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