xref: /illumos-gate/usr/src/uts/common/fs/zfs/dsl_destroy.c (revision 09409df0fea2444decd38bc1b87ab7a80d3251e5)
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 /*
50  * ds must be owned.
51  */
52 static int
53 dsl_destroy_snapshot_check_impl(dsl_dataset_t *ds, boolean_t defer)
54 {
55 	if (!dsl_dataset_is_snapshot(ds))
56 		return (SET_ERROR(EINVAL));
57 
58 	if (dsl_dataset_long_held(ds))
59 		return (SET_ERROR(EBUSY));
60 
61 	/*
62 	 * Only allow deferred destroy on pools that support it.
63 	 * NOTE: deferred destroy is only supported on snapshots.
64 	 */
65 	if (defer) {
66 		if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
67 		    SPA_VERSION_USERREFS)
68 			return (SET_ERROR(ENOTSUP));
69 		return (0);
70 	}
71 
72 	/*
73 	 * If this snapshot has an elevated user reference count,
74 	 * we can't destroy it yet.
75 	 */
76 	if (ds->ds_userrefs > 0)
77 		return (SET_ERROR(EBUSY));
78 
79 	/*
80 	 * Can't delete a branch point.
81 	 */
82 	if (ds->ds_phys->ds_num_children > 1)
83 		return (SET_ERROR(EEXIST));
84 
85 	return (0);
86 }
87 
88 static int
89 dsl_destroy_snapshot_check(void *arg, dmu_tx_t *tx)
90 {
91 	dmu_snapshots_destroy_arg_t *dsda = arg;
92 	dsl_pool_t *dp = dmu_tx_pool(tx);
93 	nvpair_t *pair;
94 	int error = 0;
95 
96 	if (!dmu_tx_is_syncing(tx))
97 		return (0);
98 
99 	for (pair = nvlist_next_nvpair(dsda->dsda_snaps, NULL);
100 	    pair != NULL; pair = nvlist_next_nvpair(dsda->dsda_snaps, pair)) {
101 		dsl_dataset_t *ds;
102 
103 		error = dsl_dataset_hold(dp, nvpair_name(pair),
104 		    FTAG, &ds);
105 
106 		/*
107 		 * If the snapshot does not exist, silently ignore it
108 		 * (it's "already destroyed").
109 		 */
110 		if (error == ENOENT)
111 			continue;
112 
113 		if (error == 0) {
114 			error = dsl_destroy_snapshot_check_impl(ds,
115 			    dsda->dsda_defer);
116 			dsl_dataset_rele(ds, FTAG);
117 		}
118 
119 		if (error == 0) {
120 			fnvlist_add_boolean(dsda->dsda_successful_snaps,
121 			    nvpair_name(pair));
122 		} else {
123 			fnvlist_add_int32(dsda->dsda_errlist,
124 			    nvpair_name(pair), error);
125 		}
126 	}
127 
128 	pair = nvlist_next_nvpair(dsda->dsda_errlist, NULL);
129 	if (pair != NULL)
130 		return (fnvpair_value_int32(pair));
131 
132 	return (0);
133 }
134 
135 struct process_old_arg {
136 	dsl_dataset_t *ds;
137 	dsl_dataset_t *ds_prev;
138 	boolean_t after_branch_point;
139 	zio_t *pio;
140 	uint64_t used, comp, uncomp;
141 };
142 
143 static int
144 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
145 {
146 	struct process_old_arg *poa = arg;
147 	dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
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 	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 
234 void
235 dsl_destroy_snapshot_sync_impl(dsl_dataset_t *ds, boolean_t defer, dmu_tx_t *tx)
236 {
237 	int err;
238 	int after_branch_point = FALSE;
239 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
240 	objset_t *mos = dp->dp_meta_objset;
241 	dsl_dataset_t *ds_prev = NULL;
242 	uint64_t obj;
243 
244 	ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
245 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
246 	ASSERT(refcount_is_zero(&ds->ds_longholds));
247 
248 	if (defer &&
249 	    (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1)) {
250 		ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
251 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
252 		ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
253 		spa_history_log_internal_ds(ds, "defer_destroy", tx, "");
254 		return;
255 	}
256 
257 	ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
258 
259 	/* We need to log before removing it from the namespace. */
260 	spa_history_log_internal_ds(ds, "destroy", tx, "");
261 
262 	dsl_scan_ds_destroyed(ds, tx);
263 
264 	obj = ds->ds_object;
265 
266 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
267 		ASSERT3P(ds->ds_prev, ==, NULL);
268 		VERIFY0(dsl_dataset_hold_obj(dp,
269 		    ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
270 		after_branch_point =
271 		    (ds_prev->ds_phys->ds_next_snap_obj != obj);
272 
273 		dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
274 		if (after_branch_point &&
275 		    ds_prev->ds_phys->ds_next_clones_obj != 0) {
276 			dsl_dataset_remove_from_next_clones(ds_prev, obj, tx);
277 			if (ds->ds_phys->ds_next_snap_obj != 0) {
278 				VERIFY0(zap_add_int(mos,
279 				    ds_prev->ds_phys->ds_next_clones_obj,
280 				    ds->ds_phys->ds_next_snap_obj, tx));
281 			}
282 		}
283 		if (!after_branch_point) {
284 			ds_prev->ds_phys->ds_next_snap_obj =
285 			    ds->ds_phys->ds_next_snap_obj;
286 		}
287 	}
288 
289 	dsl_dataset_t *ds_next;
290 	uint64_t old_unique;
291 	uint64_t used = 0, comp = 0, uncomp = 0;
292 
293 	VERIFY0(dsl_dataset_hold_obj(dp,
294 	    ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
295 	ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
296 
297 	old_unique = ds_next->ds_phys->ds_unique_bytes;
298 
299 	dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
300 	ds_next->ds_phys->ds_prev_snap_obj =
301 	    ds->ds_phys->ds_prev_snap_obj;
302 	ds_next->ds_phys->ds_prev_snap_txg =
303 	    ds->ds_phys->ds_prev_snap_txg;
304 	ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
305 	    ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
306 
307 	if (ds_next->ds_deadlist.dl_oldfmt) {
308 		process_old_deadlist(ds, ds_prev, ds_next,
309 		    after_branch_point, tx);
310 	} else {
311 		/* Adjust prev's unique space. */
312 		if (ds_prev && !after_branch_point) {
313 			dsl_deadlist_space_range(&ds_next->ds_deadlist,
314 			    ds_prev->ds_phys->ds_prev_snap_txg,
315 			    ds->ds_phys->ds_prev_snap_txg,
316 			    &used, &comp, &uncomp);
317 			ds_prev->ds_phys->ds_unique_bytes += used;
318 		}
319 
320 		/* Adjust snapused. */
321 		dsl_deadlist_space_range(&ds_next->ds_deadlist,
322 		    ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
323 		    &used, &comp, &uncomp);
324 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
325 		    -used, -comp, -uncomp, tx);
326 
327 		/* Move blocks to be freed to pool's free list. */
328 		dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
329 		    &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
330 		    tx);
331 		dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
332 		    DD_USED_HEAD, used, comp, uncomp, tx);
333 
334 		/* Merge our deadlist into next's and free it. */
335 		dsl_deadlist_merge(&ds_next->ds_deadlist,
336 		    ds->ds_phys->ds_deadlist_obj, tx);
337 	}
338 	dsl_deadlist_close(&ds->ds_deadlist);
339 	dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
340 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
341 	ds->ds_phys->ds_deadlist_obj = 0;
342 
343 	/* Collapse range in clone heads */
344 	dsl_dataset_remove_clones_key(ds,
345 	    ds->ds_phys->ds_creation_txg, tx);
346 
347 	if (dsl_dataset_is_snapshot(ds_next)) {
348 		dsl_dataset_t *ds_nextnext;
349 
350 		/*
351 		 * Update next's unique to include blocks which
352 		 * were previously shared by only this snapshot
353 		 * and it.  Those blocks will be born after the
354 		 * prev snap and before this snap, and will have
355 		 * died after the next snap and before the one
356 		 * after that (ie. be on the snap after next's
357 		 * deadlist).
358 		 */
359 		VERIFY0(dsl_dataset_hold_obj(dp,
360 		    ds_next->ds_phys->ds_next_snap_obj, FTAG, &ds_nextnext));
361 		dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
362 		    ds->ds_phys->ds_prev_snap_txg,
363 		    ds->ds_phys->ds_creation_txg,
364 		    &used, &comp, &uncomp);
365 		ds_next->ds_phys->ds_unique_bytes += used;
366 		dsl_dataset_rele(ds_nextnext, FTAG);
367 		ASSERT3P(ds_next->ds_prev, ==, NULL);
368 
369 		/* Collapse range in this head. */
370 		dsl_dataset_t *hds;
371 		VERIFY0(dsl_dataset_hold_obj(dp,
372 		    ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &hds));
373 		dsl_deadlist_remove_key(&hds->ds_deadlist,
374 		    ds->ds_phys->ds_creation_txg, tx);
375 		dsl_dataset_rele(hds, FTAG);
376 
377 	} else {
378 		ASSERT3P(ds_next->ds_prev, ==, ds);
379 		dsl_dataset_rele(ds_next->ds_prev, ds_next);
380 		ds_next->ds_prev = NULL;
381 		if (ds_prev) {
382 			VERIFY0(dsl_dataset_hold_obj(dp,
383 			    ds->ds_phys->ds_prev_snap_obj,
384 			    ds_next, &ds_next->ds_prev));
385 		}
386 
387 		dsl_dataset_recalc_head_uniq(ds_next);
388 
389 		/*
390 		 * Reduce the amount of our unconsumed refreservation
391 		 * being charged to our parent by the amount of
392 		 * new unique data we have gained.
393 		 */
394 		if (old_unique < ds_next->ds_reserved) {
395 			int64_t mrsdelta;
396 			uint64_t new_unique =
397 			    ds_next->ds_phys->ds_unique_bytes;
398 
399 			ASSERT(old_unique <= new_unique);
400 			mrsdelta = MIN(new_unique - old_unique,
401 			    ds_next->ds_reserved - old_unique);
402 			dsl_dir_diduse_space(ds->ds_dir,
403 			    DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
404 		}
405 	}
406 	dsl_dataset_rele(ds_next, FTAG);
407 
408 	/*
409 	 * This must be done after the dsl_traverse(), because it will
410 	 * re-open the objset.
411 	 */
412 	if (ds->ds_objset) {
413 		dmu_objset_evict(ds->ds_objset);
414 		ds->ds_objset = NULL;
415 	}
416 
417 	/* remove from snapshot namespace */
418 	dsl_dataset_t *ds_head;
419 	ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
420 	VERIFY0(dsl_dataset_hold_obj(dp,
421 	    ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
422 	VERIFY0(dsl_dataset_get_snapname(ds));
423 #ifdef ZFS_DEBUG
424 	{
425 		uint64_t val;
426 
427 		err = dsl_dataset_snap_lookup(ds_head,
428 		    ds->ds_snapname, &val);
429 		ASSERT0(err);
430 		ASSERT3U(val, ==, obj);
431 	}
432 #endif
433 	VERIFY0(dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx));
434 	dsl_dataset_rele(ds_head, FTAG);
435 
436 	if (ds_prev != NULL)
437 		dsl_dataset_rele(ds_prev, FTAG);
438 
439 	spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
440 
441 	if (ds->ds_phys->ds_next_clones_obj != 0) {
442 		uint64_t count;
443 		ASSERT0(zap_count(mos,
444 		    ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
445 		VERIFY0(dmu_object_free(mos,
446 		    ds->ds_phys->ds_next_clones_obj, tx));
447 	}
448 	if (ds->ds_phys->ds_props_obj != 0)
449 		VERIFY0(zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
450 	if (ds->ds_phys->ds_userrefs_obj != 0)
451 		VERIFY0(zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
452 	dsl_dir_rele(ds->ds_dir, ds);
453 	ds->ds_dir = NULL;
454 	VERIFY0(dmu_object_free(mos, obj, tx));
455 }
456 
457 static void
458 dsl_destroy_snapshot_sync(void *arg, dmu_tx_t *tx)
459 {
460 	dmu_snapshots_destroy_arg_t *dsda = arg;
461 	dsl_pool_t *dp = dmu_tx_pool(tx);
462 	nvpair_t *pair;
463 
464 	for (pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, NULL);
465 	    pair != NULL;
466 	    pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, pair)) {
467 		dsl_dataset_t *ds;
468 
469 		VERIFY0(dsl_dataset_hold(dp, nvpair_name(pair), FTAG, &ds));
470 
471 		dsl_destroy_snapshot_sync_impl(ds, dsda->dsda_defer, tx);
472 		dsl_dataset_rele(ds, FTAG);
473 	}
474 }
475 
476 /*
477  * The semantics of this function are described in the comment above
478  * lzc_destroy_snaps().  To summarize:
479  *
480  * The snapshots must all be in the same pool.
481  *
482  * Snapshots that don't exist will be silently ignored (considered to be
483  * "already deleted").
484  *
485  * On success, all snaps will be destroyed and this will return 0.
486  * On failure, no snaps will be destroyed, the errlist will be filled in,
487  * and this will return an errno.
488  */
489 int
490 dsl_destroy_snapshots_nvl(nvlist_t *snaps, boolean_t defer,
491     nvlist_t *errlist)
492 {
493 	dmu_snapshots_destroy_arg_t dsda;
494 	int error;
495 	nvpair_t *pair;
496 
497 	pair = nvlist_next_nvpair(snaps, NULL);
498 	if (pair == NULL)
499 		return (0);
500 
501 	dsda.dsda_snaps = snaps;
502 	dsda.dsda_successful_snaps = fnvlist_alloc();
503 	dsda.dsda_defer = defer;
504 	dsda.dsda_errlist = errlist;
505 
506 	error = dsl_sync_task(nvpair_name(pair),
507 	    dsl_destroy_snapshot_check, dsl_destroy_snapshot_sync,
508 	    &dsda, 0);
509 	fnvlist_free(dsda.dsda_successful_snaps);
510 
511 	return (error);
512 }
513 
514 int
515 dsl_destroy_snapshot(const char *name, boolean_t defer)
516 {
517 	int error;
518 	nvlist_t *nvl = fnvlist_alloc();
519 	nvlist_t *errlist = fnvlist_alloc();
520 
521 	fnvlist_add_boolean(nvl, name);
522 	error = dsl_destroy_snapshots_nvl(nvl, defer, errlist);
523 	fnvlist_free(errlist);
524 	fnvlist_free(nvl);
525 	return (error);
526 }
527 
528 struct killarg {
529 	dsl_dataset_t *ds;
530 	dmu_tx_t *tx;
531 };
532 
533 /* ARGSUSED */
534 static int
535 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
536     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
537 {
538 	struct killarg *ka = arg;
539 	dmu_tx_t *tx = ka->tx;
540 
541 	if (bp == NULL)
542 		return (0);
543 
544 	if (zb->zb_level == ZB_ZIL_LEVEL) {
545 		ASSERT(zilog != NULL);
546 		/*
547 		 * It's a block in the intent log.  It has no
548 		 * accounting, so just free it.
549 		 */
550 		dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
551 	} else {
552 		ASSERT(zilog == NULL);
553 		ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
554 		(void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
555 	}
556 
557 	return (0);
558 }
559 
560 static void
561 old_synchronous_dataset_destroy(dsl_dataset_t *ds, dmu_tx_t *tx)
562 {
563 	struct killarg ka;
564 
565 	/*
566 	 * Free everything that we point to (that's born after
567 	 * the previous snapshot, if we are a clone)
568 	 *
569 	 * NB: this should be very quick, because we already
570 	 * freed all the objects in open context.
571 	 */
572 	ka.ds = ds;
573 	ka.tx = tx;
574 	VERIFY0(traverse_dataset(ds,
575 	    ds->ds_phys->ds_prev_snap_txg, TRAVERSE_POST,
576 	    kill_blkptr, &ka));
577 	ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) || ds->ds_phys->ds_unique_bytes == 0);
578 }
579 
580 typedef struct dsl_destroy_head_arg {
581 	const char *ddha_name;
582 } dsl_destroy_head_arg_t;
583 
584 int
585 dsl_destroy_head_check_impl(dsl_dataset_t *ds, int expected_holds)
586 {
587 	int error;
588 	uint64_t count;
589 	objset_t *mos;
590 
591 	if (dsl_dataset_is_snapshot(ds))
592 		return (SET_ERROR(EINVAL));
593 
594 	if (refcount_count(&ds->ds_longholds) != expected_holds)
595 		return (SET_ERROR(EBUSY));
596 
597 	mos = ds->ds_dir->dd_pool->dp_meta_objset;
598 
599 	/*
600 	 * Can't delete a head dataset if there are snapshots of it.
601 	 * (Except if the only snapshots are from the branch we cloned
602 	 * from.)
603 	 */
604 	if (ds->ds_prev != NULL &&
605 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
606 		return (SET_ERROR(EBUSY));
607 
608 	/*
609 	 * Can't delete if there are children of this fs.
610 	 */
611 	error = zap_count(mos,
612 	    ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
613 	if (error != 0)
614 		return (error);
615 	if (count != 0)
616 		return (SET_ERROR(EEXIST));
617 
618 	if (dsl_dir_is_clone(ds->ds_dir) && DS_IS_DEFER_DESTROY(ds->ds_prev) &&
619 	    ds->ds_prev->ds_phys->ds_num_children == 2 &&
620 	    ds->ds_prev->ds_userrefs == 0) {
621 		/* We need to remove the origin snapshot as well. */
622 		if (!refcount_is_zero(&ds->ds_prev->ds_longholds))
623 			return (SET_ERROR(EBUSY));
624 	}
625 	return (0);
626 }
627 
628 static int
629 dsl_destroy_head_check(void *arg, dmu_tx_t *tx)
630 {
631 	dsl_destroy_head_arg_t *ddha = arg;
632 	dsl_pool_t *dp = dmu_tx_pool(tx);
633 	dsl_dataset_t *ds;
634 	int error;
635 
636 	error = dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds);
637 	if (error != 0)
638 		return (error);
639 
640 	error = dsl_destroy_head_check_impl(ds, 0);
641 	dsl_dataset_rele(ds, FTAG);
642 	return (error);
643 }
644 
645 static void
646 dsl_dir_destroy_sync(uint64_t ddobj, dmu_tx_t *tx)
647 {
648 	dsl_dir_t *dd;
649 	dsl_pool_t *dp = dmu_tx_pool(tx);
650 	objset_t *mos = dp->dp_meta_objset;
651 	dd_used_t t;
652 
653 	ASSERT(RRW_WRITE_HELD(&dmu_tx_pool(tx)->dp_config_rwlock));
654 
655 	VERIFY0(dsl_dir_hold_obj(dp, ddobj, NULL, FTAG, &dd));
656 
657 	ASSERT0(dd->dd_phys->dd_head_dataset_obj);
658 
659 	/*
660 	 * Remove our reservation. The impl() routine avoids setting the
661 	 * actual property, which would require the (already destroyed) ds.
662 	 */
663 	dsl_dir_set_reservation_sync_impl(dd, 0, tx);
664 
665 	ASSERT0(dd->dd_phys->dd_used_bytes);
666 	ASSERT0(dd->dd_phys->dd_reserved);
667 	for (t = 0; t < DD_USED_NUM; t++)
668 		ASSERT0(dd->dd_phys->dd_used_breakdown[t]);
669 
670 	VERIFY0(zap_destroy(mos, dd->dd_phys->dd_child_dir_zapobj, tx));
671 	VERIFY0(zap_destroy(mos, dd->dd_phys->dd_props_zapobj, tx));
672 	VERIFY0(dsl_deleg_destroy(mos, dd->dd_phys->dd_deleg_zapobj, tx));
673 	VERIFY0(zap_remove(mos,
674 	    dd->dd_parent->dd_phys->dd_child_dir_zapobj, dd->dd_myname, tx));
675 
676 	dsl_dir_rele(dd, FTAG);
677 	VERIFY0(dmu_object_free(mos, ddobj, tx));
678 }
679 
680 void
681 dsl_destroy_head_sync_impl(dsl_dataset_t *ds, dmu_tx_t *tx)
682 {
683 	dsl_pool_t *dp = dmu_tx_pool(tx);
684 	objset_t *mos = dp->dp_meta_objset;
685 	uint64_t obj, ddobj, prevobj = 0;
686 	boolean_t rmorigin;
687 
688 	ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
689 	ASSERT(ds->ds_prev == NULL ||
690 	    ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
691 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
692 	ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
693 
694 	/* We need to log before removing it from the namespace. */
695 	spa_history_log_internal_ds(ds, "destroy", tx, "");
696 
697 	rmorigin = (dsl_dir_is_clone(ds->ds_dir) &&
698 	    DS_IS_DEFER_DESTROY(ds->ds_prev) &&
699 	    ds->ds_prev->ds_phys->ds_num_children == 2 &&
700 	    ds->ds_prev->ds_userrefs == 0);
701 
702 	/* Remove our reservation */
703 	if (ds->ds_reserved != 0) {
704 		dsl_dataset_set_refreservation_sync_impl(ds,
705 		    (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
706 		    0, tx);
707 		ASSERT0(ds->ds_reserved);
708 	}
709 
710 	dsl_scan_ds_destroyed(ds, tx);
711 
712 	obj = ds->ds_object;
713 
714 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
715 		/* This is a clone */
716 		ASSERT(ds->ds_prev != NULL);
717 		ASSERT3U(ds->ds_prev->ds_phys->ds_next_snap_obj, !=, obj);
718 		ASSERT0(ds->ds_phys->ds_next_snap_obj);
719 
720 		dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
721 		if (ds->ds_prev->ds_phys->ds_next_clones_obj != 0) {
722 			dsl_dataset_remove_from_next_clones(ds->ds_prev,
723 			    obj, tx);
724 		}
725 
726 		ASSERT3U(ds->ds_prev->ds_phys->ds_num_children, >, 1);
727 		ds->ds_prev->ds_phys->ds_num_children--;
728 	}
729 
730 	zfeature_info_t *async_destroy =
731 	    &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY];
732 	objset_t *os;
733 
734 	/*
735 	 * Destroy the deadlist.  Unless it's a clone, the
736 	 * deadlist should be empty.  (If it's a clone, it's
737 	 * safe to ignore the deadlist contents.)
738 	 */
739 	dsl_deadlist_close(&ds->ds_deadlist);
740 	dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
741 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
742 	ds->ds_phys->ds_deadlist_obj = 0;
743 
744 	VERIFY0(dmu_objset_from_ds(ds, &os));
745 
746 	if (!spa_feature_is_enabled(dp->dp_spa, async_destroy)) {
747 		old_synchronous_dataset_destroy(ds, tx);
748 	} else {
749 		/*
750 		 * Move the bptree into the pool's list of trees to
751 		 * clean up and update space accounting information.
752 		 */
753 		uint64_t used, comp, uncomp;
754 
755 		zil_destroy_sync(dmu_objset_zil(os), tx);
756 
757 		if (!spa_feature_is_active(dp->dp_spa, async_destroy)) {
758 			dsl_scan_t *scn = dp->dp_scan;
759 
760 			spa_feature_incr(dp->dp_spa, async_destroy, tx);
761 			dp->dp_bptree_obj = bptree_alloc(mos, tx);
762 			VERIFY0(zap_add(mos,
763 			    DMU_POOL_DIRECTORY_OBJECT,
764 			    DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
765 			    &dp->dp_bptree_obj, tx));
766 			ASSERT(!scn->scn_async_destroying);
767 			scn->scn_async_destroying = B_TRUE;
768 		}
769 
770 		used = ds->ds_dir->dd_phys->dd_used_bytes;
771 		comp = ds->ds_dir->dd_phys->dd_compressed_bytes;
772 		uncomp = ds->ds_dir->dd_phys->dd_uncompressed_bytes;
773 
774 		ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
775 		    ds->ds_phys->ds_unique_bytes == used);
776 
777 		bptree_add(mos, dp->dp_bptree_obj,
778 		    &ds->ds_phys->ds_bp, ds->ds_phys->ds_prev_snap_txg,
779 		    used, comp, uncomp, tx);
780 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
781 		    -used, -comp, -uncomp, tx);
782 		dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
783 		    used, comp, uncomp, tx);
784 	}
785 
786 	if (ds->ds_prev != NULL) {
787 		if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
788 			VERIFY0(zap_remove_int(mos,
789 			    ds->ds_prev->ds_dir->dd_phys->dd_clones,
790 			    ds->ds_object, tx));
791 		}
792 		prevobj = ds->ds_prev->ds_object;
793 		dsl_dataset_rele(ds->ds_prev, ds);
794 		ds->ds_prev = NULL;
795 	}
796 
797 	/*
798 	 * This must be done after the dsl_traverse(), because it will
799 	 * re-open the objset.
800 	 */
801 	if (ds->ds_objset) {
802 		dmu_objset_evict(ds->ds_objset);
803 		ds->ds_objset = NULL;
804 	}
805 
806 	/* Erase the link in the dir */
807 	dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
808 	ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
809 	ddobj = ds->ds_dir->dd_object;
810 	ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
811 	VERIFY0(zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx));
812 
813 	spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
814 
815 	ASSERT0(ds->ds_phys->ds_next_clones_obj);
816 	ASSERT0(ds->ds_phys->ds_props_obj);
817 	ASSERT0(ds->ds_phys->ds_userrefs_obj);
818 	dsl_dir_rele(ds->ds_dir, ds);
819 	ds->ds_dir = NULL;
820 	VERIFY0(dmu_object_free(mos, obj, tx));
821 
822 	dsl_dir_destroy_sync(ddobj, tx);
823 
824 	if (rmorigin) {
825 		dsl_dataset_t *prev;
826 		VERIFY0(dsl_dataset_hold_obj(dp, prevobj, FTAG, &prev));
827 		dsl_destroy_snapshot_sync_impl(prev, B_FALSE, tx);
828 		dsl_dataset_rele(prev, FTAG);
829 	}
830 }
831 
832 static void
833 dsl_destroy_head_sync(void *arg, dmu_tx_t *tx)
834 {
835 	dsl_destroy_head_arg_t *ddha = arg;
836 	dsl_pool_t *dp = dmu_tx_pool(tx);
837 	dsl_dataset_t *ds;
838 
839 	VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
840 	dsl_destroy_head_sync_impl(ds, tx);
841 	dsl_dataset_rele(ds, FTAG);
842 }
843 
844 static void
845 dsl_destroy_head_begin_sync(void *arg, dmu_tx_t *tx)
846 {
847 	dsl_destroy_head_arg_t *ddha = arg;
848 	dsl_pool_t *dp = dmu_tx_pool(tx);
849 	dsl_dataset_t *ds;
850 
851 	VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
852 
853 	/* Mark it as inconsistent on-disk, in case we crash */
854 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
855 	ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
856 
857 	spa_history_log_internal_ds(ds, "destroy begin", tx, "");
858 	dsl_dataset_rele(ds, FTAG);
859 }
860 
861 int
862 dsl_destroy_head(const char *name)
863 {
864 	dsl_destroy_head_arg_t ddha;
865 	int error;
866 	spa_t *spa;
867 	boolean_t isenabled;
868 
869 #ifdef _KERNEL
870 	zfs_destroy_unmount_origin(name);
871 #endif
872 
873 	error = spa_open(name, &spa, FTAG);
874 	if (error != 0)
875 		return (error);
876 	isenabled = spa_feature_is_enabled(spa,
877 	    &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]);
878 	spa_close(spa, FTAG);
879 
880 	ddha.ddha_name = name;
881 
882 	if (!isenabled) {
883 		objset_t *os;
884 
885 		error = dsl_sync_task(name, dsl_destroy_head_check,
886 		    dsl_destroy_head_begin_sync, &ddha, 0);
887 		if (error != 0)
888 			return (error);
889 
890 		/*
891 		 * Head deletion is processed in one txg on old pools;
892 		 * remove the objects from open context so that the txg sync
893 		 * is not too long.
894 		 */
895 		error = dmu_objset_own(name, DMU_OST_ANY, B_FALSE, FTAG, &os);
896 		if (error == 0) {
897 			uint64_t prev_snap_txg =
898 			    dmu_objset_ds(os)->ds_phys->ds_prev_snap_txg;
899 			for (uint64_t obj = 0; error == 0;
900 			    error = dmu_object_next(os, &obj, FALSE,
901 			    prev_snap_txg))
902 				(void) dmu_free_object(os, obj);
903 			/* sync out all frees */
904 			txg_wait_synced(dmu_objset_pool(os), 0);
905 			dmu_objset_disown(os, FTAG);
906 		}
907 	}
908 
909 	return (dsl_sync_task(name, dsl_destroy_head_check,
910 	    dsl_destroy_head_sync, &ddha, 0));
911 }
912 
913 /*
914  * Note, this function is used as the callback for dmu_objset_find().  We
915  * always return 0 so that we will continue to find and process
916  * inconsistent datasets, even if we encounter an error trying to
917  * process one of them.
918  */
919 /* ARGSUSED */
920 int
921 dsl_destroy_inconsistent(const char *dsname, void *arg)
922 {
923 	objset_t *os;
924 
925 	if (dmu_objset_hold(dsname, FTAG, &os) == 0) {
926 		boolean_t inconsistent = DS_IS_INCONSISTENT(dmu_objset_ds(os));
927 		dmu_objset_rele(os, FTAG);
928 		if (inconsistent)
929 			(void) dsl_destroy_head(dsname);
930 	}
931 	return (0);
932 }
933