xref: /titanic_50/usr/src/uts/common/fs/zfs/dsl_destroy.c (revision abc79d9dd51e98eafb6fc25b4a0b4f66bef40b00)
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  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/dsl_userhold.h>
28 #include <sys/dsl_dataset.h>
29 #include <sys/dsl_synctask.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dmu_traverse.h>
34 #include <sys/dsl_scan.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/zap.h>
37 #include <sys/zfeature.h>
38 #include <sys/zfs_ioctl.h>
39 #include <sys/dsl_deleg.h>
40 
41 typedef struct dmu_snapshots_destroy_arg {
42 	nvlist_t *dsda_snaps;
43 	nvlist_t *dsda_successful_snaps;
44 	boolean_t dsda_defer;
45 	nvlist_t *dsda_errlist;
46 } dmu_snapshots_destroy_arg_t;
47 
48 /*
49  * ds must be owned.
50  */
51 static int
52 dsl_destroy_snapshot_check_impl(dsl_dataset_t *ds, boolean_t defer)
53 {
54 	if (!dsl_dataset_is_snapshot(ds))
55 		return (SET_ERROR(EINVAL));
56 
57 	if (dsl_dataset_long_held(ds))
58 		return (SET_ERROR(EBUSY));
59 
60 	/*
61 	 * Only allow deferred destroy on pools that support it.
62 	 * NOTE: deferred destroy is only supported on snapshots.
63 	 */
64 	if (defer) {
65 		if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
66 		    SPA_VERSION_USERREFS)
67 			return (SET_ERROR(ENOTSUP));
68 		return (0);
69 	}
70 
71 	/*
72 	 * If this snapshot has an elevated user reference count,
73 	 * we can't destroy it yet.
74 	 */
75 	if (ds->ds_userrefs > 0)
76 		return (SET_ERROR(EBUSY));
77 
78 	/*
79 	 * Can't delete a branch point.
80 	 */
81 	if (ds->ds_phys->ds_num_children > 1)
82 		return (SET_ERROR(EEXIST));
83 
84 	return (0);
85 }
86 
87 static int
88 dsl_destroy_snapshot_check(void *arg, dmu_tx_t *tx)
89 {
90 	dmu_snapshots_destroy_arg_t *dsda = arg;
91 	dsl_pool_t *dp = dmu_tx_pool(tx);
92 	nvpair_t *pair;
93 	int error = 0;
94 
95 	if (!dmu_tx_is_syncing(tx))
96 		return (0);
97 
98 	for (pair = nvlist_next_nvpair(dsda->dsda_snaps, NULL);
99 	    pair != NULL; pair = nvlist_next_nvpair(dsda->dsda_snaps, pair)) {
100 		dsl_dataset_t *ds;
101 
102 		error = dsl_dataset_hold(dp, nvpair_name(pair),
103 		    FTAG, &ds);
104 
105 		/*
106 		 * If the snapshot does not exist, silently ignore it
107 		 * (it's "already destroyed").
108 		 */
109 		if (error == ENOENT)
110 			continue;
111 
112 		if (error == 0) {
113 			error = dsl_destroy_snapshot_check_impl(ds,
114 			    dsda->dsda_defer);
115 			dsl_dataset_rele(ds, FTAG);
116 		}
117 
118 		if (error == 0) {
119 			fnvlist_add_boolean(dsda->dsda_successful_snaps,
120 			    nvpair_name(pair));
121 		} else {
122 			fnvlist_add_int32(dsda->dsda_errlist,
123 			    nvpair_name(pair), error);
124 		}
125 	}
126 
127 	pair = nvlist_next_nvpair(dsda->dsda_errlist, NULL);
128 	if (pair != NULL)
129 		return (fnvpair_value_int32(pair));
130 	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 	if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
148 		dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
149 		if (poa->ds_prev && !poa->after_branch_point &&
150 		    bp->blk_birth >
151 		    poa->ds_prev->ds_phys->ds_prev_snap_txg) {
152 			poa->ds_prev->ds_phys->ds_unique_bytes +=
153 			    bp_get_dsize_sync(dp->dp_spa, bp);
154 		}
155 	} else {
156 		poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
157 		poa->comp += BP_GET_PSIZE(bp);
158 		poa->uncomp += BP_GET_UCSIZE(bp);
159 		dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
160 	}
161 	return (0);
162 }
163 
164 static void
165 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
166     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
167 {
168 	struct process_old_arg poa = { 0 };
169 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
170 	objset_t *mos = dp->dp_meta_objset;
171 	uint64_t deadlist_obj;
172 
173 	ASSERT(ds->ds_deadlist.dl_oldfmt);
174 	ASSERT(ds_next->ds_deadlist.dl_oldfmt);
175 
176 	poa.ds = ds;
177 	poa.ds_prev = ds_prev;
178 	poa.after_branch_point = after_branch_point;
179 	poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
180 	VERIFY0(bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
181 	    process_old_cb, &poa, tx));
182 	VERIFY0(zio_wait(poa.pio));
183 	ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
184 
185 	/* change snapused */
186 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
187 	    -poa.used, -poa.comp, -poa.uncomp, tx);
188 
189 	/* swap next's deadlist to our deadlist */
190 	dsl_deadlist_close(&ds->ds_deadlist);
191 	dsl_deadlist_close(&ds_next->ds_deadlist);
192 	deadlist_obj = ds->ds_phys->ds_deadlist_obj;
193 	ds->ds_phys->ds_deadlist_obj = ds_next->ds_phys->ds_deadlist_obj;
194 	ds_next->ds_phys->ds_deadlist_obj = deadlist_obj;
195 	dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
196 	dsl_deadlist_open(&ds_next->ds_deadlist, mos,
197 	    ds_next->ds_phys->ds_deadlist_obj);
198 }
199 
200 static void
201 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
202 {
203 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
204 	zap_cursor_t zc;
205 	zap_attribute_t za;
206 
207 	/*
208 	 * If it is the old version, dd_clones doesn't exist so we can't
209 	 * find the clones, but dsl_deadlist_remove_key() is a no-op so it
210 	 * doesn't matter.
211 	 */
212 	if (ds->ds_dir->dd_phys->dd_clones == 0)
213 		return;
214 
215 	for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
216 	    zap_cursor_retrieve(&zc, &za) == 0;
217 	    zap_cursor_advance(&zc)) {
218 		dsl_dataset_t *clone;
219 
220 		VERIFY0(dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
221 		    za.za_first_integer, FTAG, &clone));
222 		if (clone->ds_dir->dd_origin_txg > mintxg) {
223 			dsl_deadlist_remove_key(&clone->ds_deadlist,
224 			    mintxg, tx);
225 			dsl_dataset_remove_clones_key(clone, mintxg, tx);
226 		}
227 		dsl_dataset_rele(clone, FTAG);
228 	}
229 	zap_cursor_fini(&zc);
230 }
231 
232 void
233 dsl_destroy_snapshot_sync_impl(dsl_dataset_t *ds, boolean_t defer, dmu_tx_t *tx)
234 {
235 	int err;
236 	int after_branch_point = FALSE;
237 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
238 	objset_t *mos = dp->dp_meta_objset;
239 	dsl_dataset_t *ds_prev = NULL;
240 	uint64_t obj;
241 
242 	ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
243 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
244 	ASSERT(refcount_is_zero(&ds->ds_longholds));
245 
246 	if (defer &&
247 	    (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1)) {
248 		ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
249 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
250 		ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
251 		spa_history_log_internal_ds(ds, "defer_destroy", tx, "");
252 		return;
253 	}
254 
255 	ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
256 
257 	/* We need to log before removing it from the namespace. */
258 	spa_history_log_internal_ds(ds, "destroy", tx, "");
259 
260 	dsl_scan_ds_destroyed(ds, tx);
261 
262 	obj = ds->ds_object;
263 
264 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
265 		ASSERT3P(ds->ds_prev, ==, NULL);
266 		VERIFY0(dsl_dataset_hold_obj(dp,
267 		    ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
268 		after_branch_point =
269 		    (ds_prev->ds_phys->ds_next_snap_obj != obj);
270 
271 		dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
272 		if (after_branch_point &&
273 		    ds_prev->ds_phys->ds_next_clones_obj != 0) {
274 			dsl_dataset_remove_from_next_clones(ds_prev, obj, tx);
275 			if (ds->ds_phys->ds_next_snap_obj != 0) {
276 				VERIFY0(zap_add_int(mos,
277 				    ds_prev->ds_phys->ds_next_clones_obj,
278 				    ds->ds_phys->ds_next_snap_obj, tx));
279 			}
280 		}
281 		if (!after_branch_point) {
282 			ds_prev->ds_phys->ds_next_snap_obj =
283 			    ds->ds_phys->ds_next_snap_obj;
284 		}
285 	}
286 
287 	dsl_dataset_t *ds_next;
288 	uint64_t old_unique;
289 	uint64_t used = 0, comp = 0, uncomp = 0;
290 
291 	VERIFY0(dsl_dataset_hold_obj(dp,
292 	    ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
293 	ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
294 
295 	old_unique = ds_next->ds_phys->ds_unique_bytes;
296 
297 	dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
298 	ds_next->ds_phys->ds_prev_snap_obj =
299 	    ds->ds_phys->ds_prev_snap_obj;
300 	ds_next->ds_phys->ds_prev_snap_txg =
301 	    ds->ds_phys->ds_prev_snap_txg;
302 	ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
303 	    ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
304 
305 	if (ds_next->ds_deadlist.dl_oldfmt) {
306 		process_old_deadlist(ds, ds_prev, ds_next,
307 		    after_branch_point, tx);
308 	} else {
309 		/* Adjust prev's unique space. */
310 		if (ds_prev && !after_branch_point) {
311 			dsl_deadlist_space_range(&ds_next->ds_deadlist,
312 			    ds_prev->ds_phys->ds_prev_snap_txg,
313 			    ds->ds_phys->ds_prev_snap_txg,
314 			    &used, &comp, &uncomp);
315 			ds_prev->ds_phys->ds_unique_bytes += used;
316 		}
317 
318 		/* Adjust snapused. */
319 		dsl_deadlist_space_range(&ds_next->ds_deadlist,
320 		    ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
321 		    &used, &comp, &uncomp);
322 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
323 		    -used, -comp, -uncomp, tx);
324 
325 		/* Move blocks to be freed to pool's free list. */
326 		dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
327 		    &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
328 		    tx);
329 		dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
330 		    DD_USED_HEAD, used, comp, uncomp, tx);
331 
332 		/* Merge our deadlist into next's and free it. */
333 		dsl_deadlist_merge(&ds_next->ds_deadlist,
334 		    ds->ds_phys->ds_deadlist_obj, tx);
335 	}
336 	dsl_deadlist_close(&ds->ds_deadlist);
337 	dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
338 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
339 	ds->ds_phys->ds_deadlist_obj = 0;
340 
341 	/* Collapse range in clone heads */
342 	dsl_dataset_remove_clones_key(ds,
343 	    ds->ds_phys->ds_creation_txg, tx);
344 
345 	if (dsl_dataset_is_snapshot(ds_next)) {
346 		dsl_dataset_t *ds_nextnext;
347 
348 		/*
349 		 * Update next's unique to include blocks which
350 		 * were previously shared by only this snapshot
351 		 * and it.  Those blocks will be born after the
352 		 * prev snap and before this snap, and will have
353 		 * died after the next snap and before the one
354 		 * after that (ie. be on the snap after next's
355 		 * deadlist).
356 		 */
357 		VERIFY0(dsl_dataset_hold_obj(dp,
358 		    ds_next->ds_phys->ds_next_snap_obj, FTAG, &ds_nextnext));
359 		dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
360 		    ds->ds_phys->ds_prev_snap_txg,
361 		    ds->ds_phys->ds_creation_txg,
362 		    &used, &comp, &uncomp);
363 		ds_next->ds_phys->ds_unique_bytes += used;
364 		dsl_dataset_rele(ds_nextnext, FTAG);
365 		ASSERT3P(ds_next->ds_prev, ==, NULL);
366 
367 		/* Collapse range in this head. */
368 		dsl_dataset_t *hds;
369 		VERIFY0(dsl_dataset_hold_obj(dp,
370 		    ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &hds));
371 		dsl_deadlist_remove_key(&hds->ds_deadlist,
372 		    ds->ds_phys->ds_creation_txg, tx);
373 		dsl_dataset_rele(hds, FTAG);
374 
375 	} else {
376 		ASSERT3P(ds_next->ds_prev, ==, ds);
377 		dsl_dataset_rele(ds_next->ds_prev, ds_next);
378 		ds_next->ds_prev = NULL;
379 		if (ds_prev) {
380 			VERIFY0(dsl_dataset_hold_obj(dp,
381 			    ds->ds_phys->ds_prev_snap_obj,
382 			    ds_next, &ds_next->ds_prev));
383 		}
384 
385 		dsl_dataset_recalc_head_uniq(ds_next);
386 
387 		/*
388 		 * Reduce the amount of our unconsumed refreservation
389 		 * being charged to our parent by the amount of
390 		 * new unique data we have gained.
391 		 */
392 		if (old_unique < ds_next->ds_reserved) {
393 			int64_t mrsdelta;
394 			uint64_t new_unique =
395 			    ds_next->ds_phys->ds_unique_bytes;
396 
397 			ASSERT(old_unique <= new_unique);
398 			mrsdelta = MIN(new_unique - old_unique,
399 			    ds_next->ds_reserved - old_unique);
400 			dsl_dir_diduse_space(ds->ds_dir,
401 			    DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
402 		}
403 	}
404 	dsl_dataset_rele(ds_next, FTAG);
405 
406 	/*
407 	 * This must be done after the dsl_traverse(), because it will
408 	 * re-open the objset.
409 	 */
410 	if (ds->ds_objset) {
411 		dmu_objset_evict(ds->ds_objset);
412 		ds->ds_objset = NULL;
413 	}
414 
415 	/* remove from snapshot namespace */
416 	dsl_dataset_t *ds_head;
417 	ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
418 	VERIFY0(dsl_dataset_hold_obj(dp,
419 	    ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
420 	VERIFY0(dsl_dataset_get_snapname(ds));
421 #ifdef ZFS_DEBUG
422 	{
423 		uint64_t val;
424 
425 		err = dsl_dataset_snap_lookup(ds_head,
426 		    ds->ds_snapname, &val);
427 		ASSERT0(err);
428 		ASSERT3U(val, ==, obj);
429 	}
430 #endif
431 	VERIFY0(dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx));
432 	dsl_dataset_rele(ds_head, FTAG);
433 
434 	if (ds_prev != NULL)
435 		dsl_dataset_rele(ds_prev, FTAG);
436 
437 	spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
438 
439 	if (ds->ds_phys->ds_next_clones_obj != 0) {
440 		uint64_t count;
441 		ASSERT0(zap_count(mos,
442 		    ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
443 		VERIFY0(dmu_object_free(mos,
444 		    ds->ds_phys->ds_next_clones_obj, tx));
445 	}
446 	if (ds->ds_phys->ds_props_obj != 0)
447 		VERIFY0(zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
448 	if (ds->ds_phys->ds_userrefs_obj != 0)
449 		VERIFY0(zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
450 	dsl_dir_rele(ds->ds_dir, ds);
451 	ds->ds_dir = NULL;
452 	VERIFY0(dmu_object_free(mos, obj, tx));
453 }
454 
455 static void
456 dsl_destroy_snapshot_sync(void *arg, dmu_tx_t *tx)
457 {
458 	dmu_snapshots_destroy_arg_t *dsda = arg;
459 	dsl_pool_t *dp = dmu_tx_pool(tx);
460 	nvpair_t *pair;
461 
462 	for (pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, NULL);
463 	    pair != NULL;
464 	    pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, pair)) {
465 		dsl_dataset_t *ds;
466 
467 		VERIFY0(dsl_dataset_hold(dp, nvpair_name(pair), FTAG, &ds));
468 
469 		dsl_destroy_snapshot_sync_impl(ds, dsda->dsda_defer, tx);
470 		dsl_dataset_rele(ds, FTAG);
471 	}
472 }
473 
474 /*
475  * The semantics of this function are described in the comment above
476  * lzc_destroy_snaps().  To summarize:
477  *
478  * The snapshots must all be in the same pool.
479  *
480  * Snapshots that don't exist will be silently ignored (considered to be
481  * "already deleted").
482  *
483  * On success, all snaps will be destroyed and this will return 0.
484  * On failure, no snaps will be destroyed, the errlist will be filled in,
485  * and this will return an errno.
486  */
487 int
488 dsl_destroy_snapshots_nvl(nvlist_t *snaps, boolean_t defer,
489     nvlist_t *errlist)
490 {
491 	dmu_snapshots_destroy_arg_t dsda;
492 	int error;
493 	nvpair_t *pair;
494 
495 	pair = nvlist_next_nvpair(snaps, NULL);
496 	if (pair == NULL)
497 		return (0);
498 
499 	dsda.dsda_snaps = snaps;
500 	dsda.dsda_successful_snaps = fnvlist_alloc();
501 	dsda.dsda_defer = defer;
502 	dsda.dsda_errlist = errlist;
503 
504 	error = dsl_sync_task(nvpair_name(pair),
505 	    dsl_destroy_snapshot_check, dsl_destroy_snapshot_sync,
506 	    &dsda, 0);
507 	fnvlist_free(dsda.dsda_successful_snaps);
508 
509 	return (error);
510 }
511 
512 int
513 dsl_destroy_snapshot(const char *name, boolean_t defer)
514 {
515 	int error;
516 	nvlist_t *nvl = fnvlist_alloc();
517 	nvlist_t *errlist = fnvlist_alloc();
518 
519 	fnvlist_add_boolean(nvl, name);
520 	error = dsl_destroy_snapshots_nvl(nvl, defer, errlist);
521 	fnvlist_free(errlist);
522 	fnvlist_free(nvl);
523 	return (error);
524 }
525 
526 struct killarg {
527 	dsl_dataset_t *ds;
528 	dmu_tx_t *tx;
529 };
530 
531 /* ARGSUSED */
532 static int
533 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
534     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
535 {
536 	struct killarg *ka = arg;
537 	dmu_tx_t *tx = ka->tx;
538 
539 	if (bp == NULL)
540 		return (0);
541 
542 	if (zb->zb_level == ZB_ZIL_LEVEL) {
543 		ASSERT(zilog != NULL);
544 		/*
545 		 * It's a block in the intent log.  It has no
546 		 * accounting, so just free it.
547 		 */
548 		dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
549 	} else {
550 		ASSERT(zilog == NULL);
551 		ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
552 		(void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
553 	}
554 
555 	return (0);
556 }
557 
558 static void
559 old_synchronous_dataset_destroy(dsl_dataset_t *ds, dmu_tx_t *tx)
560 {
561 	struct killarg ka;
562 
563 	/*
564 	 * Free everything that we point to (that's born after
565 	 * the previous snapshot, if we are a clone)
566 	 *
567 	 * NB: this should be very quick, because we already
568 	 * freed all the objects in open context.
569 	 */
570 	ka.ds = ds;
571 	ka.tx = tx;
572 	VERIFY0(traverse_dataset(ds,
573 	    ds->ds_phys->ds_prev_snap_txg, TRAVERSE_POST,
574 	    kill_blkptr, &ka));
575 	ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) || ds->ds_phys->ds_unique_bytes == 0);
576 }
577 
578 typedef struct dsl_destroy_head_arg {
579 	const char *ddha_name;
580 } dsl_destroy_head_arg_t;
581 
582 int
583 dsl_destroy_head_check_impl(dsl_dataset_t *ds, int expected_holds)
584 {
585 	int error;
586 	uint64_t count;
587 	objset_t *mos;
588 
589 	if (dsl_dataset_is_snapshot(ds))
590 		return (SET_ERROR(EINVAL));
591 
592 	if (refcount_count(&ds->ds_longholds) != expected_holds)
593 		return (SET_ERROR(EBUSY));
594 
595 	mos = ds->ds_dir->dd_pool->dp_meta_objset;
596 
597 	/*
598 	 * Can't delete a head dataset if there are snapshots of it.
599 	 * (Except if the only snapshots are from the branch we cloned
600 	 * from.)
601 	 */
602 	if (ds->ds_prev != NULL &&
603 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
604 		return (SET_ERROR(EBUSY));
605 
606 	/*
607 	 * Can't delete if there are children of this fs.
608 	 */
609 	error = zap_count(mos,
610 	    ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
611 	if (error != 0)
612 		return (error);
613 	if (count != 0)
614 		return (SET_ERROR(EEXIST));
615 
616 	if (dsl_dir_is_clone(ds->ds_dir) && DS_IS_DEFER_DESTROY(ds->ds_prev) &&
617 	    ds->ds_prev->ds_phys->ds_num_children == 2 &&
618 	    ds->ds_prev->ds_userrefs == 0) {
619 		/* We need to remove the origin snapshot as well. */
620 		if (!refcount_is_zero(&ds->ds_prev->ds_longholds))
621 			return (SET_ERROR(EBUSY));
622 	}
623 	return (0);
624 }
625 
626 static int
627 dsl_destroy_head_check(void *arg, dmu_tx_t *tx)
628 {
629 	dsl_destroy_head_arg_t *ddha = arg;
630 	dsl_pool_t *dp = dmu_tx_pool(tx);
631 	dsl_dataset_t *ds;
632 	int error;
633 
634 	error = dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds);
635 	if (error != 0)
636 		return (error);
637 
638 	error = dsl_destroy_head_check_impl(ds, 0);
639 	dsl_dataset_rele(ds, FTAG);
640 	return (error);
641 }
642 
643 static void
644 dsl_dir_destroy_sync(uint64_t ddobj, dmu_tx_t *tx)
645 {
646 	dsl_dir_t *dd;
647 	dsl_pool_t *dp = dmu_tx_pool(tx);
648 	objset_t *mos = dp->dp_meta_objset;
649 	dd_used_t t;
650 
651 	ASSERT(RRW_WRITE_HELD(&dmu_tx_pool(tx)->dp_config_rwlock));
652 
653 	VERIFY0(dsl_dir_hold_obj(dp, ddobj, NULL, FTAG, &dd));
654 
655 	ASSERT0(dd->dd_phys->dd_head_dataset_obj);
656 
657 	/*
658 	 * Remove our reservation. The impl() routine avoids setting the
659 	 * actual property, which would require the (already destroyed) ds.
660 	 */
661 	dsl_dir_set_reservation_sync_impl(dd, 0, tx);
662 
663 	ASSERT0(dd->dd_phys->dd_used_bytes);
664 	ASSERT0(dd->dd_phys->dd_reserved);
665 	for (t = 0; t < DD_USED_NUM; t++)
666 		ASSERT0(dd->dd_phys->dd_used_breakdown[t]);
667 
668 	VERIFY0(zap_destroy(mos, dd->dd_phys->dd_child_dir_zapobj, tx));
669 	VERIFY0(zap_destroy(mos, dd->dd_phys->dd_props_zapobj, tx));
670 	VERIFY0(dsl_deleg_destroy(mos, dd->dd_phys->dd_deleg_zapobj, tx));
671 	VERIFY0(zap_remove(mos,
672 	    dd->dd_parent->dd_phys->dd_child_dir_zapobj, dd->dd_myname, tx));
673 
674 	dsl_dir_rele(dd, FTAG);
675 	VERIFY0(dmu_object_free(mos, ddobj, tx));
676 }
677 
678 void
679 dsl_destroy_head_sync_impl(dsl_dataset_t *ds, dmu_tx_t *tx)
680 {
681 	dsl_pool_t *dp = dmu_tx_pool(tx);
682 	objset_t *mos = dp->dp_meta_objset;
683 	uint64_t obj, ddobj, prevobj = 0;
684 	boolean_t rmorigin;
685 
686 	ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
687 	ASSERT(ds->ds_prev == NULL ||
688 	    ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
689 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
690 	ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
691 
692 	/* We need to log before removing it from the namespace. */
693 	spa_history_log_internal_ds(ds, "destroy", tx, "");
694 
695 	rmorigin = (dsl_dir_is_clone(ds->ds_dir) &&
696 	    DS_IS_DEFER_DESTROY(ds->ds_prev) &&
697 	    ds->ds_prev->ds_phys->ds_num_children == 2 &&
698 	    ds->ds_prev->ds_userrefs == 0);
699 
700 	/* Remove our reservation */
701 	if (ds->ds_reserved != 0) {
702 		dsl_dataset_set_refreservation_sync_impl(ds,
703 		    (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
704 		    0, tx);
705 		ASSERT0(ds->ds_reserved);
706 	}
707 
708 	dsl_scan_ds_destroyed(ds, tx);
709 
710 	obj = ds->ds_object;
711 
712 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
713 		/* This is a clone */
714 		ASSERT(ds->ds_prev != NULL);
715 		ASSERT3U(ds->ds_prev->ds_phys->ds_next_snap_obj, !=, obj);
716 		ASSERT0(ds->ds_phys->ds_next_snap_obj);
717 
718 		dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
719 		if (ds->ds_prev->ds_phys->ds_next_clones_obj != 0) {
720 			dsl_dataset_remove_from_next_clones(ds->ds_prev,
721 			    obj, tx);
722 		}
723 
724 		ASSERT3U(ds->ds_prev->ds_phys->ds_num_children, >, 1);
725 		ds->ds_prev->ds_phys->ds_num_children--;
726 	}
727 
728 	zfeature_info_t *async_destroy =
729 	    &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY];
730 	objset_t *os;
731 
732 	/*
733 	 * Destroy the deadlist.  Unless it's a clone, the
734 	 * deadlist should be empty.  (If it's a clone, it's
735 	 * safe to ignore the deadlist contents.)
736 	 */
737 	dsl_deadlist_close(&ds->ds_deadlist);
738 	dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
739 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
740 	ds->ds_phys->ds_deadlist_obj = 0;
741 
742 	VERIFY0(dmu_objset_from_ds(ds, &os));
743 
744 	if (!spa_feature_is_enabled(dp->dp_spa, async_destroy)) {
745 		old_synchronous_dataset_destroy(ds, tx);
746 	} else {
747 		/*
748 		 * Move the bptree into the pool's list of trees to
749 		 * clean up and update space accounting information.
750 		 */
751 		uint64_t used, comp, uncomp;
752 
753 		zil_destroy_sync(dmu_objset_zil(os), tx);
754 
755 		if (!spa_feature_is_active(dp->dp_spa, async_destroy)) {
756 			spa_feature_incr(dp->dp_spa, async_destroy, tx);
757 			dp->dp_bptree_obj = bptree_alloc(mos, tx);
758 			VERIFY0(zap_add(mos,
759 			    DMU_POOL_DIRECTORY_OBJECT,
760 			    DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
761 			    &dp->dp_bptree_obj, tx));
762 		}
763 
764 		used = ds->ds_dir->dd_phys->dd_used_bytes;
765 		comp = ds->ds_dir->dd_phys->dd_compressed_bytes;
766 		uncomp = ds->ds_dir->dd_phys->dd_uncompressed_bytes;
767 
768 		ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
769 		    ds->ds_phys->ds_unique_bytes == used);
770 
771 		bptree_add(mos, dp->dp_bptree_obj,
772 		    &ds->ds_phys->ds_bp, ds->ds_phys->ds_prev_snap_txg,
773 		    used, comp, uncomp, tx);
774 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
775 		    -used, -comp, -uncomp, tx);
776 		dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
777 		    used, comp, uncomp, tx);
778 	}
779 
780 	if (ds->ds_prev != NULL) {
781 		if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
782 			VERIFY0(zap_remove_int(mos,
783 			    ds->ds_prev->ds_dir->dd_phys->dd_clones,
784 			    ds->ds_object, tx));
785 		}
786 		prevobj = ds->ds_prev->ds_object;
787 		dsl_dataset_rele(ds->ds_prev, ds);
788 		ds->ds_prev = NULL;
789 	}
790 
791 	/*
792 	 * This must be done after the dsl_traverse(), because it will
793 	 * re-open the objset.
794 	 */
795 	if (ds->ds_objset) {
796 		dmu_objset_evict(ds->ds_objset);
797 		ds->ds_objset = NULL;
798 	}
799 
800 	/* Erase the link in the dir */
801 	dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
802 	ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
803 	ddobj = ds->ds_dir->dd_object;
804 	ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
805 	VERIFY0(zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx));
806 
807 	spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
808 
809 	ASSERT0(ds->ds_phys->ds_next_clones_obj);
810 	ASSERT0(ds->ds_phys->ds_props_obj);
811 	ASSERT0(ds->ds_phys->ds_userrefs_obj);
812 	dsl_dir_rele(ds->ds_dir, ds);
813 	ds->ds_dir = NULL;
814 	VERIFY0(dmu_object_free(mos, obj, tx));
815 
816 	dsl_dir_destroy_sync(ddobj, tx);
817 
818 	if (rmorigin) {
819 		dsl_dataset_t *prev;
820 		VERIFY0(dsl_dataset_hold_obj(dp, prevobj, FTAG, &prev));
821 		dsl_destroy_snapshot_sync_impl(prev, B_FALSE, tx);
822 		dsl_dataset_rele(prev, FTAG);
823 	}
824 }
825 
826 static void
827 dsl_destroy_head_sync(void *arg, dmu_tx_t *tx)
828 {
829 	dsl_destroy_head_arg_t *ddha = arg;
830 	dsl_pool_t *dp = dmu_tx_pool(tx);
831 	dsl_dataset_t *ds;
832 
833 	VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
834 	dsl_destroy_head_sync_impl(ds, tx);
835 	dsl_dataset_rele(ds, FTAG);
836 }
837 
838 static void
839 dsl_destroy_head_begin_sync(void *arg, dmu_tx_t *tx)
840 {
841 	dsl_destroy_head_arg_t *ddha = arg;
842 	dsl_pool_t *dp = dmu_tx_pool(tx);
843 	dsl_dataset_t *ds;
844 
845 	VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
846 
847 	/* Mark it as inconsistent on-disk, in case we crash */
848 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
849 	ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
850 
851 	spa_history_log_internal_ds(ds, "destroy begin", tx, "");
852 	dsl_dataset_rele(ds, FTAG);
853 }
854 
855 int
856 dsl_destroy_head(const char *name)
857 {
858 	dsl_destroy_head_arg_t ddha;
859 	int error;
860 	spa_t *spa;
861 	boolean_t isenabled;
862 
863 #ifdef _KERNEL
864 	zfs_destroy_unmount_origin(name);
865 #endif
866 
867 	error = spa_open(name, &spa, FTAG);
868 	if (error != 0)
869 		return (error);
870 	isenabled = spa_feature_is_enabled(spa,
871 	    &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]);
872 	spa_close(spa, FTAG);
873 
874 	ddha.ddha_name = name;
875 
876 	if (!isenabled) {
877 		objset_t *os;
878 
879 		error = dsl_sync_task(name, dsl_destroy_head_check,
880 		    dsl_destroy_head_begin_sync, &ddha, 0);
881 		if (error != 0)
882 			return (error);
883 
884 		/*
885 		 * Head deletion is processed in one txg on old pools;
886 		 * remove the objects from open context so that the txg sync
887 		 * is not too long.
888 		 */
889 		error = dmu_objset_own(name, DMU_OST_ANY, B_FALSE, FTAG, &os);
890 		if (error == 0) {
891 			uint64_t prev_snap_txg =
892 			    dmu_objset_ds(os)->ds_phys->ds_prev_snap_txg;
893 			for (uint64_t obj = 0; error == 0;
894 			    error = dmu_object_next(os, &obj, FALSE,
895 			    prev_snap_txg))
896 				(void) dmu_free_object(os, obj);
897 			/* sync out all frees */
898 			txg_wait_synced(dmu_objset_pool(os), 0);
899 			dmu_objset_disown(os, FTAG);
900 		}
901 	}
902 
903 	return (dsl_sync_task(name, dsl_destroy_head_check,
904 	    dsl_destroy_head_sync, &ddha, 0));
905 }
906 
907 /*
908  * Note, this function is used as the callback for dmu_objset_find().  We
909  * always return 0 so that we will continue to find and process
910  * inconsistent datasets, even if we encounter an error trying to
911  * process one of them.
912  */
913 /* ARGSUSED */
914 int
915 dsl_destroy_inconsistent(const char *dsname, void *arg)
916 {
917 	objset_t *os;
918 
919 	if (dmu_objset_hold(dsname, FTAG, &os) == 0) {
920 		boolean_t inconsistent = DS_IS_INCONSISTENT(dmu_objset_ds(os));
921 		dmu_objset_rele(os, FTAG);
922 		if (inconsistent)
923 			(void) dsl_destroy_head(dsname);
924 	}
925 	return (0);
926 }
927