module/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) 2012, 2014 by Delphix. All rights reserved.
  24  * Copyright (c) 2013 Steven Hartland. All rights reserved.
  25  */
  26
  27 #include <sys/zfs_context.h>
  28 #include <sys/dsl_userhold.h>
  29 #include <sys/dsl_dataset.h>
  30 #include <sys/dsl_synctask.h>
  31 #include <sys/dmu_tx.h>
  32 #include <sys/dsl_pool.h>
  33 #include <sys/dsl_dir.h>
  34 #include <sys/dmu_traverse.h>
  35 #include <sys/dsl_scan.h>
  36 #include <sys/dmu_objset.h>
  37 #include <sys/zap.h>
  38 #include <sys/zfeature.h>
  39 #include <sys/zfs_ioctl.h>
  40 #include <sys/dsl_deleg.h>
  41 #include <sys/dmu_impl.h>
  42
  43 typedef struct dmu_snapshots_destroy_arg {
  44         nvlist_t *dsda_snaps;
  45         nvlist_t *dsda_successful_snaps;
  46         boolean_t dsda_defer;
  47         nvlist_t *dsda_errlist;
  48 } dmu_snapshots_destroy_arg_t;
  49
  50 int
  51 dsl_destroy_snapshot_check_impl(dsl_dataset_t *ds, boolean_t defer)
  52 {
  53         if (!dsl_dataset_is_snapshot(ds))
  54                 return (SET_ERROR(EINVAL));
  55
  56         if (dsl_dataset_long_held(ds))
  57                 return (SET_ERROR(EBUSY));
  58
  59         /*
  60          * Only allow deferred destroy on pools that support it.
  61          * NOTE: deferred destroy is only supported on snapshots.
  62          */
  63         if (defer) {
  64                 if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
  65                     SPA_VERSION_USERREFS)
  66                         return (SET_ERROR(ENOTSUP));
  67                 return (0);
  68         }
  69
  70         /*
  71          * If this snapshot has an elevated user reference count,
  72          * we can't destroy it yet.
  73          */
  74         if (ds->ds_userrefs > 0)
  75                 return (SET_ERROR(EBUSY));
  76
  77         /*
  78          * Can't delete a branch point.
  79          */
  80         if (ds->ds_phys->ds_num_children > 1)
  81                 return (SET_ERROR(EEXIST));
  82
  83         return (0);
  84 }
  85
  86 static int
  87 dsl_destroy_snapshot_check(void *arg, dmu_tx_t *tx)
  88 {
  89         dmu_snapshots_destroy_arg_t *dsda = arg;
  90         dsl_pool_t *dp = dmu_tx_pool(tx);
  91         nvpair_t *pair;
  92         int error = 0;
  93
  94         if (!dmu_tx_is_syncing(tx))
  95                 return (0);
  96
  97         for (pair = nvlist_next_nvpair(dsda->dsda_snaps, NULL);
  98             pair != NULL; pair = nvlist_next_nvpair(dsda->dsda_snaps, pair)) {
  99                 dsl_dataset_t *ds;
 100
 101                 error = dsl_dataset_hold(dp, nvpair_name(pair),
 102                     FTAG, &ds);
 103
 104                 /*
 105                  * If the snapshot does not exist, silently ignore it
 106                  * (it's "already destroyed").
 107                  */
 108                 if (error == ENOENT)
 109                         continue;
 110
 111                 if (error == 0) {
 112                         error = dsl_destroy_snapshot_check_impl(ds,
 113                             dsda->dsda_defer);
 114                         dsl_dataset_rele(ds, FTAG);
 115                 }
 116
 117                 if (error == 0) {
 118                         fnvlist_add_boolean(dsda->dsda_successful_snaps,
 119                             nvpair_name(pair));
 120                 } else {
 121                         fnvlist_add_int32(dsda->dsda_errlist,
 122                             nvpair_name(pair), error);
 123                 }
 124         }
 125
 126         pair = nvlist_next_nvpair(dsda->dsda_errlist, NULL);
 127         if (pair != NULL)
 128                 return (fnvpair_value_int32(pair));
 129
 130         return (0);
 131 }
 132
 133 struct process_old_arg {
 134         dsl_dataset_t *ds;
 135         dsl_dataset_t *ds_prev;
 136         boolean_t after_branch_point;
 137         zio_t *pio;
 138         uint64_t used, comp, uncomp;
 139 };
 140
 141 static int
 142 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
 143 {
 144         struct process_old_arg *poa = arg;
 145         dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
 146
 147         ASSERT(!BP_IS_HOLE(bp));
 148
 149         if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
 150                 dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
 151                 if (poa->ds_prev && !poa->after_branch_point &&
 152                     bp->blk_birth >
 153                     poa->ds_prev->ds_phys->ds_prev_snap_txg) {
 154                         poa->ds_prev->ds_phys->ds_unique_bytes +=
 155                             bp_get_dsize_sync(dp->dp_spa, bp);
 156                 }
 157         } else {
 158                 poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
 159                 poa->comp += BP_GET_PSIZE(bp);
 160                 poa->uncomp += BP_GET_UCSIZE(bp);
 161                 dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
 162         }
 163         return (0);
 164 }
 165
 166 static void
 167 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
 168     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
 169 {
 170         struct process_old_arg poa = { 0 };
 171         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 172         objset_t *mos = dp->dp_meta_objset;
 173         uint64_t deadlist_obj;
 174
 175         ASSERT(ds->ds_deadlist.dl_oldfmt);
 176         ASSERT(ds_next->ds_deadlist.dl_oldfmt);
 177
 178         poa.ds = ds;
 179         poa.ds_prev = ds_prev;
 180         poa.after_branch_point = after_branch_point;
 181         poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 182         VERIFY0(bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
 183             process_old_cb, &poa, tx));
 184         VERIFY0(zio_wait(poa.pio));
 185         ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
 186
 187         /* change snapused */
 188         dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
 189             -poa.used, -poa.comp, -poa.uncomp, tx);
 190
 191         /* swap next's deadlist to our deadlist */
 192         dsl_deadlist_close(&ds->ds_deadlist);
 193         dsl_deadlist_close(&ds_next->ds_deadlist);
 194         deadlist_obj = ds->ds_phys->ds_deadlist_obj;
 195         ds->ds_phys->ds_deadlist_obj = ds_next->ds_phys->ds_deadlist_obj;
 196         ds_next->ds_phys->ds_deadlist_obj = deadlist_obj;
 197         dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
 198         dsl_deadlist_open(&ds_next->ds_deadlist, mos,
 199             ds_next->ds_phys->ds_deadlist_obj);
 200 }
 201
 202 static void
 203 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
 204 {
 205         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 206         zap_cursor_t *zc;
 207         zap_attribute_t *za;
 208
 209         /*
 210          * If it is the old version, dd_clones doesn't exist so we can't
 211          * find the clones, but dsl_deadlist_remove_key() is a no-op so it
 212          * doesn't matter.
 213          */
 214         if (ds->ds_dir->dd_phys->dd_clones == 0)
 215                 return;
 216
 217         zc = kmem_alloc(sizeof (zap_cursor_t), KM_PUSHPAGE);
 218         za = kmem_alloc(sizeof (zap_attribute_t), KM_PUSHPAGE);
 219
 220         for (zap_cursor_init(zc, mos, ds->ds_dir->dd_phys->dd_clones);
 221             zap_cursor_retrieve(zc, za) == 0;
 222             zap_cursor_advance(zc)) {
 223                 dsl_dataset_t *clone;
 224
 225                 VERIFY0(dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
 226                     za->za_first_integer, FTAG, &clone));
 227                 if (clone->ds_dir->dd_origin_txg > mintxg) {
 228                         dsl_deadlist_remove_key(&clone->ds_deadlist,
 229                             mintxg, tx);
 230                         dsl_dataset_remove_clones_key(clone, mintxg, tx);
 231                 }
 232                 dsl_dataset_rele(clone, FTAG);
 233         }
 234         zap_cursor_fini(zc);
 235
 236         kmem_free(za, sizeof (zap_attribute_t));
 237         kmem_free(zc, sizeof (zap_cursor_t));
 238 }
 239
 240 void
 241 dsl_destroy_snapshot_sync_impl(dsl_dataset_t *ds, boolean_t defer, dmu_tx_t *tx)
 242 {
 243 #ifdef ZFS_DEBUG
 244         int err;
 245 #endif
 246         int after_branch_point = FALSE;
 247         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 248         objset_t *mos = dp->dp_meta_objset;
 249         dsl_dataset_t *ds_prev = NULL;
 250         uint64_t obj, old_unique, used = 0, comp = 0, uncomp = 0;
 251         dsl_dataset_t *ds_next, *ds_head, *hds;
 252
 253
 254         ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
 255         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
 256         ASSERT(refcount_is_zero(&ds->ds_longholds));
 257
 258         if (defer &&
 259             (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1)) {
 260                 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
 261                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
 262                 ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
 263                 spa_history_log_internal_ds(ds, "defer_destroy", tx, "");
 264                 return;
 265         }
 266
 267         ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 268
 269         /* We need to log before removing it from the namespace. */
 270         spa_history_log_internal_ds(ds, "destroy", tx, "");
 271
 272         dsl_scan_ds_destroyed(ds, tx);
 273
 274         obj = ds->ds_object;
 275
 276         if (ds->ds_phys->ds_prev_snap_obj != 0) {
 277                 ASSERT3P(ds->ds_prev, ==, NULL);
 278                 VERIFY0(dsl_dataset_hold_obj(dp,
 279                     ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
 280                 after_branch_point =
 281                     (ds_prev->ds_phys->ds_next_snap_obj != obj);
 282
 283                 dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
 284                 if (after_branch_point &&
 285                     ds_prev->ds_phys->ds_next_clones_obj != 0) {
 286                         dsl_dataset_remove_from_next_clones(ds_prev, obj, tx);
 287                         if (ds->ds_phys->ds_next_snap_obj != 0) {
 288                                 VERIFY0(zap_add_int(mos,
 289                                     ds_prev->ds_phys->ds_next_clones_obj,
 290                                     ds->ds_phys->ds_next_snap_obj, tx));
 291                         }
 292                 }
 293                 if (!after_branch_point) {
 294                         ds_prev->ds_phys->ds_next_snap_obj =
 295                             ds->ds_phys->ds_next_snap_obj;
 296                 }
 297         }
 298
 299         VERIFY0(dsl_dataset_hold_obj(dp,
 300             ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
 301         ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
 302
 303         old_unique = ds_next->ds_phys->ds_unique_bytes;
 304
 305         dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
 306         ds_next->ds_phys->ds_prev_snap_obj =
 307             ds->ds_phys->ds_prev_snap_obj;
 308         ds_next->ds_phys->ds_prev_snap_txg =
 309             ds->ds_phys->ds_prev_snap_txg;
 310         ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
 311             ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
 312
 313         if (ds_next->ds_deadlist.dl_oldfmt) {
 314                 process_old_deadlist(ds, ds_prev, ds_next,
 315                     after_branch_point, tx);
 316         } else {
 317                 /* Adjust prev's unique space. */
 318                 if (ds_prev && !after_branch_point) {
 319                         dsl_deadlist_space_range(&ds_next->ds_deadlist,
 320                             ds_prev->ds_phys->ds_prev_snap_txg,
 321                             ds->ds_phys->ds_prev_snap_txg,
 322                             &used, &comp, &uncomp);
 323                         ds_prev->ds_phys->ds_unique_bytes += used;
 324                 }
 325
 326                 /* Adjust snapused. */
 327                 dsl_deadlist_space_range(&ds_next->ds_deadlist,
 328                     ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
 329                     &used, &comp, &uncomp);
 330                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
 331                     -used, -comp, -uncomp, tx);
 332
 333                 /* Move blocks to be freed to pool's free list. */
 334                 dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
 335                     &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
 336                     tx);
 337                 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
 338                     DD_USED_HEAD, used, comp, uncomp, tx);
 339
 340                 /* Merge our deadlist into next's and free it. */
 341                 dsl_deadlist_merge(&ds_next->ds_deadlist,
 342                     ds->ds_phys->ds_deadlist_obj, tx);
 343         }
 344         dsl_deadlist_close(&ds->ds_deadlist);
 345         dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
 346         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 347         ds->ds_phys->ds_deadlist_obj = 0;
 348
 349         /* Collapse range in clone heads */
 350         dsl_dataset_remove_clones_key(ds,
 351             ds->ds_phys->ds_creation_txg, tx);
 352
 353         if (dsl_dataset_is_snapshot(ds_next)) {
 354                 dsl_dataset_t *ds_nextnext;
 355
 356                 /*
 357                  * Update next's unique to include blocks which
 358                  * were previously shared by only this snapshot
 359                  * and it.  Those blocks will be born after the
 360                  * prev snap and before this snap, and will have
 361                  * died after the next snap and before the one
 362                  * after that (ie. be on the snap after next's
 363                  * deadlist).
 364                  */
 365                 VERIFY0(dsl_dataset_hold_obj(dp,
 366                     ds_next->ds_phys->ds_next_snap_obj, FTAG, &ds_nextnext));
 367                 dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
 368                     ds->ds_phys->ds_prev_snap_txg,
 369                     ds->ds_phys->ds_creation_txg,
 370                     &used, &comp, &uncomp);
 371                 ds_next->ds_phys->ds_unique_bytes += used;
 372                 dsl_dataset_rele(ds_nextnext, FTAG);
 373                 ASSERT3P(ds_next->ds_prev, ==, NULL);
 374
 375                 /* Collapse range in this head. */
 376                 VERIFY0(dsl_dataset_hold_obj(dp,
 377                     ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &hds));
 378                 dsl_deadlist_remove_key(&hds->ds_deadlist,
 379                     ds->ds_phys->ds_creation_txg, tx);
 380                 dsl_dataset_rele(hds, FTAG);
 381
 382         } else {
 383                 ASSERT3P(ds_next->ds_prev, ==, ds);
 384                 dsl_dataset_rele(ds_next->ds_prev, ds_next);
 385                 ds_next->ds_prev = NULL;
 386                 if (ds_prev) {
 387                         VERIFY0(dsl_dataset_hold_obj(dp,
 388                             ds->ds_phys->ds_prev_snap_obj,
 389                             ds_next, &ds_next->ds_prev));
 390                 }
 391
 392                 dsl_dataset_recalc_head_uniq(ds_next);
 393
 394                 /*
 395                  * Reduce the amount of our unconsumed refreservation
 396                  * being charged to our parent by the amount of
 397                  * new unique data we have gained.
 398                  */
 399                 if (old_unique < ds_next->ds_reserved) {
 400                         int64_t mrsdelta;
 401                         uint64_t new_unique =
 402                             ds_next->ds_phys->ds_unique_bytes;
 403
 404                         ASSERT(old_unique <= new_unique);
 405                         mrsdelta = MIN(new_unique - old_unique,
 406                             ds_next->ds_reserved - old_unique);
 407                         dsl_dir_diduse_space(ds->ds_dir,
 408                             DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
 409                 }
 410         }
 411         dsl_dataset_rele(ds_next, FTAG);
 412
 413         /*
 414          * This must be done after the dsl_traverse(), because it will
 415          * re-open the objset.
 416          */
 417         if (ds->ds_objset) {
 418                 dmu_objset_evict(ds->ds_objset);
 419                 ds->ds_objset = NULL;
 420         }
 421
 422         /* remove from snapshot namespace */
 423         ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
 424         VERIFY0(dsl_dataset_hold_obj(dp,
 425             ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
 426         VERIFY0(dsl_dataset_get_snapname(ds));
 427 #ifdef ZFS_DEBUG
 428         {
 429                 uint64_t val;
 430
 431                 err = dsl_dataset_snap_lookup(ds_head,
 432                     ds->ds_snapname, &val);
 433                 ASSERT0(err);
 434                 ASSERT3U(val, ==, obj);
 435         }
 436 #endif
 437         VERIFY0(dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx));
 438         dsl_dataset_rele(ds_head, FTAG);
 439
 440         if (ds_prev != NULL)
 441                 dsl_dataset_rele(ds_prev, FTAG);
 442
 443         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
 444
 445         if (ds->ds_phys->ds_next_clones_obj != 0) {
 446                 ASSERTV(uint64_t count);
 447                 ASSERT0(zap_count(mos,
 448                     ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
 449                 VERIFY0(dmu_object_free(mos,
 450                     ds->ds_phys->ds_next_clones_obj, tx));
 451         }
 452         if (ds->ds_phys->ds_props_obj != 0)
 453                 VERIFY0(zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
 454         if (ds->ds_phys->ds_userrefs_obj != 0)
 455                 VERIFY0(zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
 456         dsl_dir_rele(ds->ds_dir, ds);
 457         ds->ds_dir = NULL;
 458         dmu_object_free_zapified(mos, obj, tx);
 459 }
 460
 461 static void
 462 dsl_destroy_snapshot_sync(void *arg, dmu_tx_t *tx)
 463 {
 464         dmu_snapshots_destroy_arg_t *dsda = arg;
 465         dsl_pool_t *dp = dmu_tx_pool(tx);
 466         nvpair_t *pair;
 467
 468         for (pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, NULL);
 469             pair != NULL;
 470             pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, pair)) {
 471                 dsl_dataset_t *ds;
 472
 473                 VERIFY0(dsl_dataset_hold(dp, nvpair_name(pair), FTAG, &ds));
 474
 475                 dsl_destroy_snapshot_sync_impl(ds, dsda->dsda_defer, tx);
 476                 dsl_dataset_rele(ds, FTAG);
 477         }
 478 }
 479
 480 /*
 481  * The semantics of this function are described in the comment above
 482  * lzc_destroy_snaps().  To summarize:
 483  *
 484  * The snapshots must all be in the same pool.
 485  *
 486  * Snapshots that don't exist will be silently ignored (considered to be
 487  * "already deleted").
 488  *
 489  * On success, all snaps will be destroyed and this will return 0.
 490  * On failure, no snaps will be destroyed, the errlist will be filled in,
 491  * and this will return an errno.
 492  */
 493 int
 494 dsl_destroy_snapshots_nvl(nvlist_t *snaps, boolean_t defer,
 495     nvlist_t *errlist)
 496 {
 497         dmu_snapshots_destroy_arg_t dsda;
 498         int error;
 499         nvpair_t *pair;
 500
 501         pair = nvlist_next_nvpair(snaps, NULL);
 502         if (pair == NULL)
 503                 return (0);
 504
 505         dsda.dsda_snaps = snaps;
 506         VERIFY0(nvlist_alloc(&dsda.dsda_successful_snaps,
 507             NV_UNIQUE_NAME, KM_PUSHPAGE));
 508         dsda.dsda_defer = defer;
 509         dsda.dsda_errlist = errlist;
 510
 511         error = dsl_sync_task(nvpair_name(pair),
 512             dsl_destroy_snapshot_check, dsl_destroy_snapshot_sync,
 513             &dsda, 0);
 514         fnvlist_free(dsda.dsda_successful_snaps);
 515
 516         return (error);
 517 }
 518
 519 int
 520 dsl_destroy_snapshot(const char *name, boolean_t defer)
 521 {
 522         int error;
 523         nvlist_t *nvl;
 524         nvlist_t *errlist;
 525
 526         VERIFY0(nvlist_alloc(&nvl, NV_UNIQUE_NAME, KM_PUSHPAGE));
 527         VERIFY0(nvlist_alloc(&errlist, NV_UNIQUE_NAME, KM_PUSHPAGE));
 528
 529         fnvlist_add_boolean(nvl, name);
 530         error = dsl_destroy_snapshots_nvl(nvl, defer, errlist);
 531         fnvlist_free(errlist);
 532         fnvlist_free(nvl);
 533         return (error);
 534 }
 535
 536 struct killarg {
 537         dsl_dataset_t *ds;
 538         dmu_tx_t *tx;
 539 };
 540
 541 /* ARGSUSED */
 542 static int
 543 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 544     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
 545 {
 546         struct killarg *ka = arg;
 547         dmu_tx_t *tx = ka->tx;
 548
 549         if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
 550                 return (0);
 551
 552         if (zb->zb_level == ZB_ZIL_LEVEL) {
 553                 ASSERT(zilog != NULL);
 554                 /*
 555                  * It's a block in the intent log.  It has no
 556                  * accounting, so just free it.
 557                  */
 558                 dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
 559         } else {
 560                 ASSERT(zilog == NULL);
 561                 ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
 562                 (void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
 563         }
 564
 565         return (0);
 566 }
 567
 568 static void
 569 old_synchronous_dataset_destroy(dsl_dataset_t *ds, dmu_tx_t *tx)
 570 {
 571         struct killarg ka;
 572
 573         /*
 574          * Free everything that we point to (that's born after
 575          * the previous snapshot, if we are a clone)
 576          *
 577          * NB: this should be very quick, because we already
 578          * freed all the objects in open context.
 579          */
 580         ka.ds = ds;
 581         ka.tx = tx;
 582         VERIFY0(traverse_dataset(ds,
 583             ds->ds_phys->ds_prev_snap_txg, TRAVERSE_POST,
 584             kill_blkptr, &ka));
 585         ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) || ds->ds_phys->ds_unique_bytes == 0);
 586 }
 587
 588 typedef struct dsl_destroy_head_arg {
 589         const char *ddha_name;
 590 } dsl_destroy_head_arg_t;
 591
 592 int
 593 dsl_destroy_head_check_impl(dsl_dataset_t *ds, int expected_holds)
 594 {
 595         int error;
 596         uint64_t count;
 597         objset_t *mos;
 598
 599         ASSERT(!dsl_dataset_is_snapshot(ds));
 600         if (dsl_dataset_is_snapshot(ds))
 601                 return (SET_ERROR(EINVAL));
 602
 603         if (refcount_count(&ds->ds_longholds) != expected_holds)
 604                 return (SET_ERROR(EBUSY));
 605
 606         mos = ds->ds_dir->dd_pool->dp_meta_objset;
 607
 608         /*
 609          * Can't delete a head dataset if there are snapshots of it.
 610          * (Except if the only snapshots are from the branch we cloned
 611          * from.)
 612          */
 613         if (ds->ds_prev != NULL &&
 614             ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
 615                 return (SET_ERROR(EBUSY));
 616
 617         /*
 618          * Can't delete if there are children of this fs.
 619          */
 620         error = zap_count(mos,
 621             ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
 622         if (error != 0)
 623                 return (error);
 624         if (count != 0)
 625                 return (SET_ERROR(EEXIST));
 626
 627         if (dsl_dir_is_clone(ds->ds_dir) && DS_IS_DEFER_DESTROY(ds->ds_prev) &&
 628             ds->ds_prev->ds_phys->ds_num_children == 2 &&
 629             ds->ds_prev->ds_userrefs == 0) {
 630                 /* We need to remove the origin snapshot as well. */
 631                 if (!refcount_is_zero(&ds->ds_prev->ds_longholds))
 632                         return (SET_ERROR(EBUSY));
 633         }
 634         return (0);
 635 }
 636
 637 static int
 638 dsl_destroy_head_check(void *arg, dmu_tx_t *tx)
 639 {
 640         dsl_destroy_head_arg_t *ddha = arg;
 641         dsl_pool_t *dp = dmu_tx_pool(tx);
 642         dsl_dataset_t *ds;
 643         int error;
 644
 645         error = dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds);
 646         if (error != 0)
 647                 return (error);
 648
 649         error = dsl_destroy_head_check_impl(ds, 0);
 650         dsl_dataset_rele(ds, FTAG);
 651         return (error);
 652 }
 653
 654 static void
 655 dsl_dir_destroy_sync(uint64_t ddobj, dmu_tx_t *tx)
 656 {
 657         dsl_dir_t *dd;
 658         dsl_pool_t *dp = dmu_tx_pool(tx);
 659         objset_t *mos = dp->dp_meta_objset;
 660         dd_used_t t;
 661
 662         ASSERT(RRW_WRITE_HELD(&dmu_tx_pool(tx)->dp_config_rwlock));
 663
 664         VERIFY0(dsl_dir_hold_obj(dp, ddobj, NULL, FTAG, &dd));
 665
 666         ASSERT0(dd->dd_phys->dd_head_dataset_obj);
 667
 668         /*
 669          * Remove our reservation. The impl() routine avoids setting the
 670          * actual property, which would require the (already destroyed) ds.
 671          */
 672         dsl_dir_set_reservation_sync_impl(dd, 0, tx);
 673
 674         ASSERT0(dd->dd_phys->dd_used_bytes);
 675         ASSERT0(dd->dd_phys->dd_reserved);
 676         for (t = 0; t < DD_USED_NUM; t++)
 677                 ASSERT0(dd->dd_phys->dd_used_breakdown[t]);
 678
 679         VERIFY0(zap_destroy(mos, dd->dd_phys->dd_child_dir_zapobj, tx));
 680         VERIFY0(zap_destroy(mos, dd->dd_phys->dd_props_zapobj, tx));
 681         VERIFY0(dsl_deleg_destroy(mos, dd->dd_phys->dd_deleg_zapobj, tx));
 682         VERIFY0(zap_remove(mos,
 683             dd->dd_parent->dd_phys->dd_child_dir_zapobj, dd->dd_myname, tx));
 684
 685         dsl_dir_rele(dd, FTAG);
 686         dmu_object_free_zapified(mos, ddobj, tx);
 687 }
 688
 689 void
 690 dsl_destroy_head_sync_impl(dsl_dataset_t *ds, dmu_tx_t *tx)
 691 {
 692         dsl_pool_t *dp = dmu_tx_pool(tx);
 693         objset_t *mos = dp->dp_meta_objset;
 694         uint64_t obj, ddobj, prevobj = 0;
 695         boolean_t rmorigin;
 696         objset_t *os;
 697
 698         ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 699         ASSERT(ds->ds_prev == NULL ||
 700             ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
 701         ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
 702         ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
 703
 704         /* We need to log before removing it from the namespace. */
 705         spa_history_log_internal_ds(ds, "destroy", tx, "");
 706
 707         rmorigin = (dsl_dir_is_clone(ds->ds_dir) &&
 708             DS_IS_DEFER_DESTROY(ds->ds_prev) &&
 709             ds->ds_prev->ds_phys->ds_num_children == 2 &&
 710             ds->ds_prev->ds_userrefs == 0);
 711
 712         /* Remove our reservation. */
 713         if (ds->ds_reserved != 0) {
 714                 dsl_dataset_set_refreservation_sync_impl(ds,
 715                     (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
 716                     0, tx);
 717                 ASSERT0(ds->ds_reserved);
 718         }
 719
 720         dsl_scan_ds_destroyed(ds, tx);
 721
 722         obj = ds->ds_object;
 723
 724         if (ds->ds_phys->ds_prev_snap_obj != 0) {
 725                 /* This is a clone */
 726                 ASSERT(ds->ds_prev != NULL);
 727                 ASSERT3U(ds->ds_prev->ds_phys->ds_next_snap_obj, !=, obj);
 728                 ASSERT0(ds->ds_phys->ds_next_snap_obj);
 729
 730                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
 731                 if (ds->ds_prev->ds_phys->ds_next_clones_obj != 0) {
 732                         dsl_dataset_remove_from_next_clones(ds->ds_prev,
 733                             obj, tx);
 734                 }
 735
 736                 ASSERT3U(ds->ds_prev->ds_phys->ds_num_children, >, 1);
 737                 ds->ds_prev->ds_phys->ds_num_children--;
 738         }
 739
 740         /*
 741          * Destroy the deadlist.  Unless it's a clone, the
 742          * deadlist should be empty.  (If it's a clone, it's
 743          * safe to ignore the deadlist contents.)
 744          */
 745         dsl_deadlist_close(&ds->ds_deadlist);
 746         dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
 747         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 748         ds->ds_phys->ds_deadlist_obj = 0;
 749
 750         VERIFY0(dmu_objset_from_ds(ds, &os));
 751
 752         if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
 753                 old_synchronous_dataset_destroy(ds, tx);
 754         } else {
 755                 /*
 756                  * Move the bptree into the pool's list of trees to
 757                  * clean up and update space accounting information.
 758                  */
 759                 uint64_t used, comp, uncomp;
 760
 761                 zil_destroy_sync(dmu_objset_zil(os), tx);
 762
 763                 if (!spa_feature_is_active(dp->dp_spa,
 764                     SPA_FEATURE_ASYNC_DESTROY)) {
 765                         dsl_scan_t *scn = dp->dp_scan;
 766                         spa_feature_incr(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY,
 767                             tx);
 768                         dp->dp_bptree_obj = bptree_alloc(mos, tx);
 769                         VERIFY0(zap_add(mos,
 770                             DMU_POOL_DIRECTORY_OBJECT,
 771                             DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
 772                             &dp->dp_bptree_obj, tx));
 773                         ASSERT(!scn->scn_async_destroying);
 774                         scn->scn_async_destroying = B_TRUE;
 775                 }
 776
 777                 used = ds->ds_dir->dd_phys->dd_used_bytes;
 778                 comp = ds->ds_dir->dd_phys->dd_compressed_bytes;
 779                 uncomp = ds->ds_dir->dd_phys->dd_uncompressed_bytes;
 780
 781                 ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
 782                     ds->ds_phys->ds_unique_bytes == used);
 783
 784                 bptree_add(mos, dp->dp_bptree_obj,
 785                     &ds->ds_phys->ds_bp, ds->ds_phys->ds_prev_snap_txg,
 786                     used, comp, uncomp, tx);
 787                 dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
 788                     -used, -comp, -uncomp, tx);
 789                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
 790                     used, comp, uncomp, tx);
 791         }
 792
 793         if (ds->ds_prev != NULL) {
 794                 if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
 795                         VERIFY0(zap_remove_int(mos,
 796                             ds->ds_prev->ds_dir->dd_phys->dd_clones,
 797                             ds->ds_object, tx));
 798                 }
 799                 prevobj = ds->ds_prev->ds_object;
 800                 dsl_dataset_rele(ds->ds_prev, ds);
 801                 ds->ds_prev = NULL;
 802         }
 803
 804         /*
 805          * This must be done after the dsl_traverse(), because it will
 806          * re-open the objset.
 807          */
 808         if (ds->ds_objset) {
 809                 dmu_objset_evict(ds->ds_objset);
 810                 ds->ds_objset = NULL;
 811         }
 812
 813         /* Erase the link in the dir */
 814         dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
 815         ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
 816         ddobj = ds->ds_dir->dd_object;
 817         ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
 818         VERIFY0(zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx));
 819
 820         if (ds->ds_bookmarks != 0) {
 821                 VERIFY0(zap_destroy(mos,
 822                     ds->ds_bookmarks, tx));
 823                 spa_feature_decr(dp->dp_spa, SPA_FEATURE_BOOKMARKS, tx);
 824         }
 825
 826         spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
 827
 828         ASSERT0(ds->ds_phys->ds_next_clones_obj);
 829         ASSERT0(ds->ds_phys->ds_props_obj);
 830         ASSERT0(ds->ds_phys->ds_userrefs_obj);
 831         dsl_dir_rele(ds->ds_dir, ds);
 832         ds->ds_dir = NULL;
 833         dmu_object_free_zapified(mos, obj, tx);
 834
 835         dsl_dir_destroy_sync(ddobj, tx);
 836
 837         if (rmorigin) {
 838                 dsl_dataset_t *prev;
 839                 VERIFY0(dsl_dataset_hold_obj(dp, prevobj, FTAG, &prev));
 840                 dsl_destroy_snapshot_sync_impl(prev, B_FALSE, tx);
 841                 dsl_dataset_rele(prev, FTAG);
 842         }
 843 }
 844
 845 static void
 846 dsl_destroy_head_sync(void *arg, dmu_tx_t *tx)
 847 {
 848         dsl_destroy_head_arg_t *ddha = arg;
 849         dsl_pool_t *dp = dmu_tx_pool(tx);
 850         dsl_dataset_t *ds;
 851
 852         VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
 853         dsl_destroy_head_sync_impl(ds, tx);
 854         dsl_dataset_rele(ds, FTAG);
 855 }
 856
 857 static void
 858 dsl_destroy_head_begin_sync(void *arg, dmu_tx_t *tx)
 859 {
 860         dsl_destroy_head_arg_t *ddha = arg;
 861         dsl_pool_t *dp = dmu_tx_pool(tx);
 862         dsl_dataset_t *ds;
 863
 864         VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
 865
 866         /* Mark it as inconsistent on-disk, in case we crash */
 867         dmu_buf_will_dirty(ds->ds_dbuf, tx);
 868         ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
 869
 870         spa_history_log_internal_ds(ds, "destroy begin", tx, "");
 871         dsl_dataset_rele(ds, FTAG);
 872 }
 873
 874 int
 875 dsl_destroy_head(const char *name)
 876 {
 877         dsl_destroy_head_arg_t ddha;
 878         int error;
 879         spa_t *spa;
 880         boolean_t isenabled;
 881
 882 #ifdef _KERNEL
 883         zfs_destroy_unmount_origin(name);
 884 #endif
 885
 886         error = spa_open(name, &spa, FTAG);
 887         if (error != 0)
 888                 return (error);
 889         isenabled = spa_feature_is_enabled(spa, SPA_FEATURE_ASYNC_DESTROY);
 890         spa_close(spa, FTAG);
 891
 892         ddha.ddha_name = name;
 893
 894         if (!isenabled) {
 895                 objset_t *os;
 896
 897                 error = dsl_sync_task(name, dsl_destroy_head_check,
 898                     dsl_destroy_head_begin_sync, &ddha, 0);
 899                 if (error != 0)
 900                         return (error);
 901
 902                 /*
 903                  * Head deletion is processed in one txg on old pools;
 904                  * remove the objects from open context so that the txg sync
 905                  * is not too long.
 906                  */
 907                 error = dmu_objset_own(name, DMU_OST_ANY, B_FALSE, FTAG, &os);
 908                 if (error == 0) {
 909                         uint64_t obj;
 910                         uint64_t prev_snap_txg =
 911                             dmu_objset_ds(os)->ds_phys->ds_prev_snap_txg;
 912                         for (obj = 0; error == 0;
 913                             error = dmu_object_next(os, &obj, FALSE,
 914                             prev_snap_txg))
 915                                 (void) dmu_free_long_object(os, obj);
 916                         /* sync out all frees */
 917                         txg_wait_synced(dmu_objset_pool(os), 0);
 918                         dmu_objset_disown(os, FTAG);
 919                 }
 920         }
 921
 922         return (dsl_sync_task(name, dsl_destroy_head_check,
 923             dsl_destroy_head_sync, &ddha, 0));
 924 }
 925
 926 /*
 927  * Note, this function is used as the callback for dmu_objset_find().  We
 928  * always return 0 so that we will continue to find and process
 929  * inconsistent datasets, even if we encounter an error trying to
 930  * process one of them.
 931  */
 932 /* ARGSUSED */
 933 int
 934 dsl_destroy_inconsistent(const char *dsname, void *arg)
 935 {
 936         objset_t *os;
 937
 938         if (dmu_objset_hold(dsname, FTAG, &os) == 0) {
 939                 boolean_t inconsistent = DS_IS_INCONSISTENT(dmu_objset_ds(os));
 940                 dmu_objset_rele(os, FTAG);
 941                 if (inconsistent)
 942                         (void) dsl_destroy_head(dsname);
 943         }
 944         return (0);
 945 }
 946
 947
 948 #if defined(_KERNEL) && defined(HAVE_SPL)
 949 EXPORT_SYMBOL(dsl_destroy_head);
 950 EXPORT_SYMBOL(dsl_destroy_head_sync_impl);
 951 EXPORT_SYMBOL(dsl_dataset_user_hold_check_one);
 952 EXPORT_SYMBOL(dsl_destroy_snapshot_sync_impl);
 953 EXPORT_SYMBOL(dsl_destroy_inconsistent);
 954 EXPORT_SYMBOL(dsl_dataset_user_release_tmp);
 955 EXPORT_SYMBOL(dsl_destroy_head_check_impl);
 956 #endif