xref: /titanic_50/usr/src/uts/common/fs/zfs/dsl_destroy.c (revision bbaa8b60dd95d714741fc474adad3cf710ef4efd)
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 	for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
215 	    zap_cursor_retrieve(&zc, &za) == 0;
216 	    zap_cursor_advance(&zc)) {
217 		dsl_dataset_t *clone;
218 
219 		VERIFY0(dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
220 		    za.za_first_integer, FTAG, &clone));
221 		if (clone->ds_dir->dd_origin_txg > mintxg) {
222 			dsl_deadlist_remove_key(&clone->ds_deadlist,
223 			    mintxg, tx);
224 			dsl_dataset_remove_clones_key(clone, mintxg, tx);
225 		}
226 		dsl_dataset_rele(clone, FTAG);
227 	}
228 	zap_cursor_fini(&zc);
229 }
230 
231 void
232 dsl_destroy_snapshot_sync_impl(dsl_dataset_t *ds, boolean_t defer, dmu_tx_t *tx)
233 {
234 	int err;
235 	int after_branch_point = FALSE;
236 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
237 	objset_t *mos = dp->dp_meta_objset;
238 	dsl_dataset_t *ds_prev = NULL;
239 	uint64_t obj;
240 
241 	ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
242 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
243 	ASSERT(refcount_is_zero(&ds->ds_longholds));
244 
245 	if (defer &&
246 	    (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1)) {
247 		ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
248 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
249 		ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
250 		spa_history_log_internal_ds(ds, "defer_destroy", tx, "");
251 		return;
252 	}
253 
254 	ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
255 
256 	/* We need to log before removing it from the namespace. */
257 	spa_history_log_internal_ds(ds, "destroy", tx, "");
258 
259 	dsl_scan_ds_destroyed(ds, tx);
260 
261 	obj = ds->ds_object;
262 
263 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
264 		ASSERT3P(ds->ds_prev, ==, NULL);
265 		VERIFY0(dsl_dataset_hold_obj(dp,
266 		    ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
267 		after_branch_point =
268 		    (ds_prev->ds_phys->ds_next_snap_obj != obj);
269 
270 		dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
271 		if (after_branch_point &&
272 		    ds_prev->ds_phys->ds_next_clones_obj != 0) {
273 			dsl_dataset_remove_from_next_clones(ds_prev, obj, tx);
274 			if (ds->ds_phys->ds_next_snap_obj != 0) {
275 				VERIFY0(zap_add_int(mos,
276 				    ds_prev->ds_phys->ds_next_clones_obj,
277 				    ds->ds_phys->ds_next_snap_obj, tx));
278 			}
279 		}
280 		if (!after_branch_point) {
281 			ds_prev->ds_phys->ds_next_snap_obj =
282 			    ds->ds_phys->ds_next_snap_obj;
283 		}
284 	}
285 
286 	dsl_dataset_t *ds_next;
287 	uint64_t old_unique;
288 	uint64_t used = 0, comp = 0, uncomp = 0;
289 
290 	VERIFY0(dsl_dataset_hold_obj(dp,
291 	    ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
292 	ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
293 
294 	old_unique = ds_next->ds_phys->ds_unique_bytes;
295 
296 	dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
297 	ds_next->ds_phys->ds_prev_snap_obj =
298 	    ds->ds_phys->ds_prev_snap_obj;
299 	ds_next->ds_phys->ds_prev_snap_txg =
300 	    ds->ds_phys->ds_prev_snap_txg;
301 	ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
302 	    ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
303 
304 	if (ds_next->ds_deadlist.dl_oldfmt) {
305 		process_old_deadlist(ds, ds_prev, ds_next,
306 		    after_branch_point, tx);
307 	} else {
308 		/* Adjust prev's unique space. */
309 		if (ds_prev && !after_branch_point) {
310 			dsl_deadlist_space_range(&ds_next->ds_deadlist,
311 			    ds_prev->ds_phys->ds_prev_snap_txg,
312 			    ds->ds_phys->ds_prev_snap_txg,
313 			    &used, &comp, &uncomp);
314 			ds_prev->ds_phys->ds_unique_bytes += used;
315 		}
316 
317 		/* Adjust snapused. */
318 		dsl_deadlist_space_range(&ds_next->ds_deadlist,
319 		    ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
320 		    &used, &comp, &uncomp);
321 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
322 		    -used, -comp, -uncomp, tx);
323 
324 		/* Move blocks to be freed to pool's free list. */
325 		dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
326 		    &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
327 		    tx);
328 		dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
329 		    DD_USED_HEAD, used, comp, uncomp, tx);
330 
331 		/* Merge our deadlist into next's and free it. */
332 		dsl_deadlist_merge(&ds_next->ds_deadlist,
333 		    ds->ds_phys->ds_deadlist_obj, tx);
334 	}
335 	dsl_deadlist_close(&ds->ds_deadlist);
336 	dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
337 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
338 	ds->ds_phys->ds_deadlist_obj = 0;
339 
340 	/* Collapse range in clone heads */
341 	dsl_dataset_remove_clones_key(ds,
342 	    ds->ds_phys->ds_creation_txg, tx);
343 
344 	if (dsl_dataset_is_snapshot(ds_next)) {
345 		dsl_dataset_t *ds_nextnext;
346 
347 		/*
348 		 * Update next's unique to include blocks which
349 		 * were previously shared by only this snapshot
350 		 * and it.  Those blocks will be born after the
351 		 * prev snap and before this snap, and will have
352 		 * died after the next snap and before the one
353 		 * after that (ie. be on the snap after next's
354 		 * deadlist).
355 		 */
356 		VERIFY0(dsl_dataset_hold_obj(dp,
357 		    ds_next->ds_phys->ds_next_snap_obj, FTAG, &ds_nextnext));
358 		dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
359 		    ds->ds_phys->ds_prev_snap_txg,
360 		    ds->ds_phys->ds_creation_txg,
361 		    &used, &comp, &uncomp);
362 		ds_next->ds_phys->ds_unique_bytes += used;
363 		dsl_dataset_rele(ds_nextnext, FTAG);
364 		ASSERT3P(ds_next->ds_prev, ==, NULL);
365 
366 		/* Collapse range in this head. */
367 		dsl_dataset_t *hds;
368 		VERIFY0(dsl_dataset_hold_obj(dp,
369 		    ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &hds));
370 		dsl_deadlist_remove_key(&hds->ds_deadlist,
371 		    ds->ds_phys->ds_creation_txg, tx);
372 		dsl_dataset_rele(hds, FTAG);
373 
374 	} else {
375 		ASSERT3P(ds_next->ds_prev, ==, ds);
376 		dsl_dataset_rele(ds_next->ds_prev, ds_next);
377 		ds_next->ds_prev = NULL;
378 		if (ds_prev) {
379 			VERIFY0(dsl_dataset_hold_obj(dp,
380 			    ds->ds_phys->ds_prev_snap_obj,
381 			    ds_next, &ds_next->ds_prev));
382 		}
383 
384 		dsl_dataset_recalc_head_uniq(ds_next);
385 
386 		/*
387 		 * Reduce the amount of our unconsumed refreservation
388 		 * being charged to our parent by the amount of
389 		 * new unique data we have gained.
390 		 */
391 		if (old_unique < ds_next->ds_reserved) {
392 			int64_t mrsdelta;
393 			uint64_t new_unique =
394 			    ds_next->ds_phys->ds_unique_bytes;
395 
396 			ASSERT(old_unique <= new_unique);
397 			mrsdelta = MIN(new_unique - old_unique,
398 			    ds_next->ds_reserved - old_unique);
399 			dsl_dir_diduse_space(ds->ds_dir,
400 			    DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
401 		}
402 	}
403 	dsl_dataset_rele(ds_next, FTAG);
404 
405 	/*
406 	 * This must be done after the dsl_traverse(), because it will
407 	 * re-open the objset.
408 	 */
409 	if (ds->ds_objset) {
410 		dmu_objset_evict(ds->ds_objset);
411 		ds->ds_objset = NULL;
412 	}
413 
414 	/* remove from snapshot namespace */
415 	dsl_dataset_t *ds_head;
416 	ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
417 	VERIFY0(dsl_dataset_hold_obj(dp,
418 	    ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
419 	VERIFY0(dsl_dataset_get_snapname(ds));
420 #ifdef ZFS_DEBUG
421 	{
422 		uint64_t val;
423 
424 		err = dsl_dataset_snap_lookup(ds_head,
425 		    ds->ds_snapname, &val);
426 		ASSERT0(err);
427 		ASSERT3U(val, ==, obj);
428 	}
429 #endif
430 	VERIFY0(dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx));
431 	dsl_dataset_rele(ds_head, FTAG);
432 
433 	if (ds_prev != NULL)
434 		dsl_dataset_rele(ds_prev, FTAG);
435 
436 	spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
437 
438 	if (ds->ds_phys->ds_next_clones_obj != 0) {
439 		uint64_t count;
440 		ASSERT0(zap_count(mos,
441 		    ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
442 		VERIFY0(dmu_object_free(mos,
443 		    ds->ds_phys->ds_next_clones_obj, tx));
444 	}
445 	if (ds->ds_phys->ds_props_obj != 0)
446 		VERIFY0(zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
447 	if (ds->ds_phys->ds_userrefs_obj != 0)
448 		VERIFY0(zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
449 	dsl_dir_rele(ds->ds_dir, ds);
450 	ds->ds_dir = NULL;
451 	VERIFY0(dmu_object_free(mos, obj, tx));
452 }
453 
454 static void
455 dsl_destroy_snapshot_sync(void *arg, dmu_tx_t *tx)
456 {
457 	dmu_snapshots_destroy_arg_t *dsda = arg;
458 	dsl_pool_t *dp = dmu_tx_pool(tx);
459 	nvpair_t *pair;
460 
461 	for (pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, NULL);
462 	    pair != NULL;
463 	    pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, pair)) {
464 		dsl_dataset_t *ds;
465 
466 		VERIFY0(dsl_dataset_hold(dp, nvpair_name(pair), FTAG, &ds));
467 
468 		dsl_destroy_snapshot_sync_impl(ds, dsda->dsda_defer, tx);
469 		dsl_dataset_rele(ds, FTAG);
470 	}
471 }
472 
473 /*
474  * The semantics of this function are described in the comment above
475  * lzc_destroy_snaps().  To summarize:
476  *
477  * The snapshots must all be in the same pool.
478  *
479  * Snapshots that don't exist will be silently ignored (considered to be
480  * "already deleted").
481  *
482  * On success, all snaps will be destroyed and this will return 0.
483  * On failure, no snaps will be destroyed, the errlist will be filled in,
484  * and this will return an errno.
485  */
486 int
487 dsl_destroy_snapshots_nvl(nvlist_t *snaps, boolean_t defer,
488     nvlist_t *errlist)
489 {
490 	dmu_snapshots_destroy_arg_t dsda;
491 	int error;
492 	nvpair_t *pair;
493 
494 	pair = nvlist_next_nvpair(snaps, NULL);
495 	if (pair == NULL)
496 		return (0);
497 
498 	dsda.dsda_snaps = snaps;
499 	dsda.dsda_successful_snaps = fnvlist_alloc();
500 	dsda.dsda_defer = defer;
501 	dsda.dsda_errlist = errlist;
502 
503 	error = dsl_sync_task(nvpair_name(pair),
504 	    dsl_destroy_snapshot_check, dsl_destroy_snapshot_sync,
505 	    &dsda, 0);
506 	fnvlist_free(dsda.dsda_successful_snaps);
507 
508 	return (error);
509 }
510 
511 int
512 dsl_destroy_snapshot(const char *name, boolean_t defer)
513 {
514 	int error;
515 	nvlist_t *nvl = fnvlist_alloc();
516 	nvlist_t *errlist = fnvlist_alloc();
517 
518 	fnvlist_add_boolean(nvl, name);
519 	error = dsl_destroy_snapshots_nvl(nvl, defer, errlist);
520 	fnvlist_free(errlist);
521 	fnvlist_free(nvl);
522 	return (error);
523 }
524 
525 struct killarg {
526 	dsl_dataset_t *ds;
527 	dmu_tx_t *tx;
528 };
529 
530 /* ARGSUSED */
531 static int
532 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
533     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
534 {
535 	struct killarg *ka = arg;
536 	dmu_tx_t *tx = ka->tx;
537 
538 	if (bp == NULL)
539 		return (0);
540 
541 	if (zb->zb_level == ZB_ZIL_LEVEL) {
542 		ASSERT(zilog != NULL);
543 		/*
544 		 * It's a block in the intent log.  It has no
545 		 * accounting, so just free it.
546 		 */
547 		dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
548 	} else {
549 		ASSERT(zilog == NULL);
550 		ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
551 		(void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
552 	}
553 
554 	return (0);
555 }
556 
557 static void
558 old_synchronous_dataset_destroy(dsl_dataset_t *ds, dmu_tx_t *tx)
559 {
560 	struct killarg ka;
561 
562 	/*
563 	 * Free everything that we point to (that's born after
564 	 * the previous snapshot, if we are a clone)
565 	 *
566 	 * NB: this should be very quick, because we already
567 	 * freed all the objects in open context.
568 	 */
569 	ka.ds = ds;
570 	ka.tx = tx;
571 	VERIFY0(traverse_dataset(ds,
572 	    ds->ds_phys->ds_prev_snap_txg, TRAVERSE_POST,
573 	    kill_blkptr, &ka));
574 	ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) || ds->ds_phys->ds_unique_bytes == 0);
575 }
576 
577 typedef struct dsl_destroy_head_arg {
578 	const char *ddha_name;
579 } dsl_destroy_head_arg_t;
580 
581 int
582 dsl_destroy_head_check_impl(dsl_dataset_t *ds, int expected_holds)
583 {
584 	int error;
585 	uint64_t count;
586 	objset_t *mos;
587 
588 	if (dsl_dataset_is_snapshot(ds))
589 		return (SET_ERROR(EINVAL));
590 
591 	if (refcount_count(&ds->ds_longholds) != expected_holds)
592 		return (SET_ERROR(EBUSY));
593 
594 	mos = ds->ds_dir->dd_pool->dp_meta_objset;
595 
596 	/*
597 	 * Can't delete a head dataset if there are snapshots of it.
598 	 * (Except if the only snapshots are from the branch we cloned
599 	 * from.)
600 	 */
601 	if (ds->ds_prev != NULL &&
602 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
603 		return (SET_ERROR(EBUSY));
604 
605 	/*
606 	 * Can't delete if there are children of this fs.
607 	 */
608 	error = zap_count(mos,
609 	    ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
610 	if (error != 0)
611 		return (error);
612 	if (count != 0)
613 		return (SET_ERROR(EEXIST));
614 
615 	if (dsl_dir_is_clone(ds->ds_dir) && DS_IS_DEFER_DESTROY(ds->ds_prev) &&
616 	    ds->ds_prev->ds_phys->ds_num_children == 2 &&
617 	    ds->ds_prev->ds_userrefs == 0) {
618 		/* We need to remove the origin snapshot as well. */
619 		if (!refcount_is_zero(&ds->ds_prev->ds_longholds))
620 			return (SET_ERROR(EBUSY));
621 	}
622 	return (0);
623 }
624 
625 static int
626 dsl_destroy_head_check(void *arg, dmu_tx_t *tx)
627 {
628 	dsl_destroy_head_arg_t *ddha = arg;
629 	dsl_pool_t *dp = dmu_tx_pool(tx);
630 	dsl_dataset_t *ds;
631 	int error;
632 
633 	error = dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds);
634 	if (error != 0)
635 		return (error);
636 
637 	error = dsl_destroy_head_check_impl(ds, 0);
638 	dsl_dataset_rele(ds, FTAG);
639 	return (error);
640 }
641 
642 static void
643 dsl_dir_destroy_sync(uint64_t ddobj, dmu_tx_t *tx)
644 {
645 	dsl_dir_t *dd;
646 	dsl_pool_t *dp = dmu_tx_pool(tx);
647 	objset_t *mos = dp->dp_meta_objset;
648 	dd_used_t t;
649 
650 	ASSERT(RRW_WRITE_HELD(&dmu_tx_pool(tx)->dp_config_rwlock));
651 
652 	VERIFY0(dsl_dir_hold_obj(dp, ddobj, NULL, FTAG, &dd));
653 
654 	ASSERT0(dd->dd_phys->dd_head_dataset_obj);
655 
656 	/*
657 	 * Remove our reservation. The impl() routine avoids setting the
658 	 * actual property, which would require the (already destroyed) ds.
659 	 */
660 	dsl_dir_set_reservation_sync_impl(dd, 0, tx);
661 
662 	ASSERT0(dd->dd_phys->dd_used_bytes);
663 	ASSERT0(dd->dd_phys->dd_reserved);
664 	for (t = 0; t < DD_USED_NUM; t++)
665 		ASSERT0(dd->dd_phys->dd_used_breakdown[t]);
666 
667 	VERIFY0(zap_destroy(mos, dd->dd_phys->dd_child_dir_zapobj, tx));
668 	VERIFY0(zap_destroy(mos, dd->dd_phys->dd_props_zapobj, tx));
669 	VERIFY0(dsl_deleg_destroy(mos, dd->dd_phys->dd_deleg_zapobj, tx));
670 	VERIFY0(zap_remove(mos,
671 	    dd->dd_parent->dd_phys->dd_child_dir_zapobj, dd->dd_myname, tx));
672 
673 	dsl_dir_rele(dd, FTAG);
674 	VERIFY0(dmu_object_free(mos, ddobj, tx));
675 }
676 
677 void
678 dsl_destroy_head_sync_impl(dsl_dataset_t *ds, dmu_tx_t *tx)
679 {
680 	dsl_pool_t *dp = dmu_tx_pool(tx);
681 	objset_t *mos = dp->dp_meta_objset;
682 	uint64_t obj, ddobj, prevobj = 0;
683 	boolean_t rmorigin;
684 
685 	ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
686 	ASSERT(ds->ds_prev == NULL ||
687 	    ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
688 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
689 	ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
690 
691 	/* We need to log before removing it from the namespace. */
692 	spa_history_log_internal_ds(ds, "destroy", tx, "");
693 
694 	rmorigin = (dsl_dir_is_clone(ds->ds_dir) &&
695 	    DS_IS_DEFER_DESTROY(ds->ds_prev) &&
696 	    ds->ds_prev->ds_phys->ds_num_children == 2 &&
697 	    ds->ds_prev->ds_userrefs == 0);
698 
699 	/* Remove our reservation */
700 	if (ds->ds_reserved != 0) {
701 		dsl_dataset_set_refreservation_sync_impl(ds,
702 		    (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
703 		    0, tx);
704 		ASSERT0(ds->ds_reserved);
705 	}
706 
707 	dsl_scan_ds_destroyed(ds, tx);
708 
709 	obj = ds->ds_object;
710 
711 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
712 		/* This is a clone */
713 		ASSERT(ds->ds_prev != NULL);
714 		ASSERT3U(ds->ds_prev->ds_phys->ds_next_snap_obj, !=, obj);
715 		ASSERT0(ds->ds_phys->ds_next_snap_obj);
716 
717 		dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
718 		if (ds->ds_prev->ds_phys->ds_next_clones_obj != 0) {
719 			dsl_dataset_remove_from_next_clones(ds->ds_prev,
720 			    obj, tx);
721 		}
722 
723 		ASSERT3U(ds->ds_prev->ds_phys->ds_num_children, >, 1);
724 		ds->ds_prev->ds_phys->ds_num_children--;
725 	}
726 
727 	zfeature_info_t *async_destroy =
728 	    &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY];
729 	objset_t *os;
730 
731 	/*
732 	 * Destroy the deadlist.  Unless it's a clone, the
733 	 * deadlist should be empty.  (If it's a clone, it's
734 	 * safe to ignore the deadlist contents.)
735 	 */
736 	dsl_deadlist_close(&ds->ds_deadlist);
737 	dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
738 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
739 	ds->ds_phys->ds_deadlist_obj = 0;
740 
741 	VERIFY0(dmu_objset_from_ds(ds, &os));
742 
743 	if (!spa_feature_is_enabled(dp->dp_spa, async_destroy)) {
744 		old_synchronous_dataset_destroy(ds, tx);
745 	} else {
746 		/*
747 		 * Move the bptree into the pool's list of trees to
748 		 * clean up and update space accounting information.
749 		 */
750 		uint64_t used, comp, uncomp;
751 
752 		zil_destroy_sync(dmu_objset_zil(os), tx);
753 
754 		if (!spa_feature_is_active(dp->dp_spa, async_destroy)) {
755 			dsl_scan_t *scn = dp->dp_scan;
756 
757 			spa_feature_incr(dp->dp_spa, async_destroy, tx);
758 			dp->dp_bptree_obj = bptree_alloc(mos, tx);
759 			VERIFY0(zap_add(mos,
760 			    DMU_POOL_DIRECTORY_OBJECT,
761 			    DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
762 			    &dp->dp_bptree_obj, tx));
763 			ASSERT(!scn->scn_async_destroying);
764 			scn->scn_async_destroying = B_TRUE;
765 		}
766 
767 		used = ds->ds_dir->dd_phys->dd_used_bytes;
768 		comp = ds->ds_dir->dd_phys->dd_compressed_bytes;
769 		uncomp = ds->ds_dir->dd_phys->dd_uncompressed_bytes;
770 
771 		ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
772 		    ds->ds_phys->ds_unique_bytes == used);
773 
774 		bptree_add(mos, dp->dp_bptree_obj,
775 		    &ds->ds_phys->ds_bp, ds->ds_phys->ds_prev_snap_txg,
776 		    used, comp, uncomp, tx);
777 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
778 		    -used, -comp, -uncomp, tx);
779 		dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
780 		    used, comp, uncomp, tx);
781 	}
782 
783 	if (ds->ds_prev != NULL) {
784 		if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
785 			VERIFY0(zap_remove_int(mos,
786 			    ds->ds_prev->ds_dir->dd_phys->dd_clones,
787 			    ds->ds_object, tx));
788 		}
789 		prevobj = ds->ds_prev->ds_object;
790 		dsl_dataset_rele(ds->ds_prev, ds);
791 		ds->ds_prev = NULL;
792 	}
793 
794 	/*
795 	 * This must be done after the dsl_traverse(), because it will
796 	 * re-open the objset.
797 	 */
798 	if (ds->ds_objset) {
799 		dmu_objset_evict(ds->ds_objset);
800 		ds->ds_objset = NULL;
801 	}
802 
803 	/* Erase the link in the dir */
804 	dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
805 	ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
806 	ddobj = ds->ds_dir->dd_object;
807 	ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
808 	VERIFY0(zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx));
809 
810 	spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
811 
812 	ASSERT0(ds->ds_phys->ds_next_clones_obj);
813 	ASSERT0(ds->ds_phys->ds_props_obj);
814 	ASSERT0(ds->ds_phys->ds_userrefs_obj);
815 	dsl_dir_rele(ds->ds_dir, ds);
816 	ds->ds_dir = NULL;
817 	VERIFY0(dmu_object_free(mos, obj, tx));
818 
819 	dsl_dir_destroy_sync(ddobj, tx);
820 
821 	if (rmorigin) {
822 		dsl_dataset_t *prev;
823 		VERIFY0(dsl_dataset_hold_obj(dp, prevobj, FTAG, &prev));
824 		dsl_destroy_snapshot_sync_impl(prev, B_FALSE, tx);
825 		dsl_dataset_rele(prev, FTAG);
826 	}
827 }
828 
829 static void
830 dsl_destroy_head_sync(void *arg, dmu_tx_t *tx)
831 {
832 	dsl_destroy_head_arg_t *ddha = arg;
833 	dsl_pool_t *dp = dmu_tx_pool(tx);
834 	dsl_dataset_t *ds;
835 
836 	VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
837 	dsl_destroy_head_sync_impl(ds, tx);
838 	dsl_dataset_rele(ds, FTAG);
839 }
840 
841 static void
842 dsl_destroy_head_begin_sync(void *arg, dmu_tx_t *tx)
843 {
844 	dsl_destroy_head_arg_t *ddha = arg;
845 	dsl_pool_t *dp = dmu_tx_pool(tx);
846 	dsl_dataset_t *ds;
847 
848 	VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
849 
850 	/* Mark it as inconsistent on-disk, in case we crash */
851 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
852 	ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
853 
854 	spa_history_log_internal_ds(ds, "destroy begin", tx, "");
855 	dsl_dataset_rele(ds, FTAG);
856 }
857 
858 int
859 dsl_destroy_head(const char *name)
860 {
861 	dsl_destroy_head_arg_t ddha;
862 	int error;
863 	spa_t *spa;
864 	boolean_t isenabled;
865 
866 #ifdef _KERNEL
867 	zfs_destroy_unmount_origin(name);
868 #endif
869 
870 	error = spa_open(name, &spa, FTAG);
871 	if (error != 0)
872 		return (error);
873 	isenabled = spa_feature_is_enabled(spa,
874 	    &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]);
875 	spa_close(spa, FTAG);
876 
877 	ddha.ddha_name = name;
878 
879 	if (!isenabled) {
880 		objset_t *os;
881 
882 		error = dsl_sync_task(name, dsl_destroy_head_check,
883 		    dsl_destroy_head_begin_sync, &ddha, 0);
884 		if (error != 0)
885 			return (error);
886 
887 		/*
888 		 * Head deletion is processed in one txg on old pools;
889 		 * remove the objects from open context so that the txg sync
890 		 * is not too long.
891 		 */
892 		error = dmu_objset_own(name, DMU_OST_ANY, B_FALSE, FTAG, &os);
893 		if (error == 0) {
894 			uint64_t prev_snap_txg =
895 			    dmu_objset_ds(os)->ds_phys->ds_prev_snap_txg;
896 			for (uint64_t obj = 0; error == 0;
897 			    error = dmu_object_next(os, &obj, FALSE,
898 			    prev_snap_txg))
899 				(void) dmu_free_long_object(os, obj);
900 			/* sync out all frees */
901 			txg_wait_synced(dmu_objset_pool(os), 0);
902 			dmu_objset_disown(os, FTAG);
903 		}
904 	}
905 
906 	return (dsl_sync_task(name, dsl_destroy_head_check,
907 	    dsl_destroy_head_sync, &ddha, 0));
908 }
909 
910 /*
911  * Note, this function is used as the callback for dmu_objset_find().  We
912  * always return 0 so that we will continue to find and process
913  * inconsistent datasets, even if we encounter an error trying to
914  * process one of them.
915  */
916 /* ARGSUSED */
917 int
918 dsl_destroy_inconsistent(const char *dsname, void *arg)
919 {
920 	objset_t *os;
921 
922 	if (dmu_objset_hold(dsname, FTAG, &os) == 0) {
923 		boolean_t inconsistent = DS_IS_INCONSISTENT(dmu_objset_ds(os));
924 		dmu_objset_rele(os, FTAG);
925 		if (inconsistent)
926 			(void) dsl_destroy_head(dsname);
927 	}
928 	return (0);
929 }
930