->ds_phys->ds_uncompressed_bytes, >=, uncompressed);
 212         ds->ds_phys->ds_uncompressed_bytes -= uncompressed;
 213         mutex_exit(&ds->ds_lock);
 214
 215         return (used);
 216 }
 217
 218 uint64_t
 219 dsl_dataset_prev_snap_txg(dsl_dataset_t *ds)
 220 {
 221         uint64_t trysnap = 0;
 222
 223         if (ds == NULL)
 224                 return (0);
 225         /*
 226          * The snapshot creation could fail, but that would cause an
 227          * incorrect FALSE return, which would only result in an
 228          * overestimation of the amount of space that an operation would
 229          * consume, which is OK.
 230          *
 231          * There's also a small window where we could miss a pending
 232          * snapshot, because we could set the sync task in the quiescing
 233          * phase.  So this should only be used as a guess.
 234          */
 235         if (ds->ds_trysnap_txg >
 236             spa_last_synced_txg(ds->ds_dir->dd_pool->dp_spa))
 237                 trysnap = ds->ds_trysnap_txg;
 238         return (MAX(ds->ds_phys->ds_prev_snap_txg, trysnap));
 239 }
 240
 241 boolean_t
 242 dsl_dataset_block_freeable(dsl_dataset_t *ds, const blkptr_t *bp,
 243     uint64_t blk_birth)
 244 {
 245         if (blk_birth <= dsl_dataset_prev_snap_txg(ds))
 246                 return (B_FALSE);
 247
 248         ddt_prefetch(dsl_dataset_get_spa(ds), bp);
 249
 250         return (B_TRUE);
 251 }
 252
 253 /* ARGSUSED */
 254 static void
 255 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
 256 {
 257         dsl_dataset_t *ds = dsv;
 258
 259         ASSERT(ds->ds_owner == NULL || DSL_DATASET_IS_DESTROYED(ds));
 260
 261         unique_remove(ds->ds_fsid_guid);
 262
 263         if (ds->ds_objset != NULL)
 264                 dmu_objset_evict(ds->ds_objset);
 265
 266         if (ds->ds_prev) {
 267                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 268                 ds->ds_prev = NULL;
 269         }
 270
 271         bplist_destroy(&ds->ds_pending_deadlist);
 272         if (db != NULL) {
 273                 dsl_deadlist_close(&ds->ds_deadlist);
 274         } else {
 275                 ASSERT(ds->ds_deadlist.dl_dbuf == NULL);
 276                 ASSERT(!ds->ds_deadlist.dl_oldfmt);
 277         }
 278         if (ds->ds_dir)
 279                 dsl_dir_close(ds->ds_dir, ds);
 280
 281         ASSERT(!list_link_active(&ds->ds_synced_link));
 282
 283         mutex_destroy(&ds->ds_lock);
 284         mutex_destroy(&ds->ds_recvlock);
 285         mutex_destroy(&ds->ds_opening_lock);
 286         rw_destroy(&ds->ds_rwlock);
 287         cv_destroy(&ds->ds_exclusive_cv);
 288
 289         kmem_free(ds, sizeof (dsl_dataset_t));
 290 }
 291
 292 static int
 293 dsl_dataset_get_snapname(dsl_dataset_t *ds)
 294 {
 295         dsl_dataset_phys_t *headphys;
 296         int err;
 297         dmu_buf_t *headdbuf;
 298         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 299         objset_t *mos = dp->dp_meta_objset;
 300
 301         if (ds->ds_snapname[0])
 302                 return (0);
 303         if (ds->ds_phys->ds_next_snap_obj == 0)
 304                 return (0);
 305
 306         err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
 307             FTAG, &headdbuf);
 308         if (err)
 309                 return (err);
 310         headphys = headdbuf->db_data;
 311         err = zap_value_search(dp->dp_meta_objset,
 312             headphys->ds_snapnames_zapobj, ds->ds_object, 0, ds->ds_snapname);
 313         dmu_buf_rele(headdbuf, FTAG);
 314         return (err);
 315 }
 316
 317 int
 318 dsl_dataset_snap_lookup(dsl_dataset_t *ds, const char *name, uint64_t *value)
 319 {
 320         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 321         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 322         matchtype_t mt;
 323         int err;
 324
 325         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 326                 mt = MT_FIRST;
 327         else
 328                 mt = MT_EXACT;
 329
 330         err = zap_lookup_norm(mos, snapobj, name, 8, 1,
 331             value, mt, NULL, 0, NULL);
 332         if (err == ENOTSUP && mt == MT_FIRST)
 333                 err = zap_lookup(mos, snapobj, name, 8, 1, value);
 334         return (err);
 335 }
 336
 337 static int
 338 dsl_dataset_snap_remove(dsl_dataset_t *ds, char *name, dmu_tx_t *tx)
 339 {
 340         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 341         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
 342         matchtype_t mt;
 343         int err;
 344
 345         dsl_dir_snap_cmtime_update(ds->ds_dir);
 346
 347         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 348                 mt = MT_FIRST;
 349         else
 350                 mt = MT_EXACT;
 351
 352         err = zap_remove_norm(mos, snapobj, name, mt, tx);
 353         if (err == ENOTSUP && mt == MT_FIRST)
 354                 err = zap_remove(mos, snapobj, name, tx);
 355         return (err);
 356 }
 357
 358 static int
 359 dsl_dataset_get_ref(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 360     dsl_dataset_t **dsp)
 361 {
 362         objset_t *mos = dp->dp_meta_objset;
 363         dmu_buf_t *dbuf;
 364         dsl_dataset_t *ds;
 365         int err;
 366         dmu_object_info_t doi;
 367
 368         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
 369             dsl_pool_sync_context(dp));
 370
 371         err = dmu_bonus_hold(mos, dsobj, tag, &dbuf);
 372         if (err)
 373                 return (err);
 374
 375         /* Make sure dsobj has the correct object type. */
 376         dmu_object_info_from_db(dbuf, &doi);
 377         if (doi.doi_type != DMU_OT_DSL_DATASET)
 378                 return (EINVAL);
 379
 380         ds = dmu_buf_get_user(dbuf);
 381         if (ds == NULL) {
 382                 dsl_dataset_t *winner = NULL;
 383
 384                 ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_PUSHPAGE);
 385                 ds->ds_dbuf = dbuf;
 386                 ds->ds_object = dsobj;
 387                 ds->ds_phys = dbuf->db_data;
 388                 list_link_init(&ds->ds_synced_link);
 389
 390                 mutex_init(&ds->ds_lock, NULL, MUTEX_DEFAULT, NULL);
 391                 mutex_init(&ds->ds_recvlock, NULL, MUTEX_DEFAULT, NULL);
 392                 mutex_init(&ds->ds_opening_lock, NULL, MUTEX_DEFAULT, NULL);
 393                 mutex_init(&ds->ds_sendstream_lock, NULL, MUTEX_DEFAULT, NULL);
 394
 395                 rw_init(&ds->ds_rwlock, NULL, RW_DEFAULT, NULL);
 396                 cv_init(&ds->ds_exclusive_cv, NULL, CV_DEFAULT, NULL);
 397
 398                 bplist_create(&ds->ds_pending_deadlist);
 399                 dsl_deadlist_open(&ds->ds_deadlist,
 400                     mos, ds->ds_phys->ds_deadlist_obj);
 401
 402                 list_create(&ds->ds_sendstreams, sizeof (dmu_sendarg_t),
 403                     offsetof(dmu_sendarg_t, dsa_link));
 404
 405                 if (err == 0) {
 406                         err = dsl_dir_open_obj(dp,
 407                             ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
 408                 }
 409                 if (err) {
 410                         mutex_destroy(&ds->ds_lock);
 411                         mutex_destroy(&ds->ds_recvlock);
 412                         mutex_destroy(&ds->ds_opening_lock);
 413                         rw_destroy(&ds->ds_rwlock);
 414                         cv_destroy(&ds->ds_exclusive_cv);
 415                         bplist_destroy(&ds->ds_pending_deadlist);
 416                         dsl_deadlist_close(&ds->ds_deadlist);
 417                         kmem_free(ds, sizeof (dsl_dataset_t));
 418                         dmu_buf_rele(dbuf, tag);
 419                         return (err);
 420                 }
 421
 422                 if (!dsl_dataset_is_snapshot(ds)) {
 423                         ds->ds_snapname[0] = '\0';
 424                         if (ds->ds_phys->ds_prev_snap_obj) {
 425                                 err = dsl_dataset_get_ref(dp,
 426                                     ds->ds_phys->ds_prev_snap_obj,
 427                                     ds, &ds->ds_prev);
 428                         }
 429                 } else {
 430                         if (zfs_flags & ZFS_DEBUG_SNAPNAMES)
 431                                 err = dsl_dataset_get_snapname(ds);
 432                         if (err == 0 && ds->ds_phys->ds_userrefs_obj != 0) {
 433                                 err = zap_count(
 434                                     ds->ds_dir->dd_pool->dp_meta_objset,
 435                                     ds->ds_phys->ds_userrefs_obj,
 436                                     &ds->ds_userrefs);
 437                         }
 438                 }
 439
 440                 if (err == 0 && !dsl_dataset_is_snapshot(ds)) {
 441                         /*
 442                          * In sync context, we're called with either no lock
 443                          * or with the write lock.  If we're not syncing,
 444                          * we're always called with the read lock held.
 445                          */
 446                         boolean_t need_lock =
 447                             !RW_WRITE_HELD(&dp->dp_config_rwlock) &&
 448                             dsl_pool_sync_context(dp);
 449
 450                         if (need_lock)
 451                                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 452
 453                         err = dsl_prop_get_ds(ds,
 454                             "refreservation", sizeof (uint64_t), 1,
 455                             &ds->ds_reserved, NULL);
 456                         if (err == 0) {
 457                                 err = dsl_prop_get_ds(ds,
 458                                     "refquota", sizeof (uint64_t), 1,
 459                                     &ds->ds_quota, NULL);
 460                         }
 461
 462                         if (need_lock)
 463                                 rw_exit(&dp->dp_config_rwlock);
 464                 } else {
 465                         ds->ds_reserved = ds->ds_quota = 0;
 466                 }
 467
 468                 if (err == 0) {
 469                         winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
 470                             dsl_dataset_evict);
 471                 }
 472                 if (err || winner) {
 473                         bplist_destroy(&ds->ds_pending_deadlist);
 474                         dsl_deadlist_close(&ds->ds_deadlist);
 475                         if (ds->ds_prev)
 476                                 dsl_dataset_drop_ref(ds->ds_prev, ds);
 477                         dsl_dir_close(ds->ds_dir, ds);
 478                         mutex_destroy(&ds->ds_lock);
 479                         mutex_destroy(&ds->ds_recvlock);
 480                         mutex_destroy(&ds->ds_opening_lock);
 481                         rw_destroy(&ds->ds_rwlock);
 482                         cv_destroy(&ds->ds_exclusive_cv);
 483                         kmem_free(ds, sizeof (dsl_dataset_t));
 484                         if (err) {
 485                                 dmu_buf_rele(dbuf, tag);
 486                                 return (err);
 487                         }
 488                         ds = winner;
 489                 } else {
 490                         ds->ds_fsid_guid =
 491                             unique_insert(ds->ds_phys->ds_fsid_guid);
 492                 }
 493         }
 494         ASSERT3P(ds->ds_dbuf, ==, dbuf);
 495         ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
 496         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0 ||
 497             spa_version(dp->dp_spa) < SPA_VERSION_ORIGIN ||
 498             dp->dp_origin_snap == NULL || ds == dp->dp_origin_snap);
 499         mutex_enter(&ds->ds_lock);
 500         if (!dsl_pool_sync_context(dp) && DSL_DATASET_IS_DESTROYED(ds)) {
 501                 mutex_exit(&ds->ds_lock);
 502                 dmu_buf_rele(ds->ds_dbuf, tag);
 503                 return (ENOENT);
 504         }
 505         mutex_exit(&ds->ds_lock);
 506         *dsp = ds;
 507         return (0);
 508 }
 509
 510 static int
 511 dsl_dataset_hold_ref(dsl_dataset_t *ds, void *tag)
 512 {
 513         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 514
 515         /*
 516          * In syncing context we don't want the rwlock lock: there
 517          * may be an existing writer waiting for sync phase to
 518          * finish.  We don't need to worry about such writers, since
 519          * sync phase is single-threaded, so the writer can't be
 520          * doing anything while we are active.
 521          */
 522         if (dsl_pool_sync_context(dp)) {
 523                 ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
 524                 return (0);
 525         }
 526
 527         /*
 528          * Normal users will hold the ds_rwlock as a READER until they
 529          * are finished (i.e., call dsl_dataset_rele()).  "Owners" will
 530          * drop their READER lock after they set the ds_owner field.
 531          *
 532          * If the dataset is being destroyed, the destroy thread will
 533          * obtain a WRITER lock for exclusive access after it's done its
 534          * open-context work and then change the ds_owner to
 535          * dsl_reaper once destruction is assured.  So threads
 536          * may block here temporarily, until the "destructability" of
 537          * the dataset is determined.
 538          */
 539         ASSERT(!RW_WRITE_HELD(&dp->dp_config_rwlock));
 540         mutex_enter(&ds->ds_lock);
 541         while (!rw_tryenter(&ds->ds_rwlock, RW_READER)) {
 542                 rw_exit(&dp->dp_config_rwlock);
 543                 cv_wait(&ds->ds_exclusive_cv, &ds->ds_lock);
 544                 if (DSL_DATASET_IS_DESTROYED(ds)) {
 545                         mutex_exit(&ds->ds_lock);
 546                         dsl_dataset_drop_ref(ds, tag);
 547                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 548                         return (ENOENT);
 549                 }
 550                 /*
 551                  * The dp_config_rwlock lives above the ds_lock. And
 552                  * we need to check DSL_DATASET_IS_DESTROYED() while
 553                  * holding the ds_lock, so we have to drop and reacquire
 554                  * the ds_lock here.
 555                  */
 556                 mutex_exit(&ds->ds_lock);
 557                 rw_enter(&dp->dp_config_rwlock, RW_READER);
 558                 mutex_enter(&ds->ds_lock);
 559         }
 560         mutex_exit(&ds->ds_lock);
 561         return (0);
 562 }
 563
 564 int
 565 dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 566     dsl_dataset_t **dsp)
 567 {
 568         int err = dsl_dataset_get_ref(dp, dsobj, tag, dsp);
 569
 570         if (err)
 571                 return (err);
 572         return (dsl_dataset_hold_ref(*dsp, tag));
 573 }
 574
 575 int
 576 dsl_dataset_own_obj(dsl_pool_t *dp, uint64_t dsobj, boolean_t inconsistentok,
 577     void *tag, dsl_dataset_t **dsp)
 578 {
 579         int err = dsl_dataset_hold_obj(dp, dsobj, tag, dsp);
 580         if (err)
 581                 return (err);
 582         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 583                 dsl_dataset_rele(*dsp, tag);
 584                 *dsp = NULL;
 585                 return (EBUSY);
 586         }
 587         return (0);
 588 }
 589
 590 int
 591 dsl_dataset_hold(const char *name, void *tag, dsl_dataset_t **dsp)
 592 {
 593         dsl_dir_t *dd;
 594         dsl_pool_t *dp;
 595         const char *snapname;
 596         uint64_t obj;
 597         int err = 0;
 598
 599         err = dsl_dir_open_spa(NULL, name, FTAG, &dd, &snapname);
 600         if (err)
 601                 return (err);
 602
 603         dp = dd->dd_pool;
 604         obj = dd->dd_phys->dd_head_dataset_obj;
 605         rw_enter(&dp->dp_config_rwlock, RW_READER);
 606         if (obj)
 607                 err = dsl_dataset_get_ref(dp, obj, tag, dsp);
 608         else
 609                 err = ENOENT;
 610         if (err)
 611                 goto out;
 612
 613         err = dsl_dataset_hold_ref(*dsp, tag);
 614
 615         /* we may be looking for a snapshot */
 616         if (err == 0 && snapname != NULL) {
 617                 dsl_dataset_t *ds = NULL;
 618
 619                 if (*snapname++ != '@') {
 620                         dsl_dataset_rele(*dsp, tag);
 621                         err = ENOENT;
 622                         goto out;
 623                 }
 624
 625                 dprintf("looking for snapshot '%s'\n", snapname);
 626                 err = dsl_dataset_snap_lookup(*dsp, snapname, &obj);
 627                 if (err == 0)
 628                         err = dsl_dataset_get_ref(dp, obj, tag, &ds);
 629                 dsl_dataset_rele(*dsp, tag);
 630
 631                 ASSERT3U((err == 0), ==, (ds != NULL));
 632
 633                 if (ds) {
 634                         mutex_enter(&ds->ds_lock);
 635                         if (ds->ds_snapname[0] == 0)
 636                                 (void) strlcpy(ds->ds_snapname, snapname,
 637                                     sizeof (ds->ds_snapname));
 638                         mutex_exit(&ds->ds_lock);
 639                         err = dsl_dataset_hold_ref(ds, tag);
 640                         *dsp = err ? NULL : ds;
 641                 }
 642         }
 643 out:
 644         rw_exit(&dp->dp_config_rwlock);
 645         dsl_dir_close(dd, FTAG);
 646         return (err);
 647 }
 648
 649 int
 650 dsl_dataset_own(const char *name, boolean_t inconsistentok,
 651     void *tag, dsl_dataset_t **dsp)
 652 {
 653         int err = dsl_dataset_hold(name, tag, dsp);
 654         if (err)
 655                 return (err);
 656         if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
 657                 dsl_dataset_rele(*dsp, tag);
 658                 return (EBUSY);
 659         }
 660         return (0);
 661 }
 662
 663 void
 664 dsl_dataset_name(dsl_dataset_t *ds, char *name)
 665 {
 666         if (ds == NULL) {
 667                 (void) strcpy(name, "mos");
 668         } else {
 669                 dsl_dir_name(ds->ds_dir, name);
 670                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 671                 if (ds->ds_snapname[0]) {
 672                         (void) strcat(name, "@");
 673                         /*
 674                          * We use a "recursive" mutex so that we
 675                          * can call dprintf_ds() with ds_lock held.
 676                          */
 677                         if (!MUTEX_HELD(&ds->ds_lock)) {
 678                                 mutex_enter(&ds->ds_lock);
 679                                 (void) strcat(name, ds->ds_snapname);
 680                                 mutex_exit(&ds->ds_lock);
 681                         } else {
 682                                 (void) strcat(name, ds->ds_snapname);
 683                         }
 684                 }
 685         }
 686 }
 687
 688 static int
 689 dsl_dataset_namelen(dsl_dataset_t *ds)
 690 {
 691         int result;
 692
 693         if (ds == NULL) {
 694                 result = 3;     /* "mos" */
 695         } else {
 696                 result = dsl_dir_namelen(ds->ds_dir);
 697                 VERIFY(0 == dsl_dataset_get_snapname(ds));
 698                 if (ds->ds_snapname[0]) {
 699                         ++result;       /* adding one for the @-sign */
 700                         if (!MUTEX_HELD(&ds->ds_lock)) {
 701                                 mutex_enter(&ds->ds_lock);
 702                                 result += strlen(ds->ds_snapname);
 703                                 mutex_exit(&ds->ds_lock);
 704                         } else {
 705                                 result += strlen(ds->ds_snapname);
 706                         }
 707                 }
 708         }
 709
 710         return (result);
 711 }
 712
 713 void
 714 dsl_dataset_drop_ref(dsl_dataset_t *ds, void *tag)
 715 {
 716         dmu_buf_rele(ds->ds_dbuf, tag);
 717 }
 718
 719 void
 720 dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
 721 {
 722         if (!dsl_pool_sync_context(ds->ds_dir->dd_pool)) {
 723                 rw_exit(&ds->ds_rwlock);
 724         }
 725         dsl_dataset_drop_ref(ds, tag);
 726 }
 727
 728 void
 729 dsl_dataset_disown(dsl_dataset_t *ds, void *tag)
 730 {
 731         ASSERT((ds->ds_owner == tag && ds->ds_dbuf) ||
 732             (DSL_DATASET_IS_DESTROYED(ds) && ds->ds_dbuf == NULL));
 733
 734         mutex_enter(&ds->ds_lock);
 735         ds->ds_owner = NULL;
 736         if (RW_WRITE_HELD(&ds->ds_rwlock)) {
 737                 rw_exit(&ds->ds_rwlock);
 738                 cv_broadcast(&ds->ds_exclusive_cv);
 739         }
 740         mutex_exit(&ds->ds_lock);
 741         if (ds->ds_dbuf)
 742                 dsl_dataset_drop_ref(ds, tag);
 743         else
 744                 dsl_dataset_evict(NULL, ds);
 745 }
 746
 747 boolean_t
 748 dsl_dataset_tryown(dsl_dataset_t *ds, boolean_t inconsistentok, void *tag)
 749 {
 750         boolean_t gotit = FALSE;
 751
 752         mutex_enter(&ds->ds_lock);
 753         if (ds->ds_owner == NULL &&
 754             (!DS_IS_INCONSISTENT(ds) || inconsistentok)) {
 755                 ds->ds_owner = tag;
 756                 if (!dsl_pool_sync_context(ds->ds_dir->dd_pool))
 757                         rw_exit(&ds->ds_rwlock);
 758                 gotit = TRUE;
 759         }
 760         mutex_exit(&ds->ds_lock);
 761         return (gotit);
 762 }
 763
 764 void
 765 dsl_dataset_make_exclusive(dsl_dataset_t *ds, void *owner)
 766 {
 767         ASSERT3P(owner, ==, ds->ds_owner);
 768         if (!RW_WRITE_HELD(&ds->ds_rwlock))
 769                 rw_enter(&ds->ds_rwlock, RW_WRITER);
 770 }
 771
 772 uint64_t
 773 dsl_dataset_create_sync_dd(dsl_dir_t *dd, dsl_dataset_t *origin,
 774     uint64_t flags, dmu_tx_t *tx)
 775 {
 776         dsl_pool_t *dp = dd->dd_pool;
 777         dmu_buf_t *dbuf;
 778         dsl_dataset_phys_t *dsphys;
 779         uint64_t dsobj;
 780         objset_t *mos = dp->dp_meta_objset;
 781
 782         if (origin == NULL)
 783                 origin = dp->dp_origin_snap;
 784
 785         ASSERT(origin == NULL || origin->ds_dir->dd_pool == dp);
 786         ASSERT(origin == NULL || origin->ds_phys->ds_num_children > 0);
 787         ASSERT(dmu_tx_is_syncing(tx));
 788         ASSERT(dd->dd_phys->dd_head_dataset_obj == 0);
 789
 790         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
 791             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
 792         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
 793         dmu_buf_will_dirty(dbuf, tx);
 794         dsphys = dbuf->db_data;
 795         bzero(dsphys, sizeof (dsl_dataset_phys_t));
 796         dsphys->ds_dir_obj = dd->dd_object;
 797         dsphys->ds_flags = flags;
 798         dsphys->ds_fsid_guid = unique_create();
 799         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
 800             sizeof (dsphys->ds_guid));
 801         dsphys->ds_snapnames_zapobj =
 802             zap_create_norm(mos, U8_TEXTPREP_TOUPPER, DMU_OT_DSL_DS_SNAP_MAP,
 803             DMU_OT_NONE, 0, tx);
 804         dsphys->ds_creation_time = gethrestime_sec();
 805         dsphys->ds_creation_txg = tx->tx_txg == TXG_INITIAL ? 1 : tx->tx_txg;
 806
 807         if (origin == NULL) {
 808                 dsphys->ds_deadlist_obj = dsl_deadlist_alloc(mos, tx);
 809         } else {
 810                 dsl_dataset_t *ohds;
 811
 812                 dsphys->ds_prev_snap_obj = origin->ds_object;
 813                 dsphys->ds_prev_snap_txg =
 814                     origin->ds_phys->ds_creation_txg;
 815                 dsphys->ds_referenced_bytes =
 816                     origin->ds_phys->ds_referenced_bytes;
 817                 dsphys->ds_compressed_bytes =
 818                     origin->ds_phys->ds_compressed_bytes;
 819                 dsphys->ds_uncompressed_bytes =
 820                     origin->ds_phys->ds_uncompressed_bytes;
 821                 dsphys->ds_bp = origin->ds_phys->ds_bp;
 822                 dsphys->ds_flags |= origin->ds_phys->ds_flags;
 823
 824                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
 825                 origin->ds_phys->ds_num_children++;
 826
 827                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
 828                     origin->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ohds));
 829                 dsphys->ds_deadlist_obj = dsl_deadlist_clone(&ohds->ds_deadlist,
 830                     dsphys->ds_prev_snap_txg, dsphys->ds_prev_snap_obj, tx);
 831                 dsl_dataset_rele(ohds, FTAG);
 832
 833                 if (spa_version(dp->dp_spa) >= SPA_VERSION_NEXT_CLONES) {
 834                         if (origin->ds_phys->ds_next_clones_obj == 0) {
 835                                 origin->ds_phys->ds_next_clones_obj =
 836                                     zap_create(mos,
 837                                     DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
 838                         }
 839                         VERIFY(0 == zap_add_int(mos,
 840                             origin->ds_phys->ds_next_clones_obj,
 841                             dsobj, tx));
 842                 }
 843
 844                 dmu_buf_will_dirty(dd->dd_dbuf, tx);
 845                 dd->dd_phys->dd_origin_obj = origin->ds_object;
 846                 if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
 847                         if (origin->ds_dir->dd_phys->dd_clones == 0) {
 848                                 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
 849                                 origin->ds_dir->dd_phys->dd_clones =
 850                                     zap_create(mos,
 851                                     DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
 852                         }
 853                         VERIFY3U(0, ==, zap_add_int(mos,
 854                             origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
 855                 }
 856         }
 857
 858         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
 859                 dsphys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
 860
 861         dmu_buf_rele(dbuf, FTAG);
 862
 863         dmu_buf_will_dirty(dd->dd_dbuf, tx);
 864         dd->dd_phys->dd_head_dataset_obj = dsobj;
 865
 866         return (dsobj);
 867 }
 868
 869 uint64_t
 870 dsl_dataset_create_sync(dsl_dir_t *pdd, const char *lastname,
 871     dsl_dataset_t *origin, uint64_t flags, cred_t *cr, dmu_tx_t *tx)
 872 {
 873         dsl_pool_t *dp = pdd->dd_pool;
 874         uint64_t dsobj, ddobj;
 875         dsl_dir_t *dd;
 876
 877         ASSERT(lastname[0] != '@');
 878
 879         ddobj = dsl_dir_create_sync(dp, pdd, lastname, tx);
 880         VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
 881
 882         dsobj = dsl_dataset_create_sync_dd(dd, origin, flags, tx);
 883
 884         dsl_deleg_set_create_perms(dd, tx, cr);
 885
 886         dsl_dir_close(dd, FTAG);
 887
 888         /*
 889          * If we are creating a clone, make sure we zero out any stale
 890          * data from the origin snapshots zil header.
 891          */
 892         if (origin != NULL) {
 893                 dsl_dataset_t *ds;
 894                 objset_t *os;
 895
 896                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 897                 VERIFY3U(0, ==, dmu_objset_from_ds(ds, &os));
 898                 bzero(&os->os_zil_header, sizeof (os->os_zil_header));
 899                 dsl_dataset_dirty(ds, tx);
 900                 dsl_dataset_rele(ds, FTAG);
 901         }
 902
 903         return (dsobj);
 904 }
 905
 906 /*
 907  * The snapshots must all be in the same pool.
 908  */
 909 int
 910 dmu_snapshots_destroy_nvl(nvlist_t *snaps, boolean_t defer, char *failed)
 911 {
 912         int err;
 913         dsl_sync_task_t *dst;
 914         spa_t *spa;
 915         nvpair_t *pair;
 916         dsl_sync_task_group_t *dstg;
 917
 918         pair = nvlist_next_nvpair(snaps, NULL);
 919         if (pair == NULL)
 920                 return (0);
 921
 922         err = spa_open(nvpair_name(pair), &spa, FTAG);
 923         if (err)
 924                 return (err);
 925         dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
 926
 927         for (pair = nvlist_next_nvpair(snaps, NULL); pair != NULL;
 928             pair = nvlist_next_nvpair(snaps, pair)) {
 929                 dsl_dataset_t *ds;
 930
 931                 err = dsl_dataset_own(nvpair_name(pair), B_TRUE, dstg, &ds);
 932                 if (err == 0) {
 933                         struct dsl_ds_destroyarg *dsda;
 934
 935                         dsl_dataset_make_exclusive(ds, dstg);
 936                         dsda = kmem_zalloc(sizeof (struct dsl_ds_destroyarg),
 937                             KM_SLEEP);
 938                         dsda->ds = ds;
 939                         dsda->defer = defer;
 940                         dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
 941                             dsl_dataset_destroy_sync, dsda, dstg, 0);
 942                 } else if (err == ENOENT) {
 943                         err = 0;
 944                 } else {
 945                         (void) strcpy(failed, nvpair_name(pair));
 946                         break;
 947                 }
 948         }
 949
 950         if (err == 0)
 951                 err = dsl_sync_task_group_wait(dstg);
 952
 953         for (dst = list_head(&dstg->dstg_tasks); dst;
 954             dst = list_next(&dstg->dstg_tasks, dst)) {
 955                 struct dsl_ds_destroyarg *dsda = dst->dst_arg1;
 956                 dsl_dataset_t *ds = dsda->ds;
 957
 958                 /*
 959                  * Return the file system name that triggered the error
 960                  */
 961                 if (dst->dst_err) {
 962                         dsl_dataset_name(ds, failed);
 963                 }
 964                 ASSERT3P(dsda->rm_origin, ==, NULL);
 965                 dsl_dataset_disown(ds, dstg);
 966                 kmem_free(dsda, sizeof (struct dsl_ds_destroyarg));
 967         }
 968
 969         dsl_sync_task_group_destroy(dstg);
 970         spa_close(spa, FTAG);
 971         return (err);
 972
 973 }
 974
 975 static boolean_t
 976 dsl_dataset_might_destroy_origin(dsl_dataset_t *ds)
 977 {
 978         boolean_t might_destroy = B_FALSE;
 979
 980         mutex_enter(&ds->ds_lock);
 981         if (ds->ds_phys->ds_num_children == 2 && ds->ds_userrefs == 0 &&
 982             DS_IS_DEFER_DESTROY(ds))
 983                 might_destroy = B_TRUE;
 984         mutex_exit(&ds->ds_lock);
 985
 986         return (might_destroy);
 987 }
 988
 989 /*
 990  * If we're removing a clone, and these three conditions are true:
 991  *      1) the clone's origin has no other children
 992  *      2) the clone's origin has no user references
 993  *      3) the clone's origin has been marked for deferred destruction
 994  * Then, prepare to remove the origin as part of this sync task group.
 995  */
 996 static int
 997 dsl_dataset_origin_rm_prep(struct dsl_ds_destroyarg *dsda, void *tag)
 998 {
 999         dsl_dataset_t *ds = dsda->ds;
1000         dsl_dataset_t *origin = ds->ds_prev;
1001
1002         if (dsl_dataset_might_destroy_origin(origin)) {
1003                 char *name;
1004                 int namelen;
1005                 int error;
1006
1007                 namelen = dsl_dataset_namelen(origin) + 1;
1008                 name = kmem_alloc(namelen, KM_SLEEP);
1009                 dsl_dataset_name(origin, name);
1010 #ifdef _KERNEL
1011                 error = zfs_unmount_snap(name, NULL);
1012                 if (error) {
1013                         kmem_free(name, namelen);
1014                         return (error);
1015                 }
1016 #endif
1017                 error = dsl_dataset_own(name, B_TRUE, tag, &origin);
1018                 kmem_free(name, namelen);
1019                 if (error)
1020                         return (error);
1021                 dsda->rm_origin = origin;
1022                 dsl_dataset_make_exclusive(origin, tag);
1023         }
1024
1025         return (0);
1026 }
1027
1028 /*
1029  * ds must be opened as OWNER.  On return (whether successful or not),
1030  * ds will be closed and caller can no longer dereference it.
1031  */
1032 int
1033 dsl_dataset_destroy(dsl_dataset_t *ds, void *tag, boolean_t defer)
1034 {
1035         int err;
1036         dsl_sync_task_group_t *dstg;
1037         objset_t *os;
1038         dsl_dir_t *dd;
1039         uint64_t obj;
1040         struct dsl_ds_destroyarg dsda = { 0 };
1041         dsl_dataset_t *dummy_ds;
1042
1043         dsda.ds = ds;
1044
1045         if (dsl_dataset_is_snapshot(ds)) {
1046                 /* Destroying a snapshot is simpler */
1047                 dsl_dataset_make_exclusive(ds, tag);
1048
1049                 dsda.defer = defer;
1050                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1051                     dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
1052                     &dsda, tag, 0);
1053                 ASSERT3P(dsda.rm_origin, ==, NULL);
1054                 goto out;
1055         } else if (defer) {
1056                 err = EINVAL;
1057                 goto out;
1058         }
1059
1060         dd = ds->ds_dir;
1061         dummy_ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
1062         dummy_ds->ds_dir = dd;
1063         dummy_ds->ds_object = ds->ds_object;
1064
1065         if (!spa_feature_is_enabled(dsl_dataset_get_spa(ds),
1066             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1067                 /*
1068                  * Check for errors and mark this ds as inconsistent, in
1069                  * case we crash while freeing the objects.
1070                  */
1071                 err = dsl_sync_task_do(dd->dd_pool,
1072                     dsl_dataset_destroy_begin_check,
1073                     dsl_dataset_destroy_begin_sync, ds, NULL, 0);
1074                 if (err)
1075                         goto out_free;
1076
1077                 err = dmu_objset_from_ds(ds, &os);
1078                 if (err)
1079                         goto out_free;
1080
1081                 /*
1082                  * Remove all objects while in the open context so that
1083                  * there is less work to do in the syncing context.
1084                  */
1085                 for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE,
1086                     ds->ds_phys->ds_prev_snap_txg)) {
1087                         /*
1088                          * Ignore errors, if there is not enough disk space
1089                          * we will deal with it in dsl_dataset_destroy_sync().
1090                          */
1091                         (void) dmu_free_object(os, obj);
1092                 }
1093                 if (err != ESRCH)
1094                         goto out_free;
1095
1096                 /*
1097                  * Sync out all in-flight IO.
1098                  */
1099                 txg_wait_synced(dd->dd_pool, 0);
1100
1101                 /*
1102                  * If we managed to free all the objects in open
1103                  * context, the user space accounting should be zero.
1104                  */
1105                 if (ds->ds_phys->ds_bp.blk_fill == 0 &&
1106                     dmu_objset_userused_enabled(os)) {
1107                         ASSERTV(uint64_t count);
1108
1109                         ASSERT(zap_count(os, DMU_USERUSED_OBJECT,
1110                             &count) != 0 || count == 0);
1111                         ASSERT(zap_count(os, DMU_GROUPUSED_OBJECT,
1112                             &count) != 0 || count == 0);
1113                 }
1114         }
1115
1116         rw_enter(&dd->dd_pool->dp_config_rwlock, RW_READER);
1117         err = dsl_dir_open_obj(dd->dd_pool, dd->dd_object, NULL, FTAG, &dd);
1118         rw_exit(&dd->dd_pool->dp_config_rwlock);
1119
1120         if (err)
1121                 goto out_free;
1122
1123         /*
1124          * Blow away the dsl_dir + head dataset.
1125          */
1126         dsl_dataset_make_exclusive(ds, tag);
1127         /*
1128          * If we're removing a clone, we might also need to remove its
1129          * origin.
1130          */
1131         do {
1132                 dsda.need_prep = B_FALSE;
1133                 if (dsl_dir_is_clone(dd)) {
1134                         err = dsl_dataset_origin_rm_prep(&dsda, tag);
1135                         if (err) {
1136                                 dsl_dir_close(dd, FTAG);
1137                                 goto out_free;
1138                         }
1139                 }
1140
1141                 dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
1142                 dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
1143                     dsl_dataset_destroy_sync, &dsda, tag, 0);
1144                 dsl_sync_task_create(dstg, dsl_dir_destroy_check,
1145                     dsl_dir_destroy_sync, dummy_ds, FTAG, 0);
1146                 err = dsl_sync_task_group_wait(dstg);
1147                 dsl_sync_task_group_destroy(dstg);
1148
1149                 /*
1150                  * We could be racing against 'zfs release' or 'zfs destroy -d'
1151                  * on the origin snap, in which case we can get EBUSY if we
1152                  * needed to destroy the origin snap but were not ready to
1153                  * do so.
1154                  */
1155                 if (dsda.need_prep) {
1156                         ASSERT(err == EBUSY);
1157                         ASSERT(dsl_dir_is_clone(dd));
1158                         ASSERT(dsda.rm_origin == NULL);
1159                 }
1160         } while (dsda.need_prep);
1161
1162         if (dsda.rm_origin != NULL)
1163                 dsl_dataset_disown(dsda.rm_origin, tag);
1164
1165         /* if it is successful, dsl_dir_destroy_sync will close the dd */
1166         if (err)
1167                 dsl_dir_close(dd, FTAG);
1168
1169 out_free:
1170         kmem_free(dummy_ds, sizeof (dsl_dataset_t));
1171 out:
1172         dsl_dataset_disown(ds, tag);
1173         return (err);
1174 }
1175
1176 blkptr_t *
1177 dsl_dataset_get_blkptr(dsl_dataset_t *ds)
1178 {
1179         return (&ds->ds_phys->ds_bp);
1180 }
1181
1182 void
1183 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
1184 {
1185         ASSERT(dmu_tx_is_syncing(tx));
1186         /* If it's the meta-objset, set dp_meta_rootbp */
1187         if (ds == NULL) {
1188                 tx->tx_pool->dp_meta_rootbp = *bp;
1189         } else {
1190                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1191                 ds->ds_phys->ds_bp = *bp;
1192         }
1193 }
1194
1195 spa_t *
1196 dsl_dataset_get_spa(dsl_dataset_t *ds)
1197 {
1198         return (ds->ds_dir->dd_pool->dp_spa);
1199 }
1200
1201 void
1202 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
1203 {
1204         dsl_pool_t *dp;
1205
1206         if (ds == NULL) /* this is the meta-objset */
1207                 return;
1208
1209         ASSERT(ds->ds_objset != NULL);
1210
1211         if (ds->ds_phys->ds_next_snap_obj != 0)
1212                 panic("dirtying snapshot!");
1213
1214         dp = ds->ds_dir->dd_pool;
1215
1216         if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
1217                 /* up the hold count until we can be written out */
1218                 dmu_buf_add_ref(ds->ds_dbuf, ds);
1219         }
1220 }
1221
1222 boolean_t
1223 dsl_dataset_is_dirty(dsl_dataset_t *ds)
1224 {
1225         int t;
1226
1227         for (t = 0; t < TXG_SIZE; t++) {
1228                 if (txg_list_member(&ds->ds_dir->dd_pool->dp_dirty_datasets,
1229                     ds, t))
1230                         return (B_TRUE);
1231         }
1232         return (B_FALSE);
1233 }
1234
1235 /*
1236  * The unique space in the head dataset can be calculated by subtracting
1237  * the space used in the most recent snapshot, that is still being used
1238  * in this file system, from the space currently in use.  To figure out
1239  * the space in the most recent snapshot still in use, we need to take
1240  * the total space used in the snapshot and subtract out the space that
1241  * has been freed up since the snapshot was taken.
1242  */
1243 static void
1244 dsl_dataset_recalc_head_uniq(dsl_dataset_t *ds)
1245 {
1246         uint64_t mrs_used;
1247         uint64_t dlused, dlcomp, dluncomp;
1248
1249         ASSERT(!dsl_dataset_is_snapshot(ds));
1250
1251         if (ds->ds_phys->ds_prev_snap_obj != 0)
1252                 mrs_used = ds->ds_prev->ds_phys->ds_referenced_bytes;
1253         else
1254                 mrs_used = 0;
1255
1256         dsl_deadlist_space(&ds->ds_deadlist, &dlused, &dlcomp, &dluncomp);
1257
1258         ASSERT3U(dlused, <=, mrs_used);
1259         ds->ds_phys->ds_unique_bytes =
1260             ds->ds_phys->ds_referenced_bytes - (mrs_used - dlused);
1261
1262         if (spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1263             SPA_VERSION_UNIQUE_ACCURATE)
1264                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1265 }
1266
1267 struct killarg {
1268         dsl_dataset_t *ds;
1269         dmu_tx_t *tx;
1270 };
1271
1272 /* ARGSUSED */
1273 static int
1274 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
1275     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
1276 {
1277         struct killarg *ka = arg;
1278         dmu_tx_t *tx = ka->tx;
1279
1280         if (bp == NULL)
1281                 return (0);
1282
1283         if (zb->zb_level == ZB_ZIL_LEVEL) {
1284                 ASSERT(zilog != NULL);
1285                 /*
1286                  * It's a block in the intent log.  It has no
1287                  * accounting, so just free it.
1288                  */
1289                 dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
1290         } else {
1291                 ASSERT(zilog == NULL);
1292                 ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
1293                 (void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
1294         }
1295
1296         return (0);
1297 }
1298
1299 /* ARGSUSED */
1300 static int
1301 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
1302 {
1303         dsl_dataset_t *ds = arg1;
1304         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1305         uint64_t count;
1306         int err;
1307
1308         /*
1309          * Can't delete a head dataset if there are snapshots of it.
1310          * (Except if the only snapshots are from the branch we cloned
1311          * from.)
1312          */
1313         if (ds->ds_prev != NULL &&
1314             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1315                 return (EBUSY);
1316
1317         /*
1318          * This is really a dsl_dir thing, but check it here so that
1319          * we'll be less likely to leave this dataset inconsistent &
1320          * nearly destroyed.
1321          */
1322         err = zap_count(mos, ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
1323         if (err)
1324                 return (err);
1325         if (count != 0)
1326                 return (EEXIST);
1327
1328         return (0);
1329 }
1330
1331 /* ARGSUSED */
1332 static void
1333 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1334 {
1335         dsl_dataset_t *ds = arg1;
1336         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1337
1338         /* Mark it as inconsistent on-disk, in case we crash */
1339         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1340         ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1341
1342         spa_history_log_internal(LOG_DS_DESTROY_BEGIN, dp->dp_spa, tx,
1343             "dataset = %llu", ds->ds_object);
1344 }
1345
1346 static int
1347 dsl_dataset_origin_check(struct dsl_ds_destroyarg *dsda, void *tag,
1348     dmu_tx_t *tx)
1349 {
1350         dsl_dataset_t *ds = dsda->ds;
1351         dsl_dataset_t *ds_prev = ds->ds_prev;
1352
1353         if (dsl_dataset_might_destroy_origin(ds_prev)) {
1354                 struct dsl_ds_destroyarg ndsda = {0};
1355
1356                 /*
1357                  * If we're not prepared to remove the origin, don't remove
1358                  * the clone either.
1359                  */
1360                 if (dsda->rm_origin == NULL) {
1361                         dsda->need_prep = B_TRUE;
1362                         return (EBUSY);
1363                 }
1364
1365                 ndsda.ds = ds_prev;
1366                 ndsda.is_origin_rm = B_TRUE;
1367                 return (dsl_dataset_destroy_check(&ndsda, tag, tx));
1368         }
1369
1370         /*
1371          * If we're not going to remove the origin after all,
1372          * undo the open context setup.
1373          */
1374         if (dsda->rm_origin != NULL) {
1375                 dsl_dataset_disown(dsda->rm_origin, tag);
1376                 dsda->rm_origin = NULL;
1377         }
1378
1379         return (0);
1380 }
1381
1382 /*
1383  * If you add new checks here, you may need to add
1384  * additional checks to the "temporary" case in
1385  * snapshot_check() in dmu_objset.c.
1386  */
1387 /* ARGSUSED */
1388 int
1389 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1390 {
1391         struct dsl_ds_destroyarg *dsda = arg1;
1392         dsl_dataset_t *ds = dsda->ds;
1393
1394         /* we have an owner hold, so noone else can destroy us */
1395         ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1396
1397         /*
1398          * Only allow deferred destroy on pools that support it.
1399          * NOTE: deferred destroy is only supported on snapshots.
1400          */
1401         if (dsda->defer) {
1402                 if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1403                     SPA_VERSION_USERREFS)
1404                         return (ENOTSUP);
1405                 ASSERT(dsl_dataset_is_snapshot(ds));
1406                 return (0);
1407         }
1408
1409         /*
1410          * Can't delete a head dataset if there are snapshots of it.
1411          * (Except if the only snapshots are from the branch we cloned
1412          * from.)
1413          */
1414         if (ds->ds_prev != NULL &&
1415             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1416                 return (EBUSY);
1417
1418         /*
1419          * If we made changes this txg, traverse_dsl_dataset won't find
1420          * them.  Try again.
1421          */
1422         if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1423                 return (EAGAIN);
1424
1425         if (dsl_dataset_is_snapshot(ds)) {
1426                 /*
1427                  * If this snapshot has an elevated user reference count,
1428                  * we can't destroy it yet.
1429                  */
1430                 if (ds->ds_userrefs > 0 && !dsda->releasing)
1431                         return (EBUSY);
1432
1433                 mutex_enter(&ds->ds_lock);
1434                 /*
1435                  * Can't delete a branch point. However, if we're destroying
1436                  * a clone and removing its origin due to it having a user
1437                  * hold count of 0 and having been marked for deferred destroy,
1438                  * it's OK for the origin to have a single clone.
1439                  */
1440                 if (ds->ds_phys->ds_num_children >
1441                     (dsda->is_origin_rm ? 2 : 1)) {
1442                         mutex_exit(&ds->ds_lock);
1443                         return (EEXIST);
1444                 }
1445                 mutex_exit(&ds->ds_lock);
1446         } else if (dsl_dir_is_clone(ds->ds_dir)) {
1447                 return (dsl_dataset_origin_check(dsda, arg2, tx));
1448         }
1449
1450         /* XXX we should do some i/o error checking... */
1451         return (0);
1452 }
1453
1454 struct refsarg {
1455         kmutex_t lock;
1456         boolean_t gone;
1457         kcondvar_t cv;
1458 };
1459
1460 /* ARGSUSED */
1461 static void
1462 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1463 {
1464         struct refsarg *arg = argv;
1465
1466         mutex_enter(&arg->lock);
1467         arg->gone = TRUE;
1468         cv_signal(&arg->cv);
1469         mutex_exit(&arg->lock);
1470 }
1471
1472 static void
1473 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1474 {
1475         struct refsarg arg;
1476
1477         mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1478         cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1479         arg.gone = FALSE;
1480         (void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1481             dsl_dataset_refs_gone);
1482         dmu_buf_rele(ds->ds_dbuf, tag);
1483         mutex_enter(&arg.lock);
1484         while (!arg.gone)
1485                 cv_wait(&arg.cv, &arg.lock);
1486         ASSERT(arg.gone);
1487         mutex_exit(&arg.lock);
1488         ds->ds_dbuf = NULL;
1489         ds->ds_phys = NULL;
1490         mutex_destroy(&arg.lock);
1491         cv_destroy(&arg.cv);
1492 }
1493
1494 static void
1495 remove_from_next_clones(dsl_dataset_t *ds, uint64_t obj, dmu_tx_t *tx)
1496 {
1497         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1498         int err;
1499         ASSERTV(uint64_t count);
1500
1501         ASSERT(ds->ds_phys->ds_num_children >= 2);
1502         err = zap_remove_int(mos, ds->ds_phys->ds_next_clones_obj, obj, tx);
1503         /*
1504          * The err should not be ENOENT, but a bug in a previous version
1505          * of the code could cause upgrade_clones_cb() to not set
1506          * ds_next_snap_obj when it should, leading to a missing entry.
1507          * If we knew that the pool was created after
1508          * SPA_VERSION_NEXT_CLONES, we could assert that it isn't
1509          * ENOENT.  However, at least we can check that we don't have
1510          * too many entries in the next_clones_obj even after failing to
1511          * remove this one.
1512          */
1513         if (err != ENOENT) {
1514                 VERIFY3U(err, ==, 0);
1515         }
1516         ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
1517             &count));
1518         ASSERT3U(count, <=, ds->ds_phys->ds_num_children - 2);
1519 }
1520
1521 static void
1522 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
1523 {
1524         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1525         zap_cursor_t zc;
1526         zap_attribute_t za;
1527
1528         /*
1529          * If it is the old version, dd_clones doesn't exist so we can't
1530          * find the clones, but deadlist_remove_key() is a no-op so it
1531          * doesn't matter.
1532          */
1533         if (ds->ds_dir->dd_phys->dd_clones == 0)
1534                 return;
1535
1536         for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
1537             zap_cursor_retrieve(&zc, &za) == 0;
1538             zap_cursor_advance(&zc)) {
1539                 dsl_dataset_t *clone;
1540
1541                 VERIFY3U(0, ==, dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
1542                     za.za_first_integer, FTAG, &clone));
1543                 if (clone->ds_dir->dd_origin_txg > mintxg) {
1544                         dsl_deadlist_remove_key(&clone->ds_deadlist,
1545                             mintxg, tx);
1546                         dsl_dataset_remove_clones_key(clone, mintxg, tx);
1547                 }
1548                 dsl_dataset_rele(clone, FTAG);
1549         }
1550         zap_cursor_fini(&zc);
1551 }
1552
1553 struct process_old_arg {
1554         dsl_dataset_t *ds;
1555         dsl_dataset_t *ds_prev;
1556         boolean_t after_branch_point;
1557         zio_t *pio;
1558         uint64_t used, comp, uncomp;
1559 };
1560
1561 static int
1562 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1563 {
1564         struct process_old_arg *poa = arg;
1565         dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
1566
1567         if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
1568                 dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
1569                 if (poa->ds_prev && !poa->after_branch_point &&
1570                     bp->blk_birth >
1571                     poa->ds_prev->ds_phys->ds_prev_snap_txg) {
1572                         poa->ds_prev->ds_phys->ds_unique_bytes +=
1573                             bp_get_dsize_sync(dp->dp_spa, bp);
1574                 }
1575         } else {
1576                 poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
1577                 poa->comp += BP_GET_PSIZE(bp);
1578                 poa->uncomp += BP_GET_UCSIZE(bp);
1579                 dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
1580         }
1581         return (0);
1582 }
1583
1584 static void
1585 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
1586     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
1587 {
1588         struct process_old_arg poa = { 0 };
1589         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1590         objset_t *mos = dp->dp_meta_objset;
1591
1592         ASSERT(ds->ds_deadlist.dl_oldfmt);
1593         ASSERT(ds_next->ds_deadlist.dl_oldfmt);
1594
1595         poa.ds = ds;
1596         poa.ds_prev = ds_prev;
1597         poa.after_branch_point = after_branch_point;
1598         poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1599         VERIFY3U(0, ==, bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
1600             process_old_cb, &poa, tx));
1601         VERIFY3U(zio_wait(poa.pio), ==, 0);
1602         ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
1603
1604         /* change snapused */
1605         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1606             -poa.used, -poa.comp, -poa.uncomp, tx);
1607
1608         /* swap next's deadlist to our deadlist */
1609         dsl_deadlist_close(&ds->ds_deadlist);
1610         dsl_deadlist_close(&ds_next->ds_deadlist);
1611         SWITCH64(ds_next->ds_phys->ds_deadlist_obj,
1612             ds->ds_phys->ds_deadlist_obj);
1613         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
1614         dsl_deadlist_open(&ds_next->ds_deadlist, mos,
1615             ds_next->ds_phys->ds_deadlist_obj);
1616 }
1617
1618 static int
1619 old_synchronous_dataset_destroy(dsl_dataset_t *ds, dmu_tx_t *tx)
1620 {
1621         int err;
1622         struct killarg ka;
1623
1624         /*
1625          * Free everything that we point to (that's born after
1626          * the previous snapshot, if we are a clone)
1627          *
1628          * NB: this should be very quick, because we already
1629          * freed all the objects in open context.
1630          */
1631         ka.ds = ds;
1632         ka.tx = tx;
1633         err = traverse_dataset(ds,
1634             ds->ds_phys->ds_prev_snap_txg, TRAVERSE_POST,
1635             kill_blkptr, &ka);
1636         ASSERT3U(err, ==, 0);
1637         ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) || ds->ds_phys->ds_unique_bytes == 0);
1638
1639         return (err);
1640 }
1641
1642 void
1643 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1644 {
1645         struct dsl_ds_destroyarg *dsda = arg1;
1646         dsl_dataset_t *ds = dsda->ds;
1647         int err = 0;
1648         int after_branch_point = FALSE;
1649         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1650         objset_t *mos = dp->dp_meta_objset;
1651         dsl_dataset_t *ds_prev = NULL;
1652         boolean_t wont_destroy;
1653         uint64_t obj;
1654
1655         wont_destroy = (dsda->defer &&
1656             (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1));
1657
1658         ASSERT(ds->ds_owner || wont_destroy);
1659         ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1660         ASSERT(ds->ds_prev == NULL ||
1661             ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1662         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1663
1664         if (wont_destroy) {
1665                 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1666                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1667                 ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1668                 return;
1669         }
1670
1671         /* signal any waiters that this dataset is going away */
1672         mutex_enter(&ds->ds_lock);
1673         ds->ds_owner = dsl_reaper;
1674         cv_broadcast(&ds->ds_exclusive_cv);
1675         mutex_exit(&ds->ds_lock);
1676
1677         /* Remove our reservation */
1678         if (ds->ds_reserved != 0) {
1679                 dsl_prop_setarg_t psa;
1680                 uint64_t value = 0;
1681
1682                 dsl_prop_setarg_init_uint64(&psa, "refreservation",
1683                     (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
1684                     &value);
1685                 psa.psa_effective_value = 0;    /* predict default value */
1686
1687                 dsl_dataset_set_reservation_sync(ds, &psa, tx);
1688                 ASSERT3U(ds->ds_reserved, ==, 0);
1689         }
1690
1691         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1692
1693         dsl_scan_ds_destroyed(ds, tx);
1694
1695         obj = ds->ds_object;
1696
1697         if (ds->ds_phys->ds_prev_snap_obj != 0) {
1698                 if (ds->ds_prev) {
1699                         ds_prev = ds->ds_prev;
1700                 } else {
1701                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1702                             ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1703                 }
1704                 after_branch_point =
1705                     (ds_prev->ds_phys->ds_next_snap_obj != obj);
1706
1707                 dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1708                 if (after_branch_point &&
1709                     ds_prev->ds_phys->ds_next_clones_obj != 0) {
1710                         remove_from_next_clones(ds_prev, obj, tx);
1711                         if (ds->ds_phys->ds_next_snap_obj != 0) {
1712                                 VERIFY(0 == zap_add_int(mos,
1713                                     ds_prev->ds_phys->ds_next_clones_obj,
1714                                     ds->ds_phys->ds_next_snap_obj, tx));
1715                         }
1716                 }
1717                 if (after_branch_point &&
1718                     ds->ds_phys->ds_next_snap_obj == 0) {
1719                         /* This clone is toast. */
1720                         ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1721                         ds_prev->ds_phys->ds_num_children--;
1722
1723                         /*
1724                          * If the clone's origin has no other clones, no
1725                          * user holds, and has been marked for deferred
1726                          * deletion, then we should have done the necessary
1727                          * destroy setup for it.
1728                          */
1729                         if (ds_prev->ds_phys->ds_num_children == 1 &&
1730                             ds_prev->ds_userrefs == 0 &&
1731                             DS_IS_DEFER_DESTROY(ds_prev)) {
1732                                 ASSERT3P(dsda->rm_origin, !=, NULL);
1733                         } else {
1734                                 ASSERT3P(dsda->rm_origin, ==, NULL);
1735                         }
1736                 } else if (!after_branch_point) {
1737                         ds_prev->ds_phys->ds_next_snap_obj =
1738                             ds->ds_phys->ds_next_snap_obj;
1739                 }
1740         }
1741
1742         if (dsl_dataset_is_snapshot(ds)) {
1743                 dsl_dataset_t *ds_next;
1744                 uint64_t old_unique;
1745                 uint64_t used = 0, comp = 0, uncomp = 0;
1746
1747                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1748                     ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1749                 ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1750
1751                 old_unique = ds_next->ds_phys->ds_unique_bytes;
1752
1753                 dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1754                 ds_next->ds_phys->ds_prev_snap_obj =
1755                     ds->ds_phys->ds_prev_snap_obj;
1756                 ds_next->ds_phys->ds_prev_snap_txg =
1757                     ds->ds_phys->ds_prev_snap_txg;
1758                 ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1759                     ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1760
1761
1762                 if (ds_next->ds_deadlist.dl_oldfmt) {
1763                         process_old_deadlist(ds, ds_prev, ds_next,
1764                             after_branch_point, tx);
1765                 } else {
1766                         /* Adjust prev's unique space. */
1767                         if (ds_prev && !after_branch_point) {
1768                                 dsl_deadlist_space_range(&ds_next->ds_deadlist,
1769                                     ds_prev->ds_phys->ds_prev_snap_txg,
1770                                     ds->ds_phys->ds_prev_snap_txg,
1771                                     &used, &comp, &uncomp);
1772                                 ds_prev->ds_phys->ds_unique_bytes += used;
1773                         }
1774
1775                         /* Adjust snapused. */
1776                         dsl_deadlist_space_range(&ds_next->ds_deadlist,
1777                             ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
1778                             &used, &comp, &uncomp);
1779                         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1780                             -used, -comp, -uncomp, tx);
1781
1782                         /* Move blocks to be freed to pool's free list. */
1783                         dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
1784                             &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
1785                             tx);
1786                         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
1787                             DD_USED_HEAD, used, comp, uncomp, tx);
1788
1789                         /* Merge our deadlist into next's and free it. */
1790                         dsl_deadlist_merge(&ds_next->ds_deadlist,
1791                             ds->ds_phys->ds_deadlist_obj, tx);
1792                 }
1793                 dsl_deadlist_close(&ds->ds_deadlist);
1794                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1795
1796                 /* Collapse range in clone heads */
1797                 dsl_dataset_remove_clones_key(ds,
1798                     ds->ds_phys->ds_creation_txg, tx);
1799
1800                 if (dsl_dataset_is_snapshot(ds_next)) {
1801                         dsl_dataset_t *ds_nextnext;
1802                         dsl_dataset_t *hds;
1803
1804                         /*
1805                          * Update next's unique to include blocks which
1806                          * were previously shared by only this snapshot
1807                          * and it.  Those blocks will be born after the
1808                          * prev snap and before this snap, and will have
1809                          * died after the next snap and before the one
1810                          * after that (ie. be on the snap after next's
1811                          * deadlist).
1812                          */
1813                         VERIFY(0 == dsl_dataset_hold_obj(dp,
1814                             ds_next->ds_phys->ds_next_snap_obj,
1815                             FTAG, &ds_nextnext));
1816                         dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
1817                             ds->ds_phys->ds_prev_snap_txg,
1818                             ds->ds_phys->ds_creation_txg,
1819                             &used, &comp, &uncomp);
1820                         ds_next->ds_phys->ds_unique_bytes += used;
1821                         dsl_dataset_rele(ds_nextnext, FTAG);
1822                         ASSERT3P(ds_next->ds_prev, ==, NULL);
1823
1824                         /* Collapse range in this head. */
1825                         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
1826                             ds->ds_dir->dd_phys->dd_head_dataset_obj,
1827                             FTAG, &hds));
1828                         dsl_deadlist_remove_key(&hds->ds_deadlist,
1829                             ds->ds_phys->ds_creation_txg, tx);
1830                         dsl_dataset_rele(hds, FTAG);
1831
1832                 } else {
1833                         ASSERT3P(ds_next->ds_prev, ==, ds);
1834                         dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1835                         ds_next->ds_prev = NULL;
1836                         if (ds_prev) {
1837                                 VERIFY(0 == dsl_dataset_get_ref(dp,
1838                                     ds->ds_phys->ds_prev_snap_obj,
1839                                     ds_next, &ds_next->ds_prev));
1840                         }
1841
1842                         dsl_dataset_recalc_head_uniq(ds_next);
1843
1844                         /*
1845                          * Reduce the amount of our unconsmed refreservation
1846                          * being charged to our parent by the amount of
1847                          * new unique data we have gained.
1848                          */
1849                         if (old_unique < ds_next->ds_reserved) {
1850                                 int64_t mrsdelta;
1851                                 uint64_t new_unique =
1852                                     ds_next->ds_phys->ds_unique_bytes;
1853
1854                                 ASSERT(old_unique <= new_unique);
1855                                 mrsdelta = MIN(new_unique - old_unique,
1856                                     ds_next->ds_reserved - old_unique);
1857                                 dsl_dir_diduse_space(ds->ds_dir,
1858                                     DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1859                         }
1860                 }
1861                 dsl_dataset_rele(ds_next, FTAG);
1862         } else {
1863                 zfeature_info_t *async_destroy =
1864                     &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY];
1865                 objset_t *os;
1866
1867                 /*
1868                  * There's no next snapshot, so this is a head dataset.
1869                  * Destroy the deadlist.  Unless it's a clone, the
1870                  * deadlist should be empty.  (If it's a clone, it's
1871                  * safe to ignore the deadlist contents.)
1872                  */
1873                 dsl_deadlist_close(&ds->ds_deadlist);
1874                 dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1875                 ds->ds_phys->ds_deadlist_obj = 0;
1876
1877                 VERIFY3U(0, ==, dmu_objset_from_ds(ds, &os));
1878
1879                 if (!spa_feature_is_enabled(dp->dp_spa, async_destroy)) {
1880                         err = old_synchronous_dataset_destroy(ds, tx);
1881                 } else {
1882                         /*
1883                          * Move the bptree into the pool's list of trees to
1884                          * clean up and update space accounting information.
1885                          */
1886                         uint64_t used, comp, uncomp;
1887
1888                         zil_destroy_sync(dmu_objset_zil(os), tx);
1889
1890                         if (!spa_feature_is_active(dp->dp_spa, async_destroy)) {
1891                                 spa_feature_incr(dp->dp_spa, async_destroy, tx);
1892                                 dp->dp_bptree_obj = bptree_alloc(mos, tx);
1893                                 VERIFY(zap_add(mos,
1894                                     DMU_POOL_DIRECTORY_OBJECT,
1895                                     DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
1896                                     &dp->dp_bptree_obj, tx) == 0);
1897                         }
1898
1899                         used = ds->ds_dir->dd_phys->dd_used_bytes;
1900                         comp = ds->ds_dir->dd_phys->dd_compressed_bytes;
1901                         uncomp = ds->ds_dir->dd_phys->dd_uncompressed_bytes;
1902
1903                         ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1904                             ds->ds_phys->ds_unique_bytes == used);
1905
1906                         bptree_add(mos, dp->dp_bptree_obj,
1907                             &ds->ds_phys->ds_bp, ds->ds_phys->ds_prev_snap_txg,
1908                             used, comp, uncomp, tx);
1909                         dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
1910                             -used, -comp, -uncomp, tx);
1911                         dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1912                             used, comp, uncomp, tx);
1913                 }
1914
1915                 if (ds->ds_prev != NULL) {
1916                         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
1917                                 VERIFY3U(0, ==, zap_remove_int(mos,
1918                                     ds->ds_prev->ds_dir->dd_phys->dd_clones,
1919                                     ds->ds_object, tx));
1920                         }
1921                         dsl_dataset_rele(ds->ds_prev, ds);
1922                         ds->ds_prev = ds_prev = NULL;
1923                 }
1924         }
1925
1926         /*
1927          * This must be done after the dsl_traverse(), because it will
1928          * re-open the objset.
1929          */
1930         if (ds->ds_objset) {
1931                 dmu_objset_evict(ds->ds_objset);
1932                 ds->ds_objset = NULL;
1933         }
1934
1935         if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1936                 /* Erase the link in the dir */
1937                 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1938                 ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1939                 ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1940                 err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1941                 ASSERT(err == 0);
1942         } else {
1943                 /* remove from snapshot namespace */
1944                 dsl_dataset_t *ds_head;
1945                 ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1946                 VERIFY(0 == dsl_dataset_hold_obj(dp,
1947                     ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1948                 VERIFY(0 == dsl_dataset_get_snapname(ds));
1949 #ifdef ZFS_DEBUG
1950                 {
1951                         uint64_t val;
1952
1953                         err = dsl_dataset_snap_lookup(ds_head,
1954                             ds->ds_snapname, &val);
1955                         ASSERT3U(err, ==, 0);
1956                         ASSERT3U(val, ==, obj);
1957                 }
1958 #endif
1959                 err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1960                 ASSERT(err == 0);
1961                 dsl_dataset_rele(ds_head, FTAG);
1962         }
1963
1964         if (ds_prev && ds->ds_prev != ds_prev)
1965                 dsl_dataset_rele(ds_prev, FTAG);
1966
1967         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1968         spa_history_log_internal(LOG_DS_DESTROY, dp->dp_spa, tx,
1969             "dataset = %llu", ds->ds_object);
1970
1971         if (ds->ds_phys->ds_next_clones_obj != 0) {
1972                 ASSERTV(uint64_t count);
1973                 ASSERT(0 == zap_count(mos,
1974                     ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1975                 VERIFY(0 == dmu_object_free(mos,
1976                     ds->ds_phys->ds_next_clones_obj, tx));
1977         }
1978         if (ds->ds_phys->ds_props_obj != 0)
1979                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1980         if (ds->ds_phys->ds_userrefs_obj != 0)
1981                 VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1982         dsl_dir_close(ds->ds_dir, ds);
1983         ds->ds_dir = NULL;
1984         dsl_dataset_drain_refs(ds, tag);
1985         VERIFY(0 == dmu_object_free(mos, obj, tx));
1986
1987         if (dsda->rm_origin) {
1988                 /*
1989                  * Remove the origin of the clone we just destroyed.
1990                  */
1991                 struct dsl_ds_destroyarg ndsda = {0};
1992
1993                 ndsda.ds = dsda->rm_origin;
1994                 dsl_dataset_destroy_sync(&ndsda, tag, tx);
1995         }
1996 }
1997
1998 static int
1999 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
2000 {
2001         uint64_t asize;
2002
2003         if (!dmu_tx_is_syncing(tx))
2004                 return (0);
2005
2006         /*
2007          * If there's an fs-only reservation, any blocks that might become
2008          * owned by the snapshot dataset must be accommodated by space
2009          * outside of the reservation.
2010          */
2011         ASSERT(ds->ds_reserved == 0 || DS_UNIQUE_IS_ACCURATE(ds));
2012         asize = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2013         if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
2014                 return (ENOSPC);
2015
2016         /*
2017          * Propogate any reserved space for this snapshot to other
2018          * snapshot checks in this sync group.
2019          */
2020         if (asize > 0)
2021                 dsl_dir_willuse_space(ds->ds_dir, asize, tx);
2022
2023         return (0);
2024 }
2025
2026 int
2027 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
2028 {
2029         dsl_dataset_t *ds = arg1;
2030         const char *snapname = arg2;
2031         int err;
2032         uint64_t value;
2033
2034         /*
2035          * We don't allow multiple snapshots of the same txg.  If there
2036          * is already one, try again.
2037          */
2038         if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
2039                 return (EAGAIN);
2040
2041         /*
2042          * Check for conflicting name snapshot name.
2043          */
2044         err = dsl_dataset_snap_lookup(ds, snapname, &value);
2045         if (err == 0)
2046                 return (EEXIST);
2047         if (err != ENOENT)
2048                 return (err);
2049
2050         /*
2051          * Check that the dataset's name is not too long.  Name consists
2052          * of the dataset's length + 1 for the @-sign + snapshot name's length
2053          */
2054         if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
2055                 return (ENAMETOOLONG);
2056
2057         err = dsl_dataset_snapshot_reserve_space(ds, tx);
2058         if (err)
2059                 return (err);
2060
2061         ds->ds_trysnap_txg = tx->tx_txg;
2062         return (0);
2063 }
2064
2065 void
2066 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2067 {
2068         dsl_dataset_t *ds = arg1;
2069         const char *snapname = arg2;
2070         dsl_pool_t *dp = ds->ds_dir->dd_pool;
2071         dmu_buf_t *dbuf;
2072         dsl_dataset_phys_t *dsphys;
2073         uint64_t dsobj, crtxg;
2074         objset_t *mos = dp->dp_meta_objset;
2075         int err;
2076
2077         ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
2078
2079         /*
2080          * The origin's ds_creation_txg has to be < TXG_INITIAL
2081          */
2082         if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2083                 crtxg = 1;
2084         else
2085                 crtxg = tx->tx_txg;
2086
2087         dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2088             DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2089         VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2090         dmu_buf_will_dirty(dbuf, tx);
2091         dsphys = dbuf->db_data;
2092         bzero(dsphys, sizeof (dsl_dataset_phys_t));
2093         dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2094         dsphys->ds_fsid_guid = unique_create();
2095         (void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2096             sizeof (dsphys->ds_guid));
2097         dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2098         dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2099         dsphys->ds_next_snap_obj = ds->ds_object;
2100         dsphys->ds_num_children = 1;
2101         dsphys->ds_creation_time = gethrestime_sec();
2102         dsphys->ds_creation_txg = crtxg;
2103         dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2104         dsphys->ds_referenced_bytes = ds->ds_phys->ds_referenced_bytes;
2105         dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2106         dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2107         dsphys->ds_flags = ds->ds_phys->ds_flags;
2108         dsphys->ds_bp = ds->ds_phys->ds_bp;
2109         dmu_buf_rele(dbuf, FTAG);
2110
2111         ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2112         if (ds->ds_prev) {
2113                 uint64_t next_clones_obj =
2114                     ds->ds_prev->ds_phys->ds_next_clones_obj;
2115                 ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2116                     ds->ds_object ||
2117                     ds->ds_prev->ds_phys->ds_num_children > 1);
2118                 if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2119                         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2120                         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2121                             ds->ds_prev->ds_phys->ds_creation_txg);
2122                         ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2123                 } else if (next_clones_obj != 0) {
2124                         remove_from_next_clones(ds->ds_prev,
2125                             dsphys->ds_next_snap_obj, tx);
2126                         VERIFY3U(0, ==, zap_add_int(mos,
2127                             next_clones_obj, dsobj, tx));
2128                 }
2129         }
2130
2131         /*
2132          * If we have a reference-reservation on this dataset, we will
2133          * need to increase the amount of refreservation being charged
2134          * since our unique space is going to zero.
2135          */
2136         if (ds->ds_reserved) {
2137                 int64_t delta;
2138                 ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
2139                 delta = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2140                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2141                     delta, 0, 0, tx);
2142         }
2143
2144         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2145         zfs_dbgmsg("taking snapshot %s@%s/%llu; newkey=%llu",
2146             ds->ds_dir->dd_myname, snapname, dsobj,
2147             ds->ds_phys->ds_prev_snap_txg);
2148         ds->ds_phys->ds_deadlist_obj = dsl_deadlist_clone(&ds->ds_deadlist,
2149             UINT64_MAX, ds->ds_phys->ds_prev_snap_obj, tx);
2150         dsl_deadlist_close(&ds->ds_deadlist);
2151         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
2152         dsl_deadlist_add_key(&ds->ds_deadlist,
2153             ds->ds_phys->ds_prev_snap_txg, tx);
2154
2155         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2156         ds->ds_phys->ds_prev_snap_obj = dsobj;
2157         ds->ds_phys->ds_prev_snap_txg = crtxg;
2158         ds->ds_phys->ds_unique_bytes = 0;
2159         if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2160                 ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2161
2162         err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2163             snapname, 8, 1, &dsobj, tx);
2164         ASSERT(err == 0);
2165
2166         if (ds->ds_prev)
2167                 dsl_dataset_drop_ref(ds->ds_prev, ds);
2168         VERIFY(0 == dsl_dataset_get_ref(dp,
2169             ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2170
2171         dsl_scan_ds_snapshotted(ds, tx);
2172
2173         dsl_dir_snap_cmtime_update(ds->ds_dir);
2174
2175         spa_history_log_internal(LOG_DS_SNAPSHOT, dp->dp_spa, tx,
2176             "dataset = %llu", dsobj);
2177 }
2178
2179 void
2180 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2181 {
2182         ASSERT(dmu_tx_is_syncing(tx));
2183         ASSERT(ds->ds_objset != NULL);
2184         ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2185
2186         /*
2187          * in case we had to change ds_fsid_guid when we opened it,
2188          * sync it out now.
2189          */
2190         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2191         ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2192
2193         dmu_objset_sync(ds->ds_objset, zio, tx);
2194 }
2195
2196 static void
2197 get_clones_stat(dsl_dataset_t *ds, nvlist_t *nv)
2198 {
2199         uint64_t count = 0;
2200         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
2201         zap_cursor_t zc;
2202         zap_attribute_t za;
2203         nvlist_t *propval;
2204         nvlist_t *val;
2205
2206         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2207         VERIFY(nvlist_alloc(&propval, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2208         VERIFY(nvlist_alloc(&val, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2209
2210         /*
2211          * There may me missing entries in ds_next_clones_obj
2212          * due to a bug in a previous version of the code.
2213          * Only trust it if it has the right number of entries.
2214          */
2215         if (ds->ds_phys->ds_next_clones_obj != 0) {
2216                 ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
2217                     &count));
2218         }
2219         if (count != ds->ds_phys->ds_num_children - 1) {
2220                 goto fail;
2221         }
2222         for (zap_cursor_init(&zc, mos, ds->ds_phys->ds_next_clones_obj);
2223             zap_cursor_retrieve(&zc, &za) == 0;
2224             zap_cursor_advance(&zc)) {
2225                 dsl_dataset_t *clone;
2226                 char buf[ZFS_MAXNAMELEN];
2227                 /*
2228                  * Even though we hold the dp_config_rwlock, the dataset
2229                  * may fail to open, returning ENOENT.  If there is a
2230                  * thread concurrently attempting to destroy this
2231                  * dataset, it will have the ds_rwlock held for
2232                  * RW_WRITER.  Our call to dsl_dataset_hold_obj() ->
2233                  * dsl_dataset_hold_ref() will fail its
2234                  * rw_tryenter(&ds->ds_rwlock, RW_READER), drop the
2235                  * dp_config_rwlock, and wait for the destroy progress
2236                  * and signal ds_exclusive_cv.  If the destroy was
2237                  * successful, we will see that
2238                  * DSL_DATASET_IS_DESTROYED(), and return ENOENT.
2239                  */
2240                 if (dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
2241                     za.za_first_integer, FTAG, &clone) != 0)
2242                         continue;
2243                 dsl_dir_name(clone->ds_dir, buf);
2244                 VERIFY(nvlist_add_boolean(val, buf) == 0);
2245                 dsl_dataset_rele(clone, FTAG);
2246         }
2247         zap_cursor_fini(&zc);
2248         VERIFY(nvlist_add_nvlist(propval, ZPROP_VALUE, val) == 0);
2249         VERIFY(nvlist_add_nvlist(nv, zfs_prop_to_name(ZFS_PROP_CLONES),
2250             propval) == 0);
2251 fail:
2252         nvlist_free(val);
2253         nvlist_free(propval);
2254         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2255 }
2256
2257 void
2258 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2259 {
2260         uint64_t refd, avail, uobjs, aobjs, ratio;
2261
2262         dsl_dir_stats(ds->ds_dir, nv);
2263
2264         dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2265         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2266         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2267
2268         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2269             ds->ds_phys->ds_creation_time);
2270         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2271             ds->ds_phys->ds_creation_txg);
2272         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2273             ds->ds_quota);
2274         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2275             ds->ds_reserved);
2276         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2277             ds->ds_phys->ds_guid);
2278         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_UNIQUE,
2279             ds->ds_phys->ds_unique_bytes);
2280         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_OBJSETID,
2281             ds->ds_object);
2282         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS,
2283             ds->ds_userrefs);
2284         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2285             DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2286
2287         if (ds->ds_phys->ds_prev_snap_obj != 0) {
2288                 uint64_t written, comp, uncomp;
2289                 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2290                 dsl_dataset_t *prev;
2291                 int err;
2292
2293                 rw_enter(&dp->dp_config_rwlock, RW_READER);
2294                 err = dsl_dataset_hold_obj(dp,
2295                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
2296                 rw_exit(&dp->dp_config_rwlock);
2297                 if (err == 0) {
2298                         err = dsl_dataset_space_written(prev, ds, &written,
2299                             &comp, &uncomp);
2300                         dsl_dataset_rele(prev, FTAG);
2301                         if (err == 0) {
2302                                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_WRITTEN,
2303                                     written);
2304                         }
2305                 }
2306         }
2307
2308         ratio = ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2309             (ds->ds_phys->ds_uncompressed_bytes * 100 /
2310             ds->ds_phys->ds_compressed_bytes);
2311         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRATIO, ratio);
2312
2313         if (ds->ds_phys->ds_next_snap_obj) {
2314                 /*
2315                  * This is a snapshot; override the dd's space used with
2316                  * our unique space and compression ratio.
2317                  */
2318                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2319                     ds->ds_phys->ds_unique_bytes);
2320                 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO, ratio);
2321
2322                 get_clones_stat(ds, nv);
2323         }
2324 }
2325
2326 void
2327 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2328 {
2329         stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2330         stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2331         stat->dds_guid = ds->ds_phys->ds_guid;
2332         if (ds->ds_phys->ds_next_snap_obj) {
2333                 stat->dds_is_snapshot = B_TRUE;
2334                 stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2335         } else {
2336                 stat->dds_is_snapshot = B_FALSE;
2337                 stat->dds_num_clones = 0;
2338         }
2339
2340         /* clone origin is really a dsl_dir thing... */
2341         rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2342         if (dsl_dir_is_clone(ds->ds_dir)) {
2343                 dsl_dataset_t *ods;
2344
2345                 VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2346                     ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2347                 dsl_dataset_name(ods, stat->dds_origin);
2348                 dsl_dataset_drop_ref(ods, FTAG);
2349         } else {
2350                 stat->dds_origin[0] = '\0';
2351         }
2352         rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2353 }
2354
2355 uint64_t
2356 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2357 {
2358         return (ds->ds_fsid_guid);
2359 }
2360
2361 void
2362 dsl_dataset_space(dsl_dataset_t *ds,
2363     uint64_t *refdbytesp, uint64_t *availbytesp,
2364     uint64_t *usedobjsp, uint64_t *availobjsp)
2365 {
2366         *refdbytesp = ds->ds_phys->ds_referenced_bytes;
2367         *availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2368         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2369                 *availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2370         if (ds->ds_quota != 0) {
2371                 /*
2372                  * Adjust available bytes according to refquota
2373                  */
2374                 if (*refdbytesp < ds->ds_quota)
2375                         *availbytesp = MIN(*availbytesp,
2376                             ds->ds_quota - *refdbytesp);
2377                 else
2378                         *availbytesp = 0;
2379         }
2380         *usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2381         *availobjsp = DN_MAX_OBJECT - *usedobjsp;
2382 }
2383
2384 boolean_t
2385 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2386 {
2387         ASSERTV(dsl_pool_t *dp = ds->ds_dir->dd_pool);
2388
2389         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2390             dsl_pool_sync_context(dp));
2391         if (ds->ds_prev == NULL)
2392                 return (B_FALSE);
2393         if (ds->ds_phys->ds_bp.blk_birth >
2394             ds->ds_prev->ds_phys->ds_creation_txg) {
2395                 objset_t *os, *os_prev;
2396                 /*
2397                  * It may be that only the ZIL differs, because it was
2398                  * reset in the head.  Don't count that as being
2399                  * modified.
2400                  */
2401                 if (dmu_objset_from_ds(ds, &os) != 0)
2402                         return (B_TRUE);
2403                 if (dmu_objset_from_ds(ds->ds_prev, &os_prev) != 0)
2404                         return (B_TRUE);
2405                 return (bcmp(&os->os_phys->os_meta_dnode,
2406                     &os_prev->os_phys->os_meta_dnode,
2407                     sizeof (os->os_phys->os_meta_dnode)) != 0);
2408         }
2409         return (B_FALSE);
2410 }
2411
2412 /* ARGSUSED */
2413 static int
2414 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2415 {
2416         dsl_dataset_t *ds = arg1;
2417         char *newsnapname = arg2;
2418         dsl_dir_t *dd = ds->ds_dir;
2419         dsl_dataset_t *hds;
2420         uint64_t val;
2421         int err;
2422
2423         err = dsl_dataset_hold_obj(dd->dd_pool,
2424             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2425         if (err)
2426                 return (err);
2427
2428         /* new name better not be in use */
2429         err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2430         dsl_dataset_rele(hds, FTAG);
2431
2432         if (err == 0)
2433                 err = EEXIST;
2434         else if (err == ENOENT)
2435                 err = 0;
2436
2437         /* dataset name + 1 for the "@" + the new snapshot name must fit */
2438         if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2439                 err = ENAMETOOLONG;
2440
2441         return (err);
2442 }
2443
2444 static void
2445 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2446 {
2447         dsl_dataset_t *ds = arg1;
2448         const char *newsnapname = arg2;
2449         dsl_dir_t *dd = ds->ds_dir;
2450         objset_t *mos = dd->dd_pool->dp_meta_objset;
2451         dsl_dataset_t *hds;
2452         int err;
2453
2454         ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2455
2456         VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2457             dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2458
2459         VERIFY(0 == dsl_dataset_get_snapname(ds));
2460         err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2461         ASSERT3U(err, ==, 0);
2462         mutex_enter(&ds->ds_lock);
2463         (void) strcpy(ds->ds_snapname, newsnapname);
2464         mutex_exit(&ds->ds_lock);
2465         err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2466             ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2467         ASSERT3U(err, ==, 0);
2468
2469         spa_history_log_internal(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2470             "dataset = %llu", ds->ds_object);
2471         dsl_dataset_rele(hds, FTAG);
2472 }
2473
2474 struct renamesnaparg {
2475         dsl_sync_task_group_t *dstg;
2476         char failed[MAXPATHLEN];
2477         char *oldsnap;
2478         char *newsnap;
2479 };
2480
2481 static int
2482 dsl_snapshot_rename_one(const char *name, void *arg)
2483 {
2484         struct renamesnaparg *ra = arg;
2485         dsl_dataset_t *ds = NULL;
2486         char *snapname;
2487         int err;
2488
2489         snapname = kmem_asprintf("%s@%s", name, ra->oldsnap);
2490         (void) strlcpy(ra->failed, snapname, sizeof (ra->failed));
2491
2492         /*
2493          * For recursive snapshot renames the parent won't be changing
2494          * so we just pass name for both the to/from argument.
2495          */
2496         err = zfs_secpolicy_rename_perms(snapname, snapname, CRED());
2497         if (err != 0) {
2498                 strfree(snapname);
2499                 return (err == ENOENT ? 0 : err);
2500         }
2501
2502 #ifdef _KERNEL
2503         /*
2504          * For all filesystems undergoing rename, we'll need to unmount it.
2505          */
2506         (void) zfs_unmount_snap(snapname, NULL);
2507 #endif
2508         err = dsl_dataset_hold(snapname, ra->dstg, &ds);
2509         strfree(snapname);
2510         if (err != 0)
2511                 return (err == ENOENT ? 0 : err);
2512
2513         dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2514             dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2515
2516         return (0);
2517 }
2518
2519 static int
2520 dsl_recursive_rename(char *oldname, const char *newname)
2521 {
2522         int err;
2523         struct renamesnaparg *ra;
2524         dsl_sync_task_t *dst;
2525         spa_t *spa;
2526         char *cp, *fsname = spa_strdup(oldname);
2527         int len = strlen(oldname) + 1;
2528
2529         /* truncate the snapshot name to get the fsname */
2530         cp = strchr(fsname, '@');
2531         *cp = '\0';
2532
2533         err = spa_open(fsname, &spa, FTAG);
2534         if (err) {
2535                 kmem_free(fsname, len);
2536                 return (err);
2537         }
2538         ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2539         ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2540
2541         ra->oldsnap = strchr(oldname, '@') + 1;
2542         ra->newsnap = strchr(newname, '@') + 1;
2543         *ra->failed = '\0';
2544
2545         err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2546             DS_FIND_CHILDREN);
2547         kmem_free(fsname, len);
2548
2549         if (err == 0) {
2550                 err = dsl_sync_task_group_wait(ra->dstg);
2551         }
2552
2553         for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2554             dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2555                 dsl_dataset_t *ds = dst->dst_arg1;
2556                 if (dst->dst_err) {
2557                         dsl_dir_name(ds->ds_dir, ra->failed);
2558                         (void) strlcat(ra->failed, "@", sizeof (ra->failed));
2559                         (void) strlcat(ra->failed, ra->newsnap,
2560                             sizeof (ra->failed));
2561                 }
2562                 dsl_dataset_rele(ds, ra->dstg);
2563         }
2564
2565         if (err)
2566                 (void) strlcpy(oldname, ra->failed, sizeof (ra->failed));
2567
2568         dsl_sync_task_group_destroy(ra->dstg);
2569         kmem_free(ra, sizeof (struct renamesnaparg));
2570         spa_close(spa, FTAG);
2571         return (err);
2572 }
2573
2574 static int
2575 dsl_valid_rename(const char *oldname, void *arg)
2576 {
2577         int delta = *(int *)arg;
2578
2579         if (strlen(oldname) + delta >= MAXNAMELEN)
2580                 return (ENAMETOOLONG);
2581
2582         return (0);
2583 }
2584
2585 #pragma weak dmu_objset_rename = dsl_dataset_rename
2586 int
2587 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2588 {
2589         dsl_dir_t *dd;
2590         dsl_dataset_t *ds;
2591         const char *tail;
2592         int err;
2593
2594         err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2595         if (err)
2596                 return (err);
2597
2598         if (tail == NULL) {
2599                 int delta = strlen(newname) - strlen(oldname);
2600
2601                 /* if we're growing, validate child name lengths */
2602                 if (delta > 0)
2603                         err = dmu_objset_find(oldname, dsl_valid_rename,
2604                             &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2605
2606                 if (err == 0)
2607                         err = dsl_dir_rename(dd, newname);
2608                 dsl_dir_close(dd, FTAG);
2609                 return (err);
2610         }
2611
2612         if (tail[0] != '@') {
2613                 /* the name ended in a nonexistent component */
2614                 dsl_dir_close(dd, FTAG);
2615                 return (ENOENT);
2616         }
2617
2618         dsl_dir_close(dd, FTAG);
2619
2620         /* new name must be snapshot in same filesystem */
2621         tail = strchr(newname, '@');
2622         if (tail == NULL)
2623                 return (EINVAL);
2624         tail++;
2625         if (strncmp(oldname, newname, tail - newname) != 0)
2626                 return (EXDEV);
2627
2628         if (recursive) {
2629                 err = dsl_recursive_rename(oldname, newname);
2630         } else {
2631                 err = dsl_dataset_hold(oldname, FTAG, &ds);
2632                 if (err)
2633                         return (err);
2634
2635                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2636                     dsl_dataset_snapshot_rename_check,
2637                     dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2638
2639                 dsl_dataset_rele(ds, FTAG);
2640         }
2641
2642         return (err);
2643 }
2644
2645 struct promotenode {
2646         list_node_t link;
2647         dsl_dataset_t *ds;
2648 };
2649
2650 struct promotearg {
2651         list_t shared_snaps, origin_snaps, clone_snaps;
2652         dsl_dataset_t *origin_origin;
2653         uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2654         char *err_ds;
2655 };
2656
2657 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2658
2659 static int
2660 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2661 {
2662         dsl_dataset_t *hds = arg1;
2663         struct promotearg *pa = arg2;
2664         struct promotenode *snap = list_head(&pa->shared_snaps);
2665         dsl_dataset_t *origin_ds = snap->ds;
2666         int err;
2667         uint64_t unused;
2668
2669         /* Check that it is a real clone */
2670         if (!dsl_dir_is_clone(hds->ds_dir))
2671                 return (EINVAL);
2672
2673         /* Since this is so expensive, don't do the preliminary check */
2674         if (!dmu_tx_is_syncing(tx))
2675                 return (0);
2676
2677         if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2678                 return (EXDEV);
2679
2680         /* compute origin's new unique space */
2681         snap = list_tail(&pa->clone_snaps);
2682         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2683         dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2684             origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
2685             &pa->unique, &unused, &unused);
2686
2687         /*
2688          * Walk the snapshots that we are moving
2689          *
2690          * Compute space to transfer.  Consider the incremental changes
2691          * to used for each snapshot:
2692          * (my used) = (prev's used) + (blocks born) - (blocks killed)
2693          * So each snapshot gave birth to:
2694          * (blocks born) = (my used) - (prev's used) + (blocks killed)
2695          * So a sequence would look like:
2696          * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2697          * Which simplifies to:
2698          * uN + kN + kN-1 + ... + k1 + k0
2699          * Note however, if we stop before we reach the ORIGIN we get:
2700          * uN + kN + kN-1 + ... + kM - uM-1
2701          */
2702         pa->used = origin_ds->ds_phys->ds_referenced_bytes;
2703         pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2704         pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2705         for (snap = list_head(&pa->shared_snaps); snap;
2706             snap = list_next(&pa->shared_snaps, snap)) {
2707                 uint64_t val, dlused, dlcomp, dluncomp;
2708                 dsl_dataset_t *ds = snap->ds;
2709
2710                 /* Check that the snapshot name does not conflict */
2711                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2712                 err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2713                 if (err == 0) {
2714                         err = EEXIST;
2715                         goto out;
2716                 }
2717                 if (err != ENOENT)
2718                         goto out;
2719
2720                 /* The very first snapshot does not have a deadlist */
2721                 if (ds->ds_phys->ds_prev_snap_obj == 0)
2722                         continue;
2723
2724                 dsl_deadlist_space(&ds->ds_deadlist,
2725                     &dlused, &dlcomp, &dluncomp);
2726                 pa->used += dlused;
2727                 pa->comp += dlcomp;
2728                 pa->uncomp += dluncomp;
2729         }
2730
2731         /*
2732          * If we are a clone of a clone then we never reached ORIGIN,
2733          * so we need to subtract out the clone origin's used space.
2734          */
2735         if (pa->origin_origin) {
2736                 pa->used -= pa->origin_origin->ds_phys->ds_referenced_bytes;
2737                 pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2738                 pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2739         }
2740
2741         /* Check that there is enough space here */
2742         err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2743             pa->used);
2744         if (err)
2745                 return (err);
2746
2747         /*
2748          * Compute the amounts of space that will be used by snapshots
2749          * after the promotion (for both origin and clone).  For each,
2750          * it is the amount of space that will be on all of their
2751          * deadlists (that was not born before their new origin).
2752          */
2753         if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2754                 uint64_t space;
2755
2756                 /*
2757                  * Note, typically this will not be a clone of a clone,
2758                  * so dd_origin_txg will be < TXG_INITIAL, so
2759                  * these snaplist_space() -> dsl_deadlist_space_range()
2760                  * calls will be fast because they do not have to
2761                  * iterate over all bps.
2762                  */
2763                 snap = list_head(&pa->origin_snaps);
2764                 err = snaplist_space(&pa->shared_snaps,
2765                     snap->ds->ds_dir->dd_origin_txg, &pa->cloneusedsnap);
2766                 if (err)
2767                         return (err);
2768
2769                 err = snaplist_space(&pa->clone_snaps,
2770                     snap->ds->ds_dir->dd_origin_txg, &space);
2771                 if (err)
2772                         return (err);
2773                 pa->cloneusedsnap += space;
2774         }
2775         if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2776                 err = snaplist_space(&pa->origin_snaps,
2777                     origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2778                 if (err)
2779                         return (err);
2780         }
2781
2782         return (0);
2783 out:
2784         pa->err_ds =  snap->ds->ds_snapname;
2785         return (err);
2786 }
2787
2788 static void
2789 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2790 {
2791         dsl_dataset_t *hds = arg1;
2792         struct promotearg *pa = arg2;
2793         struct promotenode *snap = list_head(&pa->shared_snaps);
2794         dsl_dataset_t *origin_ds = snap->ds;
2795         dsl_dataset_t *origin_head;
2796         dsl_dir_t *dd = hds->ds_dir;
2797         dsl_pool_t *dp = hds->ds_dir->dd_pool;
2798         dsl_dir_t *odd = NULL;
2799         uint64_t oldnext_obj;
2800         int64_t delta;
2801
2802         ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2803
2804         snap = list_head(&pa->origin_snaps);
2805         origin_head = snap->ds;
2806
2807         /*
2808          * We need to explicitly open odd, since origin_ds's dd will be
2809          * changing.
2810          */
2811         VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2812             NULL, FTAG, &odd));
2813
2814         /* change origin's next snap */
2815         dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2816         oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2817         snap = list_tail(&pa->clone_snaps);
2818         ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2819         origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2820
2821         /* change the origin's next clone */
2822         if (origin_ds->ds_phys->ds_next_clones_obj) {
2823                 remove_from_next_clones(origin_ds, snap->ds->ds_object, tx);
2824                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2825                     origin_ds->ds_phys->ds_next_clones_obj,
2826                     oldnext_obj, tx));
2827         }
2828
2829         /* change origin */
2830         dmu_buf_will_dirty(dd->dd_dbuf, tx);
2831         ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2832         dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2833         dd->dd_origin_txg = origin_head->ds_dir->dd_origin_txg;
2834         dmu_buf_will_dirty(odd->dd_dbuf, tx);
2835         odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2836         origin_head->ds_dir->dd_origin_txg =
2837             origin_ds->ds_phys->ds_creation_txg;
2838
2839         /* change dd_clone entries */
2840         if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2841                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2842                     odd->dd_phys->dd_clones, hds->ds_object, tx));
2843                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2844                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2845                     hds->ds_object, tx));
2846
2847                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2848                     pa->origin_origin->ds_dir->dd_phys->dd_clones,
2849                     origin_head->ds_object, tx));
2850                 if (dd->dd_phys->dd_clones == 0) {
2851                         dd->dd_phys->dd_clones = zap_create(dp->dp_meta_objset,
2852                             DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
2853                 }
2854                 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2855                     dd->dd_phys->dd_clones, origin_head->ds_object, tx));
2856
2857         }
2858
2859         /* move snapshots to this dir */
2860         for (snap = list_head(&pa->shared_snaps); snap;
2861             snap = list_next(&pa->shared_snaps, snap)) {
2862                 dsl_dataset_t *ds = snap->ds;
2863
2864                 /* unregister props as dsl_dir is changing */
2865                 if (ds->ds_objset) {
2866                         dmu_objset_evict(ds->ds_objset);
2867                         ds->ds_objset = NULL;
2868                 }
2869                 /* move snap name entry */
2870                 VERIFY(0 == dsl_dataset_get_snapname(ds));
2871                 VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2872                     ds->ds_snapname, tx));
2873                 VERIFY(0 == zap_add(dp->dp_meta_objset,
2874                     hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2875                     8, 1, &ds->ds_object, tx));
2876
2877                 /* change containing dsl_dir */
2878                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2879                 ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2880                 ds->ds_phys->ds_dir_obj = dd->dd_object;
2881                 ASSERT3P(ds->ds_dir, ==, odd);
2882                 dsl_dir_close(ds->ds_dir, ds);
2883                 VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2884                     NULL, ds, &ds->ds_dir));
2885
2886                 /* move any clone references */
2887                 if (ds->ds_phys->ds_next_clones_obj &&
2888                     spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2889                         zap_cursor_t zc;
2890                         zap_attribute_t za;
2891
2892                         for (zap_cursor_init(&zc, dp->dp_meta_objset,
2893                             ds->ds_phys->ds_next_clones_obj);
2894                             zap_cursor_retrieve(&zc, &za) == 0;
2895                             zap_cursor_advance(&zc)) {
2896                                 dsl_dataset_t *cnds;
2897                                 uint64_t o;
2898
2899                                 if (za.za_first_integer == oldnext_obj) {
2900                                         /*
2901                                          * We've already moved the
2902                                          * origin's reference.
2903                                          */
2904                                         continue;
2905                                 }
2906
2907                                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
2908                                     za.za_first_integer, FTAG, &cnds));
2909                                 o = cnds->ds_dir->dd_phys->dd_head_dataset_obj;
2910
2911                                 VERIFY3U(zap_remove_int(dp->dp_meta_objset,
2912                                     odd->dd_phys->dd_clones, o, tx), ==, 0);
2913                                 VERIFY3U(zap_add_int(dp->dp_meta_objset,
2914                                     dd->dd_phys->dd_clones, o, tx), ==, 0);
2915                                 dsl_dataset_rele(cnds, FTAG);
2916                         }
2917                         zap_cursor_fini(&zc);
2918                 }
2919
2920                 ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2921         }
2922
2923         /*
2924          * Change space accounting.
2925          * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2926          * both be valid, or both be 0 (resulting in delta == 0).  This
2927          * is true for each of {clone,origin} independently.
2928          */
2929
2930         delta = pa->cloneusedsnap -
2931             dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2932         ASSERT3S(delta, >=, 0);
2933         ASSERT3U(pa->used, >=, delta);
2934         dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2935         dsl_dir_diduse_space(dd, DD_USED_HEAD,
2936             pa->used - delta, pa->comp, pa->uncomp, tx);
2937
2938         delta = pa->originusedsnap -
2939             odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2940         ASSERT3S(delta, <=, 0);
2941         ASSERT3U(pa->used, >=, -delta);
2942         dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2943         dsl_dir_diduse_space(odd, DD_USED_HEAD,
2944             -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2945
2946         origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2947
2948         /* log history record */
2949         spa_history_log_internal(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2950             "dataset = %llu", hds->ds_object);
2951
2952         dsl_dir_close(odd, FTAG);
2953 }
2954
2955 static char *snaplist_tag = "snaplist";
2956 /*
2957  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2958  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2959  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2960  * snapshots back to this dataset's origin.
2961  */
2962 static int
2963 snaplist_make(dsl_pool_t *dp, boolean_t own,
2964     uint64_t first_obj, uint64_t last_obj, list_t *l)
2965 {
2966         uint64_t obj = last_obj;
2967
2968         ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2969
2970         list_create(l, sizeof (struct promotenode),
2971             offsetof(struct promotenode, link));
2972
2973         while (obj != first_obj) {
2974                 dsl_dataset_t *ds;
2975                 struct promotenode *snap;
2976                 int err;
2977
2978                 if (own) {
2979                         err = dsl_dataset_own_obj(dp, obj,
2980                             0, snaplist_tag, &ds);
2981                         if (err == 0)
2982                                 dsl_dataset_make_exclusive(ds, snaplist_tag);
2983                 } else {
2984                         err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2985                 }
2986                 if (err == ENOENT) {
2987                         /* lost race with snapshot destroy */
2988                         struct promotenode *last = list_tail(l);
2989                         ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2990                         obj = last->ds->ds_phys->ds_prev_snap_obj;
2991                         continue;
2992                 } else if (err) {
2993                         return (err);
2994                 }
2995
2996                 if (first_obj == 0)
2997                         first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2998
2999                 snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
3000                 snap->ds = ds;
3001                 list_insert_tail(l, snap);
3002                 obj = ds->ds_phys->ds_prev_snap_obj;
3003         }
3004
3005         return (0);
3006 }
3007
3008 static int
3009 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
3010 {
3011         struct promotenode *snap;
3012
3013         *spacep = 0;
3014         for (snap = list_head(l); snap; snap = list_next(l, snap)) {
3015                 uint64_t used, comp, uncomp;
3016                 dsl_deadlist_space_range(&snap->ds->ds_deadlist,
3017                     mintxg, UINT64_MAX, &used, &comp, &uncomp);
3018                 *spacep += used;
3019         }
3020         return (0);
3021 }
3022
3023 static void
3024 snaplist_destroy(list_t *l, boolean_t own)
3025 {
3026         struct promotenode *snap;
3027
3028         if (!l || !list_link_active(&l->list_head))
3029                 return;
3030
3031         while ((snap = list_tail(l)) != NULL) {
3032                 list_remove(l, snap);
3033                 if (own)
3034                         dsl_dataset_disown(snap->ds, snaplist_tag);
3035                 else
3036                         dsl_dataset_rele(snap->ds, snaplist_tag);
3037                 kmem_free(snap, sizeof (struct promotenode));
3038         }
3039         list_destroy(l);
3040 }
3041
3042 /*
3043  * Promote a clone.  Nomenclature note:
3044  * "clone" or "cds": the original clone which is being promoted
3045  * "origin" or "ods": the snapshot which is originally clone's origin
3046  * "origin head" or "ohds": the dataset which is the head
3047  * (filesystem/volume) for the origin
3048  * "origin origin": the origin of the origin's filesystem (typically
3049  * NULL, indicating that the clone is not a clone of a clone).
3050  */
3051 int
3052 dsl_dataset_promote(const char *name, char *conflsnap)
3053 {
3054         dsl_dataset_t *ds;
3055         dsl_dir_t *dd;
3056         dsl_pool_t *dp;
3057         dmu_object_info_t doi;
3058         struct promotearg pa;
3059         struct promotenode *snap;
3060         int err;
3061
3062         bzero(&pa, sizeof(struct promotearg));
3063         err = dsl_dataset_hold(name, FTAG, &ds);
3064         if (err)
3065                 return (err);
3066         dd = ds->ds_dir;
3067         dp = dd->dd_pool;
3068
3069         err = dmu_object_info(dp->dp_meta_objset,
3070             ds->ds_phys->ds_snapnames_zapobj, &doi);
3071         if (err) {
3072                 dsl_dataset_rele(ds, FTAG);
3073                 return (err);
3074         }
3075
3076         if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
3077                 dsl_dataset_rele(ds, FTAG);
3078                 return (EINVAL);
3079         }
3080
3081         /*
3082          * We are going to inherit all the snapshots taken before our
3083          * origin (i.e., our new origin will be our parent's origin).
3084          * Take ownership of them so that we can rename them into our
3085          * namespace.
3086          */
3087         rw_enter(&dp->dp_config_rwlock, RW_READER);
3088
3089         err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
3090             &pa.shared_snaps);
3091         if (err != 0)
3092                 goto out;
3093
3094         err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
3095         if (err != 0)
3096                 goto out;
3097
3098         snap = list_head(&pa.shared_snaps);
3099         ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
3100         err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
3101             snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
3102         if (err != 0)
3103                 goto out;
3104
3105         if (snap->ds->ds_dir->dd_phys->dd_origin_obj != 0) {
3106                 err = dsl_dataset_hold_obj(dp,
3107                     snap->ds->ds_dir->dd_phys->dd_origin_obj,
3108                     FTAG, &pa.origin_origin);
3109                 if (err != 0)
3110                         goto out;
3111         }
3112
3113 out:
3114         rw_exit(&dp->dp_config_rwlock);
3115
3116         /*
3117          * Add in 128x the snapnames zapobj size, since we will be moving
3118          * a bunch of snapnames to the promoted ds, and dirtying their
3119          * bonus buffers.
3120          */
3121         if (err == 0) {
3122                 err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
3123                     dsl_dataset_promote_sync, ds, &pa,
3124                     2 + 2 * doi.doi_physical_blocks_512);
3125                 if (err && pa.err_ds && conflsnap)
3126                         (void) strncpy(conflsnap, pa.err_ds, MAXNAMELEN);
3127         }
3128
3129         snaplist_destroy(&pa.shared_snaps, B_TRUE);
3130         snaplist_destroy(&pa.clone_snaps, B_FALSE);
3131         snaplist_destroy(&pa.origin_snaps, B_FALSE);
3132         if (pa.origin_origin)
3133                 dsl_dataset_rele(pa.origin_origin, FTAG);
3134         dsl_dataset_rele(ds, FTAG);
3135         return (err);
3136 }
3137
3138 struct cloneswaparg {
3139         dsl_dataset_t *cds; /* clone dataset */
3140         dsl_dataset_t *ohds; /* origin's head dataset */
3141         boolean_t force;
3142         int64_t unused_refres_delta; /* change in unconsumed refreservation */
3143 };
3144
3145 /* ARGSUSED */
3146 static int
3147 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
3148 {
3149         struct cloneswaparg *csa = arg1;
3150
3151         /* they should both be heads */
3152         if (dsl_dataset_is_snapshot(csa->cds) ||
3153             dsl_dataset_is_snapshot(csa->ohds))
3154                 return (EINVAL);
3155
3156         /* the branch point should be just before them */
3157         if (csa->cds->ds_prev != csa->ohds->ds_prev)
3158                 return (EINVAL);
3159
3160         /* cds should be the clone (unless they are unrelated) */
3161         if (csa->cds->ds_prev != NULL &&
3162             csa->cds->ds_prev != csa->cds->ds_dir->dd_pool->dp_origin_snap &&
3163             csa->ohds->ds_object !=
3164             csa->cds->ds_prev->ds_phys->ds_next_snap_obj)
3165                 return (EINVAL);
3166
3167         /* the clone should be a child of the origin */
3168         if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
3169                 return (EINVAL);
3170
3171         /* ohds shouldn't be modified unless 'force' */
3172         if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
3173                 return (ETXTBSY);
3174
3175         /* adjust amount of any unconsumed refreservation */
3176         csa->unused_refres_delta =
3177             (int64_t)MIN(csa->ohds->ds_reserved,
3178             csa->ohds->ds_phys->ds_unique_bytes) -
3179             (int64_t)MIN(csa->ohds->ds_reserved,
3180             csa->cds->ds_phys->ds_unique_bytes);
3181
3182         if (csa->unused_refres_delta > 0 &&
3183             csa->unused_refres_delta >
3184             dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3185                 return (ENOSPC);
3186
3187         if (csa->ohds->ds_quota != 0 &&
3188             csa->cds->ds_phys->ds_unique_bytes > csa->ohds->ds_quota)
3189                 return (EDQUOT);
3190
3191         return (0);
3192 }
3193
3194 /* ARGSUSED */
3195 static void
3196 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3197 {
3198         struct cloneswaparg *csa = arg1;
3199         dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3200
3201         ASSERT(csa->cds->ds_reserved == 0);
3202         ASSERT(csa->ohds->ds_quota == 0 ||
3203             csa->cds->ds_phys->ds_unique_bytes <= csa->ohds->ds_quota);
3204
3205         dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3206         dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3207
3208         if (csa->cds->ds_objset != NULL) {
3209                 dmu_objset_evict(csa->cds->ds_objset);
3210                 csa->cds->ds_objset = NULL;
3211         }
3212
3213         if (csa->ohds->ds_objset != NULL) {
3214                 dmu_objset_evict(csa->ohds->ds_objset);
3215                 csa->ohds->ds_objset = NULL;
3216         }
3217
3218         /*
3219          * Reset origin's unique bytes, if it exists.
3220          */
3221         if (csa->cds->ds_prev) {
3222                 dsl_dataset_t *origin = csa->cds->ds_prev;
3223                 uint64_t comp, uncomp;
3224
3225                 dmu_buf_will_dirty(origin->ds_dbuf, tx);
3226                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3227                     origin->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3228                     &origin->ds_phys->ds_unique_bytes, &comp, &uncomp);
3229         }
3230
3231         /* swap blkptrs */
3232         {
3233                 blkptr_t tmp;
3234                 tmp = csa->ohds->ds_phys->ds_bp;
3235                 csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3236                 csa->cds->ds_phys->ds_bp = tmp;
3237         }
3238
3239         /* set dd_*_bytes */
3240         {
3241                 int64_t dused, dcomp, duncomp;
3242                 uint64_t cdl_used, cdl_comp, cdl_uncomp;
3243                 uint64_t odl_used, odl_comp, odl_uncomp;
3244
3245                 ASSERT3U(csa->cds->ds_dir->dd_phys->
3246                     dd_used_breakdown[DD_USED_SNAP], ==, 0);
3247
3248                 dsl_deadlist_space(&csa->cds->ds_deadlist,
3249                     &cdl_used, &cdl_comp, &cdl_uncomp);
3250                 dsl_deadlist_space(&csa->ohds->ds_deadlist,
3251                     &odl_used, &odl_comp, &odl_uncomp);
3252
3253                 dused = csa->cds->ds_phys->ds_referenced_bytes + cdl_used -
3254                     (csa->ohds->ds_phys->ds_referenced_bytes + odl_used);
3255                 dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3256                     (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3257                 duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3258                     cdl_uncomp -
3259                     (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3260
3261                 dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3262                     dused, dcomp, duncomp, tx);
3263                 dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3264                     -dused, -dcomp, -duncomp, tx);
3265
3266                 /*
3267                  * The difference in the space used by snapshots is the
3268                  * difference in snapshot space due to the head's
3269                  * deadlist (since that's the only thing that's
3270                  * changing that affects the snapused).
3271                  */
3272                 dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3273                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3274                     &cdl_used, &cdl_comp, &cdl_uncomp);
3275                 dsl_deadlist_space_range(&csa->ohds->ds_deadlist,
3276                     csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3277                     &odl_used, &odl_comp, &odl_uncomp);
3278                 dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3279                     DD_USED_HEAD, DD_USED_SNAP, tx);
3280         }
3281
3282         /* swap ds_*_bytes */
3283         SWITCH64(csa->ohds->ds_phys->ds_referenced_bytes,
3284             csa->cds->ds_phys->ds_referenced_bytes);
3285         SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3286             csa->cds->ds_phys->ds_compressed_bytes);
3287         SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3288             csa->cds->ds_phys->ds_uncompressed_bytes);
3289         SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3290             csa->cds->ds_phys->ds_unique_bytes);
3291
3292         /* apply any parent delta for change in unconsumed refreservation */
3293         dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3294             csa->unused_refres_delta, 0, 0, tx);
3295
3296         /*
3297          * Swap deadlists.
3298          */
3299         dsl_deadlist_close(&csa->cds->ds_deadlist);
3300         dsl_deadlist_close(&csa->ohds->ds_deadlist);
3301         SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3302             csa->cds->ds_phys->ds_deadlist_obj);
3303         dsl_deadlist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3304             csa->cds->ds_phys->ds_deadlist_obj);
3305         dsl_deadlist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3306             csa->ohds->ds_phys->ds_deadlist_obj);
3307
3308         dsl_scan_ds_clone_swapped(csa->ohds, csa->cds, tx);
3309 }
3310
3311 /*
3312  * Swap 'clone' with its origin head datasets.  Used at the end of "zfs
3313  * recv" into an existing fs to swizzle the file system to the new
3314  * version, and by "zfs rollback".  Can also be used to swap two
3315  * independent head datasets if neither has any snapshots.
3316  */
3317 int
3318 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3319     boolean_t force)
3320 {
3321         struct cloneswaparg csa;
3322         int error;
3323
3324         ASSERT(clone->ds_owner);
3325         ASSERT(origin_head->ds_owner);
3326 retry:
3327         /*
3328          * Need exclusive access for the swap. If we're swapping these
3329          * datasets back after an error, we already hold the locks.
3330          */
3331         if (!RW_WRITE_HELD(&clone->ds_rwlock))
3332                 rw_enter(&clone->ds_rwlock, RW_WRITER);
3333         if (!RW_WRITE_HELD(&origin_head->ds_rwlock) &&
3334             !rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3335                 rw_exit(&clone->ds_rwlock);
3336                 rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3337                 if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3338                         rw_exit(&origin_head->ds_rwlock);
3339                         goto retry;
3340                 }
3341         }
3342         csa.cds = clone;
3343         csa.ohds = origin_head;
3344         csa.force = force;
3345         error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3346             dsl_dataset_clone_swap_check,
3347             dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3348         return (error);
3349 }
3350
3351 /*
3352  * Given a pool name and a dataset object number in that pool,
3353  * return the name of that dataset.
3354  */
3355 int
3356 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3357 {
3358         spa_t *spa;
3359         dsl_pool_t *dp;
3360         dsl_dataset_t *ds;
3361         int error;
3362
3363         if ((error = spa_open(pname, &spa, FTAG)) != 0)
3364                 return (error);
3365         dp = spa_get_dsl(spa);
3366         rw_enter(&dp->dp_config_rwlock, RW_READER);
3367         if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3368                 dsl_dataset_name(ds, buf);
3369                 dsl_dataset_rele(ds, FTAG);
3370         }
3371         rw_exit(&dp->dp_config_rwlock);
3372         spa_close(spa, FTAG);
3373
3374         return (error);
3375 }
3376
3377 int
3378 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3379     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3380 {
3381         int error = 0;
3382
3383         ASSERT3S(asize, >, 0);
3384
3385         /*
3386          * *ref_rsrv is the portion of asize that will come from any
3387          * unconsumed refreservation space.
3388          */
3389         *ref_rsrv = 0;
3390
3391         mutex_enter(&ds->ds_lock);
3392         /*
3393          * Make a space adjustment for reserved bytes.
3394          */
3395         if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3396                 ASSERT3U(*used, >=,
3397                     ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3398                 *used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3399                 *ref_rsrv =
3400                     asize - MIN(asize, parent_delta(ds, asize + inflight));
3401         }
3402
3403         if (!check_quota || ds->ds_quota == 0) {
3404                 mutex_exit(&ds->ds_lock);
3405                 return (0);
3406         }
3407         /*
3408          * If they are requesting more space, and our current estimate
3409          * is over quota, they get to try again unless the actual
3410          * on-disk is over quota and there are no pending changes (which
3411          * may free up space for us).
3412          */
3413         if (ds->ds_phys->ds_referenced_bytes + inflight >= ds->ds_quota) {
3414                 if (inflight > 0 ||
3415                     ds->ds_phys->ds_referenced_bytes < ds->ds_quota)
3416                         error = ERESTART;
3417                 else
3418                         error = EDQUOT;
3419
3420                 DMU_TX_STAT_BUMP(dmu_tx_quota);
3421         }
3422         mutex_exit(&ds->ds_lock);
3423
3424         return (error);
3425 }
3426
3427 /* ARGSUSED */
3428 static int
3429 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3430 {
3431         dsl_dataset_t *ds = arg1;
3432         dsl_prop_setarg_t *psa = arg2;
3433         int err;
3434
3435         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3436                 return (ENOTSUP);
3437
3438         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3439                 return (err);
3440
3441         if (psa->psa_effective_value == 0)
3442                 return (0);
3443
3444         if (psa->psa_effective_value < ds->ds_phys->ds_referenced_bytes ||
3445             psa->psa_effective_value < ds->ds_reserved)
3446                 return (ENOSPC);
3447
3448         return (0);
3449 }
3450
3451 extern void dsl_prop_set_sync(void *, void *, dmu_tx_t *);
3452
3453 void
3454 dsl_dataset_set_quota_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3455 {
3456         dsl_dataset_t *ds = arg1;
3457         dsl_prop_setarg_t *psa = arg2;
3458         uint64_t effective_value = psa->psa_effective_value;
3459
3460         dsl_prop_set_sync(ds, psa, tx);
3461         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3462
3463         if (ds->ds_quota != effective_value) {
3464                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3465                 ds->ds_quota = effective_value;
3466         }
3467 }
3468
3469 int
3470 dsl_dataset_set_quota(const char *dsname, zprop_source_t source, uint64_t quota)
3471 {
3472         dsl_dataset_t *ds;
3473         dsl_prop_setarg_t psa;
3474         int err;
3475
3476         dsl_prop_setarg_init_uint64(&psa, "refquota", source, &quota);
3477
3478         err = dsl_dataset_hold(dsname, FTAG, &ds);
3479         if (err)
3480                 return (err);
3481
3482         /*
3483          * If someone removes a file, then tries to set the quota, we
3484          * want to make sure the file freeing takes effect.
3485          */
3486         txg_wait_open(ds->ds_dir->dd_pool, 0);
3487
3488         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3489             dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3490             ds, &psa, 0);
3491
3492         dsl_dataset_rele(ds, FTAG);
3493         return (err);
3494 }
3495
3496 static int
3497 dsl_dataset_set_reservation_check(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;
3502         uint64_t unique;
3503         int err;
3504
3505         if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3506             SPA_VERSION_REFRESERVATION)
3507                 return (ENOTSUP);
3508
3509         if (dsl_dataset_is_snapshot(ds))
3510                 return (EINVAL);
3511
3512         if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3513                 return (err);
3514
3515         effective_value = psa->psa_effective_value;
3516
3517         /*
3518          * If we are doing the preliminary check in open context, the
3519          * space estimates may be inaccurate.
3520          */
3521         if (!dmu_tx_is_syncing(tx))
3522                 return (0);
3523
3524         mutex_enter(&ds->ds_lock);
3525         if (!DS_UNIQUE_IS_ACCURATE(ds))
3526                 dsl_dataset_recalc_head_uniq(ds);
3527         unique = ds->ds_phys->ds_unique_bytes;
3528         mutex_exit(&ds->ds_lock);
3529
3530         if (MAX(unique, effective_value) > MAX(unique, ds->ds_reserved)) {
3531                 uint64_t delta = MAX(unique, effective_value) -
3532                     MAX(unique, ds->ds_reserved);
3533
3534                 if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3535                         return (ENOSPC);
3536                 if (ds->ds_quota > 0 &&
3537                     effective_value > ds->ds_quota)
3538                         return (ENOSPC);
3539         }
3540
3541         return (0);
3542 }
3543
3544 static void
3545 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3546 {
3547         dsl_dataset_t *ds = arg1;
3548         dsl_prop_setarg_t *psa = arg2;
3549         uint64_t effective_value = psa->psa_effective_value;
3550         uint64_t unique;
3551         int64_t delta;
3552
3553         dsl_prop_set_sync(ds, psa, tx);
3554         DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3555
3556         dmu_buf_will_dirty(ds->ds_dbuf, tx);
3557
3558         mutex_enter(&ds->ds_dir->dd_lock);
3559         mutex_enter(&ds->ds_lock);
3560         ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
3561         unique = ds->ds_phys->ds_unique_bytes;
3562         delta = MAX(0, (int64_t)(effective_value - unique)) -
3563             MAX(0, (int64_t)(ds->ds_reserved - unique));
3564         ds->ds_reserved = effective_value;
3565         mutex_exit(&ds->ds_lock);
3566
3567         dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3568         mutex_exit(&ds->ds_dir->dd_lock);
3569 }
3570
3571 int
3572 dsl_dataset_set_reservation(const char *dsname, zprop_source_t source,
3573     uint64_t reservation)
3574 {
3575         dsl_dataset_t *ds;
3576         dsl_prop_setarg_t psa;
3577         int err;
3578
3579         dsl_prop_setarg_init_uint64(&psa, "refreservation", source,
3580             &reservation);
3581
3582         err = dsl_dataset_hold(dsname, FTAG, &ds);
3583         if (err)
3584                 return (err);
3585
3586         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3587             dsl_dataset_set_reservation_check,
3588             dsl_dataset_set_reservation_sync, ds, &psa, 0);
3589
3590         dsl_dataset_rele(ds, FTAG);
3591         return (err);
3592 }
3593
3594 typedef struct zfs_hold_cleanup_arg {
3595         dsl_pool_t *dp;
3596         uint64_t dsobj;
3597         char htag[MAXNAMELEN];
3598 } zfs_hold_cleanup_arg_t;
3599
3600 static void
3601 dsl_dataset_user_release_onexit(void *arg)
3602 {
3603         zfs_hold_cleanup_arg_t *ca = arg;
3604
3605         (void) dsl_dataset_user_release_tmp(ca->dp, ca->dsobj, ca->htag,
3606             B_TRUE);
3607         kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t));
3608 }
3609
3610 void
3611 dsl_register_onexit_hold_cleanup(dsl_dataset_t *ds, const char *htag,
3612     minor_t minor)
3613 {
3614         zfs_hold_cleanup_arg_t *ca;
3615
3616         ca = kmem_alloc(sizeof (zfs_hold_cleanup_arg_t), KM_SLEEP);
3617         ca->dp = ds->ds_dir->dd_pool;
3618         ca->dsobj = ds->ds_object;
3619         (void) strlcpy(ca->htag, htag, sizeof (ca->htag));
3620         VERIFY3U(0, ==, zfs_onexit_add_cb(minor,
3621             dsl_dataset_user_release_onexit, ca, NULL));
3622 }
3623
3624 /*
3625  * If you add new checks here, you may need to add
3626  * additional checks to the "temporary" case in
3627  * snapshot_check() in dmu_objset.c.
3628  */
3629 static int
3630 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3631 {
3632         dsl_dataset_t *ds = arg1;
3633         struct dsl_ds_holdarg *ha = arg2;
3634         char *htag = ha->htag;
3635         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3636         int error = 0;
3637
3638         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3639                 return (ENOTSUP);
3640
3641         if (!dsl_dataset_is_snapshot(ds))
3642                 return (EINVAL);
3643
3644         /* tags must be unique */
3645         mutex_enter(&ds->ds_lock);
3646         if (ds->ds_phys->ds_userrefs_obj) {
3647                 error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3648                     8, 1, tx);
3649                 if (error == 0)
3650                         error = EEXIST;
3651                 else if (error == ENOENT)
3652                         error = 0;
3653         }
3654         mutex_exit(&ds->ds_lock);
3655
3656         if (error == 0 && ha->temphold &&
3657             strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
3658                 error = E2BIG;
3659
3660         return (error);
3661 }
3662
3663 void
3664 dsl_dataset_user_hold_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3665 {
3666         dsl_dataset_t *ds = arg1;
3667         struct dsl_ds_holdarg *ha = arg2;
3668         char *htag = ha->htag;
3669         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3670         objset_t *mos = dp->dp_meta_objset;
3671         uint64_t now = gethrestime_sec();
3672         uint64_t zapobj;
3673
3674         mutex_enter(&ds->ds_lock);
3675         if (ds->ds_phys->ds_userrefs_obj == 0) {
3676                 /*
3677                  * This is the first user hold for this dataset.  Create
3678                  * the userrefs zap object.
3679                  */
3680                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3681                 zapobj = ds->ds_phys->ds_userrefs_obj =
3682                     zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3683         } else {
3684                 zapobj = ds->ds_phys->ds_userrefs_obj;
3685         }
3686         ds->ds_userrefs++;
3687         mutex_exit(&ds->ds_lock);
3688
3689         VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3690
3691         if (ha->temphold) {
3692                 VERIFY(0 == dsl_pool_user_hold(dp, ds->ds_object,
3693                     htag, &now, tx));
3694         }
3695
3696         spa_history_log_internal(LOG_DS_USER_HOLD,
3697             dp->dp_spa, tx, "<%s> temp = %d dataset = %llu", htag,
3698             (int)ha->temphold, ds->ds_object);
3699 }
3700
3701 static int
3702 dsl_dataset_user_hold_one(const char *dsname, void *arg)
3703 {
3704         struct dsl_ds_holdarg *ha = arg;
3705         dsl_dataset_t *ds;
3706         int error;
3707         char *name;
3708
3709         /* alloc a buffer to hold dsname@snapname plus terminating NULL */
3710         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3711         error = dsl_dataset_hold(name, ha->dstg, &ds);
3712         strfree(name);
3713         if (error == 0) {
3714                 ha->gotone = B_TRUE;
3715                 dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3716                     dsl_dataset_user_hold_sync, ds, ha, 0);
3717         } else if (error == ENOENT && ha->recursive) {
3718                 error = 0;
3719         } else {
3720                 (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3721         }
3722         return (error);
3723 }
3724
3725 int
3726 dsl_dataset_user_hold_for_send(dsl_dataset_t *ds, char *htag,
3727     boolean_t temphold)
3728 {
3729         struct dsl_ds_holdarg *ha;
3730         int error;
3731
3732         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3733         ha->htag = htag;
3734         ha->temphold = temphold;
3735         error = dsl_sync_task_do(ds->ds_dir->dd_pool,
3736             dsl_dataset_user_hold_check, dsl_dataset_user_hold_sync,
3737             ds, ha, 0);
3738         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3739
3740         return (error);
3741 }
3742
3743 int
3744 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3745     boolean_t recursive, boolean_t temphold, int cleanup_fd)
3746 {
3747         struct dsl_ds_holdarg *ha;
3748         dsl_sync_task_t *dst;
3749         spa_t *spa;
3750         int error;
3751         minor_t minor = 0;
3752
3753         if (cleanup_fd != -1) {
3754                 /* Currently we only support cleanup-on-exit of tempholds. */
3755                 if (!temphold)
3756                         return (EINVAL);
3757                 error = zfs_onexit_fd_hold(cleanup_fd, &minor);
3758                 if (error)
3759                         return (error);
3760         }
3761
3762         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3763
3764         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3765
3766         error = spa_open(dsname, &spa, FTAG);
3767         if (error) {
3768                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3769                 if (cleanup_fd != -1)
3770                         zfs_onexit_fd_rele(cleanup_fd);
3771                 return (error);
3772         }
3773
3774         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3775         ha->htag = htag;
3776         ha->snapname = snapname;
3777         ha->recursive = recursive;
3778         ha->temphold = temphold;
3779
3780         if (recursive) {
3781                 error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3782                     ha, DS_FIND_CHILDREN);
3783         } else {
3784                 error = dsl_dataset_user_hold_one(dsname, ha);
3785         }
3786         if (error == 0)
3787                 error = dsl_sync_task_group_wait(ha->dstg);
3788
3789         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3790             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3791                 dsl_dataset_t *ds = dst->dst_arg1;
3792
3793                 if (dst->dst_err) {
3794                         dsl_dataset_name(ds, ha->failed);
3795                         *strchr(ha->failed, '@') = '\0';
3796                 } else if (error == 0 && minor != 0 && temphold) {
3797                         /*
3798                          * If this hold is to be released upon process exit,
3799                          * register that action now.
3800                          */
3801                         dsl_register_onexit_hold_cleanup(ds, htag, minor);
3802                 }
3803                 dsl_dataset_rele(ds, ha->dstg);
3804         }
3805
3806         if (error == 0 && recursive && !ha->gotone)
3807                 error = ENOENT;
3808
3809         if (error)
3810                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3811
3812         dsl_sync_task_group_destroy(ha->dstg);
3813
3814         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3815         spa_close(spa, FTAG);
3816         if (cleanup_fd != -1)
3817                 zfs_onexit_fd_rele(cleanup_fd);
3818         return (error);
3819 }
3820
3821 struct dsl_ds_releasearg {
3822         dsl_dataset_t *ds;
3823         const char *htag;
3824         boolean_t own;          /* do we own or just hold ds? */
3825 };
3826
3827 static int
3828 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3829     boolean_t *might_destroy)
3830 {
3831         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3832         uint64_t zapobj;
3833         uint64_t tmp;
3834         int error;
3835
3836         *might_destroy = B_FALSE;
3837
3838         mutex_enter(&ds->ds_lock);
3839         zapobj = ds->ds_phys->ds_userrefs_obj;
3840         if (zapobj == 0) {
3841                 /* The tag can't possibly exist */
3842                 mutex_exit(&ds->ds_lock);
3843                 return (ESRCH);
3844         }
3845
3846         /* Make sure the tag exists */
3847         error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3848         if (error) {
3849                 mutex_exit(&ds->ds_lock);
3850                 if (error == ENOENT)
3851                         error = ESRCH;
3852                 return (error);
3853         }
3854
3855         if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3856             DS_IS_DEFER_DESTROY(ds))
3857                 *might_destroy = B_TRUE;
3858
3859         mutex_exit(&ds->ds_lock);
3860         return (0);
3861 }
3862
3863 static int
3864 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3865 {
3866         struct dsl_ds_releasearg *ra = arg1;
3867         dsl_dataset_t *ds = ra->ds;
3868         boolean_t might_destroy;
3869         int error;
3870
3871         if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3872                 return (ENOTSUP);
3873
3874         error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3875         if (error)
3876                 return (error);
3877
3878         if (might_destroy) {
3879                 struct dsl_ds_destroyarg dsda = {0};
3880
3881                 if (dmu_tx_is_syncing(tx)) {
3882                         /*
3883                          * If we're not prepared to remove the snapshot,
3884                          * we can't allow the release to happen right now.
3885                          */
3886                         if (!ra->own)
3887                                 return (EBUSY);
3888                 }
3889                 dsda.ds = ds;
3890                 dsda.releasing = B_TRUE;
3891                 return (dsl_dataset_destroy_check(&dsda, tag, tx));
3892         }
3893
3894         return (0);
3895 }
3896
3897 static void
3898 dsl_dataset_user_release_sync(void *arg1, void *tag, dmu_tx_t *tx)
3899 {
3900         struct dsl_ds_releasearg *ra = arg1;
3901         dsl_dataset_t *ds = ra->ds;
3902         dsl_pool_t *dp = ds->ds_dir->dd_pool;
3903         objset_t *mos = dp->dp_meta_objset;
3904         uint64_t zapobj;
3905         uint64_t dsobj = ds->ds_object;
3906         uint64_t refs;
3907         int error;
3908
3909         mutex_enter(&ds->ds_lock);
3910         ds->ds_userrefs--;
3911         refs = ds->ds_userrefs;
3912         mutex_exit(&ds->ds_lock);
3913         error = dsl_pool_user_release(dp, ds->ds_object, ra->htag, tx);
3914         VERIFY(error == 0 || error == ENOENT);
3915         zapobj = ds->ds_phys->ds_userrefs_obj;
3916         VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3917
3918         spa_history_log_internal(LOG_DS_USER_RELEASE,
3919             dp->dp_spa, tx, "<%s> %lld dataset = %llu",
3920             ra->htag, (longlong_t)refs, dsobj);
3921
3922         if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3923             DS_IS_DEFER_DESTROY(ds)) {
3924                 struct dsl_ds_destroyarg dsda = {0};
3925
3926                 ASSERT(ra->own);
3927                 dsda.ds = ds;
3928                 dsda.releasing = B_TRUE;
3929                 /* We already did the destroy_check */
3930                 dsl_dataset_destroy_sync(&dsda, tag, tx);
3931         }
3932 }
3933
3934 static int
3935 dsl_dataset_user_release_one(const char *dsname, void *arg)
3936 {
3937         struct dsl_ds_holdarg *ha = arg;
3938         struct dsl_ds_releasearg *ra;
3939         dsl_dataset_t *ds;
3940         int error;
3941         void *dtag = ha->dstg;
3942         char *name;
3943         boolean_t own = B_FALSE;
3944         boolean_t might_destroy;
3945
3946         /* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3947         name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3948         error = dsl_dataset_hold(name, dtag, &ds);
3949         strfree(name);
3950         if (error == ENOENT && ha->recursive)
3951                 return (0);
3952         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3953         if (error)
3954                 return (error);
3955
3956         ha->gotone = B_TRUE;
3957
3958         ASSERT(dsl_dataset_is_snapshot(ds));
3959
3960         error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3961         if (error) {
3962                 dsl_dataset_rele(ds, dtag);
3963                 return (error);
3964         }
3965
3966         if (might_destroy) {
3967 #ifdef _KERNEL
3968                 name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3969                 error = zfs_unmount_snap(name, NULL);
3970                 strfree(name);
3971                 if (error) {
3972                         dsl_dataset_rele(ds, dtag);
3973                         return (error);
3974                 }
3975 #endif
3976                 if (!dsl_dataset_tryown(ds, B_TRUE, dtag)) {
3977                         dsl_dataset_rele(ds, dtag);
3978                         return (EBUSY);
3979                 } else {
3980                         own = B_TRUE;
3981                         dsl_dataset_make_exclusive(ds, dtag);
3982                 }
3983         }
3984
3985         ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3986         ra->ds = ds;
3987         ra->htag = ha->htag;
3988         ra->own = own;
3989         dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3990             dsl_dataset_user_release_sync, ra, dtag, 0);
3991
3992         return (0);
3993 }
3994
3995 int
3996 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3997     boolean_t recursive)
3998 {
3999         struct dsl_ds_holdarg *ha;
4000         dsl_sync_task_t *dst;
4001         spa_t *spa;
4002         int error;
4003
4004 top:
4005         ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
4006
4007         (void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
4008
4009         error = spa_open(dsname, &spa, FTAG);
4010         if (error) {
4011                 kmem_free(ha, sizeof (struct dsl_ds_holdarg));
4012                 return (error);
4013         }
4014
4015         ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
4016         ha->htag = htag;
4017         ha->snapname = snapname;
4018         ha->recursive = recursive;
4019         if (recursive) {
4020                 error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
4021                     ha, DS_FIND_CHILDREN);
4022         } else {
4023                 error = dsl_dataset_user_release_one(dsname, ha);
4024         }
4025         if (error == 0)
4026                 error = dsl_sync_task_group_wait(ha->dstg);
4027
4028         for (dst = list_head(&ha->dstg->dstg_tasks); dst;
4029             dst = list_next(&ha->dstg->dstg_tasks, dst)) {
4030                 struct dsl_ds_releasearg *ra = dst->dst_arg1;
4031                 dsl_dataset_t *ds = ra->ds;
4032
4033                 if (dst->dst_err)
4034                         dsl_dataset_name(ds, ha->failed);
4035
4036                 if (ra->own)
4037                         dsl_dataset_disown(ds, ha->dstg);
4038                 else
4039                         dsl_dataset_rele(ds, ha->dstg);
4040
4041                 kmem_free(ra, sizeof (struct dsl_ds_releasearg));
4042         }
4043
4044         if (error == 0 && recursive && !ha->gotone)
4045                 error = ENOENT;
4046
4047         if (error && error != EBUSY)
4048                 (void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
4049
4050         dsl_sync_task_group_destroy(ha->dstg);
4051         kmem_free(ha, sizeof (struct dsl_ds_holdarg));
4052         spa_close(spa, FTAG);
4053
4054         /*
4055          * We can get EBUSY if we were racing with deferred destroy and
4056          * dsl_dataset_user_release_check() hadn't done the necessary
4057          * open context setup.  We can also get EBUSY if we're racing
4058          * with destroy and that thread is the ds_owner.  Either way
4059          * the busy condition should be transient, and we should retry
4060          * the release operation.
4061          */
4062         if (error == EBUSY)
4063                 goto top;
4064
4065         return (error);
4066 }
4067
4068 /*
4069  * Called at spa_load time (with retry == B_FALSE) to release a stale
4070  * temporary user hold. Also called by the onexit code (with retry == B_TRUE).
4071  */
4072 int
4073 dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, char *htag,
4074     boolean_t retry)
4075 {
4076         dsl_dataset_t *ds;
4077         char *snap;
4078         char *name;
4079         int namelen;
4080         int error;
4081
4082         do {
4083                 rw_enter(&dp->dp_config_rwlock, RW_READER);
4084                 error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
4085                 rw_exit(&dp->dp_config_rwlock);
4086                 if (error)
4087                         return (error);
4088                 namelen = dsl_dataset_namelen(ds)+1;
4089                 name = kmem_alloc(namelen, KM_SLEEP);
4090                 dsl_dataset_name(ds, name);
4091                 dsl_dataset_rele(ds, FTAG);
4092
4093                 snap = strchr(name, '@');
4094                 *snap = '\0';
4095                 ++snap;
4096                 error = dsl_dataset_user_release(name, snap, htag, B_FALSE);
4097                 kmem_free(name, namelen);
4098
4099                 /*
4100                  * The object can't have been destroyed because we have a hold,
4101                  * but it might have been renamed, resulting in ENOENT.  Retry
4102                  * if we've been requested to do so.
4103                  *
4104                  * It would be nice if we could use the dsobj all the way
4105                  * through and avoid ENOENT entirely.  But we might need to
4106                  * unmount the snapshot, and there's currently no way to lookup
4107                  * a vfsp using a ZFS object id.
4108                  */
4109         } while ((error == ENOENT) && retry);
4110
4111         return (error);
4112 }
4113
4114 int
4115 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
4116 {
4117         dsl_dataset_t *ds;
4118         int err;
4119
4120         err = dsl_dataset_hold(dsname, FTAG, &ds);
4121         if (err)
4122                 return (err);
4123
4124         VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
4125         if (ds->ds_phys->ds_userrefs_obj != 0) {
4126                 zap_attribute_t *za;
4127                 zap_cursor_t zc;
4128
4129                 za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
4130                 for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
4131                     ds->ds_phys->ds_userrefs_obj);
4132                     zap_cursor_retrieve(&zc, za) == 0;
4133                     zap_cursor_advance(&zc)) {
4134                         VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
4135                             za->za_first_integer));
4136                 }
4137                 zap_cursor_fini(&zc);
4138                 kmem_free(za, sizeof (zap_attribute_t));
4139         }
4140         dsl_dataset_rele(ds, FTAG);
4141         return (0);
4142 }
4143
4144 /*
4145  * Note, this function is used as the callback for dmu_objset_find().  We
4146  * always return 0 so that we will continue to find and process
4147  * inconsistent datasets, even if we encounter an error trying to
4148  * process one of them.
4149  */
4150 /* ARGSUSED */
4151 int
4152 dsl_destroy_inconsistent(const char *dsname, void *arg)
4153 {
4154         dsl_dataset_t *ds;
4155
4156         if (dsl_dataset_own(dsname, B_TRUE, FTAG, &ds) == 0) {
4157                 if (DS_IS_INCONSISTENT(ds))
4158                         (void) dsl_dataset_destroy(ds, FTAG, B_FALSE);
4159                 else
4160                         dsl_dataset_disown(ds, FTAG);
4161         }
4162         return (0);
4163 }
4164
4165
4166 /*
4167  * Return (in *usedp) the amount of space written in new that is not
4168  * present in oldsnap.  New may be a snapshot or the head.  Old must be
4169  * a snapshot before new, in new's filesystem (or its origin).  If not then
4170  * fail and return EINVAL.
4171  *
4172  * The written space is calculated by considering two components:  First, we
4173  * ignore any freed space, and calculate the written as new's used space
4174  * minus old's used space.  Next, we add in the amount of space that was freed
4175  * between the two snapshots, thus reducing new's used space relative to old's.
4176  * Specifically, this is the space that was born before old->ds_creation_txg,
4177  * and freed before new (ie. on new's deadlist or a previous deadlist).
4178  *
4179  * space freed                         [---------------------]
4180  * snapshots                       ---O-------O--------O-------O------
4181  *                                         oldsnap            new
4182  */
4183 int
4184 dsl_dataset_space_written(dsl_dataset_t *oldsnap, dsl_dataset_t *new,
4185     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
4186 {
4187         int err = 0;
4188         uint64_t snapobj;
4189         dsl_pool_t *dp = new->ds_dir->dd_pool;
4190
4191         *usedp = 0;
4192         *usedp += new->ds_phys->ds_referenced_bytes;
4193         *usedp -= oldsnap->ds_phys->ds_referenced_bytes;
4194
4195         *compp = 0;
4196         *compp += new->ds_phys->ds_compressed_bytes;
4197         *compp -= oldsnap->ds_phys->ds_compressed_bytes;
4198
4199         *uncompp = 0;
4200         *uncompp += new->ds_phys->ds_uncompressed_bytes;
4201         *uncompp -= oldsnap->ds_phys->ds_uncompressed_bytes;
4202
4203         rw_enter(&dp->dp_config_rwlock, RW_READER);
4204         snapobj = new->ds_object;
4205         while (snapobj != oldsnap->ds_object) {
4206                 dsl_dataset_t *snap;
4207                 uint64_t used, comp, uncomp;
4208
4209                 if (snapobj == new->ds_object) {
4210                         snap = new;
4211                 } else {
4212                         err = dsl_dataset_hold_obj(dp, snapobj, FTAG, &snap);
4213                         if (err != 0)
4214                                 break;
4215                 }
4216
4217                 if (snap->ds_phys->ds_prev_snap_txg ==
4218                     oldsnap->ds_phys->ds_creation_txg) {
4219                         /*
4220                          * The blocks in the deadlist can not be born after
4221                          * ds_prev_snap_txg, so get the whole deadlist space,
4222                          * which is more efficient (especially for old-format
4223                          * deadlists).  Unfortunately the deadlist code
4224                          * doesn't have enough information to make this
4225                          * optimization itself.
4226                          */
4227                         dsl_deadlist_space(&snap->ds_deadlist,
4228                             &used, &comp, &uncomp);
4229                 } else {
4230                         dsl_deadlist_space_range(&snap->ds_deadlist,
4231                             0, oldsnap->ds_phys->ds_creation_txg,
4232                             &used, &comp, &uncomp);
4233                 }
4234                 *usedp += used;
4235                 *compp += comp;
4236                 *uncompp += uncomp;
4237
4238                 /*
4239                  * If we get to the beginning of the chain of snapshots
4240                  * (ds_prev_snap_obj == 0) before oldsnap, then oldsnap
4241                  * was not a snapshot of/before new.
4242                  */
4243                 snapobj = snap->ds_phys->ds_prev_snap_obj;
4244                 if (snap != new)
4245                         dsl_dataset_rele(snap, FTAG);
4246                 if (snapobj == 0) {
4247                         err = EINVAL;
4248                         break;
4249                 }
4250
4251         }
4252         rw_exit(&dp->dp_config_rwlock);
4253         return (err);
4254 }
4255
4256 /*
4257  * Return (in *usedp) the amount of space that will be reclaimed if firstsnap,
4258  * lastsnap, and all snapshots in between are deleted.
4259  *
4260  * blocks that would be freed            [---------------------------]
4261  * snapshots                       ---O-------O--------O-------O--------O
4262  *                                        firstsnap        lastsnap
4263  *
4264  * This is the set of blocks that were born after the snap before firstsnap,
4265  * (birth > firstsnap->prev_snap_txg) and died before the snap after the
4266  * last snap (ie, is on lastsnap->ds_next->ds_deadlist or an earlier deadlist).
4267  * We calculate this by iterating over the relevant deadlists (from the snap
4268  * after lastsnap, backward to the snap after firstsnap), summing up the
4269  * space on the deadlist that was born after the snap before firstsnap.
4270  */
4271 int
4272 dsl_dataset_space_wouldfree(dsl_dataset_t *firstsnap,
4273     dsl_dataset_t *lastsnap,
4274     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
4275 {
4276         int err = 0;
4277         uint64_t snapobj;
4278         dsl_pool_t *dp = firstsnap->ds_dir->dd_pool;
4279
4280         ASSERT(dsl_dataset_is_snapshot(firstsnap));
4281         ASSERT(dsl_dataset_is_snapshot(lastsnap));
4282
4283         /*
4284          * Check that the snapshots are in the same dsl_dir, and firstsnap
4285          * is before lastsnap.
4286          */
4287         if (firstsnap->ds_dir != lastsnap->ds_dir ||
4288             firstsnap->ds_phys->ds_creation_txg >
4289             lastsnap->ds_phys->ds_creation_txg)
4290                 return (EINVAL);
4291
4292         *usedp = *compp = *uncompp = 0;
4293
4294         rw_enter(&dp->dp_config_rwlock, RW_READER);
4295         snapobj = lastsnap->ds_phys->ds_next_snap_obj;
4296         while (snapobj != firstsnap->ds_object) {
4297                 dsl_dataset_t *ds;
4298                 uint64_t used, comp, uncomp;
4299
4300                 err = dsl_dataset_hold_obj(dp, snapobj, FTAG, &ds);
4301                 if (err != 0)
4302                         break;
4303
4304                 dsl_deadlist_space_range(&ds->ds_deadlist,
4305                     firstsnap->ds_phys->ds_prev_snap_txg, UINT64_MAX,
4306                     &used, &comp, &uncomp);
4307                 *usedp += used;
4308                 *compp += comp;
4309                 *uncompp += uncomp;
4310
4311                 snapobj = ds->ds_phys->ds_prev_snap_obj;
4312                 ASSERT3U(snapobj, !=, 0);
4313                 dsl_dataset_rele(ds, FTAG);
4314         }
4315         rw_exit(&dp->dp_config_rwlock);
4316         return (err);
4317 }
4318
4319 #if defined(_KERNEL) && defined(HAVE_SPL)
4320 EXPORT_SYMBOL(dmu_snapshots_destroy_nvl);
4321 EXPORT_SYMBOL(dsl_dataset_hold);
4322 EXPORT_SYMBOL(dsl_dataset_hold_obj);
4323 EXPORT_SYMBOL(dsl_dataset_own);
4324 EXPORT_SYMBOL(dsl_dataset_own_obj);
4325 EXPORT_SYMBOL(dsl_dataset_name);
4326 EXPORT_SYMBOL(dsl_dataset_rele);
4327 EXPORT_SYMBOL(dsl_dataset_disown);
4328 EXPORT_SYMBOL(dsl_dataset_drop_ref);
4329 EXPORT_SYMBOL(dsl_dataset_tryown);
4330 EXPORT_SYMBOL(dsl_dataset_make_exclusive);
4331 EXPORT_SYMBOL(dsl_dataset_create_sync);
4332 EXPORT_SYMBOL(dsl_dataset_create_sync_dd);
4333 EXPORT_SYMBOL(dsl_dataset_destroy);
4334 EXPORT_SYMBOL(dsl_dataset_destroy_check);
4335 EXPORT_SYMBOL(dsl_dataset_destroy_sync);
4336 EXPORT_SYMBOL(dsl_dataset_snapshot_check);
4337 EXPORT_SYMBOL(dsl_dataset_snapshot_sync);
4338 EXPORT_SYMBOL(dsl_dataset_rename);
4339 EXPORT_SYMBOL(dsl_dataset_promote);
4340 EXPORT_SYMBOL(dsl_dataset_clone_swap);
4341 EXPORT_SYMBOL(dsl_dataset_user_hold);
4342 EXPORT_SYMBOL(dsl_dataset_user_release);
4343 EXPORT_SYMBOL(dsl_dataset_user_release_tmp);
4344 EXPORT_SYMBOL(dsl_dataset_get_holds);
4345 EXPORT_SYMBOL(dsl_dataset_get_blkptr);
4346 EXPORT_SYMBOL(dsl_dataset_set_blkptr);
4347 EXPORT_SYMBOL(dsl_dataset_get_spa);
4348 EXPORT_SYMBOL(dsl_dataset_modified_since_lastsnap);
4349 EXPORT_SYMBOL(dsl_dataset_space_written);
4350 EXPORT_SYMBOL(dsl_dataset_space_wouldfree);
4351 EXPORT_SYMBOL(dsl_dataset_sync);
4352 EXPORT_SYMBOL(dsl_dataset_block_born);
4353 EXPORT_SYMBOL(dsl_dataset_block_kill);
4354 EXPORT_SYMBOL(dsl_dataset_block_freeable);
4355 EXPORT_SYMBOL(dsl_dataset_prev_snap_txg);
4356 EXPORT_SYMBOL(dsl_dataset_dirty);
4357 EXPORT_SYMBOL(dsl_dataset_stats);
4358 EXPORT_SYMBOL(dsl_dataset_fast_stat);
4359 EXPORT_SYMBOL(dsl_dataset_space);
4360 EXPORT_SYMBOL(dsl_dataset_fsid_guid);
4361 EXPORT_SYMBOL(dsl_dsobj_to_dsname);
4362 EXPORT_SYMBOL(dsl_dataset_check_quota);
4363 EXPORT_SYMBOL(dsl_dataset_set_quota);
4364 EXPORT_SYMBOL(dsl_dataset_set_quota_sync);
4365 EXPORT_SYMBOL(dsl_dataset_set_reservation);
4366 EXPORT_SYMBOL(dsl_destroy_inconsistent);
4367 #endif