sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_destroy.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) 2013 by Delphix. All rights reserved.
  24  * Copyright (c) 2013 Steven Hartland. All rights reserved.
  25  * Copyright (c) 2013 by Joyent, Inc. All rights reserved.
  26  */
  27
  28 #include <sys/zfs_context.h>
  29 #include <sys/dsl_userhold.h>
  30 #include <sys/dsl_dataset.h>
  31 #include <sys/dsl_synctask.h>
  32 #include <sys/dmu_tx.h>
  33 #include <sys/dsl_pool.h>
  34 #include <sys/dsl_dir.h>
  35 #include <sys/dmu_traverse.h>
  36 #include <sys/dsl_scan.h>
  37 #include <sys/dmu_objset.h>
  38 #include <sys/zap.h>
  39 #include <sys/zfeature.h>
  40 #include <sys/zfs_ioctl.h>
  41 #include <sys/dsl_deleg.h>
  42 #include <sys/dmu_impl.h>
  43
  44 typedef struct dmu_snapshots_destroy_arg {
  45         nvlist_t *dsda_snaps;
  46         nvlist_t *dsda_successful_snaps;
  47         boolean_t dsda_defer;
  48         nvlist_t *dsda_errlist;
  49 } dmu_snapshots_destroy_arg_t;
  50
  51 int
  52 dsl_destroy_snapshot_check_impl(dsl_dataset_t *ds, boolean_t defer)
  53 {
  54         if (!dsl_dataset_is_snapshot(ds))
  55                 return (SET_ERROR(EINVAL));
  56
  57         if (dsl_dataset_long_held(ds))
  58                 return (SET_ERROR(EBUSY));
  59
  60         /*
  61          * Only allow deferred destroy on pools that support it.
  62          * NOTE: deferred destroy is only supported on snapshots.
  63          */
  64         if (defer) {
  65                 if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
  66                     SPA_VERSION_USERREFS)
  67                         return (SET_ERROR(ENOTSUP));
  68                 return (0);
  69         }
  70
  71         /*
  72          * If this snapshot has an elevated user reference count,
  73          * we can't destroy it yet.
  74          */
  75         if (ds->ds_userrefs > 0)
  76                 return (SET_ERROR(EBUSY));
  77
  78         /*
  79          * Can't delete a branch point.
  80          */
  81         if (ds->ds_phys->ds_num_children > 1)
  82                 return (SET_ERROR(EEXIST));
  83
  84         return (0);
  85 }
  86
  87 static int
  88 dsl_destroy_snapshot_check(void *arg, dmu_tx_t *tx)
  89 {
  90         dmu_snapshots_destroy_arg_t *dsda = arg;
  91         dsl_pool_t *dp = dmu_tx_pool(tx);
  92         nvpair_t *pair;
  93         int error = 0;
  94
  95         if (!dmu_tx_is_syncing(tx))
  96                 return (0);
  97
  98         for (pair = nvlist_next_nvpair(dsda->dsda_snaps, NULL);
  99             pair != NULL; pair = nvlist_next_nvpair(dsda->dsda_snaps, pair)) {
 100                 dsl_dataset_t *ds;
 101
 102                 error = dsl_dataset_hold(dp, nvpair_name(pair),
 103                     FTAG, &ds);
 104
 105                 /*
 106                  * If the snapshot does not exist, silently ignore it
 107                  * (it's "already destroyed").
 108                  */
 109                 if (error == ENOENT)
 110                         continue;
 111
 112                 if (error == 0) {
 113                         error = dsl_destroy_snapshot_check_impl(ds,
 114                             dsda->dsda_defer);
 115                         dsl_dataset_rele(ds, FTAG);
 116                 }
 117
 118                 if (error == 0) {
 119                         fnvlist_add_boolean(dsda->dsda_successful_snaps,
 120                             nvpair_name(pair));
 121                 } else {
 122                         fnvlist_add_int32(dsda->dsda_errlist,
 123                             nvpair_name(pair), error);
 124                 }
 125         }
 126
 127         pair = nvlist_next_nvpair(dsda->dsda_errlist, NULL);
 128         if (pair != NULL)
 129                 return (fnvpair_value_int32(pair));
 130
 131         return (0);
 132 }
 133
 134 struct process_old_arg {
 135         dsl_dataset_t *ds;
 136         dsl_dataset_t *ds_prev;
 137         boolean_t after_branch_point;
 138         zio_t *pio;
 139         uint64_t used, comp, uncomp;
 140 };
 141
 142 static int
 143 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
 144 {
 145         struct process_old_arg *poa = arg;
 146         dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
 147
 148         ASSERT(!BP_IS_HOLE(bp));
 149
 150         if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
 151                 dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
 152                 if (poa->ds_prev && !poa->after_branch_point &&
 153                     bp->blk_birth >
 154                     poa->ds_prev->ds_phys->ds_prev_snap_txg) {
 155                         poa->ds_prev->ds_phys->ds_unique_bytes +=
 156                             bp_get_dsize_sync(dp->dp_spa, bp);
 157                 }
 158         } else {
 159                 poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
 160                 poa->comp += BP_GET_PSIZE(bp);
 161                 poa->uncomp += BP_GET_UCSIZE(bp);
 162                 dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
 163         }
 164         return (0);
 165 }
 166
 167 static void
 168 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
 169     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
 170 {
 171         struct process_old_arg poa = { 0 };
 172         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 173         objset_t *mos = dp->dp_meta_objset;
 174         uint64_t deadlist_obj;
 175
 176         ASSERT(ds->ds_deadlist.dl_oldfmt);
 177         ASSERT(ds_next->ds_deadlist.dl_oldfmt);
 178
 179         poa.ds = ds;
 180         poa.ds_prev = ds_prev;
 181         poa.after_branch_point = after_branch_point;
 182         poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 183         VERIFY0(bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
 184             process_old_cb, &poa, tx));
 185         VERIFY0(zio_wait(poa.pio));
 186         ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
 187
 188         /* change snapused */
 189         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
 190             -poa.used, -poa.comp, -poa.uncomp, tx);
 191
 192         /* swap next's deadlist to our deadlist */
 193         dsl_deadlist_close(&ds->ds_deadlist);
 194         dsl_deadlist_close(&ds_next->ds_deadlist);
 195         deadlist_obj = ds->ds_phys->ds_deadlist_obj;
 196         ds->ds_phys->ds_deadlist_obj = ds_next->ds_phys->ds_deadlist_obj;
 197         ds_next->ds_phys->ds_deadlist_obj = deadlist_obj;
 198         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
 199         dsl_deadlist_open(&ds_next->ds_deadlist, mos,
 200             ds_next->ds_phys->ds_deadlist_obj);
 201 }
 202
 203 static void
 204 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
 205 {
 206         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 207         zap_cursor_t zc;
 208         zap_attribute_t za;
 209
 210         /*
 211          * If it is the old version, dd_clones doesn't exist so we can't
 212          * find the clones, but dsl_deadlist_remove_key() is a no-op so it
 213          * doesn't matter.
 214          */
 215         if (ds->ds_dir->dd_phys->dd_clones == 0)
 216                 return;
 217
 218         for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
 219             zap_cursor_retrieve(&zc, &za) == 0;
 220             zap_cursor_advance(&zc)) {
 221                 dsl_dataset_t *clone;
 222
 223                 VERIFY0(dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
 224                     za.za_first_integer, FTAG, &clone));
 225                 if (clone->ds_dir->dd_origin_txg > mintxg) {
 226                         dsl_deadlist_remove_key(&clone->ds_deadlist,
 227                             mintxg, tx);
 228                         dsl_dataset_remove_clones_key(clone, mintxg, tx);
 229                 }
 230                 dsl_dataset_rele(clone, FTAG);
 231         }
 232         zap_cursor_fini(&zc);
 233 }
 234
 235 void
 236 dsl_destroy_snapshot_sync_impl(dsl_dataset_t *ds, boolean_t defer, dmu_tx_t *tx)
 237 {
 238         int err;
 239         int after_branch_point = FALSE;
 240         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 241         objset_t *mos = dp->dp_meta_objset;
 242         dsl_dataset_t *ds_prev = NULL;
 243         uint64_t obj;
 244
 245         ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
 246         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
 247         ASSERT(refcount_is_zero(&ds->ds_longholds));
 248
 249         if (defer &&
 250             (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1)) {
 251                 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
 252                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
 253                 ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
 254                 spa_history_log_internal_ds(ds, "defer_destroy", tx, "");
 255                 return;
 256         }
 257
 258         ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 259
 260         /* We need to log before removing it from the namespace. */
 261         spa_history_log_internal_ds(ds, "destroy", tx, "");
 262
 263         dsl_scan_ds_destroyed(ds, tx);
 264
 265         obj = ds->ds_object;
 266
 267         if (ds->ds_phys->ds_prev_snap_obj != 0) {
 268                 ASSERT3P(ds->ds_prev, ==, NULL);
 269                 VERIFY0(dsl_dataset_hold_obj(dp,
 270                     ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
 271                 after_branch_point =
 272                     (ds_prev->ds_phys->ds_next_snap_obj != obj);
 273
 274                 dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
 275                 if (after_branch_point &&
 276                     ds_prev->ds_phys->ds_next_clones_obj != 0) {
 277                         dsl_dataset_remove_from_next_clones(ds_prev, obj, tx);
 278                         if (ds->ds_phys->ds_next_snap_obj != 0) {
 279                                 VERIFY0(zap_add_int(mos,
 280                                     ds_prev->ds_phys->ds_next_clones_obj,
 281                                     ds->ds_phys->ds_next_snap_obj, tx));
 282                         }
 283                 }
 284                 if (!after_branch_point) {
 285                         ds_prev->ds_phys->ds_next_snap_obj =
 286                             ds->ds_phys->ds_next_snap_obj;
 287                 }
 288         }
 289
 290         dsl_dataset_t *ds_next;
 291         uint64_t old_unique;
 292         uint64_t used = 0, comp = 0, uncomp = 0;
 293
 294         VERIFY0(dsl_dataset_hold_obj(dp,
 295             ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
 296         ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
 297
 298         old_unique = ds_next->ds_phys->ds_unique_bytes;
 299
 300         dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
 301         ds_next->ds_phys->ds_prev_snap_obj =
 302             ds->ds_phys->ds_prev_snap_obj;
 303         ds_next->ds_phys->ds_prev_snap_txg =
 304             ds->ds_phys->ds_prev_snap_txg;
 305         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
 306             ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
 307
 308         if (ds_next->ds_deadlist.dl_oldfmt) {
 309                 process_old_deadlist(ds, ds_prev, ds_next,
 310                     after_branch_point, tx);
 311         } else {
 312                 /* Adjust prev's unique space. */
 313                 if (ds_prev && !after_branch_point) {
 314                         dsl_deadlist_space_range(&ds_next->ds_deadlist,
 315                             ds_prev->ds_phys->ds_prev_snap_txg,
 316                             ds->ds_phys->ds_prev_snap_txg,
 317                             &used, &comp, &uncomp);
 318                         ds_prev->ds_phys->ds_unique_bytes += used;
 319                 }
 320
 321                 /* Adjust snapused. */
 322                 dsl_deadlist_space_range(&ds_next->ds_deadlist,
 323                     ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
 324                     &used, &comp, &uncomp);
 325                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
 326                     -used, -comp, -uncomp, tx);
 327
 328                 /* Move blocks to be freed to pool's free list. */
 329                 dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
 330                     &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
 331                     tx);
 332                 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
 333                     DD_USED_HEAD, used, comp, uncomp, tx);
 334
 335                 /* Merge our deadlist into next's and free it. */
 336                 dsl_deadlist_merge(&ds_next->ds_deadlist,
 337                     ds->ds_phys->ds_deadlist_obj, tx);
 338         }
 339         dsl_deadlist_close(&ds->ds_deadlist);
 340         dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
 341         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 342         ds->ds_phys->ds_deadlist_obj = 0;
 343
 344         /* Collapse range in clone heads */
 345         dsl_dataset_remove_clones_key(ds,
 346             ds->ds_phys->ds_creation_txg, tx);
 347
 348         if (dsl_dataset_is_snapshot(ds_next)) {
 349                 dsl_dataset_t *ds_nextnext;
 350
 351                 /*
 352                  * Update next's unique to include blocks which
 353                  * were previously shared by only this snapshot
 354                  * and it.  Those blocks will be born after the
 355                  * prev snap and before this snap, and will have
 356                  * died after the next snap and before the one
 357                  * after that (ie. be on the snap after next's
 358                  * deadlist).
 359                  */
 360                 VERIFY0(dsl_dataset_hold_obj(dp,
 361                     ds_next->ds_phys->ds_next_snap_obj, FTAG, &ds_nextnext));
 362                 dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
 363                     ds->ds_phys->ds_prev_snap_txg,
 364                     ds->ds_phys->ds_creation_txg,
 365                     &used, &comp, &uncomp);
 366                 ds_next->ds_phys->ds_unique_bytes += used;
 367                 dsl_dataset_rele(ds_nextnext, FTAG);
 368                 ASSERT3P(ds_next->ds_prev, ==, NULL);
 369
 370                 /* Collapse range in this head. */
 371                 dsl_dataset_t *hds;
 372                 VERIFY0(dsl_dataset_hold_obj(dp,
 373                     ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &hds));
 374                 dsl_deadlist_remove_key(&hds->ds_deadlist,
 375                     ds->ds_phys->ds_creation_txg, tx);
 376                 dsl_dataset_rele(hds, FTAG);
 377
 378         } else {
 379                 ASSERT3P(ds_next->ds_prev, ==, ds);
 380                 dsl_dataset_rele(ds_next->ds_prev, ds_next);
 381                 ds_next->ds_prev = NULL;
 382                 if (ds_prev) {
 383                         VERIFY0(dsl_dataset_hold_obj(dp,
 384                             ds->ds_phys->ds_prev_snap_obj,
 385                             ds_next, &ds_next->ds_prev));
 386                 }
 387
 388                 dsl_dataset_recalc_head_uniq(ds_next);
 389
 390                 /*
 391                  * Reduce the amount of our unconsumed refreservation
 392                  * being charged to our parent by the amount of
 393                  * new unique data we have gained.
 394                  */
 395                 if (old_unique < ds_next->ds_reserved) {
 396                         int64_t mrsdelta;
 397                         uint64_t new_unique =
 398                             ds_next->ds_phys->ds_unique_bytes;
 399
 400                         ASSERT(old_unique <= new_unique);
 401                         mrsdelta = MIN(new_unique - old_unique,
 402                             ds_next->ds_reserved - old_unique);
 403                         dsl_dir_diduse_space(ds->ds_dir,
 404                             DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
 405                 }
 406         }
 407         dsl_dataset_rele(ds_next, FTAG);
 408
 409         /*
 410          * This must be done after the dsl_traverse(), because it will
 411          * re-open the objset.
 412          */
 413         if (ds->ds_objset) {
 414                 dmu_objset_evict(ds->ds_objset);
 415                 ds->ds_objset = NULL;
 416         }
 417
 418         /* remove from snapshot namespace */
 419         dsl_dataset_t *ds_head;
 420         ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
 421         VERIFY0(dsl_dataset_hold_obj(dp,
 422             ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
 423         VERIFY0(dsl_dataset_get_snapname(ds));
 424 #ifdef ZFS_DEBUG
 425         {
 426                 uint64_t val;
 427
 428                 err = dsl_dataset_snap_lookup(ds_head,
 429                     ds->ds_snapname, &val);
 430                 ASSERT0(err);
 431                 ASSERT3U(val, ==, obj);
 432         }
 433 #endif
 434         VERIFY0(dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx, B_TRUE));
 435         dsl_dataset_rele(ds_head, FTAG);
 436
 437         if (ds_prev != NULL)
 438                 dsl_dataset_rele(ds_prev, FTAG);
 439
 440         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
 441
 442         if (ds->ds_phys->ds_next_clones_obj != 0) {
 443                 uint64_t count;
 444                 ASSERT0(zap_count(mos,
 445                     ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
 446                 VERIFY0(dmu_object_free(mos,
 447                     ds->ds_phys->ds_next_clones_obj, tx));
 448         }
 449         if (ds->ds_phys->ds_props_obj != 0)
 450                 VERIFY0(zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
 451         if (ds->ds_phys->ds_userrefs_obj != 0)
 452                 VERIFY0(zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
 453         dsl_dir_rele(ds->ds_dir, ds);
 454         ds->ds_dir = NULL;
 455         dmu_object_free_zapified(mos, obj, tx);
 456 }
 457
 458 static void
 459 dsl_destroy_snapshot_sync(void *arg, dmu_tx_t *tx)
 460 {
 461         dmu_snapshots_destroy_arg_t *dsda = arg;
 462         dsl_pool_t *dp = dmu_tx_pool(tx);
 463         nvpair_t *pair;
 464
 465         for (pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, NULL);
 466             pair != NULL;
 467             pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, pair)) {
 468                 dsl_dataset_t *ds;
 469
 470                 VERIFY0(dsl_dataset_hold(dp, nvpair_name(pair), FTAG, &ds));
 471
 472                 dsl_destroy_snapshot_sync_impl(ds, dsda->dsda_defer, tx);
 473                 dsl_dataset_rele(ds, FTAG);
 474         }
 475 }
 476
 477 /*
 478  * The semantics of this function are described in the comment above
 479  * lzc_destroy_snaps().  To summarize:
 480  *
 481  * The snapshots must all be in the same pool.
 482  *
 483  * Snapshots that don't exist will be silently ignored (considered to be
 484  * "already deleted").
 485  *
 486  * On success, all snaps will be destroyed and this will return 0.
 487  * On failure, no snaps will be destroyed, the errlist will be filled in,
 488  * and this will return an errno.
 489  */
 490 int
 491 dsl_destroy_snapshots_nvl(nvlist_t *snaps, boolean_t defer,
 492     nvlist_t *errlist)
 493 {
 494         dmu_snapshots_destroy_arg_t dsda;
 495         int error;
 496         nvpair_t *pair;
 497
 498         pair = nvlist_next_nvpair(snaps, NULL);
 499         if (pair == NULL)
 500                 return (0);
 501
 502         dsda.dsda_snaps = snaps;
 503         dsda.dsda_successful_snaps = fnvlist_alloc();
 504         dsda.dsda_defer = defer;
 505         dsda.dsda_errlist = errlist;
 506
 507         error = dsl_sync_task(nvpair_name(pair),
 508             dsl_destroy_snapshot_check, dsl_destroy_snapshot_sync,
 509             &dsda, 0);
 510         fnvlist_free(dsda.dsda_successful_snaps);
 511
 512         return (error);
 513 }
 514
 515 int
 516 dsl_destroy_snapshot(const char *name, boolean_t defer)
 517 {
 518         int error;
 519         nvlist_t *nvl = fnvlist_alloc();
 520         nvlist_t *errlist = fnvlist_alloc();
 521
 522         fnvlist_add_boolean(nvl, name);
 523         error = dsl_destroy_snapshots_nvl(nvl, defer, errlist);
 524         fnvlist_free(errlist);
 525         fnvlist_free(nvl);
 526         return (error);
 527 }
 528
 529 struct killarg {
 530         dsl_dataset_t *ds;
 531         dmu_tx_t *tx;
 532 };
 533
 534 /* ARGSUSED */
 535 static int
 536 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 537     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
 538 {
 539         struct killarg *ka = arg;
 540         dmu_tx_t *tx = ka->tx;
 541
 542         if (BP_IS_HOLE(bp))
 543                 return (0);
 544
 545         if (zb->zb_level == ZB_ZIL_LEVEL) {
 546                 ASSERT(zilog != NULL);
 547                 /*
 548                  * It's a block in the intent log.  It has no
 549                  * accounting, so just free it.
 550                  */
 551                 dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
 552         } else {
 553                 ASSERT(zilog == NULL);
 554                 ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
 555                 (void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
 556         }
 557
 558         return (0);
 559 }
 560
 561 static void
 562 old_synchronous_dataset_destroy(dsl_dataset_t *ds, dmu_tx_t *tx)
 563 {
 564         struct killarg ka;
 565
 566         /*
 567          * Free everything that we point to (that's born after
 568          * the previous snapshot, if we are a clone)
 569          *
 570          * NB: this should be very quick, because we already
 571          * freed all the objects in open context.
 572          */
 573         ka.ds = ds;
 574         ka.tx = tx;
 575         VERIFY0(traverse_dataset(ds,
 576             ds->ds_phys->ds_prev_snap_txg, TRAVERSE_POST,
 577             kill_blkptr, &ka));
 578         ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) || ds->ds_phys->ds_unique_bytes == 0);
 579 }
 580
 581 typedef struct dsl_destroy_head_arg {
 582         const char *ddha_name;
 583 } dsl_destroy_head_arg_t;
 584
 585 int
 586 dsl_destroy_head_check_impl(dsl_dataset_t *ds, int expected_holds)
 587 {
 588         int error;
 589         uint64_t count;
 590         objset_t *mos;
 591
 592         if (dsl_dataset_is_snapshot(ds))
 593                 return (SET_ERROR(EINVAL));
 594
 595         if (refcount_count(&ds->ds_longholds) != expected_holds)
 596                 return (SET_ERROR(EBUSY));
 597
 598         mos = ds->ds_dir->dd_pool->dp_meta_objset;
 599
 600         /*
 601          * Can't delete a head dataset if there are snapshots of it.
 602          * (Except if the only snapshots are from the branch we cloned
 603          * from.)
 604          */
 605         if (ds->ds_prev != NULL &&
 606             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
 607                 return (SET_ERROR(EBUSY));
 608
 609         /*
 610          * Can't delete if there are children of this fs.
 611          */
 612         error = zap_count(mos,
 613             ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
 614         if (error != 0)
 615                 return (error);
 616         if (count != 0)
 617                 return (SET_ERROR(EEXIST));
 618
 619         if (dsl_dir_is_clone(ds->ds_dir) && DS_IS_DEFER_DESTROY(ds->ds_prev) &&
 620             ds->ds_prev->ds_phys->ds_num_children == 2 &&
 621             ds->ds_prev->ds_userrefs == 0) {
 622                 /* We need to remove the origin snapshot as well. */
 623                 if (!refcount_is_zero(&ds->ds_prev->ds_longholds))
 624                         return (SET_ERROR(EBUSY));
 625         }
 626         return (0);
 627 }
 628
 629 static int
 630 dsl_destroy_head_check(void *arg, dmu_tx_t *tx)
 631 {
 632         dsl_destroy_head_arg_t *ddha = arg;
 633         dsl_pool_t *dp = dmu_tx_pool(tx);
 634         dsl_dataset_t *ds;
 635         int error;
 636
 637         error = dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds);
 638         if (error != 0)
 639                 return (error);
 640
 641         error = dsl_destroy_head_check_impl(ds, 0);
 642         dsl_dataset_rele(ds, FTAG);
 643         return (error);
 644 }
 645
 646 static void
 647 dsl_dir_destroy_sync(uint64_t ddobj, dmu_tx_t *tx)
 648 {
 649         dsl_dir_t *dd;
 650         dsl_pool_t *dp = dmu_tx_pool(tx);
 651         objset_t *mos = dp->dp_meta_objset;
 652         dd_used_t t;
 653
 654         ASSERT(RRW_WRITE_HELD(&dmu_tx_pool(tx)->dp_config_rwlock));
 655
 656         VERIFY0(dsl_dir_hold_obj(dp, ddobj, NULL, FTAG, &dd));
 657
 658         ASSERT0(dd->dd_phys->dd_head_dataset_obj);
 659
 660         /*
 661          * Decrement the filesystem count for all parent filesystems.
 662          *
 663          * When we receive an incremental stream into a filesystem that already
 664          * exists, a temporary clone is created.  We never count this temporary
 665          * clone, whose name begins with a '%'.
 666          */
 667         if (dd->dd_myname[0] != '%' && dd->dd_parent != NULL)
 668                 dsl_fs_ss_count_adjust(dd->dd_parent, -1,
 669                     DD_FIELD_FILESYSTEM_COUNT, tx);
 670
 671         /*
 672          * Remove our reservation. The impl() routine avoids setting the
 673          * actual property, which would require the (already destroyed) ds.
 674          */
 675         dsl_dir_set_reservation_sync_impl(dd, 0, tx);
 676
 677         ASSERT0(dd->dd_phys->dd_used_bytes);
 678         ASSERT0(dd->dd_phys->dd_reserved);
 679         for (t = 0; t < DD_USED_NUM; t++)
 680                 ASSERT0(dd->dd_phys->dd_used_breakdown[t]);
 681
 682         VERIFY0(zap_destroy(mos, dd->dd_phys->dd_child_dir_zapobj, tx));
 683         VERIFY0(zap_destroy(mos, dd->dd_phys->dd_props_zapobj, tx));
 684         VERIFY0(dsl_deleg_destroy(mos, dd->dd_phys->dd_deleg_zapobj, tx));
 685         VERIFY0(zap_remove(mos,
 686             dd->dd_parent->dd_phys->dd_child_dir_zapobj, dd->dd_myname, tx));
 687
 688         dsl_dir_rele(dd, FTAG);
 689         dmu_object_free_zapified(mos, ddobj, tx);
 690 }
 691
 692 void
 693 dsl_destroy_head_sync_impl(dsl_dataset_t *ds, dmu_tx_t *tx)
 694 {
 695         dsl_pool_t *dp = dmu_tx_pool(tx);
 696         objset_t *mos = dp->dp_meta_objset;
 697         uint64_t obj, ddobj, prevobj = 0;
 698         boolean_t rmorigin;
 699
 700         ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 701         ASSERT(ds->ds_prev == NULL ||
 702             ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
 703         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
 704         ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
 705
 706         /* We need to log before removing it from the namespace. */
 707         spa_history_log_internal_ds(ds, "destroy", tx, "");
 708
 709         rmorigin = (dsl_dir_is_clone(ds->ds_dir) &&
 710             DS_IS_DEFER_DESTROY(ds->ds_prev) &&
 711             ds->ds_prev->ds_phys->ds_num_children == 2 &&
 712             ds->ds_prev->ds_userrefs == 0);
 713
 714         /* Remove our reservation */
 715         if (ds->ds_reserved != 0) {
 716                 dsl_dataset_set_refreservation_sync_impl(ds,
 717                     (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
 718                     0, tx);
 719                 ASSERT0(ds->ds_reserved);
 720         }
 721
 722         dsl_scan_ds_destroyed(ds, tx);
 723
 724         obj = ds->ds_object;
 725
 726         if (ds->ds_phys->ds_prev_snap_obj != 0) {
 727                 /* This is a clone */
 728                 ASSERT(ds->ds_prev != NULL);
 729                 ASSERT3U(ds->ds_prev->ds_phys->ds_next_snap_obj, !=, obj);
 730                 ASSERT0(ds->ds_phys->ds_next_snap_obj);
 731
 732                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
 733                 if (ds->ds_prev->ds_phys->ds_next_clones_obj != 0) {
 734                         dsl_dataset_remove_from_next_clones(ds->ds_prev,
 735                             obj, tx);
 736                 }
 737
 738                 ASSERT3U(ds->ds_prev->ds_phys->ds_num_children, >, 1);
 739                 ds->ds_prev->ds_phys->ds_num_children--;
 740         }
 741
 742         /*
 743          * Destroy the deadlist.  Unless it's a clone, the
 744          * deadlist should be empty.  (If it's a clone, it's
 745          * safe to ignore the deadlist contents.)
 746          */
 747         dsl_deadlist_close(&ds->ds_deadlist);
 748         dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
 749         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 750         ds->ds_phys->ds_deadlist_obj = 0;
 751
 752         objset_t *os;
 753         VERIFY0(dmu_objset_from_ds(ds, &os));
 754
 755         if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
 756                 old_synchronous_dataset_destroy(ds, tx);
 757         } else {
 758                 /*
 759                  * Move the bptree into the pool's list of trees to
 760                  * clean up and update space accounting information.
 761                  */
 762                 uint64_t used, comp, uncomp;
 763
 764                 zil_destroy_sync(dmu_objset_zil(os), tx);
 765
 766                 if (!spa_feature_is_active(dp->dp_spa,
 767                     SPA_FEATURE_ASYNC_DESTROY)) {
 768                         dsl_scan_t *scn = dp->dp_scan;
 769                         spa_feature_incr(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY,
 770                             tx);
 771                         dp->dp_bptree_obj = bptree_alloc(mos, tx);
 772                         VERIFY0(zap_add(mos,
 773                             DMU_POOL_DIRECTORY_OBJECT,
 774                             DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
 775                             &dp->dp_bptree_obj, tx));
 776                         ASSERT(!scn->scn_async_destroying);
 777                         scn->scn_async_destroying = B_TRUE;
 778                 }
 779
 780                 used = ds->ds_dir->dd_phys->dd_used_bytes;
 781                 comp = ds->ds_dir->dd_phys->dd_compressed_bytes;
 782                 uncomp = ds->ds_dir->dd_phys->dd_uncompressed_bytes;
 783
 784                 ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
 785                     ds->ds_phys->ds_unique_bytes == used);
 786
 787                 bptree_add(mos, dp->dp_bptree_obj,
 788                     &ds->ds_phys->ds_bp, ds->ds_phys->ds_prev_snap_txg,
 789                     used, comp, uncomp, tx);
 790                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
 791                     -used, -comp, -uncomp, tx);
 792                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
 793                     used, comp, uncomp, tx);
 794         }
 795
 796         if (ds->ds_prev != NULL) {
 797                 if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
 798                         VERIFY0(zap_remove_int(mos,
 799                             ds->ds_prev->ds_dir->dd_phys->dd_clones,
 800                             ds->ds_object, tx));
 801                 }
 802                 prevobj = ds->ds_prev->ds_object;
 803                 dsl_dataset_rele(ds->ds_prev, ds);
 804                 ds->ds_prev = NULL;
 805         }
 806
 807         /*
 808          * This must be done after the dsl_traverse(), because it will
 809          * re-open the objset.
 810          */
 811         if (ds->ds_objset) {
 812                 dmu_objset_evict(ds->ds_objset);
 813                 ds->ds_objset = NULL;
 814         }
 815
 816         /* Erase the link in the dir */
 817         dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
 818         ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
 819         ddobj = ds->ds_dir->dd_object;
 820         ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
 821         VERIFY0(zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx));
 822
 823         if (ds->ds_bookmarks != 0) {
 824                 VERIFY0(zap_destroy(mos,
 825                     ds->ds_bookmarks, tx));
 826                 spa_feature_decr(dp->dp_spa, SPA_FEATURE_BOOKMARKS, tx);
 827         }
 828
 829         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
 830
 831         ASSERT0(ds->ds_phys->ds_next_clones_obj);
 832         ASSERT0(ds->ds_phys->ds_props_obj);
 833         ASSERT0(ds->ds_phys->ds_userrefs_obj);
 834         dsl_dir_rele(ds->ds_dir, ds);
 835         ds->ds_dir = NULL;
 836         dmu_object_free_zapified(mos, obj, tx);
 837
 838         dsl_dir_destroy_sync(ddobj, tx);
 839
 840         if (rmorigin) {
 841                 dsl_dataset_t *prev;
 842                 VERIFY0(dsl_dataset_hold_obj(dp, prevobj, FTAG, &prev));
 843                 dsl_destroy_snapshot_sync_impl(prev, B_FALSE, tx);
 844                 dsl_dataset_rele(prev, FTAG);
 845         }
 846 }
 847
 848 static void
 849 dsl_destroy_head_sync(void *arg, dmu_tx_t *tx)
 850 {
 851         dsl_destroy_head_arg_t *ddha = arg;
 852         dsl_pool_t *dp = dmu_tx_pool(tx);
 853         dsl_dataset_t *ds;
 854
 855         VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
 856         dsl_destroy_head_sync_impl(ds, tx);
 857         dsl_dataset_rele(ds, FTAG);
 858 }
 859
 860 static void
 861 dsl_destroy_head_begin_sync(void *arg, dmu_tx_t *tx)
 862 {
 863         dsl_destroy_head_arg_t *ddha = arg;
 864         dsl_pool_t *dp = dmu_tx_pool(tx);
 865         dsl_dataset_t *ds;
 866
 867         VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
 868
 869         /* Mark it as inconsistent on-disk, in case we crash */
 870         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 871         ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
 872
 873         spa_history_log_internal_ds(ds, "destroy begin", tx, "");
 874         dsl_dataset_rele(ds, FTAG);
 875 }
 876
 877 int
 878 dsl_destroy_head(const char *name)
 879 {
 880         dsl_destroy_head_arg_t ddha;
 881         int error;
 882         spa_t *spa;
 883         boolean_t isenabled;
 884
 885 #ifdef _KERNEL
 886         zfs_destroy_unmount_origin(name);
 887 #endif
 888
 889         error = spa_open(name, &spa, FTAG);
 890         if (error != 0)
 891                 return (error);
 892         isenabled = spa_feature_is_enabled(spa, SPA_FEATURE_ASYNC_DESTROY);
 893         spa_close(spa, FTAG);
 894
 895         ddha.ddha_name = name;
 896
 897         if (!isenabled) {
 898                 objset_t *os;
 899
 900                 error = dsl_sync_task(name, dsl_destroy_head_check,
 901                     dsl_destroy_head_begin_sync, &ddha, 0);
 902                 if (error != 0)
 903                         return (error);
 904
 905                 /*
 906                  * Head deletion is processed in one txg on old pools;
 907                  * remove the objects from open context so that the txg sync
 908                  * is not too long.
 909                  */
 910                 error = dmu_objset_own(name, DMU_OST_ANY, B_FALSE, FTAG, &os);
 911                 if (error == 0) {
 912                         uint64_t prev_snap_txg =
 913                             dmu_objset_ds(os)->ds_phys->ds_prev_snap_txg;
 914                         for (uint64_t obj = 0; error == 0;
 915                             error = dmu_object_next(os, &obj, FALSE,
 916                             prev_snap_txg))
 917                                 (void) dmu_free_long_object(os, obj);
 918                         /* sync out all frees */
 919                         txg_wait_synced(dmu_objset_pool(os), 0);
 920                         dmu_objset_disown(os, FTAG);
 921                 }
 922         }
 923
 924         return (dsl_sync_task(name, dsl_destroy_head_check,
 925             dsl_destroy_head_sync, &ddha, 0));
 926 }
 927
 928 /*
 929  * Note, this function is used as the callback for dmu_objset_find().  We
 930  * always return 0 so that we will continue to find and process
 931  * inconsistent datasets, even if we encounter an error trying to
 932  * process one of them.
 933  */
 934 /* ARGSUSED */
 935 int
 936 dsl_destroy_inconsistent(const char *dsname, void *arg)
 937 {
 938         objset_t *os;
 939
 940         if (dmu_objset_hold(dsname, FTAG, &os) == 0) {
 941                 boolean_t inconsistent = DS_IS_INCONSISTENT(dmu_objset_ds(os));
 942                 dmu_objset_rele(os, FTAG);
 943                 if (inconsistent)
 944                         (void) dsl_destroy_head(dsname);
 945         }
 946         return (0);
 947 }